Microsoft Confidential Computing – When Agents Work Inside The Seal, Attestation Becomes The Control

Publication Record

Evidence current through: Thursday, November 19, 2026, at 12:45 a.m. UTC (7:45 p.m. ET and 4:45 p.m. PT on Wednesday, November 18).

Microsoft’s Ignite summary of November 17, 2026, calls all three of its confidential computing items previews. Its detailed post of the same day calls only the clean room a preview, and this note follows that more exact reading, at the status Microsoft gave by the time above. Developments after that time are not reflected. Times in the text are given in UTC.

Research Note In Two MinutesNote synopsis · AI drafted · Analyst reviewed

Drafted by AI from the full research note and reviewed by a Greyhound Research analyst before publication. The analysis and conclusions are Greyhound Research’s own.

At Microsoft Ignite on November 17, 2026, Microsoft set out three confidential computing items. Confidential computing keeps data sealed while a computer works on it, so the cloud operator cannot read it. VR72, the first, is built to stretch that seal across a rack of GPUs, the chips that run large AI models. A second puts a smaller sealed space inside a virtual machine, alone or in containers on Microsoft’s Kubernetes service. The third is a clean room where several companies can analyze pooled data without seeing each other’s records. Microsoft’s summary calls all three previews. Its detailed post the same day is more exact. VR72 was a technology showcase, and the smaller sealed spaces are planned for a future release, so only the clean room, first previewed in June, is in preview. None of the three has a published price, region list, or general availability date.

Greyhound Research’s position is that Microsoft has shown how much bigger the sealed room can be, and the harder question is now who sits inside it. The seal keeps the operator out. It does not keep the customer’s own agent out, because the agent is the one reading the data. Buyers should judge these offerings by their proof, known as attestation, more than by their encryption. Greyhound Research reads this as a familiar stage for a security feature. Sealed computing is moving from a premium option toward a standard expectation, and as that happens buyers tend to stop asking whether it is present and start asking how its evidence is checked. The more exact status makes that advice firmer. There is less to sign today and more time to set terms.

Three findings carry the argument. First, Microsoft has not said whether the agent memory and cached context its own services keep sit inside the seal. Second, researchers have forged the hardware’s proof in the lab. Microsoft’s Ignite demonstration checked that proof with a local verifier, while Google already offers a verifier it does not own. Third, the clean room’s real risk is what an approved question may return, so it leaks by answer rather than by breach.

The note sets out a ten-move Greyhound CXO Runbook, with seven exhibits, six tables, and six questions for Microsoft. None of the moves waits on Microsoft. The first three are simple to state. Map which agent data sits inside the seal. Name who verifies the proof, and how often. Write the clean room’s allowed answers into the agreement before any data goes in.

The evidence has limits. This note first read the status from Microsoft’s summary and now follows Microsoft’s more exact detailed post, read on November 19. The 2025 attacks needed hands on the server, which large data centers make much harder. Greyhound Research names the signals that would change its view within twelve months. Until they arrive, the advice holds: pilot on what is generally available, and commit regulated data only when the proof, its checker, and the agent’s memory are documented.

In This Research NoteSection guide · AI drafted · Analyst reviewed

Sixteen sections in eight parts, each with a short summary. Select a title to go straight to that section. Each section opens with a Greyhound Spotlight, the one fact in it a reader would not expect, and eleven end with a block named The Deeper Read, which holds the fuller argument and the passages it rests on. Select any line in that block to open its full text.

  1. Part 1What Ignite Changed
  2. 1.1What Microsoft Announced: One Restated Preview, Two Items Still In Development

    Microsoft’s summary calls all three items previews. Its detailed post calls VR72 a showcase and the smaller sealed spaces a future release, which leaves the clean room, first previewed in June, as the one preview. None has a price, region list, or release date.

  3. 1.2The Seal And The Reader: What Confidential Computing Keeps Out

    The seal keeps the cloud operator, its administrators, and other tenants out. It does not keep out the customer’s own agent, which is inside by design.

  4. Part 2The Proof
  5. 2.1Attestation: The Proof Matters More Than The Lock

    Encryption while data is in use is now standard on the three large clouds. What separates offers is how the proof is produced, checked, kept, and refreshed as models and tools change.

  6. 2.2Who Checks The Proof: Independent Verifiers And Microsoft’s Open Question

    Google lets customers check the proof with a verifier it does not own. AWS and NVIDIA run their own. Microsoft’s VR72 demonstration used a local verifier, and no verifier is named for the enclaves.

  7. 2.3When The Seal Was Opened: Three Lab Attacks In One Year

    In 2025 researchers forged the hardware’s proof with cheap equipment and hands on the server. Large data centers make that much harder, and the lesson is about checking the proof.

  8. Part 3Agents Inside The Seal
  9. 3.1Agent Memory: The Notes Left On The Desk

    An agent’s work passes through prompts, retrieved documents, model computation, memory, cached context, tool calls, and outputs. Microsoft places only the computation clearly inside the seal.

  10. 3.2The Clean Room: Leaks By Answer, Not By Breach

    The clean room keeps each party’s raw data from the others. Its risk is what an approved question may return, especially when an agent writes the questions.

  11. Part 4The Platform Contest
  12. 4.1Three Clouds Compared: Sealed Compute Is Becoming Standard

    Microsoft, AWS, and Google all sell sealed computing and confidential GPUs in some form. Microsoft leads on rack scale and was early. Google leads on independent verification.

  13. Part 5The Risks And The Counter-Argument
  14. 5.1Risks To Watch: Where A Sealed Room Can Still Leak

    Five risks, from unsealed agent memory to proofs nobody re-checks, each with an early warning and an owner.

  15. 5.2The Strongest Objection: The Seal Is Enough For Now

    The fair counter-case: rack scale removes the risk committee’s top objection, Microsoft was early, and enclaves narrow what must be trusted. It holds for the provider question and falls short on agents.

  16. Part 6What To Do, By Role
  17. 6.1The Greyhound CXO Runbook: Ten Moves Before Regulated Data Goes In

    Ten moves for the CISO, the data protection officer, the chief risk officer, platform engineering, enterprise architecture, procurement, and the CIO, each naming the assumption it corrects and who decides.

  18. 6.2Questions For Microsoft: What Buyers Will Ask Next

    Six questions only Microsoft can answer, on agent memory, the verifier, refresh, the clean room, physical security, and terms. Each is tied to a Runbook move.

  19. Part 7The Final Greyhound Standpoint
  20. 7.1Confidential Computing: The Final Greyhound Standpoint

    The seal is set to grow, the reader is still inside, and the proof is the control. The limits of the evidence, and the signals that would change the view within twelve months.

  21. Part 8Appendix, Sources, And Disclosures
  22. 8.1Appendix: The Evidence Register, Source By Source

    Every source behind this research note, with when it was read and where it is used. Passages quoted in The Deeper Read are not repeated here.

  23. 8.2Sources And References: Every Document Behind Every Claim

    Thirty-four sources in six classes, from Microsoft’s announcements and documentation to AWS, Google, NVIDIA, security research, and the research this note builds on.

  24. 8.3Disclosures: Independence, Financial, Intended Use, Methodology, Authoring, And Copyright

    Independence first, then the Financial Disclosure, intended use, research method and its limits, authoring, and copyright terms.

Where To Start, By RoleReader guide · AI drafted · Analyst reviewed

Part 1What Ignite Changed

What Microsoft Announced: One Restated Preview, Two Items Still In Development

Greyhound SpotlightThe only Ignite confidential computing item that Microsoft’s detailed post calls a preview is the clean room, and that preview opened in June 2026.

Microsoft’s Ignite summary, the Book of News, lists three confidential computing offerings and says all three are now in preview. A preview is a test release without full service terms. Microsoft’s detailed post on the same day, by Vikas Bhatia, gives a more exact status. It describes VR72 as a technology showcase on one server with four GPUs, and says VR72 confidential virtual machines and Kata-based confidential containers on AKS remain in development. The smaller sealed spaces are planned for a future release. Only the clean room is described as a limited preview. No price, region list, or release date is published for any of the three.

The two readings differ, and this note gives the detailed post more weight. It is the more specific document, it is written for technical buyers, and it describes what was actually shown. Greyhound Research’s first reading, on November 18, followed the summary’s wording that three offerings were in preview. The detailed post, read on November 19, shows that two of the three are not yet in preview, so the view has moved from three previews to one. As a result, the advice has become firmer. For a CIO, Ignite is a direction to plan for, and only the clean room can be tried at all, on test-only terms.

For buyers, the gap between the two documents is itself information. A summary written for the whole market tends to round statuses up, and a post written for engineers states them exactly, so procurement should plan from the engineers’ version. What is new is the ambition rather than the product: sealing a rack of GPUs is a change in scale from sealing one machine. Greyhound Research expects the first general release to arrive in a few regions, much as the confidential GPU machines did in 2024. The less obvious effect falls on budgets. With no price published, finance teams cannot yet compare sealed capacity with ordinary GPU capacity or write the business case.

Table 1 sets the three Ignite items beside the confidential computing products a buyer can sign today. Generally available means fully released, with support and service terms.

Greyhound Research · Table 1 · Status ledger

One preview at Ignite, two items in development, and no prices

Microsoft’s Ignite 2026 confidential computing items, in both of Microsoft’s readings, beside the confidential computing products a buyer can sign today, November 19, 2026.

Greyhound Research’s judgment: sign only on the generally available rows, and plan the three Ignite items as roadmap until terms are published.

Announced at Ignite, in developmentIn previewGenerally availableContext
OfferingStatusWhat Microsoft has writtenNot yet written
VR72 multi-GPU confidential computingShowcase, in development“protect large models that extend beyond a single GPU”Summary: preview. Detailed post: a bare-metal showcase on one node with four GPUs, with VR72 confidential VMs in developmentPrice, regions, release date, verifier
Azure Confidential Enclaves and AKS Kata CoCo confidential containersIn development“a smaller trusted execution environment within a broader virtual machine”Summary: preview. Detailed post: planned for a future releasePrice, regions, release date, verifier
Azure Confidential Clean RoomLimited preview, restatedMultiparty analytics with each party’s raw data protectedPreview opened in June 2026. Still a limited preview on November 17Price, regions, release date, agent-written queries
Confidential VMs with NVIDIA H100 GPUsGenerally availableProtected space spans the VM and the attached GPUSince September 24, 2024, East US2 and West Europe at launchAgent memory inside the boundary
Confidential VMs on Azure LocalGenerally availableAMD SEV-SNP VMs on customer-controlled infrastructureIntel TDX VMs and confidential GPUs there in previewAgent memory inside the boundary
Intel TDX confidential worker nodes on AKSGenerally availableWith Azure Container LinuxIn preview with Ubuntu 26.04, per Microsoft, November 17, 2026Agent memory inside the boundary
AMD v6 confidential compute for PostgreSQLGenerally available“helps protect data in use”Not part of the Ignite confidential computing section
Confidential containers on AKSPreview in documentationAMD SEV-SNP, Azure Linux onlyDocumentation updated September 18, 2026How the Ignite preview differs from the documented one
Against this: Microsoft was the first cloud to offer confidential computing with NVIDIA H100 GPUs, so the Ignite items build on a product already in use.
Source: Microsoft Book of News and Microsoft Community Hub (Bhatia, November 17, 2026), Official Microsoft Blog (Phillips, October 20, 2026), Microsoft Community Hub (September 24, 2024), Microsoft Learn.

Table 1 lists Microsoft’s confidential computing items by status, and shows that of the three Ignite items only the clean room is in preview, beside four generally available products, with no price for any Ignite item. Microsoft’s summary calls all three previews. Its detailed post calls VR72 a bare-metal showcase on one node with four GPUs, with VR72 confidential virtual machines in development, and plans the enclaves and Kata containers for a future release. Azure Confidential Clean Room is a limited preview that restates a June 2026 preview. Generally available: confidential virtual machines with NVIDIA H100 GPUs since September 24, 2024, AMD SEV-SNP confidential virtual machines on Azure Local, Intel TDX confidential worker nodes on AKS with Azure Container Linux, and AMD v6 confidential compute for PostgreSQL. Context: confidential containers on AKS, a documented preview. Each row names what Microsoft has not yet written. Statuses are as Microsoft published them through November 19, 2026.

Table 1. One preview at Ignite, two items in development, and no prices. Data as of November 19, 2026. © Greyhound Research.

Only the generally available rows can be signed now. Confidential virtual machines with NVIDIA H100 GPUs have been generally available since September 2024, in two regions at launch, and Microsoft’s November post adds Intel TDX confidential worker nodes on AKS. Of the three Ignite items above them, the clean room restates a June preview and the other two are still in development. Budgets for regulated AI belong on the generally available rows until Microsoft publishes terms for the rest.

Greyhound Standpoint:

At Greyhound Research, we believe Ignite marks the point at which Microsoft set out how confidential computing moves from a niche control to an option at scale, and the question buyers put to it has moved too. Until now the question was whether the cloud operator could read the data. Once sealed capacity reaches the largest models, the question becomes whether the buyer can prove what ran on it, and that is a matter of evidence rather than encryption.

Microsoft’s summary calls all three items previews, while its detailed post calls VR72 a showcase, places the smaller sealed spaces in a future release, and leaves the clean room as the only preview. None carries a price, a region list, or a release date. Enterprise technology has a long record of features whose terms and even names changed before general availability, and buyers who planned on keynote slides have often had to plan again. That is not a criticism of Microsoft, whose detailed post is candid and whose confidential GPU machines have been generally available since 2024. It is a reason to separate direction from commitment. The view would change quickly if Microsoft opened VR72 and the enclaves with published terms.

Security and procurement leaders should treat Ignite as a roadmap. Pilots belong on the generally available confidential GPU machines this quarter, and any regulated production commitment should wait, at the next architecture review, for published prices, regions, and release dates.

The Deeper Read · About 4 minutes

Select a question to open the full answer

1What exactly did Microsoft announce?

In Microsoft’s summary. The Book of News places a section headed “Azure expands confidential computing for protecting highly sensitive data” in its Azure chapter. It says the goal is “protecting not only where data is stored and how it moves, but also how it is processed,” states that three offerings are now in preview, and names VR72 multi-GPU confidential computing, Azure Confidential Enclaves with AKS Kata CoCo confidential containers, and Azure Confidential Clean Room. The section links to a longer Microsoft post.

In Microsoft’s detailed post. Vikas Bhatia’s November 17 post on Microsoft’s Tech Community says the VR72 work starts with confidential bare-metal instances, shown at Ignite on one compute node with four Rubin GPUs. It says VR72 confidential virtual machines, AKS and Kata-based confidential containers, and Microsoft Foundry integration remain in development, and that the enclave and per-pod container experiences are planned for a future release. Azure Confidential Clean Rooms, it says, are in limited preview, and no prices are given.

What sits elsewhere on the page. The only confidential computing item called generally available in the Book of News is in its PostgreSQL section, where AMD v6 confidential compute helps protect data in use. Its confidential computing section does not mention agents, Foundry, the agent sandboxes, or Context Cache.

In the source’s own words

  • Microsoft’s Book of News, Azure confidential computingprotecting not only where data is stored and how it moves, but also how it is processed.Source
  • Microsoft’s Book of News, Azure confidential computingThree offerings are now in previewSource
  • Microsoft’s Vikas Bhatia, Tech Community, November 17A confidential bare-metal technology showcase on NVIDIA Vera Rubin NVL72, with a single-node, four-GPU AI demonstration.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17VR72 confidential VMs, AKS and Kata-based confidential containers, and Microsoft Foundry integration remain in development.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17Both capabilities are under active development and planned for a future release.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17Azure Confidential Clean Rooms, in limited preview, applies this thinking across organizations.Source
  • Microsoft’s Book of News, PostgreSQLAMD v6 confidential compute, now generally available, helps protect data in use.Source
2Why does this note follow the detailed post?

Two readings. Read alone, the Book of News supports three previews. Read beside the detailed post, it is a summary that groups a showcase, two items in development, and one limited preview under one word.

The weight given. This note follows the detailed post because it is the more specific of the two, it names what was demonstrated, and it agrees with Microsoft’s clean room documentation. Table 1 records both readings for each item.

What the gap between documents tells a buyer. When a summary and a technical post disagree, the technical post is usually the one a contract will later match. Buyers can take that as a working rule: plan on the summary, sign on the detail. Here the rule moves two of the three items from a pilot plan to a watch list, and leaves the clean room as the only one a buyer can try, on test-only terms.

3What can a buyer sign today?

Confidential GPUs since 2024. Microsoft announced general availability of confidential virtual machines with NVIDIA H100 GPUs on September 24, 2024, in East US2 and West Europe. Its current size documentation says the protected space spans the confidential virtual machine and the attached GPU. Confidential virtual machines on Azure Local, Microsoft’s software for customer-run sites, are generally available, according to an October 20 post by Douglas Phillips, President and Chief Technology Officer of Microsoft Specialized Clouds. Microsoft’s November 17 post adds that the generally available Azure Local machines use AMD SEV-SNP on supported AMD EPYC Genoa processors, with Intel TDX machines and confidential GPUs there in preview, and that Intel TDX confidential worker nodes on AKS are generally available with Azure Container Linux.

The clean room’s history. Microsoft announced the preview of multiparty analytics in Azure Confidential Clean Rooms in June 2026. Its October post called the service a limited preview, the product documentation, updated June 17, says the same, and the November 17 post still does. The test-only terms in that documentation have not changed since June 17.

In the source’s own words

  • Microsoft Learn, NCCadsH100v5 sizesIn this VM SKU, Trusted Execution Environment (TEE) spans confidential VM on the CPU and attached GPU, enabling secure offload of data, models, and computation to the GPU.Source
  • Microsoft Community Hub, September 24, 2024Azure NCC H100 v5 virtual machines are currently available in East US2 and West Europe regions.Source
  • The Official Microsoft Blog, Phillips on October 20Confidential virtual machines on Azure Local are generally available,Source
  • Microsoft Learn, Azure Confidential Clean RoomsMultiparty analytics with Azure Confidential Clean Rooms is currently in limited preview.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17now generally available with Azure Container Linux and in preview with Ubuntu 26.04.Source
4Which chip sits behind VR72?

Named in the detailed post. The Book of News says VR72 will protect large models that extend beyond a single GPU, without naming the chip. Microsoft’s detailed post names it as NVIDIA Vera Rubin NVL72. NVIDIA calls Vera Rubin NVL72 the first rack-scale platform to deliver NVIDIA Confidential Computing, and Microsoft says it delivered its first production Vera Rubin systems in August 2026. This note first gave the pairing as Greyhound Research’s inference, and it is now Microsoft’s own statement.

A company figure, not used. NVIDIA says its confidential computing runs at up to 98% of the performance of unsecured deployments, and adds elsewhere on the page that actual results depend on the workload, platform, and configuration. The figure is NVIDIA’s own and is not used in this note’s analysis.

In the source’s own words

  • Microsoft’s Book of News, Azure confidential computingopen a path to protect large models that extend beyond a single GPU.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17Our work with NVIDIA on Vera Rubin NVL72, or VR72, starts with confidential bare-metal instances.Source
  • NVIDIA Newsroom, January 5, 2026Vera Rubin NVL72 is the first rack-scale platform to deliver NVIDIA Confidential ComputingSource
  • Microsoft Azure Infrastructure Blog, November 12In August, we delivered the first production Vera Rubin systems in our datacenters.Source

↑ Back to the section guide

Part 1What Ignite Changed

The Seal And The Reader: What Confidential Computing Keeps Out

Greyhound SpotlightConfidential computing seals the envelope against the courier. An AI agent working inside the envelope is the reader.

Confidential computing protects data while a computer is using it. The chip keeps a sealed space, called a trusted execution environment, that even the cloud operator cannot look into. That answers the question risk committees ask most about public cloud: can the provider see our data? It does not answer a newer question. What does our own agent do with the data it is allowed to see?

This changes what the old cloud objection is about. For more than a decade, boards asked of public cloud whether the provider could look, and contracts and audits grew up around that question. Confidential computing gives it a technical answer, which is real progress. It also moves the remaining risk somewhere less familiar. Once the provider is shut out, the main party left with access is software the buyer chose and switched on. Greyhound Research expects many risk registers to mark the provider risk as reduced and leave the agent risk unowned. That reallocation is easy to miss, because a control that answers one question well tends to be credited with answering the next one too.

Exhibit 1 sets the parties who might read sensitive data against the three Ignite items. A tenant is another customer sharing the same hardware. The seal keeps out everyone except the parties the customer lets in, and the agent is let in by design.

Greyhound Research · Exhibit 1 · Who the seal keeps out

The seal keeps everyone out except the reader

Parties who might read sensitive data, against the three confidential computing items Microsoft set out at Ignite 2026.

The seal settles who else can look. What the agent may do still needs the buyer’s own controls.

Kept out by the sealInside by designNot applicableVR VR72 · EN Enclaves and Kata containers · CR Clean room
Who might read the data
VR72 multi-GPU
Enclaves and Kata containers
Clean room
1The cloud operator and its administrators
VRKept out
ENKept out
CRKept out
2Other tenants on the same hardware
VRKept out
ENKept out
CRKept out
3Other parties to the analysis
VRNot applicable
ENNot applicable
CRKept out
4Software outside the smaller sealed space
VRNot applicable
ENKept out
CRNot applicable
5The customer’s own agent
VRInside by design
ENInside by design
CRInside by design
Against this: Inside the seal, the customer can still pair the agent with identity, narrow rights, and approval, which the seal does not replace.
Source: Greyhound Research analytical model, not measured data. Greyhound Research reading of Microsoft’s Book of News and Microsoft Learn.

Exhibit 1 is a matrix of five parties who might read sensitive data against three Microsoft confidential computing items from Ignite 2026, and shows that the seal keeps every party out except the customer’s own agent, which is inside by design. The cloud operator and its administrators, and other tenants on the same hardware, are kept out by all three. Other parties to an analysis are kept out by the clean room. Software outside the smaller sealed space is kept out by the enclaves and Kata containers. The customer’s own agent is inside by design in all three, shown as a tinted row. The marks are Greyhound Research analysis, not measured data.

Exhibit 1. The seal keeps everyone out except the reader. Current through November 19, 2026. © Greyhound Research.

Every offering keeps the operator out, and the clean room also keeps the other collaborators out. None of them keeps the customer’s own agent out, because the agent is the workload being protected. The exhibit is Greyhound Research’s reading, not a vendor’s claim. It shows that the seal settles one question: who else can look. What the agent may do stays with the controls the buyer already owns.

Greyhound Standpoint:

Greyhound Research’s view is that confidential computing answers the provider question well and leaves the agent question where it was. The seal decides who else can look. It does not decide what the workload inside may do, and an agent is the workload. Whatever the agent is permitted to read, it reads in the clear, and whatever it is permitted to do, it can do from inside a fully sealed machine.

Practitioners have seen this shape before. When server virtualization arrived, isolation between machines held up well, and the problems that followed came largely from what administrators and applications were allowed to do inside them. Isolation and authorization are different controls, and each platform shift tempts buyers to let the newer one stand in for the older one. Microsoft deserves credit for showing how to take the seal to the largest models, and its own post says the seal does not replace permissions. The limit is one of design, not of quality, because the technology does what it was built to do. The view would change if Microsoft tied agent identity and permissions to the sealed boundary in its documentation.

Security leaders should pair every sealed workload with the agent controls it does not replace. Each agent needs its own identity, a written list of what it may read and call, and human approval before it acts on regulated data. Those terms belong in the pilot gate, not after it.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1How does the seal work, and what does each offering change?

The mechanism. A trusted execution environment encrypts data in memory while it is being processed, and the hardware signs a statement, called attestation, that genuine chips are running the expected code. Microsoft’s confidential containers on AKS, for example, use AMD SEV-SNP, a processor feature that encrypts a virtual machine’s memory, and shield workloads from untrusted parties, including customer and tenant administrators.

Three directions. VR72 is built to stretch the boundary across a rack of GPUs, though the Ignite showcase used one server with four. Confidential Enclaves and Kata containers shrink it to a smaller box inside a virtual machine, so less has to be trusted. The clean room shares one sealed space between several parties, and Microsoft says it protects each participant’s raw data from the other collaborators and from cloud operators.

In the source’s own words

  • Microsoft’s Book of News, Azure confidential computinga smaller trusted execution environment within a broader virtual machine.Source
  • Microsoft’s Book of News, Azure confidential computingThis will help businesses narrow what must be trusted and verifiedSource
  • Microsoft Learn, confidential containers on AKSSecurity designs that assume least privilege and highest isolation shielding from all untrusted parties, including customer and tenant admins.Source
2Why is the agent inside by design?

The workload is the reader. An agent that answers a question must read the data the question concerns. Sealing the machine keeps others from watching it read, and it leaves the agent’s choices untouched. Greyhound Research made the same point about agents in its draft note on Azure agent infrastructure: once an agent has a machine of its own, what it may read, change, and spend decides what it can do.

Microsoft agrees. Microsoft’s detailed post says that confidential computing addresses protection during processing and does not replace application permissions or the careful handling of logs and tool calls.

Identity sets the rights. Greyhound Research called Entra Agent ID a policy control plane for agent-based computing in May 2025, and that identity layer, not the seal, is where an agent’s rights are set.

What Microsoft’s own caveat tells a buyer. A vendor that states the limit of its product in its launch post gives buyers a usable sentence. The line that confidential computing does not replace application permissions can go straight into an internal design standard, so that no sealed workload is approved without its agent controls. It also signals that Microsoft expects the identity layer, not the seal, to carry an agent’s rights.

In the source’s own words

  • Microsoft’s Vikas Bhatia, Tech Community, November 17Confidential computing addresses protection during processing.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17Neither replaces application permissions, careful handling of logs and tool callsSource

↑ Back to the section guide

Part 2The Proof

Attestation: The Proof Matters More Than The Lock

Greyhound SpotlightAttestation proves what was running at one moment. An agent whose model, tools, or prompts change next week is no longer the thing that was proved.

Every large cloud now encrypts data while it is in use, so the lock no longer sets one provider apart. The proof does. Before a bank lets sensitive data or keys into a sealed space, something has to show that the space is genuine and running the right code. That proof is attestation, and auditors will lean on it.

What changes for buyers is the kind of evidence they hold. Encryption is a setting that is on or off, and an auditor can confirm it once. Attestation is a record, produced each time a check runs, and its worth depends on being kept, dated, and matched to the workload it describes. That turns confidential computing from a security purchase into an evidence discipline, closer to how banks record who approved each payment. Greyhound Research expects supervisors in regulated sectors to ask, in time, for the proof itself rather than for a statement that sealing was switched on. The cost few budgets show is the work of keeping and reading these records.

Exhibit 2 follows the proof through five steps, from the chip measuring what it runs to the proof being refreshed. Secure key release means the keys to the data are handed over only after the proof checks out.

Greyhound Research · Exhibit 2 · The path of the proof

Microsoft has written four steps and left out the fifth

Five steps that turn a sealed space into evidence a buyer can rely on, read against Microsoft’s documentation, November 19, 2026.

A proof taken once cannot vouch for an agent that changes. The refresh rule is the step to settle first.

Written by MicrosoftPartly writtenNot written
1Measure WrittenThe chip records what code it loadsWritten. AKS confidential containers reflect container layers and runtime settings in the hardware report.Open. Which measurements the Ignite items take.
2Sign WrittenThe chip signs that record with its own keyWritten. Microsoft signs the guest operating system so customers can verify it.Open. Lab attacks in 2025 extracted signing keys with physical access.
3Verify Partly writtenA checker confirms the signed recordWritten. Azure Attestation checks SGX, SEV-SNP, and TPM platforms. The clean room lets every party verify.Open. The VR72 demonstration used a local verifier. No verifier is named for the enclaves.
4Release keys WrittenKeys to the data are handed over only after the checkWritten. Workload attestation adds assurance before data or keys are released.Open. How release works across a rack of GPUs.
5Refresh Not writtenThe check is repeated when the agent changesWritten. None in the documents read.Open. Whether a model or tool update forces a new check.
Against this: Microsoft’s attestation service is mature and already used for secure key release, so the first four steps rest on a working foundation.
Source: Greyhound Research analysis of Microsoft Learn (Azure Attestation, confidential containers on AKS, Azure Confidential Clean Rooms), the Official Microsoft Blog (Phillips, October 20, 2026), and Microsoft Community Hub (Bhatia, November 17, 2026). The steps and grades are Greyhound Research’s.

Exhibit 2 sets out five steps of attestation as a numbered list, and shows that Microsoft has written the first four for its existing services and not the fifth. Step one, measure, is written for AKS confidential containers. Step two, sign, is written, with the caveat that lab attacks in 2025 extracted signing keys. Step three, verify, is partly written: Azure Attestation checks SGX, SEV-SNP, and TPM platforms and the clean room lets every party verify, while the VR72 demonstration used a local verifier and no verifier is named for the enclaves. Step four, release keys, is written. Step five, refresh when the agent changes, is not written in any document read. The steps and grades are Greyhound Research analysis.

Exhibit 2. Microsoft has written four steps and left out the fifth. Current through November 19, 2026. © Greyhound Research.

Microsoft has written the first steps for its existing services. Its attestation service checks the proof, and its documents say keys can be released only after the check. The last step is the one an agent needs most. Agents change as models, tools, and prompts are updated. No Microsoft document read says how often the proof is refreshed, or whether an update forces a new check.

Greyhound Standpoint:

Greyhound Research holds that attestation, not encryption, is now the substance of confidential computing. All three large clouds encrypt data while it is in use, so encryption no longer separates one offer from another. What separates them is the proof: whether it covers the code that matters, who checks it, and whether it is renewed when the workload changes. A proof that cannot answer those three questions gives an auditor a signature without the context to rely on it.

The refresh question is where agents differ from earlier workloads. Security engineers have long known a class of flaw called time-of-check to time-of-use, in which a check passes and the thing checked then changes before it is used. An agent whose model, tools, or instructions are updated after start-up reproduces that flaw at the scale of a deployment. Microsoft has done careful work on the first steps, with an attestation service and documented release of keys only after a check. The documents read do not address refresh, which is not proof that it is missing. A published refresh policy would settle the point.

Chief information security officers should ask for the verifier and the refresh rule in writing before any pilot. A model or tool update should trigger a fresh check before an agent touches regulated data, and the proof should be retained as evidence for as long as the data it protected.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1What does Microsoft’s attestation service do?

A unified checker. Microsoft describes Azure Attestation as a unified solution for remotely verifying a platform and the code running inside it. Its documentation lists Intel SGX enclaves, virtualization-based security enclaves, trusted platform modules, trusted launch, and confidential virtual machines on AMD SEV-SNP. The page, last updated June 29, 2026, does not list Intel TDX or GPUs, and it gives no price or region list.

Keys after the proof. Microsoft’s October sovereignty post says that workload attestation adds assurance before sensitive data or keys are released. AKS confidential containers reflect the container layers and runtime settings in the hardware report, and Microsoft signs the guest operating system so customers can verify it.

In the source’s own words

  • Microsoft Learn, Azure Attestation overviewMicrosoft Azure Attestation is a unified solution for remotely verifying the trustworthiness of a platform and integrity of the binaries running inside it.Source
  • The Official Microsoft Blog, Phillips on October 20Workload attestation adds assurance before sensitive data or keys are released.Source
  • Microsoft Learn, confidential containers on AKSMicrosoft signs the guest OS and container runtime environment so you can verify them through attestation.Source
2Why does refresh matter for agents?

One moment, not a month. Attestation measures what is loaded when a sealed space starts or when a check is requested. An agent’s behavior depends on its model, its tools, and its instructions, and all three change over a deployment’s life. Microsoft’s Ignite documents do not say whether attestation for the new items covers those changes, or how often it is repeated. Greyhound Research made a related point about agent memory in August 2025, in Transforming AI Economics with Procedural Memory, which held that memory without layered governance undermines trust faster than it delivers savings.

Many layers, one proof. Greyhound Research has described confidential AI as a multilayer architecture that spans hardware, cloud, accelerator, attestation, and model serving. A proof taken of one layer at one moment cannot stand for all of them over a deployment’s life.

What the silence on refresh tells a buyer. The attestation documents describe what is checked, not how often. For a workload fixed at start-up, that is enough. For an agent, it means the buyer’s change process, not the attestation service, decides whether a proof is still current. That makes the change log as important as the proof itself, because it is what shows an auditor which version of the agent the proof described.

In the source’s own words

  • Greyhound Research, procedural memory notewithout layered governance controls, memory undermines trust faster than it delivers savings.Source
  • Greyhound Research, published commentary, 2026Confidential AI is a multilayer architecture that spans hardware, cloud, accelerator, attestation, and model serving,Source

↑ Back to the section guide

Part 2The Proof

Who Checks The Proof: Independent Verifiers And Microsoft’s Open Question

Greyhound SpotlightSince June 2026, Google customers can have the proof checked by Intel’s service rather than Google’s. Microsoft’s VR72 demonstration at Ignite checked the proof with a local verifier.

A proof is only as good as the party that inspects it. If the cloud provider both runs the sealed space and checks the proof, the customer is still trusting the provider. An independent checker splits those roles. That matters to regulators, because it lets a bank show that no single company vouched for its own seal.

Independent checking is an old idea in security, but it is new to this layer. Buyers already accept outside audits of data centers and cloud controls, and attestation is where that habit has been slowest to arrive. Google’s move shows that the separation can be offered as a product, which makes it a fair question to put to any provider. Greyhound Research expects Microsoft to describe its verifier choices as VR72 and the enclaves approach release, and buyers can shape that answer by asking now. One consequence is commercial. A verifier the buyer picks stays with the buyer, which ties the record of proof less to one cloud and can lower the cost of a later move.

Table 2 sets out who checks the proof on each platform read for this note, and whether that checker belongs to the cloud provider. A verifier is the service that reads the signed proof and confirms it.

Greyhound Research · Table 2 · Verifier ledger

Only Google lets an outside party check the proof

Who verifies the attestation on each platform read for this note, and whether that checker belongs to the provider, November 19, 2026.

Only Google offers a checker it does not own today. Make a named verifier a condition of any regulated pilot.

A checker the provider does not ownThe provider’s own checkerNot stated
PlatformWho checks the proofWhat is writtenIndependent of the provider
Microsoft, VR72 and the enclavesNot yet namedThe VR72 demonstration used a local verifier, and remote attestation is called an architectural direction. No verifier named for the enclavesUnknown
Microsoft, clean roomEach participantEvery party can verify the code and hardware independentlyYes, by design
Microsoft, existing servicesAzure AttestationSGX, VBS, SEV-SNP confidential VMs, TPM, trusted launchNo
Google, Confidential SpaceIntel Trust Authority or Google Cloud AttestationIntel Trust Authority generally available since June 23, 2026Yes, Intel
AWS, Nitro Enclaves and SandboxesNitro Hypervisor and Nitro Secure ModuleKeys released only to code that matches its measurementsNo
NVIDIA, GPUsNVIDIA Remote Attestation ServiceChecks GPUs against trusted reference values and issues signed tokensNo, NVIDIA’s own
Against this: The clean room already lets each party check the proof for itself, the strongest verification statement among Microsoft’s three Ignite items.
Source: Microsoft Learn (Azure Attestation, Azure Confidential Clean Rooms), Microsoft Community Hub (Bhatia, November 17, 2026), Google Cloud Blog (June 23, 2026) and Documentation, AWS Documentation and Compute Blog (November 10, 2026), NVIDIA Documentation.

Table 2 lists who checks the proof on six platforms, and shows that only Google offers a verifier it does not own, Intel Trust Authority, generally available since June 23, 2026. For VR72, Microsoft’s Ignite demonstration used a local verifier, and Microsoft calls remote attestation an architectural direction. No verifier is named for the enclaves. Microsoft’s clean room lets each participant verify independently. Microsoft’s existing services use Azure Attestation. AWS uses the Nitro Hypervisor and Nitro Secure Module, and releases keys only to matching code. NVIDIA runs its own Remote Attestation Service for GPUs. The independence column is Greyhound Research’s reading of each provider’s documents.

Table 2. Only Google lets an outside party check the proof. Data as of November 19, 2026. © Greyhound Research.

Google is the only provider in the table that offers a verifier it does not own, through Intel Trust Authority. AWS, NVIDIA, and Microsoft’s existing services run their own checkers, while Microsoft’s clean room lets each participant check. For VR72, Microsoft says the Ignite demonstration used a local verifier and calls remote attestation an architectural direction. No verifier is named for the enclaves. A buyer cannot yet tell whether VR72 will use Microsoft’s service, NVIDIA’s, or one the customer chooses.

Greyhound Standpoint:

Greyhound Research’s reading is that an independent verifier is becoming the feature regulated buyers ask about first, and that Google has set the current bar. When the same company runs the sealed space and checks the proof of it, the customer’s assurance still rests on that company’s word. A verifier the provider does not own separates the two roles, which is what allows a bank to show a supervisor that no single party vouched for its own seal.

The pattern is an old one. Financial audit rests on the auditor being independent of the audited, and the hardware security modules banks have relied on for decades earn trust through validation against a published standard by laboratories the vendor does not control. Microsoft has a mature attestation service, and its clean room already lets each participant verify the proof independently. Its post is candid that the VR72 demonstration used a local verifier, and no verifier is named for the enclaves. Neither is yet a product, so this reflects timing more than intent, and buyers can still shape it. A named, customer-selectable verifier for each item would close the gap.

Before any bank or insurer signs, procurement and the CISO should obtain Microsoft’s answer in writing, naming the verifier for each item, whether the customer may choose its own, and where the proof is kept and for how long. That answer belongs in the due diligence file at contract review.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1How does each verifier work?

Google and Intel. On June 23, 2026, Google announced that Intel Trust Authority is generally available as an independent attestation verifier for Confidential Space. Its documentation, updated November 10, lists Google Cloud Attestation, which runs in the workload’s region, and Intel Trust Authority, which runs in a region the customer defines.

AWS. The Nitro Hypervisor generates a signed attestation document for each Nitro Enclave, and AWS’s key service can release keys only to enclaves whose measurements match. AWS says its new Nitro Sandboxes use the Nitro Secure Module for an attestation document in the same style.

NVIDIA. NVIDIA’s Remote Attestation Service verifies its GPUs and platforms against trusted reference values and issues signed tokens. The service is NVIDIA’s, and its page says NVIDIA will collect the IP addresses of callers.

In the source’s own words

  • Google Cloud Blog, June 23, 2026We are pleased to announce that Intel Trust Authority (ITA) is now generally available as an independent attestation verifier service for Confidential Space.Source
  • Google Cloud Blog, June 23, 2026By decoupling attestation verification from the cloud service provider, customers benefit from enhanced transparency, stronger assurance, and a more robust trust model.Source
  • AWS Documentation, Nitro Enclaves attestationAn enclave's attestation document is generated by the Nitro Hypervisor.Source
  • AWS Compute Blog, November 10, 2026AWS KMS releases your secrets only to workloads that prove they run the code you approved.Source
  • NVIDIA documentation, Remote Attestation Serviceis a cloud-based service that verifies the integrity and authenticity of NVIDIA GPUs and platforms.Source
2Where does Microsoft’s clean room stand on verification?

Each participant can check. The clean room’s documentation says cryptographic remote attestation at each step lets every participant independently verify that the clean room runs known code on genuine confidential hardware. That is the strongest verification statement among the three Ignite items.

VR72’s demonstration. Microsoft’s detailed post says the Ignite system collected CPU and GPU evidence and used a minimal local verifier to check signatures, token binding, and expected GPU measurements before inference was shown. It calls remote attestation an architectural direction. In Greyhound Research’s reading, a local verifier demonstrates the method, not the independent check a supervisor would expect. For the enclaves, no Microsoft document read names a verifier.

What “architectural direction” tells a buyer. The phrase is candid. It says remote attestation is intended for VR72 rather than built. Read beside the clean room’s documentation, it shows that Microsoft already knows how to let each party check the proof, so the open question for VR72 and the enclaves is a product decision more than a technical one. That is the kind of decision customer requests can still influence.

In the source’s own words

  • Microsoft Learn, Azure Confidential Clean RoomsCryptographic remote attestation at each step forms the cornerstone of the service, letting every participant independently verify that the clean room is running known and attested code on genuine confidential hardware.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17uses a minimal local verifier to check signatures, token binding and expected GPU measurementsSource
  • Microsoft’s Vikas Bhatia, Tech Community, November 17Remote attestation is an architectural direction, while the Ignite demonstration uses a local verifier.Source

↑ Back to the section guide

Part 2The Proof

When The Seal Was Opened: Three Lab Attacks In One Year

Greyhound SpotlightResearchers forged the proof that a sealed space is genuine with under USD 1,000 of equipment. They needed physical access to the server.

Three research teams opened the seal in 2025. Each placed a small device between the processor and its memory, and read or redirected what passed between them. One of them, TEE.Fail, extracted the keys that sign the proof and used them to forge it. None of the attacks has been reported in the wild, and each needed someone at the machine.

The attacks change the reading of the threat, not the case for the technology. Sealed computing was designed to keep out remote attackers and privileged insiders, and the vendors’ reply that physical attacks fall outside that design matches how they have long scoped it. What is new is how cheap the equipment has become, which moves physical access from a concern about well-funded attackers to one about anyone who can reach the server. Greyhound Research expects buyers to give more weight to where hardware sits, and to ask for site-level evidence alongside the proof. A quieter effect falls on sovereign and edge sites, where workloads moved for control can end up with weaker physical custody than a hyperscale hall.

Exhibit 3 places the three attacks on a 2025 calendar, from the first private warning to vendors to public release. An interposer is a device fitted between a chip and its memory to watch the traffic.

Greyhound Research · Exhibit 3 · The 2025 attack calendar

Three lab attacks reached the seal in one year

Private disclosures to vendors and public releases of three memory-interposer attacks on trusted execution environments, 2025.

The lab attacks are a reason to verify more carefully, not to step back. Weigh where the hardware sits as heavily as the chip.

A disclosure or publication fell in this monthPrivate means told to vendors before release
JanQuiet
FebPrivateBattering RAM told to Intel and AMD
MarQuiet
AprQuiet
MayQuiet
JunQuiet
JulQuiet
AugQuiet
SepPublicBattering RAM embargo ends September 30
OctPublicWireTap public October 2. TEE.Fail public October 28
NovQuiet
DecQuiet
Battering RAM: interposer under USD 50WireTap and TEE.Fail: under USD 1,000All three: physical access to the server
Against this: None of the attacks has been reported in the wild, and hyperscale data centers have physical controls that make them far harder.
Source: Battering RAM, WireTap, and TEE.fail researchers’ sites, and press coverage of WireTap (October 2, 2025).

Exhibit 3 places three 2025 attacks on trusted execution environments on a month grid, and shows three months with a private disclosure or public release. In February, Battering RAM was disclosed to Intel and AMD. In September the Battering RAM embargo ended on September 30. In October, WireTap became public on October 2 and TEE.Fail on October 28, the date of the Intel guidance its researchers link. Notes beneath say Battering RAM’s interposer cost under USD 50, WireTap and TEE.Fail under USD 1,000, and that all three needed physical access to the server.

Exhibit 3. Three lab attacks reached the seal in one year. Data as of November 19, 2026. © Greyhound Research.

Battering RAM reached Intel and AMD in February, months before it went public, and both vendors answered that physical attacks fall outside what their products protect against. That answer is fair for a guarded data center and less reassuring for a colocation site, an edge cabinet, or a supplier’s rack. The exhibit records dates, not damage. Its lesson is that the proof needs a checker the buyer trusts, because the proof itself can be forged.

Greyhound Standpoint:

In Greyhound Research’s assessment, the 2025 attacks are a lesson about verification and physical custody, not a reason to step back from confidential computing. The mechanism explains why. Each attack placed a device between processor and memory, and the most serious extracted the keys that sign the proof. A forged proof does more harm than a broken seal, because it tells every check further along that all is well.

Hardware security has followed this arc before. Physical attacks on memory and chips have a long research history, and chip vendors have long drawn their threat models to exclude them, leaving physical protection to whoever runs the building. That is a reasonable division of labor in a guarded hyperscale data center, where access is controlled and logged, and none of these attacks has been reported in the wild. It is less reassuring for a colocation hall, an edge cabinet, or a partner’s rack, where the buyer often knows less about who can reach the hardware. The evidence is laboratory evidence and should be weighed as such. The view would ease if future chips protected the memory bus, and would firm if an exploitation outside the lab were reported.

Risk teams should add three questions to the next architecture review: where each sealed workload physically runs, who can reach that hardware, and how quickly the chosen verifier would stop accepting a signing key known to be compromised.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1What exactly did each attack do?

TEE.Fail. The researchers from Georgia Tech, Purdue, and van Schaik, LLC built a DDR5 memory interposer for under USD 1,000, extracted attestation keys from Intel TDX and AMD SEV-SNP systems, which also undermines Intel SGX attestation, and showed how those keys can compromise NVIDIA’s GPU confidential computing. Their site links the security guidance Intel published on October 28, 2025, records AMD’s view that interposer attacks are out of scope for SEV-SNP, and says the method has not, as far as the team knows, been exploited in the wild.

WireTap. A team from Purdue and Georgia Tech extracted an SGX attestation key from DDR4 servers with a setup costing under USD 1,000, for the ACM CCS 2025 conference. The attack became public on October 2, 2025.

Battering RAM. Researchers from KU Leuven, the University of Birmingham, and Durham University built an interposer for less than USD 50 that redirects protected memory on Intel SGX and AMD SEV-SNP. Vendors were told in February 2025 under an embargo to September 30, 2025, and Intel and AMD said that physical attacks are out of scope.

In the source’s own words

  • TEE.fail researchers’ siteit should cost you under $1000 to build a setup similar to ours.Source
  • TEE.fail researchers’ siteAs far as we know, this has not been exploited in the wild.Source
  • WireTap researchers’ sitewe are able to extract an SGX secret attestation key from a machine in fully trusted statusSource
  • Battering RAM researchers’ siterequires only brief one-time physical access, which is realistic in cloud environments,Source
2How did the vendors respond?

Out of scope. The TEE.Fail researchers report that AMD considers interposer attacks out of scope for SEV-SNP and does not plan firmware updates, and that Intel’s only mitigation is physical security. Intel treats WireTap as outside its threat model. Both vendors acknowledged Battering RAM and said physical attacks on memory are out of scope for their current products.

What “out of scope” tells a buyer. The vendors’ answer is accurate and narrow. It places physical protection with whoever runs the building, so the buyer’s assurance rests partly on the cloud provider’s site controls. That moves the question from the chip vendor to Microsoft, and from the product sheet to the data center audit reports a buyer can request.

In the source’s own words

  • WireTap researchers’ siteWireTap is considered by Intel to be outside the threat model, as SGX offers no protections against physical attacks.Source
  • TEE.fail researchers’ siteAMD considers interposer attacks to be out of scope for SEV-SNP.Source

↑ Back to the section guide

Part 3Agents Inside The Seal

Agent Memory: The Notes Left On The Desk

Greyhound SpotlightMicrosoft has not said whether the agent memory and cached context its own services keep sit inside the new seal. If they sit outside, the seal protects the arithmetic and leaves the notes on the desk.

The most sensitive thing in an agent system is often what the agent remembers. A request travels well beyond the model. It picks up documents from search, keeps working notes, saves context for reuse, calls tools, and writes answers. Each of those steps can hold customer data. Microsoft’s summary ties confidential computing to large models and smaller sealed spaces, not to agents. Its detailed post goes further. It names retrieved context and agent memory among what databases hold, and lists Foundry integration as still in development.

For a data protection officer, this changes the shape of the data map. Earlier AI pilots were often a model and a data source, and privacy reviews were built for that pair. Agents add stores that the service creates, and some persist after the task ends. Microsoft’s detailed post names agent memory among what databases hold, which suggests it sees the issue, and Greyhound Research expects the boundary for managed agent services to be documented in a later release. Until then, the gap lands on deletion requests. A company asked to erase personal data needs an answer that covers cached context and working memory, and today that answer cannot be read from Microsoft’s documents.

Exhibit 4 follows one agent request along its path and marks where Microsoft’s documents place the seal. Cached context is text the model has already processed and keeps for reuse. Retrieval means fetching documents to answer a question.

Greyhound Research · Exhibit 4 · Where the seal sits on an agent request

Only the model’s arithmetic is clearly sealed

One agent request from prompt to answer, marked where Microsoft’s documents place the confidential boundary, November 19, 2026.

Only the arithmetic is clearly sealed. Keep regulated data out of every step the documents do not cover.

Inside the seal, per MicrosoftMemory and cache, not statedOther steps, not stated
Prompt and retrievalNot statedInstructions and documents fetched to answer the request.
Model computationInside the sealData, models, and computation on the confidential GPU, per Microsoft.
Memory and cacheNot stated for Microsoft’s servicesContext Cache keeps processed text for 1 to 30 days after last use. A confidential database can hold memory the customer runs.
Tools and outputNot statedCalls to other systems, answers, and logs.
Against this: Not stated is not the same as outside. Microsoft may document a wider boundary when VR72 and the enclaves reach general availability.
Source: Greyhound Research analysis of Microsoft’s Book of News, Microsoft Community Hub (Bhatia, November 17, 2026), and Microsoft Learn (NCCadsH100v5 and the Azure CLI reference), with Context Cache as read in Greyhound Research’s draft note on Azure agent infrastructure.

Exhibit 4 follows one agent request along a bar from request to answer, in six steps: prompt and instructions, retrieved documents, model computation, working memory and cached context, tool calls, and output and logs. A marker labeled the seal sits over model computation, the only step Microsoft’s documents place inside the confidential boundary. Every other step is marked not stated. A strip beneath groups the steps into four and notes that Context Cache keeps processed text for 1 to 30 days after last use, and that a confidential database can hold memory the customer runs. Memory held by Microsoft’s own agent services is not stated. Not stated is not the same as outside. The mapping is Greyhound Research analysis.

Exhibit 4. Only the model’s arithmetic is clearly sealed. Current through November 19, 2026. © Greyhound Research.

Only the model’s computation is clearly inside the seal in Microsoft’s documents. Every other step is not stated, which is not the same as outside. A customer can keep agent memory in a database it runs on a confidential virtual machine. Microsoft’s own Context Cache, covered in Greyhound Research’s draft note on Azure agent infrastructure, keeps processed prompt text in Azure storage for up to 30 days after last use. No document read places it inside a confidential boundary.

Greyhound Standpoint:

Greyhound Research’s position is that agent memory, cached context, and retrieved documents are the real test of confidential computing for agents. The model’s arithmetic is brief. The material around it lasts. Retrieved documents, working notes, cached prompts, and tool outputs can rest in storage for days, and Microsoft’s Context Cache keeps processed text for up to 30 days after last use. If those places sit outside the seal, the longest-lived copies of sensitive data are the least protected.

Each generation of enterprise systems has learned that sensitive data escapes through the places nobody mapped: temporary files, logs, backups, and caches. The main store tends to be protected first and the copies around it later. Agents add new places of the same kind. Microsoft places computation inside the seal by name and points to confidential databases for memory customers run, but is silent on its own agent stores. Not stated is not the same as outside, and Microsoft may well document a wider boundary. Until it does, the prudent assumption is that these stores are not sealed. Documentation placing agent memory and cached context inside the boundary would change this view more than any other single disclosure.

Data protection officers should map every place an agent’s data rests during a task, mark which places are sealed, and keep regulated data out of the rest. Microsoft should be asked to document the boundary before any regulated agent pilot passes its production gate.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1Which parts of an agent’s work does Microsoft place inside the seal?

Computation, by name. The Book of News ties VR72 to protecting large models that extend beyond a single GPU, and the confidential GPU documentation says the protected space covers data, models, and computation offloaded to the GPU. Agents, Foundry, the agent sandboxes, and Context Cache go unmentioned in the Book of News section, and the detailed post lists Foundry integration as still in development.

Memory the customer runs. Microsoft’s detailed post says that databases hold the business records, retrieved context, and agent memory that make AI systems useful, and offers SQL Server on confidential virtual machines, with the customer in control of the database. That can seal memory the customer runs itself. It does not speak to memory held by Microsoft’s managed agent services.

Memory, not stated. Microsoft’s Azure CLI reference sets a time to live of 1 to 30 days for a Context Cache container and deletes content not accessed within it. Greyhound Research’s draft note on Azure agent infrastructure reads Context Cache as Azure Storage resources linked to a model deployment, and found no document placing it inside a confidential boundary. Greyhound Research suggests a separate encryption key for each agent’s memory, so that destroying the key erases what the agent remembered.

What “not stated” tells a buyer. Microsoft’s documents seal the computation by name and offer a sealed database for memory the customer runs. Together they point to a workable design today: keep agent memory in a store the customer controls on confidential machines, and leave managed caches out of regulated work until their boundary is documented. That costs some convenience, and it keeps the data map accurate.

In the source’s own words

  • Microsoft Learn, NCCadsH100v5 sizesenabling secure offload of data, models, and computation to the GPU.Source
  • Microsoft’s Vikas Bhatia, Tech Community, November 17Databases hold the business records, retrieved context and agent memory that make those systems useful.Source
  • Microsoft Learn, Azure CLI reference for Context Cache containersBlobs in the container that have not been accessed within this number of days will be automatically deleted.Source
2What does Greyhound Research’s earlier work say?

Privacy travels with the workload. Greyhound Research wrote in May 2025 that confidential containers let enterprises process sensitive data without exposing it, even during runtime, and reported from its Greyhound CIO Pulse 2025 research that only 28% of enterprises then used confidential computing. Its April 2025 note on TCS SovereignSecure Cloud held that sovereign cloud offers compliance flexibility rather than replacing hyperscalers.

The custody test. Greyhound Research said in December 2025 that when a cloud provider has any ability to decrypt customer data, the data is no longer truly sovereign. That test has not changed. Confidential computing answers more of it while data is being processed, and agent memory kept outside the seal is where it can still fall short.

In the source’s own words

  • Greyhound Research, IBM Think 2025 noteonly 28% of enterprises currently use confidential computingSource
  • Greyhound Research, published commentary, December 1, 2025the data is no longer truly sovereign.Source

↑ Back to the section guide

Part 3Agents Inside The Seal

The Clean Room: Leaks By Answer, Not By Breach

Greyhound SpotlightMicrosoft tells clean room users not to put personal or regulated data into the preview. The service is for testing only.

A clean room lets several companies pool data that none of them can see. Banks might check for fraud across accounts, or hospitals pool trial results. The seal protects the raw records. Answers still come out. If the agreed questions are loose, or an agent is allowed to ask open ones, the answers can reveal what the seal was meant to hide.

Clean rooms answer a long-standing commercial need. Banks, insurers, and hospitals have wanted to pool data for fraud, risk, and research for years, and legal teams have often slowed the effort over who could see what. Sealed hardware removes much of that objection, which is why Greyhound Research expects interest to rise once the service leaves preview. The harder work then moves to the agreement. Data-sharing terms were mostly written for people running known reports, and few anticipate software that composes its own questions. The effect that gets less attention is on liability. When an answer reveals too much, the parties will need to know whose approval let it out, and the agreement is where that is settled.

Exhibit 5 sets out the clean room’s controls in three groups: what Microsoft has written, what the parties must decide, and what is not yet stated. A post-filter drops results drawn from too few records to be safe.

Greyhound Research · Exhibit 5 · The clean room’s controls

Strong controls for analysts, none stated for agents

Azure Confidential Clean Rooms controls as documented, the terms parties must write, and what is not stated, November 19, 2026.

The walls are sound. The risk is in the answers, so write the allowed questions before any data goes in.

Written by MicrosoftControls in the documentationApproval. Every party whose data a query touches must approve it.Column lists. Each party lists the columns queries may use.Minimum counts. Results from groups that are too small are dropped.Written
For the parties to decideTerms for the agreementAllowed questions. Which queries the room may run at all.Group sizes. The minimum count each party will accept.Retention. What results may be kept, and by whom.The buyer’s to write
Not stated by MicrosoftGaps for agent-written queriesQuery volume. No limit on how many approved queries can run.Query patterns. No check on answers that combine to reveal more.Agents. No rule for queries an agent drafts.Not stated
Against this: Microsoft’s documentation tells customers to keep personal and regulated data out of the preview, which limits exposure while the gaps remain.
Source: Greyhound Research analysis of Microsoft Learn (Azure Confidential Clean Rooms, updated June 17, 2026). The grouping is Greyhound Research’s.

Exhibit 5 groups the clean room’s controls into three cards. Written by Microsoft: approval by every party whose data a query touches, column lists that limit what queries may use, and minimum counts that drop results from small groups. For the parties to decide: the allowed questions, the minimum group size, and the retention of results. Not stated by Microsoft: limits on query volume, checks on answers that combine to reveal more, and rules for agent-written queries. The grouping is Greyhound Research analysis.

Exhibit 5. Strong controls for analysts, none stated for agents. Current through November 19, 2026. © Greyhound Research.

Microsoft has written real controls. Every party whose data a question touches must approve it, each party lists the columns others may use, and small groups can be dropped from results. What it has not stated is how those controls work when an agent drafts the questions. The leak a buyer should plan for is by answer, not by breach.

Greyhound Standpoint:

Greyhound Research’s judgment is that the clean room leaks by answer, not by breach. The seal protects the raw records, and Microsoft’s approval and filtering controls are a sound base. A clean room exists to release answers, though, and answers carry information. A set of approved aggregate questions, each harmless alone, can together single out one customer or one patient. The risk sits in the terms of the room more than in its walls.

Statistical agencies have managed this for generations, suppressing small cells in published tables and limiting repeated queries, because aggregates leak when they are combined. Microsoft’s minimum counts follow that practice, and its rule that every affected party approves each query is a real safeguard. What the documents read do not address is pace. Approval built for analysts assumes a person drafts each query and others read it. An agent can draft hundreds, and careful review becomes harder as volume rises. Microsoft advises against personal or regulated data in the preview, which is responsible. A documented limit on the number and pattern of approved queries, agent-written ones included, would change this view.

Chief risk officers, with legal counsel, should write the allowed questions, minimum group sizes, query limits, and the role of agents into the data-sharing agreement before any data goes in. Until the service leaves preview, it should hold no personal or regulated data, as Microsoft itself advises.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1What controls has Microsoft written for the clean room?

Approval, columns, and minimum counts. The documentation says all collaborators whose datasets a query references must approve it, each dataset declares an allowedFields list of columns queries may use, and post-filters drop aggregated groups below a minimum count. Whoever composes the query sets these, and the others review and approve them.

Testing only. The same page says the limited preview should not process personal data or data subject to legal or regulatory compliance requirements, and should not be used in production.

In the source’s own words

  • Microsoft Learn, Azure Confidential Clean RoomsAll collaborators whose datasets are referenced by a query must approve it.Source
  • Microsoft Learn, Azure Confidential Clean Roomsa minimum count under which aggregated groups in the output are droppedSource
  • Microsoft Learn, Azure Confidential Clean RoomsCustomers should not use the preview to process personal data or other data that is subject to legal or regulatory compliance requirements.Source
2What changes when an agent writes the questions?

Questions at machine speed. An approval process built for analysts assumes a person drafts each query and the others read it. Agents can draft many queries quickly, each innocent alone, whose answers together reveal more than any one of them. No Microsoft document read describes limits on the number or pattern of approved queries, or on agent-written queries.

Google’s equivalent. Google says Confidential Space lets several parties pool data for training and inference while each party’s data stays protected from the others and from Google. Its documentation separates the workload operator, who has no access to the data, from the data collaborators who own it.

What the two designs tell a buyer. Microsoft and Google both separate the parties who own the data from the party who runs the work. The Microsoft documents read describe no limits for questions written by software, and the Google passage quoted here does not address them either. On this evidence the gap is not unique to one provider, so the agreement, more than the choice of cloud, is where buyers should close it.

In the source’s own words

  • Microsoft’s Book of News, Azure confidential computingfor multiparty analytics while helping protect each participant's sensitive raw dataSource
  • Google Cloud documentation, Confidential Space overviewData collaborators can't access each other's data or modify the workload code.Source

↑ Back to the section guide

Part 4The Platform Contest

Three Clouds Compared: Sealed Compute Is Becoming Standard

Greyhound SpotlightAWS made Nitro Sandboxes broadly available on November 10, 2026, a week before Ignite, at no extra charge.

Sealed computing is becoming a standard expectation rather than a premium feature. All three large clouds now sell it, and each puts the seal in a different place. For a buyer, that means confidential computing should become a checkbox on regulated AI tenders, written in by buyers rather than required by regulators, and the comparison moves to scale, proof, and price.

Three leaders on three different layers is a better position for buyers than one leader on all of them. It gives procurement a credible comparison and room to negotiate, because no provider is ahead on every point that matters. It also means the choice rests on which layer a buyer weighs most: scale for the largest models, independent proof, or everyday isolation for containers. Greyhound Research expects each provider to close the gaps the others have exposed, so today’s leads may narrow within a few release cycles. A second-order effect touches architecture. Each cloud places the seal differently, and a design built on one provider’s model may not carry over cleanly, which raises the cost of moving later.

Table 3 compares Microsoft, AWS, and Google layer by layer on each vendor’s own documents. Sealed containers are containers that each run inside their own hardware-protected space.

Greyhound Research · Table 3 · Three clouds, layer by layer

Ahead on scale, behind on independent proof

Microsoft, AWS, and Google compared layer by layer on each vendor’s own documents, November 19, 2026.

Each cloud leads on a different layer. Weigh the verifier row most heavily when comparing offers.

Generally availablePreviewNot statedNot assessed
LayerMicrosoftAWSGoogle
Sealed virtual machinesGenerally available Confidential VMs, also on Azure LocalGenerally available AMD SEV-SNP, also on Dedicated HostsGenerally available Confidential VMs
Confidential GPUsGenerally available H100 since September 2024VR72 rack scale shown, in developmentNot assessed Not found in the documents readGenerally available Hopper in Confidential SpaceBlackwell G4 in preview
Sealed containersPreview Confidential containers on AKSPer-pod Kata and enclaves in developmentBroadly available Nitro Sandboxes, no extra chargePreview Confidential GKE Nodes on G4
Independent verifierNot stated Local verifier in the VR72 demonstrationOwn Nitro Hypervisor and Secure ModuleGenerally available Intel Trust Authority
Multiparty analysisLimited preview Confidential Clean RoomsNot assessed Outside this note’s evidenceGenerally available Confidential Space
Agent memory inside the sealNot statedNot statedNot stated
Against this: Microsoft was first to offer confidential H100 GPUs and first to bring up a Vera Rubin NVL72 system for validation.
Source: Microsoft Book of News, Community Hub (Bhatia, November 17, 2026), Learn, and blogs, AWS Compute Blog (November 10, 2026), What’s New (July 2, 2026), and documentation, Google Cloud Blog (June 23, 2026) and documentation.

Table 3 compares Microsoft, AWS, and Google on six layers, and shows Microsoft ahead on confidential GPUs, Google ahead on independent verification, and no provider documenting agent memory inside the seal. Sealed virtual machines are generally available on all three. Confidential GPUs are generally available on Microsoft since September 2024, with VR72 rack scale shown and still in development, and on Google in Confidential Space, with AWS not assessed. Sealed containers are in preview on Microsoft and Google, with Microsoft’s per-pod Kata containers and enclaves in development, and broadly available on AWS as Nitro Sandboxes. An independent verifier is generally available only on Google, through Intel Trust Authority, and Microsoft’s VR72 demonstration used a local verifier. Agent memory inside the seal is not stated by any of the three. Two cells are not assessed.

Table 3. Ahead on scale, behind on independent proof. Data as of November 19, 2026. © Greyhound Research.

Microsoft leads on scale, with confidential GPUs on sale since 2024 and a rack-scale showcase now. Google leads on independent proof. AWS has made sealed containers a default building block with an attested mode. None of the three ties its seal to agent memory in the documents read. Two cells are not assessed, so the table does not rank AWS on confidential GPUs or multiparty analysis. A buyer comparing offers should weigh the verifier row most heavily.

Greyhound Standpoint:

Greyhound Research concludes that sealed computing is becoming a standard expectation, and that the contest is moving from whether a cloud offers it to how well its proof can be checked. Microsoft was early and is ahead on scale, with confidential GPUs on sale since 2024 and the first rack-scale showcase, still in development. Google is ahead on independent verification. AWS has made hardware isolation an everyday building block at no extra charge. Each leads on a different layer, and none ties its seal to agent memory in the documents read.

The pattern is familiar from encryption of stored data, which moved within a few years from a selling point to a default across the large clouds. Once a security feature becomes standard, buyers stop paying for its presence and start examining its evidence: key custody, audit logs, and who can see what. Confidential computing is entering that phase. Two cells in the comparison are not assessed, so this is not a ranking, and one rival announcement could reorder the verifier row. Within twelve months, sealed computing should be a standard line in regulated AI tenders, set by buyers rather than by regulation.

Enterprise architects should revise tender criteria now, weighting proof and verifier above encryption, and favoring the provider whose attestation can be checked by a party the buyer chooses. The next infrastructure renewal is the natural point to apply them.

The Deeper Read · About 2 minutes

Select a question to open the full answer

1What exactly does each rival offer?

AWS. AWS added AMD SEV-SNP on EC2 Dedicated Hosts on July 2, 2026, for confidential workloads on servers dedicated to one customer. Nitro Sandboxes give each container the boundary of a hardware-isolated virtual machine, are broadly available in commercial Regions at no additional charge, and share the isolation technology of Nitro Enclaves. AI agents that write and run code can run each session in a separate sandbox.

Google. Google made Confidential Space support for NVIDIA Hopper GPUs generally available on June 23, 2026, and previewed Confidential G4 virtual machines and Confidential GKE Nodes with NVIDIA RTX PRO 6000 Blackwell GPUs. A related Google article titled as a general availability announcement for Confidential G4 was seen but not read, so this note keeps the preview status Google’s June post gives.

In the source’s own words

  • AWS What’s New, July 2, 2026enabling customers to run their confidential computing workloads on physical servers fully dedicated to their use.Source
  • AWS Compute Blog, November 10, 2026Nitro Sandboxes share the same Nitro System isolation technology that we have operated for years as Nitro Enclaves.Source
  • AWS Compute Blog, November 10, 2026AI agents that write and run their own code can run each session in a separate sandbox.Source
  • Google Cloud Blog, June 23, 2026Confidential Space support for NVIDIA Hopper GPUs is now generally available.Source
2Will regulation make it mandatory?

Buyers, not regulators. Greyhound Research wrote in December 2025 that client-side encryption, confidential computing, and external key management were no longer optional but baseline requirements for public sector contracts in high-compliance markets. It has also held that regulation alone does not make confidential computing mandatory. Both views stand. The requirement is arriving through tenders, first in high-compliance public-sector contracts and then in regulated AI work more widely. This note’s twelve-month expectation is a judgment about buyers, not a forecast of new law.

In the source’s own words

  • Greyhound Research, published commentary, December 1, 2025baseline requirements for public sector contracts in high-compliance markets.Source
3Is Microsoft really early?

First with confidential GPUs. NVIDIA said in September 2024 that Azure was the first cloud provider to offer confidential computing with NVIDIA H100 GPUs, a claim Microsoft’s own announcement of the same day supports with general availability in two regions. Microsoft also became the first cloud to bring up a Vera Rubin NVL72 system for validation, in March 2026.

What being early tells a buyer. An early start usually brings operating experience, reference customers, and a longer record of fixes, all of which help a risk review. It does not settle who leads on the next layer. Microsoft’s firsts are in bringing up new hardware, and the layer buyers now ask about is verification, where being first to the hardware does not by itself confer a lead.

In the source’s own words

  • Microsoft Azure Infrastructure Blog, November 12In March, Microsoft became the first cloud to bring up an NVIDIA Vera Rubin NVL72 system for validation.Source

↑ Back to the section guide

Part 5The Risks And The Counter-Argument

Risks To Watch: Where A Sealed Room Can Still Leak

Greyhound SpotlightNone of the five risks needs anyone to break the seal. Each happens beside it.

Greyhound Research sees five risks in sealed computing for agents. Most sit beside the seal rather than inside it, and each shows first inside the buyer’s own organization, where it can be caught early.

These risks differ from the ones most security programs track. Security teams have long organized around attackers and breaches, with controls that detect an intrusion. Here the exposure comes from a gap between what the buyer assumes is sealed and what the documents cover, so no alert fires when it opens. That places the first line of defense in review work, with data maps, contracts, and site lists checked against each other. Greyhound Research expects internal audit to find these gaps before security tooling does. The point most coverage misses is that none of the five is a technical failure, so none will appear in a vendor’s incident report, and a buyer who watches only those reports will miss them.

Table 4 sets each risk beside its early warning and its owner. Every early warning can be checked without help from Microsoft.

Greyhound Research · Table 4 · Risk register

Every risk sits beside the seal, not inside it

Five risks in sealed computing for agents, each with its early warning and its owner.

Every risk sits beside the seal. Check all five warnings in the buyer’s own records this quarter.

RiskEarly warning in the buyer’s own estateOwner
Unsealed agent memoryAgent memory, cached context, or retrieved documents rest outside the seal.A data map with no seal marked against memory or cache.Data protection officer
An unchecked proofThe proof is taken but no one the buyer trusts verifies it.A pilot plan that names no verifier.Chief information security officer
A stale proofThe proof is taken once and the agent’s model or tools change later.A change process with no fresh check after updates.Chief information security officer
A loose clean roomApproved questions return answers that, together, reveal protected records.A data-sharing agreement without allowed questions or group sizes.Chief risk officer
Hands on the hardwareA sealed workload runs where outsiders can reach the server.A site list that includes colocation or edge cabinets.Head of platform engineering
Against this: This is a judgment about order, not a prediction of harm.
Source: Greyhound Research analytical model, not measured data. Greyhound Research analysis of the evidence in this research note.

Table 4 lists five risks in sealed computing for agents, each with an early warning in the buyer’s own estate and an owner. Unsealed agent memory, warned by a data map with no seal marked, owned by the data protection officer. An unchecked proof, warned by a pilot plan that names no verifier, owned by the CISO. A stale proof, warned by a change process with no fresh check, owned by the CISO. A loose clean room, warned by an agreement without allowed questions or group sizes, owned by the chief risk officer. Hands on the hardware, warned by a site list with colocation or edge cabinets, owned by platform engineering. The risks are Greyhound Research analysis.

Table 4. Every risk sits beside the seal, not inside it. Current through November 19, 2026. © Greyhound Research.

The five share a cause: an assumption that sealing the machine seals everything the agent touches. Each warning sits in a document the buyer already holds, such as a data map, a contract, a site list, or a test plan. The table names owners, not likelihoods, and ranks none of the risks above another. None of the five needs Microsoft’s help to check.

Greyhound Standpoint:

Greyhound Research sees one root under all five risks: the assumption that sealing the machine seals everything the agent touches. None requires anyone to break the seal. Each sits beside it, in a memory store, a proof nobody checks again, a loose agreement, or a site where the hardware can be reached. That is why each is cheaper to prevent than to unwind. Before production, fixing one is a design decision. After production, it can mean moving data, reopening contracts, and explaining to auditors why the record has a gap.

Large technology programs have tended to go wrong less often at the headline control than at the joints around it, where one team assumed another had the matter in hand. Outsourcing contracts learned to name responsibility at each handover for that reason, and sealed agent programs need the same discipline. The evidence here supports an order of work, not a forecast. The table names owners rather than likelihoods, and nothing read suggests any of these risks has caused harm in practice. If Microsoft documented agent memory and verification for VR72 and the enclaves before they ship, two of the five would shrink considerably.

Each early warning sits in a document the buyer already holds. The owners named in the Runbook that follows can check all five this quarter, and the results belong on the agenda of the next board risk review, before any regulated pilot moves to production.

↑ Back to the section guide

Part 5The Risks And The Counter-Argument

The Strongest Objection: The Seal Is Enough For Now

Greyhound SpotlightThe strongest objection to this note is that sealing the largest models removes the objection risk committees raise most often. That part stands, once VR72 reaches customers.

The objection deserves its full weight. For years, risk committees have refused to send sensitive data to the largest models in a public cloud, because the provider could in principle read it. Rack-scale sealing, once it ships, answers that refusal. Microsoft was early to this field, and its smaller enclaves narrow what must be trusted, which is the right engineering direction. On this view, the agent questions can wait.

The objection has weight because it matches how risk committees work. They tend to clear one concern at a time, and the provider concern has been the one holding back regulated AI on the largest models. Removing it will let projects move, and Greyhound Research expects some regulated firms to plan sealed model capacity once VR72 has published terms. The difficulty is sequence. A committee that clears the provider question may not reopen it when agents arrive on the same machines. What tends to go unnoticed is that the approval itself becomes a precedent. Once a sealed platform is cleared for models, agent projects may be approved on the strength of that earlier decision.

Exhibit 6 tests each part of the objection against the evidence. Three parts stand. The claim that agent questions can wait falls short on timing.

Greyhound Research · Exhibit 6 · The objection held to the evidence

Three parts of the objection stand, and one falls

The fair counter-case to this research note, part by part, against the evidence gathered for each.

The objection holds for the provider question and falls short on timing for agents. Settle both at one review.

StandsFalls
Rack scale removes the top objectionStandsSealing the largest models answers the provider question, once VR72 ships. At Ignite it was a showcase.
Microsoft was earlyStandsConfidential H100 GPUs on sale since 2024, and the first Vera Rubin NVL72 bring-up.
Enclaves narrow what must be trustedStandsA smaller sealed space inside a virtual machine is the right direction, planned for a future release.
Agent questions can waitFalls on timingThe boundary for VR72 and the enclaves is still being drawn, and an unstated boundary becomes the default.
Source: Greyhound Research analytical model, not measured data. Greyhound Research analysis of Microsoft’s Book of News, Microsoft Community Hub (including Bhatia, November 17, 2026), and NVIDIA’s blog.

Exhibit 6 holds four parts of the strongest objection to this research note against the evidence, as four cards, and shows that three stand and one falls on timing. Rack scale removes the top objection of risk committees: stands, once VR72 ships, since at Ignite it was a showcase. Microsoft was early, with confidential H100 GPUs since 2024 and the first Vera Rubin NVL72 bring-up: stands. Enclaves narrow what must be trusted: stands, as a direction planned for a future release. Agent questions can wait: falls on timing, because the boundary for VR72 and the enclaves is still being drawn and an unstated boundary becomes the default. The verdicts are Greyhound Research analysis.

Exhibit 6. Three parts of the objection stand, and one falls. Current through November 19, 2026. © Greyhound Research.

Greyhound Standpoint:

Greyhound Research accepts the objection for the provider question and does not accept it for agents. Sealing the largest models will be a real step. It answers the concern risk committees have raised most often about public cloud, Microsoft earned its early position, and smaller enclaves narrow what must be trusted. Buyers should welcome all three.

Where the objection falls short is timing. Agents will run on the same sealed hardware that models run on, and the boundary documented at launch tends to become the boundary customers live with. Platform history is consistent on this point. Defaults set at general availability shape contracts, reference architectures, and audit expectations for years, and they are much harder to change once customers have built on them. With VR72 and the enclaves still in development, that boundary is still being drawn, which makes now the cheapest moment to settle agent memory, the verifier, and the clean room’s answers. Leaving them for a later release risks inheriting whatever the first release left unstated. The objection would win the timing argument if Microsoft documented those three matters before VR72 and the enclaves reach general availability, and Greyhound Research would say so.

CIOs and CISOs should take the bigger seal and the agent questions to the same architecture review, approving sealed capacity for models while making agent use conditional on the three answers being in writing.

The Deeper Read · About 1 minute

Select a question to open the full answer

1Why does the objection fall short on timing?

Agents arrive on the same hardware. The same Ignite announced agent infrastructure on Azure, and Greyhound Research’s draft note on it found that the authority an agent holds, not its machine, decides what it can do. Regulated buyers who adopt sealed GPUs for models this year will run agents on them next year. A boundary that leaves agent memory unstated today becomes the default then.

The rival case. AWS lets agents that write code run each session in a separate sandbox, and its attested mode releases secrets only to approved code. That is isolation and proof. Neither AWS’s post nor Microsoft’s Ignite documents place the memory of a managed agent service inside the seal.

What the rival case tells a buyer. AWS pairs each agent session with its own sandbox and its own proof, which shows isolation for each agent becoming a design pattern. Neither provider’s documents read place the memory of a managed agent service inside the seal. The timing argument therefore applies across providers, and the memory question belongs in every cloud’s due diligence, not only Microsoft’s.

In the source’s own words

  • AWS Compute Blog, November 10, 2026They use the Nitro-based Nitro Secure Module (NSM) to gain an attestation document in the same style as Nitro Enclaves.Source

↑ Back to the section guide

Part 6What To Do, By Role

The Greyhound CXO Runbook: Ten Moves Before Regulated Data Goes In

Greyhound SpotlightEvery move uses a document the buyer already controls: a data map, a test plan, a site list, a data-sharing agreement, a contract, a budget, a design standard, a reference architecture, or a risk register.

Ten moves follow from the evidence, at least one for each leader who will own part of a sealed agent program. Each names the assumption it corrects, the evidence behind it, and the executive who decides. None waits for Microsoft, because each uses controls the buyer already holds.

Sequence matters more than speed here. The data map comes first because the later moves depend on it: a verifier clause, a site list, or a clean room agreement written before the map exists may protect the wrong places. The moves also cost little, since they rely on staff and documents rather than new tools. Greyhound Research expects each to need revision as Microsoft publishes terms, and giving every document a review date keeps that revision routine. The gain that is easy to overlook is leverage. A buyer that arrives at a negotiation with its own map and its own clauses sets the agenda, and the same documents serve as well when AWS or Google is the counterparty.

Table 5 sets the ten moves beside the assumption each corrects, its evidence, and its owner. Two of the ten finish only when Microsoft publishes its terms.

Greyhound Research · Table 5 · The Greyhound CXO Runbook by owner

Ten moves, and none waits on Microsoft to start

Ten moves for the leaders who will own part of a sealed agent program, with the assumption each corrects, its evidence, and its owner.

Start all ten this quarter. Only the verifier and signing moves wait on Microsoft to finish.

Starts on controls the buyer already holdsCompletes when Microsoft answers

Select a move to open its full text in The Deeper Read below

MoveThe assumption it correctsEvidence anchorThe decision sits with
1/Map which agent data sits inside the sealA sealed machine seals everything the agent touchesExhibit 4Data protection officer, with the CISO
2/Name the verifier and the refresh rule before any pilotA signed proof is a checked proofExhibit 2, Table 2CISO
3/Write the clean room’s allowed answers into the agreementA sealed room cannot leakExhibit 5Chief risk officer, with legal counsel
4/Pilot on what is generally availableKeynote items are ready to plan onTable 1Head of platform engineering
5/Write attestation evidence into the contractEncryption terms are enoughTable 2, Exhibit 3Head of procurement, with the CISO
6/Budget for proof, not only for sealed capacityThe seal is the differentiatorTable 3CIO, with the CFO
7/Approve no sealed workload without its agent controlsSealing a workload governs what its agent may doExhibit 1CISO, with the identity lead
8/Test erasure in every agent store before personal data goes inExpiry after disuse answers an erasure requestExhibit 4Data protection officer
9/Record which cloud’s seal each design depends onA sealed design moves between clouds as easily as its workloadTable 3Head of enterprise architecture, with the head of platform engineering
10/Keep the approval for models and the approval for agents separateClearing the provider question clears agents on the same machinesExhibit 6Chief risk officer, with the CIO
Source: Greyhound Research analysis of the evidence in this research note.

Table 5 sets out the ten moves of the Greyhound CXO Runbook, each with the assumption it corrects, the exhibit or table it rests on, and the executive with whom the decision sits, and shows that none waits on Microsoft to start while two complete only when Microsoft answers. The moves are: map which agent data sits inside the seal, name the verifier and the refresh rule, write the clean room’s allowed answers into the agreement, pilot on what is generally available, write attestation evidence into the contract, budget for proof, approve no sealed workload without its agent controls, test erasure in every agent store, record which cloud’s seal each design depends on, and keep the approval for models and the approval for agents separate. The moves are Greyhound Research analysis.

Table 5. Ten moves, and none waits on Microsoft to start. Current through November 19, 2026. © Greyhound Research.

The two that finish later, on the verifier and on signing, wait on answers and terms Microsoft has not yet published. Those answers are what the questions that follow ask for, and each question maps to a move a buyer can make meanwhile. Every other move rests on an exhibit or table in this note and names who decides. The remaining eight can be finished with controls the buyer already holds.

The Deeper Read · About 5 minutes

Select a move to open its full text

1Data protection officers: map which agent data sits inside the seal.

Exhibit 4 shows that Microsoft places only the model’s computation clearly inside the seal. This move corrects the belief that a sealed machine seals everything the agent touches. Its action is to list every place an agent’s data rests during a task, including retrieved documents, working memory, cached context, tool calls, and logs, mark which are sealed, and keep regulated data out of the rest. The decision sits with the data protection officer, together with the CISO.

2CISOs: name the verifier and the refresh rule before any pilot.

Exhibit 2 and Table 2 show a proof checked once, by a checker Microsoft has not named for the new items beyond a local verifier in the VR72 demonstration. This move corrects the assumption that a signed proof is a checked proof. Its action is to ask Microsoft in writing which verifier checks each preview and whether the customer can choose its own, and to require a fresh check whenever a model or tool changes. The decision sits with the CISO.

3Chief risk officers: write the clean room’s allowed answers into the agreement.

Exhibit 5 shows approval and filtering controls written for analysts, with nothing stated for agent-written queries. This move corrects the belief that a sealed room cannot leak. Its action is to write the allowed questions, minimum group sizes, and limits on agent-written queries into the data-sharing agreement before any data goes in, and to keep personal data out while the service is in preview. The decision sits with the chief risk officer, together with legal counsel.

4Platform engineering leads: pilot on what is generally available.

Table 1 shows confidential H100 machines generally available since 2024 and, among the Ignite items, one limited preview and two still in development, none with terms. This move corrects the habit of planning on keynote items. Its action is to run this quarter’s pilot on the generally available confidential GPU machines, with non-production data, and to record where each sealed workload physically runs. The decision sits with the head of platform engineering.

5Procurement: write attestation evidence into the contract.

Table 2 and Exhibit 3 show that the proof can be forged in the lab and that its checker decides its worth. This move corrects the assumption that encryption terms are enough. Its action is to make the verifier, the refresh rule, and the retention of attestation evidence signed terms, and to hold any production commitment until Microsoft publishes prices, regions, and release dates. Greyhound Research’s November 15 note on the Microsoft 365 outage held that consolidating on one suite remains right while its standard terms fall short, and the same reasoning applies to sealed computing. The decision sits with the head of procurement, together with the CISO.

6CIOs: budget for proof, not only for sealed capacity.

Table 3 shows sealed computing becoming standard on all three clouds. This move corrects the belief that the seal is the differentiator. Its action is to budget for the people and tools that check, store, and audit proofs, and to compare providers on verification, not on encryption. The decision sits with the CIO, together with the CFO.

What budgeting for proof means in practice. Proof work is ongoing rather than a one-time purchase. Each check produces evidence that must be stored, reviewed, and kept for as long as the data it protected, so the cost grows with the number of sealed workloads and agents. Placing it in the run budget rather than a project budget keeps it funded after the pilot team moves on.

7CISOs: approve no sealed workload without the agent controls the seal does not replace.

Exhibit 1 shows that all three Ignite items keep the cloud operator and other tenants out, and that the customer’s own agent is inside by design in every case. This move corrects the assumption that sealing a workload also governs what its agent may do. Its action is to add one rule to the security design standard the CISO already owns: no sealed workload passes the pilot gate until each agent on it has its own identity, a written list of what it may read and call, and human approval before it acts on regulated data. Microsoft’s detailed post supplies the wording, since it says confidential computing does not replace application permissions or the careful handling of logs and tool calls. Written into the standard, that sentence becomes a gate condition internal audit can test. The decision sits with the CISO, together with the identity lead, before the first pilot gate.

8Data protection officers: test erasure in every agent store before personal data goes in.

Exhibit 4 marks memory and cache as not stated for Microsoft’s services, and Microsoft’s Azure CLI reference deletes Context Cache content that has not been accessed within a time to live of 1 to 30 days. This move corrects the assumption that expiry after disuse answers an erasure request. Its action is to take each store named in the data map of the first move and run a deletion request through it end to end, from working memory and cached context to tool outputs and logs, before any personal data reaches an agent. Where a store cannot show the result, agent memory can move to a database the customer runs on confidential machines, with a separate encryption key for each agent’s memory, so that destroying the key erases what the agent remembered. Managed caches stay out of personal data work until their deletion path is documented. The decision sits with the data protection officer, at the production gate.

9Enterprise architects: record which cloud’s seal each design depends on.

Table 3 shows Microsoft, AWS, and Google each ahead on a different layer, and this note finds that each cloud places the seal differently, so a design built on one provider’s model may not carry over cleanly. This move corrects the assumption that a sealed design moves between clouds as easily as the workload it protects. Its action is to add a short entry to the reference architecture for every sealed design, naming the layer it depends on, from sealed virtual machines and confidential GPUs to sealed containers, the verifier, and multiparty analysis, and recording where its proof is kept. Where a choice of verifier exists, the entry should favor one the buyer picks, because that verifier stays with the buyer and ties the record of proof less to one cloud. The decision sits with the head of enterprise architecture, together with the head of platform engineering, at the next architecture review.

10Chief risk officers: keep the approval for models and the approval for agents as separate decisions.

Exhibit 6 shows three parts of the strongest objection standing and one falling on timing, because agents will run on the same sealed hardware as models while the boundary for VR72 and the enclaves is still being drawn. This move corrects the assumption that clearing the provider question also clears agent use on the same machines. Its action is to record provider risk and agent risk as separate entries in the risk register, each with its own owner, so that an approval for sealed model capacity cannot later be read as an approval for agents. Each agent project then returns as a decision of its own, conditional on agent memory, the verifier, and the clean room’s allowed answers being settled in writing. Without that split, a committee that clears the provider question may not reopen it when agents arrive. The decision sits with the chief risk officer, together with the CIO, at the next board risk review.

↑ Back to the section guide

Part 6What To Do, By Role

Questions For Microsoft: What Buyers Will Ask Next

Greyhound SpotlightFive of the six questions have no answer at all in what Microsoft has published, including how often the proof is refreshed.

Every finding in this note rests on what Microsoft has published, and Microsoft alone can close the gaps it names. Six questions follow, in the order the evidence raises them. Each gives what Microsoft has said so far, what the answer would settle, and the Runbook move it maps to. No move waits on them, and an account team can answer each from documents Microsoft already holds.

Six written questions carry more weight than a briefing. In a meeting, a gap tends to be answered with a roadmap, and the buyer leaves with an impression rather than a record. In writing, each answer can be filed with the risk assessment, tested by internal audit, and cited if the product later differs. Written questions also give account teams something concrete to escalate, which tends to speed a reply. Greyhound Research expects answers on regions and prices to arrive with general release, and the answer on agent memory to depend on the design of Microsoft’s managed agent services, which could take longer. Where answers are partial, the open parts set the limits of any pilot and should be recorded.

Table 6 tracks the six questions against what Microsoft has said so far. One is answered in part, and five not at all.

Greyhound Research · Table 6 · Questions for Microsoft, tracked

One question answered in part, five not at all

The six questions only Microsoft can answer, against what it had published by November 19, 2026.

Put all six in writing now. Answers on agent memory and the verifier would move this view most.

Stated in partNot stated

Select a question to open its full text in The Deeper Read below

QuestionWhat Microsoft has said so farThe Runbook move it settles
Q1Do agent memory, cached context, and retrieved documents sit inside the seal?Not stated Agent memory is named among what databases hold. Foundry integration is in development.Runbook move 1
Q2Who verifies the proof for VR72, the enclaves, and the clean room?Stated in part Clean room participants verify. The VR72 demonstration used a local verifier. Not stated for the enclaves.Runbook move 2
Q3How often is the proof refreshed, and does a model or tool change force a new check?Not stated No refresh rule in the documents read.Runbook move 2
Q4How are agent-written queries governed in the clean room?Not stated Approval, column lists, and minimum counts are documented for queries in general.Runbook move 3
Q5Where will the new items physically run, and who can reach the hardware?Not stated No region list for any of the three items.Runbook move 4
Q6What will the three items cost, where, and from when?Not stated One limited preview and two items in development, with no price, region list, or release date.Runbook move 5
Against this: None of the six is a trick question. Each asks Microsoft to write down what its product does or will do at release.
Source: Greyhound Research analysis of Microsoft’s Book of News, Microsoft Community Hub (Bhatia, November 17, 2026), and Microsoft Learn.

Table 6 tracks six questions only Microsoft can answer against what it had published by November 19, 2026, and shows one answered in part and five not stated at all. Answered in part: who verifies the proof, since the clean room lets every participant verify and the VR72 demonstration used a local verifier. Not stated: whether agent memory sits inside the seal, how often the proof is refreshed, how agent-written clean room queries are governed, where the new items physically run, and their price, regions, and release dates. Each question names the Runbook move it settles. The grading is Greyhound Research analysis.

Table 6. One question answered in part, five not at all. Current through November 19, 2026. © Greyhound Research.

Two answers would change this note’s reading most. One is documentation placing agent memory and cached context inside the confidential boundary. The other is a named, customer-selectable verifier for the new items. Either would move Microsoft from early to leading on confidential computing for agents. None of the six is a trick question. Each asks Microsoft to write down what its product does or will do at release.

The Deeper Read · About 2 minutes

Select a question to open its full text

Q1Do agent memory, cached context, and retrieved documents sit inside the seal?

Said so far. Not stated. Microsoft’s summary does not mention agents, Foundry, or Context Cache. Its detailed post names agent memory among what databases hold and lists Foundry integration as still in development, without placing agent memory inside the seal. The answer settles. Whether the data map of the first Runbook move can mark agent memory as sealed. Evidence that would settle it. Documentation of the confidential boundary for agent workloads. Greyhound Research would update. Exhibit 4.

Q2Who verifies the proof for VR72, the enclaves, and the clean room?

Said so far. In part. The clean room lets every participant verify the proof independently. For VR72, the Ignite demonstration used a local verifier, and Microsoft calls remote attestation an architectural direction. For the enclaves, not stated. The answer settles. Whether the customer can choose a verifier Microsoft does not own, under the second Runbook move. Evidence that would settle it. Attestation documentation for each item. Greyhound Research would update. Table 2.

Why “in part” matters. The partial answer already sets a floor. In the clean room, Microsoft has shown that every participant can check the proof. A buyer can reasonably ask that VR72 and the enclaves meet the same standard before general release, and can record any shortfall as a condition of the pilot.

Q3How often is the proof refreshed, and does a model or tool change force a new check?

Said so far. Not stated. The answer settles. Whether attestation covers an agent after its model, tools, or instructions change, under the second Runbook move. Evidence that would settle it. A documented refresh policy. Greyhound Research would update. Exhibit 2.

Q4How are agent-written queries governed in the clean room?

Said so far. Not stated. Approval, column lists, and minimum counts are documented for queries in general. The answer settles. Whether an agent can draft queries, and under what limits, under the third Runbook move. Evidence that would settle it. Clean room documentation on automated or agent-written queries. Greyhound Research would update. Exhibit 5.

Q5Where will the new items physically run, and who can reach the hardware?

Said so far. Not stated. The answer settles. How much weight the physical-access caveat of the 2025 attacks carries for each item, under the fourth Runbook move. Evidence that would settle it. A region list and a statement of physical controls. Greyhound Research would update. Exhibit 3.

Q6What will the three items cost, where, and from when?

Said so far. Not stated. One is a limited preview, and two remain in development. The answer settles. When the signing terms of the fifth Runbook move can be met. Evidence that would settle it. Published prices, regions, and release dates. Greyhound Research would update. Table 1.

↑ Back to the section guide

Part 7The Final Greyhound Standpoint

Confidential Computing: The Final Greyhound Standpoint

Greyhound SpotlightFour signals within twelve months would move this position, and Microsoft controls all four.

Greyhound Research’s final judgment is that Ignite 2026 showed how much bigger Microsoft’s sealed room can become and left the harder question with the buyer. The seal keeps the cloud operator out, and Microsoft has shown it reaching across GPUs, which sets confidential computing on the path from a niche control to an option at scale. Inside the seal, the customer’s own agent still reads what it is given and acts on it. When agents work inside the seal, attestation becomes the control: proof of what ran, checked by someone the buyer trusts, and refreshed when the agent changes.

Credit is due on several counts. On sealing, Microsoft was early and is ahead on scale. Confidential H100 machines have been on sale since 2024, VR72 is designed to extend the seal to the largest models, and the enclaves will narrow what must be trusted, which is the right engineering direction. Google is ahead on proof, with a verifier it does not own, and AWS has made isolation a default building block. No provider documents agent memory inside the seal. The clean room is a sound idea whose risk lies in the answers it returns.

The shape of this moment is familiar. Security features in enterprise technology tend to follow a common arc, from premium option to default to a question of evidence. Encryption of stored data followed it, and confidential computing is now moving from the first stage to the second. Within twelve months it should be a standard line in regulated AI tenders, set by buyers, not regulators, and the questions will turn to who checks the proof, how often, and where the agent’s memory sits. Buyers who set those terms early will negotiate from documents. Those who wait will negotiate from defaults.

The evidence has limits. This note first read the Ignite items as three previews, following Microsoft’s summary, and now follows its detailed post, which shows one limited preview and two items in development, none with a price, region list, or release date. That makes the advice firmer, with less to sign and more time to set terms. The 2025 attacks needed physical access, and none has been reported in the wild. Much of what this note calls missing is not stated, rather than shown to be absent. Statuses are as each company published them through November 19, 2026.

Four signals would move this position within twelve months. They are VR72 confidential virtual machines opened to customers with a release date, published prices against ordinary GPU capacity, an attestation service customers can verify without trusting Microsoft alone, and documentation placing agent memory and cached context inside the boundary.

Exhibit 7 sets those signals side by side with what each would change. All four sit with Microsoft.

Greyhound Research · Exhibit 7 · Signals that would move the position

Four signals in twelve months would move this position

What would move Greyhound Research’s judgment within twelve months, and what each would change.

Of the four signals, documented agent memory inside the boundary would move this view furthest.

Signal Microsoft controlsEvent outside its control
Agent memory inside the boundaryDocumentation placing agent memory and cached context inside the confidential boundary.Would settle the reader’s notes, the question that matters most. Controlled by: Microsoft.An independent verifierAn attestation service customers can verify without trusting Microsoft alone.Would close the gap with Google on proof. Controlled by: Microsoft.VR72 open to customersConfidential VMs on VR72 in preview, then generally available with a date.Would turn the showcase into a product to sign. Controlled by: Microsoft.Published pricesPrices for the Ignite items against ordinary GPU capacity.Would let regulated buyers budget for sealed AI. Controlled by: Microsoft.Outside Microsoft’s controlAn announcement on sealed computing or attestation at a rival’s conference could reset the comparison.
Source: Greyhound Research analytical model, not measured data. Greyhound Research analysis.

Exhibit 7 lists the four signals that would move Greyhound Research’s position within twelve months, each controlled by Microsoft, and one event outside its control. Documentation placing agent memory and cached context inside the boundary would settle the question that matters most. An independent verifier would close the gap with Google on proof. VR72 confidential virtual machines opened to customers with a release date would turn the showcase into a product to sign. Published prices would let regulated buyers budget. An announcement at a rival’s conference could reset the comparison. The signals are Greyhound Research analysis.

Exhibit 7. Four signals in twelve months would move this position. Current through November 19, 2026. © Greyhound Research.

One event outside Microsoft’s control could reset the comparison: an announcement on sealed computing or attestation at a rival’s conference. Of the four signals, documented agent memory inside the boundary would matter most, because it decides whether the seal covers the reader’s notes.

Chief information security officers in regulated sectors have reason to welcome the bigger seal and to stay exacting. Before any pilot, they should obtain Microsoft’s written answers on which agent data stays inside the seal, who verifies the proof, and how cached prompts are deleted on request. Procurement should then write the verifier, the refresh rule, and evidence retention into the contract at signing. Until the signals arrive, companies should pilot on what is generally available, check the proof, and keep regulated data out of anything the seal does not cover.

↑ Back to the section guide

Part 8Appendix, Sources, And Disclosures

Appendix: The Evidence Register, Source By Source

Greyhound Research sets out here the evidence behind every fact, figure, and quotation in this research note, one drop-down for each class of source, so that any sentence can be checked against the words it rests on without leaving the page. Each row gives the source, when it was read, the part of the source the words sit in, what it says, and where in this research note that material is used. Times of reading are in UTC. Every source was first read on November 18, 2026, between 10:05 and 12:00 UTC, and re-read against the live page on November 19, 2026, between 00:10 and 00:45 UTC. Sources added on November 19 were first read then. Where a passage was summarized and not quoted, the row says so. A passage quoted in The Deeper Read of a section is not repeated here, and the source’s entry names the section that quotes it.

Microsoft announcements, blogs, and documentation11 sources, 47 extracts
Source, read, and usedWhere in the sourceWhat it says
Microsoft (2026a)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, The Seal And The Reader, Agent Memory, The Clean Room. Table 1. Exhibits 1 and 4Quoted in The Deeper Read of What Microsoft Announced, The Seal And The Reader, The Clean RoomAzure expands confidential computing[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same section[quoted in The Deeper Read]
Same sectionThe three offerings are VR72 multi-GPU confidential computing, Azure Confidential Enclaves and AKS Kata CoCo confidential containers, and Azure Confidential Clean Room (summarized, not quoted).
PostgreSQL[quoted in The Deeper Read]
Reading noteThe confidential computing section does not mention agents, Foundry, the agent sandboxes, or Context Cache. Its "Learn more" link is a short link to a longer Microsoft post. The short link could not be followed by the tools used, so the note cites the detailed Tech Community post of the same day directly.
Bhatia, V. (2026)First read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, The Seal And The Reader, Who Checks The Proof, Agent Memory, Questions For Microsoft. Tables 1, 2, 3, and 6. Exhibits 2, 4, 6, and 7Quoted in The Deeper Read of What Microsoft Announced, Who Checks The Proof, The Seal And The Reader, Agent MemoryHighlights[quoted in The Deeper Read]
Confidential AI[quoted in The Deeper Read]
Confidential AI[quoted in The Deeper Read]
Confidential AI[quoted in The Deeper Read]
Confidential AI[quoted in The Deeper Read]
Confidential AI[quoted in The Deeper Read]
Confidential AI[quoted in The Deeper Read] (fragment).
Confidential AKS[quoted in The Deeper Read] The two are confidential enclaves and per-pod isolation in confidential containers on AKS, built on upcoming Confidential Child Capable VMs.
Confidential AKSIntel TDX worker nodes on AKS: [quoted in The Deeper Read]
Azure LocalConfidential VMs with AMD SEV-SNP generally available on supported AMD EPYC Genoa processors. Intel TDX confidential VMs and confidential GPUs on Azure Local in preview (summarized, not quoted).
Data foundation[quoted in The Deeper Read]
Data foundation[quoted in The Deeper Read]
Reading noteThe post gives no prices and no region lists, names no customers, and does not mention Azure Attestation by name. The page shows the author as a Microsoft employee without a job title. Its title contains a dash, given as a colon in Sources And References.
Microsoft (2026b)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft AnnouncedPageAnnounces the preview of the multiparty analytics feature of Azure Confidential Clean Rooms (summarized, not quoted).
Reading noteThe page body did not render when read. The date of June 2 is from the blog index. The page shows a modified date of June 4.
Microsoft (2024)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Three Clouds Compared, Confidential Computing. Tables 1 and 3Quoted in The Deeper Read of What Microsoft AnnouncedAnnouncement[quoted in The Deeper Read]
AnnouncementGeneral availability of Azure confidential virtual machines with NVIDIA H100 GPUs, September 24, 2024 (summarized, not quoted).
Phillips, D. (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Attestation. Table 1. Exhibit 2Quoted in The Deeper Read of What Microsoft Announced, AttestationPost[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
PostAzure Confidential Clean Rooms are described as now in limited preview (summarized, not quoted).
Sickle, B. (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Three Clouds Compared. Exhibit 6Quoted in The Deeper Read of What Microsoft Announced, Three Clouds ComparedPost[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
Reading noteThe post does not mention confidential computing.
Microsoft (2026c)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Who Checks The Proof, The Clean Room. Table 2. Exhibit 5Quoted in The Deeper Read of What Microsoft Announced, Who Checks The Proof, The Clean RoomStatus[quoted in The Deeper Read]
Attestation[quoted in The Deeper Read]
Governance[quoted in The Deeper Read]
Output[quoted in The Deeper Read]
Status[quoted in The Deeper Read]
FieldsEach contributed dataset declares an allowedFields list of the columns queries may use (summarized, not quoted).
Microsoft (2026d)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Attestation, Who Checks The Proof. Exhibit 2. Table 2Quoted in The Deeper Read of AttestationOverview[quoted in The Deeper Read]
Reading noteThe page lists SGX, VBS, TPM, trusted launch, and AMD SEV-SNP confidential VMs. It does not list Intel TDX or GPUs, or give prices or regions. It shows a last-updated date of June 29, 2026, and metadata of September 10, 2026.
Microsoft (2026e)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in The Seal And The Reader, Attestation. Table 1. Exhibit 2Quoted in The Deeper Read of The Seal And The Reader, AttestationIsolation[quoted in The Deeper Read]
Auditability[quoted in The Deeper Read]
OverviewConfidential containers on AKS use AMD SEV-SNP, support Azure Linux only, and are in preview (summarized, not quoted).
Microsoft (2026f)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft Announced, Agent Memory. Exhibit 4Quoted in The Deeper Read of What Microsoft AnnouncedOverview[quoted in The Deeper Read]
Reading noteThe page does not state a release status or list regions.
Microsoft (n.d.)First read November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent Memory. Exhibit 4Quoted in The Deeper Read of Agent Memory--time-to-live[quoted in The Deeper Read]
--time-to-liveThe time to live is set in days, from 1 to 30, and defaults to 1 day (summarized, not quoted).
AWS, Google, and NVIDIA9 sources, 21 extracts
Source, read, and usedWhere in the sourceWhat it says
Amazon Web Services (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Three Clouds Compared. Table 3Quoted in The Deeper Read of Three Clouds ComparedAnnouncement[quoted in The Deeper Read]
Amazon Web Services (n.d.)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Who Checks The Proof. Table 2Quoted in The Deeper Read of Who Checks The ProofAttestation[quoted in The Deeper Read]
Key releaseAWS KMS can accept attestation documents and limit key actions to enclaves whose measurements match (summarized, not quoted).
Graf, A. (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Who Checks The Proof, Three Clouds Compared, The Strongest Objection. Tables 2 and 3Quoted in The Deeper Read of Who Checks The Proof, Three Clouds Compared, The Strongest ObjectionPost[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
AvailabilityBroadly available in commercial Regions worldwide at no additional charge (summarized, not quoted).
Google Cloud (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Who Checks The Proof, The Clean Room. Table 2Quoted in The Deeper Read of The Clean RoomRoles[quoted in The Deeper Read]
AttestationGoogle Cloud Attestation runs in the workload's region. Intel Trust Authority runs in a region the customer defines (summarized, not quoted).
Lugani, S. and Narjala, R. (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Who Checks The Proof, The Clean Room, Three Clouds Compared, Confidential Computing. Tables 2 and 3Quoted in The Deeper Read of Who Checks The Proof, Three Clouds ComparedPost[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
Post[quoted in The Deeper Read]
HeadingConfidential G4 VMs with NVIDIA RTX PRO 6000 Blackwell GPUs are in preview (summarized, not quoted).
Reading noteA related Google article titled as a general availability announcement for Confidential G4 was listed on the page and not read.
Su, C. (2024)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Three Clouds Compared. Exhibit 6PostAzure is described as the first cloud provider to offer confidential computing with NVIDIA H100 GPUs (summarized, not quoted).
NVIDIA (2026)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft AnnouncedQuoted in The Deeper Read of What Microsoft AnnouncedRelease[quoted in The Deeper Read]
ReleaseMicrosoft is named among the first clouds to deploy Vera Rubin-based instances in 2026 (summarized, not quoted).
NVIDIA (n.d.a)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in What Microsoft AnnouncedPageConfidential computing at up to 98% of the performance of unsecured deployments, with results that depend on workload, platform, and configuration (summarized, not quoted). Company figure, not used in the analysis.
NVIDIA (n.d.b)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Who Checks The Proof. Table 2Quoted in The Deeper Read of Who Checks The ProofIntroduction[quoted in The Deeper Read]
IntroductionIssues cryptographically signed Entity Attestation Token format tokens (summarized, not quoted).
Security research3 sources, 10 extracts
Source, read, and usedWhere in the sourceWhat it says
Chuang, J. et al. (2025)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in When The Seal Was Opened. Exhibit 3Quoted in The Deeper Read of When The Seal Was OpenedFAQ[quoted in The Deeper Read]
FAQ[quoted in The Deeper Read]
FAQ[quoted in The Deeper Read]
AbstractExtracted attestation keys can be used to compromise NVIDIA GPU confidential computing (summarized, not quoted).
LinksThe site links Intel’s security announcement INTEL-2025-10-28-001, dated October 28, 2025, and AMD bulletin AMD-SB-3040 (summarized, not quoted).
De Meulemeester, J. et al. (2025)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in When The Seal Was Opened. Exhibit 3Quoted in The Deeper Read of When The Seal Was OpenedSite[quoted in The Deeper Read]
SiteInterposer built for less than USD 50. Vendors told in February 2025 under an embargo to September 30, 2025. Paper at IEEE S&P, May 2026 (summarized, not quoted).
Seto, A. et al. (2025)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in When The Seal Was Opened. Exhibit 3Quoted in The Deeper Read of When The Seal Was OpenedSite[quoted in The Deeper Read]
FAQ[quoted in The Deeper Read]
Reading noteThe site gives the venue, ACM CCS 2025, and no release date.
Media3 sources, 3 extracts
Source, read, and usedWhere in the sourceWhat it says
Arghire, I. (2025)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in When The Seal Was Opened. Exhibit 3ArticleWireTap reported on October 2, 2025, with an interposer for less than USD 1,000 (summarized, not quoted).
Kovacs, E. (2025)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in When The Seal Was OpenedArticleBattering RAM reported on October 1, 2025. Intel and AMD said physical attacks are out of scope (summarized, not quoted).
Toulas, B. (2025)Read November 18, 2026, between 10:05 and 12:00 UTC. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in no section. Kept as a reference. The TEE.Fail text now rests on the researchers’ siteArticleTEE.Fail disclosed to Intel in April, NVIDIA in June, and AMD in August 2025, and reported on October 28, 2025 (summarized, not quoted). Not relied on in the text.
Greyhound Research8 sources, 12 extracts
Source, read, and usedWhere in the sourceWhat it says
Greyhound Research (2025a)Read November 18, 2026, before 12:00 UTC, on greyhoundresearch.com. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent MemoryGreyhound Standpoint"sovereign cloud isn't about displacing hyperscalers" (fragment).
Greyhound Research (2025b)Read November 18, 2026, before 12:00 UTC, on greyhoundresearch.com. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent MemoryQuoted in The Deeper Read of Agent MemoryGreyhound Pulse heading[quoted in The Deeper Read]
NoteConfidential containers let enterprises process sensitive data without exposing it, even during runtime (summarized, not quoted).
Greyhound Research (2025c)Read November 18, 2026, before 12:00 UTC, on greyhoundresearch.com. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in The Seal And The ReaderNoteEntra Agent ID as a policy control plane for agent-based computing (summarized, not quoted). Listed as research this note builds on.
Greyhound Research (2025d)Read November 18, 2026, before 12:00 UTC, on greyhoundresearch.com. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in AttestationQuoted in The Deeper Read of AttestationGreyhound Standpoint[quoted in The Deeper Read]
Greyhound Research (2025e)First read November 19, 2026, between 00:10 and 00:45 UTCUsed in Agent Memory, Three Clouds ComparedQuoted in The Deeper Read of Agent Memory, Three Clouds ComparedPublished comment[quoted in The Deeper Read]
Published comment[quoted in The Deeper Read]
Reading noteGreyhound Research comment published on December 1, 2025. Client-side encryption, confidential computing, and external key management are described as no longer optional (summarized, not quoted).
Greyhound Research (n.d.)First read November 19, 2026, between 00:10 and 00:45 UTCUsed in AttestationQuoted in The Deeper Read of AttestationPublished comment[quoted in The Deeper Read]
Reading noteGreyhound Research comment published in 2026. The page shows no date. It also says privacy now has to travel with the workload (summarized, not quoted).
Greyhound Research (2026a)Read November 18, 2026, before 12:00 UTC, on greyhoundresearch.com. Re-read November 19, 2026, between 00:10 and 00:45 UTCUsed in The Greyhound CXO RunbookNoteConsolidation on one suite remains right while its standard terms fall short (summarized, not quoted). Listed as research this note builds on.
Greyhound Research (2026b)Draft read November 18, 2026. Not yet published. Its Context Cache reading was checked against the Azure CLI reference on November 19Used in The Seal And The Reader, Agent Memory, The Strongest Objection. Exhibit 4DraftContext Cache is read as Azure Storage resources linked to a model deployment, with a time-to-live of 1 to 30 days (summarized, not quoted).

↑ Back to the section guide

Part 8Appendix, Sources, And Disclosures

Sources And References: Every Document Behind Every Claim

Every fact, figure, and quotation in this research note traces to one of the sources below, presented in Harvard reference format and grouped by class. Each source is also linked in the sentence that carries it. Microsoft’s, AWS’s, Google’s, and NVIDIA’s documents are each company’s account of its own products, prices, and terms, and prove what the company said rather than that the thing is true. The security research records what its authors found, and the news reports what their authors were told. Standing pages that carry no publication date are marked n.d. All sources were read on November 18, 2026, and re-read on November 19, 2026.

Microsoft announcements and blogs

Bhatia, V. (2026) What’s New in Azure Confidential Computing: Ignite 2026. Microsoft Community Hub, November 17. Available at: https://techcommunity.microsoft.com/blog/azureconfidentialcomputingblog/what%E2%80%99s-new-in-azure-confidential-computing-%E2%80%94-ignite-2026/4573909

Microsoft (2026a) Microsoft Ignite 2026 | Book of News. Microsoft News, November 17. Available at: https://news.microsoft.com/events/ignite/2026/book-of-news/

Microsoft (2026b) Preview of multiparty analytics with Azure Confidential Clean Rooms. Microsoft Community Hub, June 2. Available at: https://techcommunity.microsoft.com/blog/azureconfidentialcomputingblog/preview-of-multiparty-analytics-with-azure-confidential-clean-rooms/4523274

Microsoft (2024) General Availability: Azure confidential VMs with NVIDIA H100 Tensor Core GPUs. Microsoft Community Hub, September 24. Available at: https://techcommunity.microsoft.com/t5/azure-confidential-computing/general-availability-azure-confidential-vms-with-nvidia-h100/ba-p/4242644

Phillips, D. (2026) Advancing digital sovereignty with greater control and choice. The Official Microsoft Blog, October 20. Available at: https://blogs.microsoft.com/blog/2026/10/20/advancing-digital-sovereignty-with-greater-control-and-choice/

Sickle, B. (2026) Join NVIDIA and Microsoft at Ignite 2026. Microsoft Azure Infrastructure Blog, November 12. Available at: https://techcommunity.microsoft.com/blog/azureinfrastructureblog/join-nvidia-and-microsoft-at-ignite-2026/4573012

Microsoft documentation

Microsoft (2026c) Azure Confidential Clean Rooms. Microsoft Learn, updated June 17. Available at: https://learn.microsoft.com/en-us/azure/confidential-computing/confidential-clean-rooms

Microsoft (2026d) Azure Attestation overview. Microsoft Learn, updated June 29. Available at: https://learn.microsoft.com/en-us/azure/attestation/overview

Microsoft (2026e) Confidential containers (preview) with Azure Kubernetes Service (AKS). Microsoft Learn, updated September 18. Available at: https://learn.microsoft.com/en-us/azure/aks/confidential-containers-overview

Microsoft (2026f) NCCadsH100v5 sizes series. Microsoft Learn, updated July 28. Available at: https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/gpu-accelerated/nccadsh100v5-series

Microsoft (n.d.) az storage context-cache container. Microsoft Learn, Azure CLI reference. Available at: https://learn.microsoft.com/en-us/cli/azure/storage/context-cache/container?view=azure-cli-latest

AWS, Google, and NVIDIA

Amazon Web Services (2026) Amazon EC2 Dedicated Hosts now support AMD SEV-SNP. AWS What’s New, July 2. Available at: https://aws.amazon.com/about-aws/whats-new/2026/07/ec2-amd-sev-snp-dedicated-hosts/

Amazon Web Services (n.d.) Cryptographic attestation. AWS Nitro Enclaves User Guide. Available at: https://docs.aws.amazon.com/enclaves/latest/user/set-up-attestation.html

Graf, A. (2026) Nitro Sandboxes: Hardware isolation for every EKS pod, ECS task, and EC2 workload. AWS Compute Blog, November 10. Available at: https://aws.amazon.com/blogs/compute/nitro-sandboxes-hardware-isolation-for-every-eks-pod-ecs-task-and-ec2-workload/

Google Cloud (2026) Confidential Space overview. Google Cloud Documentation, updated November 10. Available at: https://docs.cloud.google.com/confidential-computing/confidential-space/docs/confidential-space-overview

Lugani, S. and Narjala, R. (2026) Verifiable, private AI: Google Cloud expands Confidential Computing frontiers. Google Cloud Blog, June 23. Available at: https://cloud.google.com/blog/products/identity-security/verifiable-trust-in-the-ai-era-whats-new-in-confidential-computing

Su, C. (2024) Now in General Availability: NVIDIA H100 GPUs in Microsoft Azure Confidential Virtual Machines. NVIDIA Blog, September 24. Available at: https://blogs.nvidia.com/blog/azure-confidential-vm-h100-general-availability

NVIDIA (2026) NVIDIA Kicks Off the Next Generation of AI With Rubin. NVIDIA Newsroom, January 5. Available at: https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer

NVIDIA (n.d.a) AI Security with Confidential Computing. NVIDIA. Available at: https://www.nvidia.com/en-us/data-center/solutions/confidential-computing/

NVIDIA (n.d.b) NVIDIA Remote Attestation Service (NRAS) introduction. NVIDIA Documentation. Available at: https://docs.nvidia.com/attestation/cloud-services/latest/nras/nras_introduction.html

Security research

Chuang, J., Seto, A., Berrios, N., van Schaik, S., Garman, C. and Genkin, D. (2025) TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition. Available at: https://tee.fail/

De Meulemeester, J., Oswald, D., Verbauwhede, I. and Van Bulck, J. (2025) Battering RAM. Available at: https://batteringram.eu

Seto, A., Duran, O. K., Amer, S., Chuang, J., van Schaik, S., Genkin, D. and Garman, C. (2025) WireTap: Breaking Server SGX via DRAM Bus Interposition. ACM CCS 2025. Available at: https://wiretap.fail

Media

Arghire, I. (2025) WireTap Attack Breaks Intel SGX Security. SecurityWeek, October 2. Available at: https://www.securityweek.com/wiretap-attack-breaks-intel-sgx-security/

Kovacs, E. (2025) Battering RAM attack breaks Intel and AMD security tech with $50 device. SecurityWeek, October 1. Available at: https://www.securityweek.com/battering-ram-attack-breaks-intel-and-amd-security-tech-with-50-device/

Toulas, B. (2025) TEE.Fail attack breaks confidential computing on Intel, AMD, NVIDIA CPUs. BleepingComputer, October 28. Available at: https://www.bleepingcomputer.com/news/security/teefail-attack-breaks-confidential-computing-on-intel-amd-nvidia-cpus/

Research this note builds on

Greyhound Research (2025a) TCS SovereignSecure Cloud: A Real Advance, Not a Magic Pill. Research note, April 27. Available at: https://greyhoundresearch.com/tcs-sovereignsecure-cloud-a-real-advance-not-a-magic-pill/

Greyhound Research (2025b) IBM Think 2025: Infrastructure for AI Scale – LinuxONE Emperor 5 Redefines Secure AI Ops. Research note, May 8. Available at: https://greyhoundresearch.com/ibm-think-2025-infrastructure-for-ai-scale-linuxone-emperor-5-redefines-secure-ai-ops/

Greyhound Research (2025c) Microsoft Build 2025: Making AI Agents More Capable and Secure. Research note, May 22. Available at: https://greyhoundresearch.com/microsoft-build-2025-making-ai-agents-more-capable-and-secure/

Greyhound Research (2025d) Transforming AI Economics with Procedural Memory. Research note, August 30. Available at: https://greyhoundresearch.com/transforming-ai-economics-with-procedural-memory/

Greyhound Research (2025e) Published comment on end-to-end encryption and government data sovereignty. December 1. Available at: https://www.cio.com/article/4098582/end-to-end-encryption-is-next-frontline-in-governments-data-sovereignty-war-with-hyperscalers.html

Greyhound Research (2026a) Microsoft 365 Outage – When The Suite Goes Dark, Restoration Order Becomes The Control. Research note, November 15. Available at: https://greyhoundresearch.com/microsoft-365-outage-when-the-suite-goes-dark-restoration-order-becomes-the-control/

Greyhound Research (2026b) Microsoft Azure Agent Infrastructure – When Agents Get Their Own Machines, Authority Becomes The Control. Research note, November 2026. Available at: https://greyhoundresearch.com/microsoft-azure-agent-infrastructure-when-agents-get-their-own-machines-authority-becomes-the-control/

Greyhound Research (n.d.) Published comment on agentic AI and confidential computing. Available at: https://www.darkreading.com/endpoint-security/agentic-ai-challenges-progress-in-confidential-computing

↑ Back to the section guide

Part 8Appendix, Sources, And Disclosures

The evidence in this research note is current through November 19, 2026. It was first gathered by 12:00 UTC on November 18, the day after Microsoft’s Ignite announcements, and every source was re-read against the live page on November 19, between 00:10 and 00:45 UTC. Microsoft’s detailed post on the confidential computing items could not be read by the first cutoff. It was read on November 19 and changed the status reading in this note, as the first section explains. Any written response received from an organization named in it after that time is dated where it appears.

Independence Disclosure

Research under the Greyhound Research brand is independent, and no organization holds a right of approval over its content or publication. It is not commissioned, sponsored, or funded by any organization it covers. Greyhound Research may send a research note to companies named in it for factual review before publication. Any such review is confined to the correction of fact, and no viewpoint, conclusion, or recommendation is guided by those companies. All insights and recommendations are Greyhound Research’s alone, and editorial control remains entirely with Greyhound Research. Greyhound Consilium is the B2B thought-leadership arm of The House Of Greyhound and a sister brand to Greyhound Research. It builds on Greyhound Research’s independent research through collaborations that produce thought-leadership content, research, and practical toolkits for decision-makers. Every Greyhound Consilium publication names its partner on the face of the document. This research note is published under the Greyhound Research brand.

Financial Disclosure

Greyhound Research undertakes advisory and consulting engagements with enterprise buyers and technology providers, and no such engagement confers control over its research findings or its decision to publish them. For this research note, Greyhound Research has no financial interest in, and no commercial relationship with, Microsoft, Amazon Web Services, Google, Anthropic, or OpenAI to declare. Oracle is a client of Greyhound Research. Oracle is not a subject of this research note.

Intended Use

This research note is written for chief information officers, chief information security officers, data protection officers, chief risk officers, heads of platform engineering and enterprise architects, procurement and vendor management leaders, board members, and the press. It reads Microsoft’s confidential computing announcements at Microsoft Ignite 2026, the documentation beside them, the published documentation of AWS, Google, and NVIDIA, and published security research, to inform decisions on sealed computing, attestation, agent data, clean rooms, and contracts. This research note is not investment advice, a recommendation to buy, sell, or hold any security, or a view on any company’s share price or valuation. It is not legal advice, and its reading of contract and data protection matters is to be confirmed with counsel. Its assessments rest on the documents listed in Sources And References as they stood on November 19, 2026, and can change as the companies named publish more. Product availability, status, regions, features, and prices are described as each company published them through November 19, 2026, so confirm them with the vendor before acting on them.

Research Methodology

Greyhound Research’s analysis combines lived experience with public sources. The lived experience is years of continuously tracking the market, Greyhound Fieldnotes from ongoing advisory engagements with enterprise buyers and with technology providers, and Greyhound Pulse, the firm’s market trackers. Lived experience forms the judgment and never stands in for evidence. The public sources are the thirty-four documents listed in Sources And References, in six classes, and every factual claim is verified against the document that supports it. The passages relied on are quoted in The Deeper Read of the section that uses them, and the other words taken from each source, the passages summarized rather than quoted and the reading notes, are set out in the appendix. A finding that something is not stated or not documented is bounded to the documents the text names, as they stood when re-read between 00:10 and 00:45 UTC on November 19, 2026. Statuses are given as each company states them, and where two Microsoft documents differ, both are reported.

Microsoft’s detailed Ignite post on the confidential computing items, published on its Tech Community on November 17, could not be read by the first cutoff and was read on November 19. Where it and the Book of News differ on status, this note reports both and follows the detailed post, as the first section explains. The link between VR72 and NVIDIA Vera Rubin NVL72, first drawn as Greyhound Research’s inference, is now stated in that post. NVIDIA’s performance figure for confidential computing is a company claim and is not used in the analysis. The Greyhound Pulse figure on confidential computing use is cited as published in Greyhound Research’s research note of May 8, 2025. No composite figure is computed. No Greyhound Fieldnotes are quoted, and no factual claim in this research note rests on unpublished material. Earlier Greyhound Research comment is cited only where it was published.

The reading of who the seal keeps out, the attestation steps, the agent data path, the clean room controls, the readings in Exhibits 1 to 7 and Tables 1 to 6, the risks, the Greyhound CXO Runbook, the questions for Microsoft, the signals, and every term proposed for a contract are Greyhound Research’s analysis, not statements by Microsoft or any other company. Quotations are checked against the live source before publication.

Authoring Disclosure

This material reflects the analysis, viewpoint and editorial judgment of Greyhound Research. The position presented is developed from Greyhound Research’s research, advisory work and accumulated institutional knowledge. Depending on the assignment, AI may be used to retrieve relevant material; structure, draft or polish prose from analyst-provided concepts, dictation, notes, rough drafts and source material; support secondary research and factual cross-checking; and assist with editing, condensation and consistency with Greyhound Research’s house style. AI does not independently establish Greyhound Research’s viewpoint, analytical position or conclusions. Material factual claims are verified against source evidence before release. Greyhound Research reviews and approves the final material and retains responsibility for it.

All content in this document and on the Greyhound Research website is protected by copyright law and may not be reproduced, distributed, transmitted, displayed, published, or broadcast without the prior written permission of Greyhound Research or, in the case of third-party materials, the prior written consent of the copyright owner of that content. You may not alter, delete, obscure, or conceal any trademark, copyright, or other notice appearing in any Greyhound Research content. Quotation of up to one hundred words is permitted where it carries attribution to Greyhound Research and a link to the published page. Reproduction, redistribution or republication beyond that extent, in whole or in part, by email or in any medium including websites, newsletters and intranets, requires prior written permission, which may be requested at the address below. The published version of each content piece on the Greyhound Research website carries sharing tools for readers who wish to share it.

For questions arising from these disclosures, including licensing requests, contact the Greyhound Research Community Relations Team at connect@thofgr.com. For media inquiries, write to media@thofgr.com.

↑ Back to the section guide

Analyst In Focus: Sanchit Vir Gogia

Sanchit Vir Gogia, or SVG as he is popularly known, is a globally recognized technology analyst, innovation strategist, digital consultant and board advisor. SVG is the Chief Analyst at Greyhound Research (part of The House Of Greyhound), a Global, Award-Winning Technology Research, Advisory, Consulting & Education firm. Greyhound Research works closely with global organizations, their CxOs and the Board of Directors on Technology & Digital Transformation decisions.


Discover more from Greyhound Research

Subscribe to get the latest posts sent to your email.

Discover more from Greyhound Research

Subscribe now to keep reading and get access to the full archive.

Continue reading