Oracle OCI And Multicloud – When Rivals Run Inside Oracle’s Cloud, The Contract Becomes The Control

What AI World Changed: Multicloud Turns Inside Out

Oracle used AI World to bring its rivals inside its own cloud. Google’s Tensor Processing Units are coming to Oracle Cloud Infrastructure, “marking the first time Google TPUs will be offered on a hyperscale cloud platform outside Google,” Oracle’s release said. Snowflake will run “as a native, first-party service on Oracle Cloud Infrastructure (OCI),” in another, bought and billed through Oracle. And Oracle plans to deploy OpenAI’s own Jalapeño inference accelerators “to expand efficient inference capacity for OpenAI,” wrote Karan Batta, senior vice president, product, Oracle Cloud Infrastructure.

Cerebras, SambaNova, Intel and AMD chips are coming too, beside NVIDIA’s newest systems, which Oracle announced as generally available, and Positron’s, which it says are available now. “Bringing Google TPUs to OCI gives customers access to multiple leading AI compute architectures in one cloud,” said Mahesh Thiagarajan, president, Oracle Cloud Infrastructure.

The other direction moved too. Oracle’s database now runs in Google Cloud’s region in Mexico, and “Oracle AI Database@Google Cloud is available in 17 Google Cloud regions across Asia Pacific, Europe, and the Americas,” Oracle said. Oracle also announced plans “to obtain FedRAMP Class D (formerly FedRAMP High) certification” for its database inside the government clouds of AWS, Microsoft and Google.

Greyhound Research reads these announcements as the point at which Oracle completed a turn it began years ago. Multicloud used to mean spreading a company’s risk across several clouds and several contracts. Oracle now offers the reverse as well: many vendors’ technology on one Oracle contract, and Oracle’s database inside every major cloud. The architecture no longer sets the boundaries between vendors. The contract does.

The releases write the commercial path clearly. For Snowflake, “Oracle will manage pricing, discounts, and billing, providing a single commercial relationship for the service.” What they leave to the buyer are five terms that matter once a rival’s technology runs inside Oracle’s cloud: who supports it, where it runs, what the money buys and where it stops, when it arrives, and how a workload leaves.

Greyhound Research’s coverage of Oracle AI World has found that review becomes the control when agents draft the setup, that fixing does when AI finds the flaws faster, and that ownership does when agents reach every function. When rivals run inside Oracle’s cloud, the contract becomes the control.

The timing matters because buyers are already under strain. Greyhound CIO Pulse 2026 finds more than half of CIOs naming energy cost and availability as their single biggest barrier to AI scale. An offer that brings more chips, more capacity and more partners into one cloud speaks directly to that constraint. It also asks buyers to accept new dependencies at the moment many are trying to reduce them.

The test for buyers is whether every rival service they plan to buy through Oracle comes with a named first responder, a written region, a credit plan that survives the year, a date they can plan on, and a way out they have tried. Passing it depends less on the silicon than on terms the buyer writes before signing.

↑ Back to the section guide

What Oracle Announced: Rivals Inside OCI, And Oracle Inside Every Cloud

Greyhound Research counts the infrastructure and multicloud announcements Oracle made at AI World by what each one commits to, in Oracle’s own words. On October 27 alone, Oracle issued releases on Google’s TPUs, Snowflake, Cerebras, NVIDIA, Intel, its Acceleron network, its database in Mexico and in the US government clouds, and a customer, ADT.

Batta’s blog the same day tied them together under one idea. “Our multicloud work follows the same idea,” he wrote, after setting out “More ways to choose the right compute.” Six OCI engineering blogs added the detail, for NVIDIA, NVIDIA’s Vera CPUs, AMD, Positron, and two on networking, from first principles and across content delivery and application security. SambaNova issued its own release, with a quotation from Batta. Exhibit 1 maps both directions of the announcements on one page, with the database’s earlier presence in other clouds shown as context.

Read by status, the announcements fall into three groups. Available now, in Oracle’s words, are NVIDIA’s Vera Rubin NVL72 systems and Vera CPUs, Positron’s Atlas inference instances, Oracle’s database in Google Cloud’s Mexico region, and, in limited availability, a cost tool for AI spending across providers. Dated for 2027 are Google’s TPUs and Cerebras inference. Undated are Snowflake, Intel, AMD, SambaNova, OpenAI’s Jalapeño accelerators, NVIDIA’s next eight-GPU system, and the government clouds. Table 1 sets them out.

The wording matters as much as the grouping. NVIDIA’s release carries the subheading “NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPUs Are Now Generally Available on OCI,” while Oracle’s blog on the Vera CPUs says “NVIDIA Vera CPUs on OCI will be available in select regions this quarter.” Both are Oracle’s words, and a buyer planning capacity needs the second. The only site any release names is Abilene, where “OpenAI has trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs,” and that is OpenAI’s capacity on an earlier generation, not a region a customer can choose for Vera Rubin.

↑ Back to the section guide

Eight Chip Makers: Choice In One Cloud, On A Calendar

Greyhound Research reads the compute announcements as one offer with three speeds. Eight makers’ chips are on, or coming to, OCI: NVIDIA, Positron, Google, Cerebras, SambaNova, Intel, AMD and OpenAI. “AI inference demands choice,” Thiagarajan said of Cerebras. The choice is real, and it arrives in stages that a capacity plan has to respect.

Available now. NVIDIA’s Vera Rubin NVL72 systems are generally available, and Oracle calls itself “the first hyperscaler to bring Vera CPUs to market” in its engineering blog. OpenAI is among the early users of the Vera CPUs. “We are bringing Vera into production on OCI to support these workloads at scale,” said Uday Ruddarraju, chief technology officer of compute at OpenAI. Positron’s Atlas instances are “available now,” Oracle says, aimed at “smaller workhorse models like GPT-OSS, Qwen 3.5 and Gemma 4.”

Dated for 2027. Google’s eighth-generation chips come in two designs, “TPU 8t for training and TPU 8i for inference,” and “Google TPUs on OCI will be available in calendar year 2027, ramping over a multi-year period to meet customer needs.” Batta’s blog adds that the ramp runs “to gigawatt scale over multiple years.” Cerebras inference “is planned to be available in calendar year 2027,” building on Oracle’s existing inference capabilities, “which include Dedicated AI Clusters in OCI Enterprise AI.”

Undated. Intel’s collaboration covers “planned OCI X14 offerings expected to be based on next-gen Intel server processor codenamed Diamond Rapids.” AMD’s E7 instances are “Upcoming,” in Batta’s word, and Oracle’s blog on AMD’s Helios racks asks readers to “discuss a suitable configuration, supported software, and deployment availability.” SambaNova says “Oracle plans to host the systems in its data centers and provide the capacity to customers.” OpenAI’s Jalapeño accelerators serve OpenAI’s own inference, not a customer’s.

Behind the calendar sits a constraint no buyer controls. Clay Magouyrk, chief executive of Oracle, told diginomica that “inferencing is not going to be 2MW sites everywhere, but we’re delivering a lot of 50MW, 100MW, 200MW sites,” and that “There is only so much memory in the world.” Buyers feel the same pressure from the other side. Greyhound CIO Pulse 2026, across more than 1,000 enterprise technology leaders, finds 55 percent of CIOs citing energy cost and availability as their single biggest barrier to AI scale, up from 47 percent a year earlier.

↑ Back to the section guide

The Layers Beneath The Chips: Network, Storage And Operations

Greyhound Research reads the least noticed announcements of AI World as the ones that make many chips behave like one cloud. Oracle’s networking release said “Oracle today announced new capabilities in Oracle Acceleron,” and that “New capabilities span AI data movement, distributed AI communication, and secure multi-tenant AI infrastructure.” Its four named capabilities are the plumbing beneath every chip in Table 1, and none carries a date.

In Oracle’s words, the High-Performance Fabric “Helps customers deploy inference, storage, and HPC workloads with greater performance through RDMA connectivity.” The NVMe option “Helps customers support AI, database, and other data-intensive workloads with NVMe block storage,” which Oracle’s networking blog says “Improves NVMe block storage performance by up to 3x.” The Oracle Collective Communications Library “Helps customers keep distributed AI workloads running when network paths fail,” and the front-end multiplanar network “Helps customers improve resiliency across compute infrastructure by bringing multiplanar networking.”

“AI infrastructure can be unforgiving when it comes to network friction,” said Pradeep Vincent, executive vice president, Compute and AI Infrastructure, Oracle Cloud Infrastructure. “With Oracle Acceleron, we are building a network that can respond by rerouting traffic.” For a buyer running training jobs across thousands of accelerators, a network that reroutes rather than stops is the difference between a delayed job and a failed one.

The common layer matters most where the chips differ. Oracle’s networking blog says its communications library supports the plugin interfaces of both NVIDIA’s NCCL and AMD’s RCCL. AMD’s newsroom says “Oracle Acceleron converged NICs incorporate AMD Pensando data processing units (DPUs) for AMD CPUs and GPUs on OCI.” NVIDIA’s release says “OCI Superclusters combine NVIDIA NVL72 rack-scale systems, Oracle Acceleron networking with NVIDIA Spectrum-X.” One network beneath several chip families is the part of the offer most likely to lower the cost of moving between them.

The same layer carries the security that sharing requires. NVIDIA’s release says “Oracle Acceleron provides hardware-based isolation for workloads running untrusted, AI-generated code.” At Oracle’s investor day, Thiagarajan said “a 512 GPU cluster can be spliced across 32 customers with complete security,” and described “bring your own hardware,” “to bring any hardware rack directly and place it on our network with the security properties.” A cloud that shares capacity this finely, and admits outside hardware onto its network, makes the support and placement terms in this note more important, not less.

Operations tooling arrived in the same weeks. Oracle introduced Cluster Health Checks “to help you identify unhealthy GPU nodes.” Its Kubernetes engine can now “coordinate maintenance event repair for eligible bare metal GPU worker nodes,” Oracle’s release note says, and Managed Karpenter for the same engine is in preview: “To join the preview for OKE Managed Karpenter, please reach out to your Oracle account team,” Oracle’s blog says. The health checks and the repair are documented for Oracle’s GPU nodes, and Karpenter for Kubernetes nodes of any shape. Whether any of the three reaches TPUs, Cerebras or SambaNova systems is not stated.

Partners appear at the edges too. “We are launching Oracle Cloud Infrastructure Global CDN Connect,” Oracle’s networking blog says, with Cloudflare, Fastly, Netskrt and Mainstreaming, and “Zero OCI data egress charges on eligible cache fetches in supported regions.” OCI App Defense is a “planned, next-generation web application and API protection service,” for which Oracle “entered a non-binding MOU to collaborate with F5.” And Batta wrote that Oracle is “planning to bring native, managed ONTAP capabilities” from NetApp. Each is useful, and each adds a partner to the support line.

↑ Back to the section guide

OpenAI Inside OCI: Tenant, Chip Maker And Supplier

Greyhound Research reads OpenAI as the clearest case of a rival living inside Oracle’s cloud, because it appears there in three roles at once. As a tenant, OpenAI runs some of the largest workloads on OCI. “In Abilene, OpenAI has trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs,” NVIDIA’s release says. “As our frontier models advance, we need infrastructure that can keep pace,” said Sreekanth Pothanis, engineering leader, compute infra, at OpenAI.

As a chip maker, OpenAI is putting its own silicon into Oracle’s data centers. “We plan to deploy OpenAI’s Jalapeño inference accelerators, developed with Broadcom,” Batta wrote, “to expand efficient inference capacity for OpenAI.” That capacity serves OpenAI’s own inference. No document describes it as capacity a customer can reserve, and it carries no date.

As a supplier, OpenAI sells to Oracle’s customers through Oracle’s contract. In June, OpenAI announced that customers would be able to “apply eligible Oracle Universal Credits toward OpenAI models and Codex through OCI,” with “Availability will begin in the coming weeks.” At AI World, Batta wrote that Oracle is “also bringing OpenAI models, ChatGPT, and Codex closer to Oracle applications, data, and infrastructure,” and Oracle’s OpenAI release said its AI Workspace “is designed to connect Oracle Fusion Applications, Oracle NetSuite, Oracle AI Database,” “Oracle Cloud Infrastructure (OCI), and other enterprise tools, and data sources.”

Open-weight models complete the picture. “Whether they are an open weights model like gpt-oss, that is part of what we deliver in Fusion Applications,” said Chris Leone, executive vice president, Applications Development, at Oracle, at the investor day. Oracle’s documentation lists OpenAI’s open-weight gpt-oss models on OCI Generative AI, and added regions for them in March, and Positron’s new instances are aimed at “smaller workhorse models like GPT-OSS.”

The three roles do not answer the buyer’s question about place. OpenAI’s training runs in Abilene, its open-weight models run where OCI Generative AI documents them, and its frontier models bought on Oracle credits run in a location neither OpenAI’s June announcement nor Oracle’s releases state. Greyhound Research’s earlier notes covered OpenAI inside Oracle’s applications and security services. This section asks only the infrastructure question, and it is unanswered.

↑ Back to the section guide

Snowflake Inside OCI: One Bill, And What It Leaves Out

Greyhound Research reads Snowflake’s arrival as the clearest test of the new model, because it puts a rival’s data platform on Oracle’s paper. Oracle’s release says Snowflake AI Data Cloud “will be available” on OCI, “with supported sources including Oracle Fusion Cloud Applications, Oracle NetSuite,” and “Oracle Autonomous AI Database, MySQL HeatWave, and Oracle NoSQL database.”

The commercial terms are precise. “Customers will be able to purchase, provision, and consume Snowflake on OCI directly through Oracle.” “Oracle will manage pricing, discounts, and billing, providing a single commercial relationship for the service.” “Customers will also be able to use Oracle Universal Credits to consume Snowflake on OCI.” The rationale is stated too: native transfer “while reducing cross-cloud data movement and associated egress costs.”

“By making Snowflake available as a native OCI service, we’re giving customers more choice,” said Magouyrk. Sridhar Ramaswamy, chief executive of Snowflake, said “Snowflake is excited to be the flagship AI data platform partner running as a first party offering.” Snowflake’s own blog, by Christian Kleinerman, executive vice president of product, describes the same arrangement in the conditional: “Customers would contract directly with Oracle to purchase Snowflake, simplifying contracting requirements.”

The blog also says where the work happens after purchase. “After provisioning a Snowflake account through the OCI control plane,” it says, “customers would manage its warehouses, databases, data pipelines and access controls.” Snowflake on OCI “is planned to support Snowflake’s cross-cloud data sharing, replication, failover,” and those capabilities are “subject to technical validation and availability.”

What neither document states is who answers when the service fails. The release does not say whether Oracle or Snowflake takes the first call, under whose service level the service runs, or who owns the root cause of an incident. Neither document gives a date, a region or a price. The blog notes that its forward-looking statements “are not commitments to deliver any product offerings.”

↑ Back to the section guide

The Five Terms: What Oracle’s Releases Leave To The Buyer

Greyhound Research tested every AI World offer against the five terms a buyer would need before signing for a rival’s technology inside Oracle’s cloud. Support asks who answers when it fails. Place asks where it runs. Money asks what pays for it and where that money stops. Time asks when it arrives. Exit asks how a workload leaves it. The test reads only what Oracle and its partners published, so a term not stated is a term the buyer will have to write.

The pattern is consistent. For the rival services, Oracle’s documents write the money, and sometimes the date. They name no first responder for any of them, name no region for any of the new chips or for Snowflake, and describe a way out for one, Snowflake, as a plan. For Oracle’s own instances, the support is Oracle’s, and the region is still unnamed. Table 2 sets out the reading, item by item.

Each missing term has a cost a buyer can name. Without support, an outage becomes a dispute about whose part failed. Without place, a regulated dataset can be processed somewhere its owner never approved. Without the money’s limits, a commitment can be spent on services that have not arrived, or billed at list price after a renewal is missed. Without time, a capacity plan rests on a year with no quarter. Without exit, the first serious negotiation happens at renewal, when the buyer has the least room to move.

Greyhound Research read each term only where a document states it. A status quoted in Table 2 is Oracle’s or its partner’s wording. Where a cell says “Not stated”, Greyhound Research searched the release, the blogs and the documentation named in Sources And References and found no statement on that term as of October 31, 2026. That is a finding about the documents, not about Oracle’s intentions, and Oracle may well answer each term in the order itself.

None of these gaps is unusual for a release. A release announces, and a contract commits. What is new is how much now depends on the contract. When the chips, the data platform and the models come from five or six companies and arrive on one Oracle order, the order is the only document that can say who does what. The sections that follow take the five terms one at a time.

↑ Back to the section guide

Support: Who Answers When A Rival Fails Inside OCI

Greyhound Research reads support as the term most changed by the new model, because each rival service now has two companies standing behind it and no document saying which stands first. Google’s TPUs show the shape. “OCI will provide the power, cooling, and datacenter networking to deploy Google TPUs at scale within its infrastructure,” the release says. Google designs the chips. A training job that fails could fail in either company’s part, and the release does not say who takes the call.

SambaNova’s systems follow the same pattern. “Oracle plans to host the systems in its data centers and provide the capacity to customers,” SambaNova’s release says, and Batta’s quotation in it is about bringing “SambaNova’s RDU chips to OCI to give AI builders another efficient way to run models at scale.” Cerebras inference is planned for 2027, building on Oracle’s capabilities “which include Dedicated AI Clusters in OCI Enterprise AI.” Each is a rival’s system in Oracle’s building, and none of the three documents describes the support between them.

Snowflake is the sharpest case, because the buyer works in both companies’ tools. The account is provisioned “through the OCI control plane,” and then “customers would manage its warehouses, databases, data pipelines and access controls,” in Snowflake’s words. A query that fails could fail in Snowflake’s engine, in OCI’s network, or in the integration between Oracle’s sources and Snowflake. A buyer needs to know who opens the incident, who owns its root cause, and whose service level pays out.

The government plan uses the phrase buyers will most want defined. Oracle’s release says the options “help simplify procurement, provide a unified customer experience and support model,” and names no party for the support. Oracle’s database runs on its own Exadata infrastructure inside each hyperscaler’s government cloud, so two companies stand behind every workload. A unified model is the right goal, and the document that defines it has not yet been published.

Greyhound Research proposes that a support schedule for a rival service bought through Oracle should say seven things. It should name the first responder for every incident, whichever company’s part has failed. It should map each company’s severity levels to one shared scale, so that a severity one for the buyer is a severity one for both. It should state the service level that applies and who pays the service credit. It should name who owns the root cause and the written report that follows. It should give the escalation path between the two companies, with named roles. It should require notice when either company changes its part of the service. And it should say what happens to support if the partnership itself changes.

None of this is exotic. Buyers already write such schedules for systems integrators and managed service providers. What is new is writing one for a cloud provider about a rival that lives inside it.

↑ Back to the section guide

Money: One Contract, Many Vendors

Greyhound Research reads the money at AI World as a single wallet growing new pockets. Snowflake on OCI is payable with Oracle Universal Credits. At the event, Oracle also said it “has increased the portion of a customer’s eligible Universal Credits commitment that can be applied to qualifying Marketplace purchases,” including partner AI offerings, “with eligible UCs available for qualifying purchases.” The same release says “Oracle is expanding Marketplace purchasing to India and Japan,” with no date.

Two earlier pockets complete the wallet. In June, OpenAI said customers could “apply eligible Oracle Universal Credits toward OpenAI models and Codex through OCI.” And Oracle’s Multicloud Universal Credits, announced at AI World 2025, are “usable across Oracle Cloud Infrastructure (OCI), Oracle AI Database@AWS, Oracle AI Database@Azure, and Oracle AI Database@Google Cloud.” One Oracle commitment can now reach Snowflake, OpenAI, Marketplace partners and Oracle’s database in three other clouds.

The wallet has edges, and Oracle’s own page draws them. “MUC supports USD only.” The credits are “Not available in Non-commercial, Government, Alloy datacenters,” and are “Intended for direct Oracle customers.” If the contract is not renewed, “any continued Oracle AI Database@AWS, Oracle AI Database@Azure, or Oracle AI Database@Google Cloud consumption is billed at list price.” And “Oracle invoices you for any unused credits at the end of each annual period.”

The Mexico release adds a second route for the same database. Customers can buy Oracle’s database “directly through Google Cloud Marketplace,” and Oracle says the purchase options “allow purchases to apply toward existing Google Cloud commitments,” while “Oracle Multicloud Universal Credits offer additional purchasing flexibility.” The same workload can draw down an Oracle commitment or a Google one, and a buyer holding both will want to decide which, before the year-end shows the answer.

Oracle has also started building the instrument a buyer will need. Its Cost Advisor for AI, now in limited availability in Ashburn, lets customers “See costs from all your connected AI providers in one dashboard to compare spending and make informed decisions,” and “Supported providers are OCI Generative AI, OpenAI, and Anthropic,” Oracle’s release note says. A view across providers is the right tool for a wallet that pays many of them.

Buyers have been wary of the opposite risk. In Greyhound CIO Pulse 2025, published in May 2025, 48 percent of AI leads said single-provider dependence was now their top infrastructure risk. One Oracle contract that buys many vendors’ technology answers part of that worry and sharpens another: the technology is diversified, and the commitment is not.

The wallet also changes the negotiation. Oracle says it “will manage pricing, discounts, and billing” for Snowflake on OCI, so the discount on a rival’s service is now negotiated with Oracle, inside the Oracle commitment, rather than with the rival. A larger commitment can earn better terms across all of it. It also means a single renewal now prices several vendors’ services at once.

Greyhound Research suggests a CFO model several numbers before signing. One is the share of the commitment planned for rival services, and the share of that planned for services not yet available. Another is the exposure at the annual true-up if those services arrive late. A third is the list-price exposure on Database@ use if the multicloud credits are not renewed. And the last is the currency exposure on a commitment that Oracle’s page says “supports USD only” when the buyer reports in another currency.

↑ Back to the section guide

Oracle Inside Every Cloud: The Database Travels Further

Greyhound Research reads the second direction of AI World as the database following the customer, wherever the customer’s other systems already are. In Mexico, Oracle announced “the general availability of Oracle AI Database@Google Cloud” in Google Cloud’s Mexico region, “helping keep Oracle data in Mexico to support data residency, sovereignty, and compliance requirements,” with Exadata Database Service, Autonomous AI Database and Autonomous AI Lakehouse.

A customer spoke for it. “Oracle AI Database@Google Cloud is helping us modernize our core banking environment,” said Rafael Martínez Meza, director, IT Infrastructure and Production, at Banco Actinver, a bank serving over 170,000 clients in Mexico. “Oracle AI Database@Google Cloud helps customers migrate and modernize on-premises Oracle workloads in Google Cloud,” said Ana Laura Garcia, senior director, tech sales, Latin America, at Oracle.

ADT shows why a buyer moves the database rather than the rest. ADT “moved critical applications to Oracle AI Database@Google Cloud,” the release says, including its billing, tax and order management applications, and “The move also brings data from those applications closer to ADT’s Google BigQuery data warehouse,” “making it more accessible for planned AI projects using Gemini Enterprise.” ADT runs Oracle’s Maximum Availability Architecture “across the three Google Cloud regions it currently uses.”

The ADT case shows what travels with the database and what does not. The data and the database service travel. The AI that will use them is the host cloud’s: ADT plans its AI projects on Gemini Enterprise, beside Oracle data in Google Cloud. Oracle has also made the database easier to watch across clouds, with a Datadog integration for “Oracle AI Database@AWS, Oracle AI Database@Azure, and Oracle AI Database@Google Cloud.”

Large estates now mix both directions. At Oracle’s investor day on October 28, Sanjay Apte, head of Oracle cloud services at Vodafone, described a dedicated OCI estate with “six regions spread across three different countries,” and added: “And that’s where we are using Oracle Database@Azure, Oracle Database@AWS.” One customer, three clouds, each with its own support line and its own commitments.

↑ Back to the section guide

Place: Where Each Service Runs, At The OCI Layer

Greyhound Research has already mapped where the reasoning runs for Fusion’s AI, in its research report of October. This section goes one layer down, to OCI itself, where the new chips and services will run. At that layer, the AI World releases are almost silent on place. No release names a region for Google’s TPUs, Cerebras, SambaNova, Positron, Intel, AMD or Snowflake.

Two places are named, and neither is a customer choice. NVIDIA’s release names Abilene, where “OpenAI has trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs,” an earlier generation built for one tenant. Oracle’s Vera CPU blog names “select regions,” without listing them. The release adds a control that matters wherever the work runs: “Oracle Acceleron provides hardware-based isolation for workloads running untrusted, AI-generated code.”

Oracle’s documentation shows how placement works for the AI models already on OCI, and it is candid. OCI Generative AI “now supports access to on-demand models through routing profiles in 30 OC1 commercial regions,” its release note says, so that customers can “Call supported models processed in other regions.” Its regions page lists eleven regions that process models locally and says “if the Calling Region and Destination Region are not the same, then a cross-region call will be made.”

The same documentation shows that a model reached through OCI may not run in OCI at all. For Google’s Gemini models, Oracle’s model regions page says they “are hosted externally by Google,” so “a call to a Google Gemini 2.5 Pro model (through the OCI Generative AI service) results in a call to a Google location.” Processing takes place “within a Google Americas location” for US regions, “within a Google EU location” for Frankfurt, and “only within Google’s Brazil location” for Brazil. For Osaka and Hyderabad, it “can take place within any Google location globally.”

Greyhound Research reads placement as three questions, not one. Where does the data rest? Where is it processed? And whose infrastructure processes it? Oracle’s Gemini notes answer all three for one family of models. For every new service announced at AI World, the releases answer none of them yet.

That is the model for every new service. A buyer can reach a service from one region, have it run in another, and in some cases have it run in another company’s cloud. For OpenAI models bought on Oracle credits, neither OpenAI’s June announcement nor Oracle’s releases say where the models run. Cost Advisor for AI is in limited availability in Ashburn only. Table 4 sets out what is written, service by service.

↑ Back to the section guide

Government Clouds: The Database Goes First, The AI Separately

Greyhound Research reads the government plan as the multicloud model at its most demanding, because every term must also satisfy an authorization. Oracle announced plans “to obtain FedRAMP Class D (formerly FedRAMP High) certification for” “Oracle AI Database@AWS, Oracle AI Database@Azure, and Oracle AI Database@Google Cloud,” “in AWS GovCloud (US), Microsoft Azure Government, and Google Cloud Assured Workloads.” It names no date.

The plan covers “regulated Oracle AI Database workloads using Oracle Exadata Database Service on Dedicated Exadata Infrastructure,” bought “through their respective cloud marketplaces or qualified channel partners,” and “without having to manage a separate environment or billing account.” “By bringing Oracle AI Database services to government cloud environments, we’re working to remove those barriers,” said Nathan Thomas, senior vice president, product management, at Oracle.

Each hyperscaler spoke for its part. “Oracle AI Database@AWS on AWS GovCloud (US) will be built to support FedRAMP High and Impact Levels 4 and 5,” said G2 Krishnamoorthy, vice president, database services, at AWS. “Expanding Oracle AI Database services to Azure Government will give our joint customers greater flexibility,” said Brett Tanzer, vice president, Azure product management, at Microsoft. Raj Pai, vice president, product management, databases, at Google Cloud, said the collaboration “will help government organizations modernize regulated database workloads.”

The same executives also pitch their own AI beside Oracle’s data: Krishnamoorthy names “services like Amazon Bedrock,” Tanzer “Microsoft AI and analytics services,” and Pai “powerful agentic capabilities like Gemini.” The release does not say which of those AI services are available inside the government clouds, or under which authorization. As at ADT, the AI beside the database would be the host cloud’s, and in a government cloud its availability is a separate question from the database’s.

Two Oracle documents already draw lines a public sector buyer should read with the plan. Oracle’s Fusion Service Descriptions, which Greyhound Research reported on in October, say “Certain commercial third-party models, including commercial OpenAI models, are not supported in U.S. Government cloud services.” And Oracle’s Multicloud Universal Credits are “Not available in Non-commercial, Government, Alloy datacenters.” The database is heading into government clouds. Oracle’s credits, and some of the models it sells commercially, are not yet going with it.

↑ Back to the section guide

Time And Exit: Buying A Calendar, Planning The Way Out

Greyhound Research reads time and exit together, because a buyer who signs for a date also needs a way out if the date moves. Two chips have a year: Google’s TPUs, “ramping over a multi-year period,” and Cerebras, “planned to be available in calendar year 2027.” Snowflake “will be available,” with no date in Oracle’s release or in Snowflake’s blog. Intel, AMD, SambaNova and NVIDIA’s next eight-GPU system have no date.

Greyhound Research proposes that a buyer can sign for a calendar on four conditions. The order should state the planned quarter and region for each item. It should say what the buyer may move to if the date slips, among shapes available now. It should hold a price, or the method for setting one. And it should let the buyer reduce or redirect the commitment if an item does not arrive. None of these is described in the releases, so each is the buyer’s to ask for.

Exit is less written still. No release describes how a model trained or tuned for Google’s TPUs moves to another chip, inside OCI or outside it. For inference, Positron’s instances offer OpenAI-compatible interfaces, Oracle says, which lowers the cost of moving a workload on or off them. For Snowflake, the way out is planned: cross-cloud sharing, replication and failover, “subject to technical validation and availability.”

The common layers described earlier help at the margin. A communications library that supports both NVIDIA’s and AMD’s plugin interfaces means a distributed job need not be rebuilt around a new network when it moves between those two chip families. Neither Oracle’s release nor its blog describes the same path for TPUs, Cerebras or SambaNova systems.

The money has an exit cost of its own. If a Multicloud Universal Credits contract is not renewed, Database@ use is “billed at list price,” and unused credits are invoiced each year. Greyhound Research argued in July that buyers should negotiate the exit before entry at the application layer. The same discipline now applies one layer down, to chips, data platforms and the commitment that buys them. It is also the position Greyhound Research took on Oracle’s AI infrastructure contracts in September 2025, where pricing and conversion rights “must all be spelled out in detail.” AI World does not change that position. It adds the vendors it applies to.

↑ Back to the section guide

Risks To Watch: Where One Contract Can Go Wrong

Greyhound Research sees six risks in the new model, each with an early warning a buyer can watch for. None argues against buying a rival’s technology through Oracle. Each argues for writing the term before the risk arrives.

The first is an incident with no owner. A failure on Snowflake or a TPU cluster that each company attributes to the other costs time in the hours that matter most. The early warning is a support case that moves between Oracle and a partner more than once.

The second is a capacity plan built on a 2027 date. The early warning is a 2027 chip that still has no region or price by the middle of 2027. A credit concentrated on roadmap items comes next, and its early warning is commitment allocated to services that are not yet available when the annual true-up approaches.

A placement assumed but not written is another. Its early warning is a model or service found processing outside the region the buyer believed it ran in. A government workload whose AI layer is not authorized is a further risk, with an early warning of an AI service pitched beside the database but absent from the authorization in that realm.

The last is an exit never tested. Its early warning is a renewal date approaching with no workload ever moved off a rival’s chip or platform, even in a test.

↑ Back to the section guide

The Greyhound CXO Runbook: Six Moves, None Waiting For Oracle

Six moves follow from the evidence, written for the leaders who will each own part of a multicloud estate bought through Oracle. Each names who decides. None waits for Oracle, because the terms in each move are the buyer’s to ask for and to sign.

1/ CIOs: plan on what is available or dated, and treat the rest as options.

Take Table 1 and plan only on the five items available now and the two with a year. For each undated item, ask the Oracle account team in writing for the planned quarter, the regions and the price, and keep it out of next year’s capacity plan until it has all three. Ask too which of Oracle’s common layers, from its Acceleron network to its GPU health checks, reach each chip you plan to use. The decision sits with the CIO, together with the head of infrastructure.

2/ Heads of infrastructure: buy the 2027 calendar with terms for a slip.

For Google’s TPUs, Cerebras or any undated chip, put four terms in the order: the planned quarter and region, the shapes available now that the buyer may move to if the date slips, a price or the method for setting one, and the right to reduce or redirect the commitment if the item does not arrive. The decision sits with the head of infrastructure, together with procurement.

3/ Vendor management: write the support boundary before any rival service runs.

For Snowflake on OCI, the TPUs and any Database@ deployment, write the incident path into the order or a document it references: who takes the first call, whose service level applies, who owns root cause, how service credits are paid, and how Oracle and the partner escalate to each other. A rival service without a written incident path should not carry production data. The decision sits with the head of vendor management, together with the CIO.

4/ CFOs: manage the Oracle commitment as a portfolio.

Decide how much of the commitment may go to rival services, starting with Snowflake, and check that each planned service is eligible in your realm and currency. Model the annual true-up for unused credits, the list-price exposure at renewal, and which of your Oracle and Google commitments each Database@ purchase draws down. Use a cross-provider cost view from the first month. The decision sits with the CFO, together with procurement.

5/ CISOs: get placement in writing, service by service.

For each chip, service and model, get the region where it runs and where its processing may happen in writing before data or models move, starting with regulated workloads and government clouds. Where a model reached through OCI is hosted by another company, as Oracle documents for Gemini, record who hosts it and where processing may take place, and decide whether that is acceptable for the data concerned. Treat open-weight models with documented regions and frontier models bought on Oracle credits as two different placement decisions. The decision sits with the CISO, together with the data protection officer.

6/ Chief data officers: test the way out before scale.

Before any rival service carries production at scale, move one workload out in a test: a model off a TPU in a pilot, data replicated out of Snowflake on OCI, and an inference workload moved between instances. Price the renewal of the database commitment before the first year ends. The decision sits with the chief data officer, together with the CIO.

Sequenced in Exhibit 11 against the calendar of the announcements, four of the six moves start now.

↑ Back to the section guide

Questions For Oracle: What Buyers Will Ask Next

Q1 Buyers need to know who answers. Snowflake on OCI will be bought, provisioned and billed through Oracle and managed in Snowflake’s own tools, and Google’s TPUs will run on Oracle’s power, cooling and network. Who takes the first call on an incident on each, whose service level applies, who owns root cause, and will Oracle publish the support model for first-party partner services before they are available?

Q2 Placement decides what data may go where. No AI World release or blog names a customer region for Google’s TPUs, Cerebras, SambaNova, Positron or Snowflake, and Oracle’s Vera CPU blog says “select regions.” Which regions will host each, and will Oracle publish them in the same form as its documentation for OCI Generative AI, which says where each model is processed and who hosts it?

Q3 Capacity plans need dates. Google’s TPUs and Cerebras have a year. Snowflake, Intel’s X14, AMD’s E7, SambaNova and NVIDIA’s HGX Rubin NVL8 do not. What are the planned quarters, and will Oracle reconcile “Now Generally Available” with “this quarter” in “select regions” for the Vera CPUs?

Q4 The same credit now buys several vendors. OpenAI models are payable with eligible Universal Credits, and neither OpenAI’s June announcement nor Oracle’s AI World releases say where models bought that way through OCI run. Where do OpenAI models bought on Oracle credits run, by customer region, and is that location a term the customer can set?

Q5 The wallet has edges. Multicloud Universal Credits support “USD only” and are “Not available in Non-commercial, Government, Alloy datacenters.” Will Oracle support other currencies and the Government and Alloy realms, and what happens to Snowflake and other partner services bought on credits when an order lapses?

Q6 Government buyers need the whole stack authorized. The government plan names no date, describes “a unified customer experience and support model” without naming its parties, and pitches each hyperscaler’s AI beside the database. When will each Database@ service hold FedRAMP Class D, which AI services beside it are authorized in each government cloud, and who supports the combination?

Q7 A way out is part of the price. No release describes how a model trained for Google’s TPUs moves to another chip, and Snowflake’s replication and failover on OCI are “subject to technical validation and availability.” What portability will Oracle commit for workloads on rival silicon and on Snowflake on OCI, and will it be written into the order?

Q8 One estate needs common layers under every chip. Oracle’s Cluster Health Checks and its Kubernetes repair of GPU nodes are documented for its own GPU nodes, and the new Acceleron capabilities carry no date. Which of these layers reach Google’s TPUs, Cerebras and SambaNova systems, when will each Acceleron capability be available in each region, and who supports the partner services at the edge, from App Defense with F5 to managed NetApp ONTAP?

Read together in Table 5, each of the eight questions concerns a term Oracle’s releases leave open, and each maps to a Runbook move a buyer can make meanwhile.

↑ Back to the section guide

OCI And Multicloud: The Final Greyhound Standpoint

Greyhound Research’s final judgment is that Oracle has turned multicloud inside out, and that AI World completed the turn. Rivals’ chips and a rival’s data platform are coming into OCI on Oracle’s contract, and Oracle’s database has gone further into rivals’ clouds, now in seventeen Google Cloud regions, with the US government clouds in plan. Two chip makers’ systems are available now, two have a year, and four do not.

The strongest technical argument for the model sits beneath the chips: one Oracle network that speaks to more than one chip family, isolation for code no one has reviewed, and operations tools that find a failing node before a job does. The clearest example of the model is OpenAI, which is at once a tenant of OCI, a maker of chips planned for it, and a supplier to Oracle’s customers on Oracle’s credits.

Read from the order form up, the decision is more contractual than the releases suggest. The releases write the commercial path clearly, how to buy and who bills. They leave five terms to the buyer: who answers when a rival’s service fails, where it runs, what the commitment buys and where it stops, when it arrives, and how a workload leaves. Oracle’s documentation shows it can write such terms candidly, as its notes on where Gemini is processed show.

Greyhound Research’s recommendation to Oracle follows from Oracle’s own documents. The Gemini notes show that Oracle can say who hosts a model and where it is processed, region by region. The same candor belongs in one published schedule for every first-party partner service, answering the five terms before the service is sold: who takes the first call and owns root cause, which regions host it, what the credits buy and where they stop, the planned quarter, and the path out. The cost to Oracle is a negotiation it already has with every buyer, held once in public rather than many times in private. The return is larger. Oracle has won the commercial argument for bringing rivals inside its cloud, and the first hyperscaler to publish the operational terms for rivals’ technology inside its cloud sets the standard the others have to meet. Until then, every buyer will write that schedule alone, and each will write it differently.

The position would strengthen with a published support model for first-party partner services, regions for each new chip and for Snowflake, planned quarters for the undated items, a stated location for OpenAI models bought on credits, an authorization date and an AI list for each government cloud, and written portability for rival silicon and Snowflake. It would weaken if incidents on rival services went unowned, if 2027 dates slipped without terms, if credits sat on unavailable services at the true-up, or if a model ran somewhere its buyer did not know.

Of the ten signals in Exhibit 12, four would show up first in a buyer’s own estate rather than in anything Oracle publishes.

Until then, the buyers who gain most will be those who treat the contract as the control. They will plan on what is available or dated, write the incident path before a rival service runs, manage the commitment as a portfolio, get placement in writing service by service, and test the way out before scale.

Oracle has supplied the choice and written the bill. Until Oracle publishes the rest, the contract is the control, and the buyer holds the pen.

↑ Back to the section guide

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. Everything in the first four classes is Oracle’s own account of itself and proves what Oracle said rather than that the thing is true. Partner announcements are each partner’s account of itself. Standing pages that carry no publication date are marked n.d. All sources were accessed on October 31, 2026.

Oracle announcements

Oracle Corporation (2026a) Oracle Brings Google TPUs to Oracle Cloud Infrastructure. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-brings-google-tpus-to-oracle-cloud-infrastructure-2026-10-27/

Oracle Corporation (2026b) Oracle Expands Native AI and Data Services on OCI with Snowflake. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-expands-native-ai-and-data-services-on-oci-with-snowflake-2026-10-27/

Oracle Corporation (2026c) Cerebras and Oracle Team Up to Bring Cerebras AI Inference to Oracle Cloud Infrastructure. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-cerebras-and-oracle-team-up-to-bring-cerebras-ai-inference-to-oracle-cloud-infrastructure-2026-10-27/

Oracle Corporation (2026d) NVIDIA and Oracle Expand NVIDIA Vera Rubin to Multi-Gigawatt Scale and Bring NVIDIA Vera CPU to Oracle Cloud Infrastructure. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-nvidia-and-oracle-expand-nvidia-vera-rubin-to-multi-gigawatt-scale-2026-10-27/

Oracle Corporation (2026e) Intel and Oracle Plan Expanded Collaboration to Support Enterprise and Agentic AI on Oracle Cloud Infrastructure. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-intel-and-oracle-plan-expanded-collaboration-to-support-enterprise-and-agentic-ai-on-oracle-cloud-infrastructure-2026-10-27/

Oracle Corporation (2026f) Oracle Unveils New Cloud Networking Innovations for AI. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-unveils-new-cloud-networking-innovations-for-ai-2026-10-27/

Oracle Corporation (2026g) Oracle AI Database@Google Cloud Now Available in Mexico. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-ai-database-google-cloud-now-available-in-mexico-2026-10-27/

Oracle Corporation (2026h) Oracle Plans to Expand Multicloud AI Database Services to US Government and Public Sector Workloads. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-plans-to-expand-multicloud-ai-database-services-to-us-government-and-public-sector-workloads-2026-10-27/

Oracle Corporation (2026i) ADT Secures AI Innovation with Oracle. Press release, October 27, via PR Newswire. Available at: https://www.prnewswire.com/news-releases/adt-secures-ai-innovation-with-oracle-302920193.html

Oracle Corporation (2026j) Oracle Expands Marketplace to Unlock New Partner Growth Opportunities. Press release, October 23. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-expands-marketplace-to-unlock-new-partner-growth-opportunities-2026-10-23/

Oracle Corporation (2026u) Oracle Reinvents Enterprise Computing for the AI Era with OpenAI. Press release, October 27. Available at: https://www.oracle.com/news/announcement/ai-world-oracle-reinvents-enterprise-computing-for-the-ai-era-with-openai-2026-10-27/

Oracle Corporation (2025) OCI redefines top-level AI performance with Zettascale10 and Acceleron. Web page, October 16. Available at: https://www.oracle.com/cloud/ai-world-oci-roundup/

Oracle blogs

Batta, K. (2026) Building the Next Decade of Oracle Cloud Infrastructure. Oracle announcement blog, October 27. Available at: https://www.oracle.com/news/announcement/blog/ai-world-building-the-next-decade-of-oracle-cloud-infrastructure-2026-10-27/

Oracle Corporation (2026k) Announcing General Availability of NVIDIA Vera Rubin NVL72 and Vera CPUs on OCI; Introducing Rubin NVL8 GPUs on OCI, and New Agentic AI Solutions. Oracle Cloud Infrastructure blog, October 27. Available at: https://blogs.oracle.com/cloud-infrastructure/nvidia-vera-rubin-vera-cpus-oci

Oracle Corporation (2026l) Announcing General Availability of OCI Compute with NVIDIA Vera CPUs. Oracle Cloud Infrastructure blog, October 27. Available at: https://blogs.oracle.com/cloud-infrastructure/nvidia-vera-cpus-oci

Oracle Corporation (2026m) Oracle Cloud Infrastructure and AMD Helios Build an Open Foundation for AI at Scale. Oracle Cloud Infrastructure blog, October 27. Available at: https://blogs.oracle.com/cloud-infrastructure/amd-helios-oci-ai-infrastructure

Oracle Corporation (2026n) Announcing General Availability of Positron Atlas on Oracle Cloud Infrastructure: Enabling the Next Generation of Price-Performance Inference at Scale. Oracle Cloud Infrastructure blog, October 27. Available at: https://blogs.oracle.com/cloud-infrastructure/positron-atlas-oci-ai-inference

Oracle Corporation (2026o) First Principles: Oracle Acceleron Networking for AI at Scale. Oracle Cloud Infrastructure blog, October 27. Available at: https://blogs.oracle.com/cloud-infrastructure/first-principles-networking-for-ai-scale

Oracle Corporation (2026p) OCI Networking Advances: Powering AI, Global Content Delivery, and App Security. Oracle Cloud Infrastructure blog, October 27. Available at: https://blogs.oracle.com/cloud-infrastructure/oci-networking-advances-powering-ai

Oracle Corporation (2026q) Extend observability for Oracle AI Database across clouds with Datadog. Oracle Cloud Infrastructure blog, October 26. Available at: https://blogs.oracle.com/cloud-infrastructure/oracle-ai-database-multicloud-datadog

Oracle Corporation (2026v) Managed Karpenter for OCI Kubernetes Engine Preview. Oracle Cloud Infrastructure blog, October 26. Available at: https://blogs.oracle.com/cloud-infrastructure/managed-karpenter-for-oke-preview

Oracle documentation

Oracle Corporation (2026r) Global Generative AI availability. OCI Generative AI release note, October 20. Available at: https://docs.oracle.com/iaas/releasenotes/generative-ai/expanded-on-demand-model-access.htm

Oracle Corporation (2026s) Oracle Cost Advisor for AI in Limited Availability. OCI Generative AI release note, October 27. Available at: https://docs.oracle.com/iaas/releasenotes/generative-ai/cost-advisor-limited-availability.htm

Oracle Corporation (2026w) Validate GPU cluster health with Cluster Health Checks. OCI Compute release note, October 22. Available at: https://docs.oracle.com/iaas/releasenotes/compute/cluster-health-checks.htm

Oracle Corporation (2026x) OCI Kubernetes Engine (OKE) support for GPU maintenance event repair. OCI Kubernetes Engine release note, October 21. Available at: https://docs.oracle.com/iaas/releasenotes/conteng/conteng_node-maintenance-event-repair.htm

Oracle Corporation (2026y) New Regions for OpenAI gpt-oss Models in OCI Generative AI. OCI Generative AI release note, March 4. Available at: https://docs.oracle.com/iaas/releasenotes/generative-ai/gpt-oss-regions.htm

Oracle Corporation (n.d.a) Generative AI Regions. OCI Generative AI documentation. Available at: https://docs.oracle.com/iaas/Content/generative-ai/regions.htm

Oracle Corporation (n.d.b) Generative AI Models by Region, including its Notes for External Calls. OCI Generative AI documentation. Available at: https://docs.oracle.com/en-us/iaas/Content/generative-ai/model-endpoint-regions.htm

Oracle Corporation (n.d.c) Multicloud Universal Credits (MUC). Oracle Multicloud documentation. Available at: https://docs.oracle.com/iaas/Content/multicloud/universal-credits.htm

Oracle Corporation (2026z) Oracle Fusion Service Descriptions, v102026. Effective October 20. Available at: https://www.oracle.com/contracts/docs/oracle-fusion-cloud-service-desc-1843611.pdf

Oracle investor day

Oracle Corporation (2026t) Oracle (ORCL) Transcript: Investor Day 2026. Stock Analysis, October 28. Available at: https://stockanalysis.com/stocks/orcl/transcripts/780561-investor-day-2026/

Partner announcements

Kleinerman, C. (2026) Snowflake and Oracle Expand Their Partnership to Bring the Snowflake AI Data Cloud to OCI. Snowflake blog, October 27. Available at: https://www.snowflake.com/en/blog/snowflake-oracle-partnership-ai-data-cloud-oci/

AMD (2026) AMD Expands AI and Cloud Infrastructure Collaboration with Oracle. AMD newsroom, October 27. Available at: https://newsroom.amd.com/news/amd-expands-ai-cloud-infrastructure-oracle/

OpenAI (2026) Access OpenAI models and Codex through your Oracle cloud commitment. OpenAI, June 10. Available at: https://openai.com/index/openai-on-oracle-cloud

SambaNova Systems (2026) SambaNova and Oracle Collaborate to Bring High-Performance Inference to Oracle Cloud Infrastructure. Press release, October 27. Available at: https://sambanova.ai/press/sambanova-and-oracle-collaborate-to-bring-high-performance-inference-to-oracle-cloud-infrastructure

Named media

Lauchlan, S. (2026) Oracle AI World 26 - building data centers? Look, it’s complicated, says Oracle co-CEO Clay Magouyrk, but it’s under control. diginomica, October 29. Available at: https://diginomica.com/oracle-ai-world-26-building-data-centers-look-its-complicated-says-oracle-co-ceo-clay-magouyrk-its

Research this note builds on

Greyhound Research (2025a) Oracle’s $40B AI Investment: A Game Changer. Commentary, May 27, including findings from Greyhound CIO Pulse 2025. Available at: https://greyhoundresearch.com/oracles-40b-ai-investment-a-game-changer/

Greyhound Research (2025b) Backlog Isn’t Bank Balance: A Reality Check On Oracle’s AI Story. Sprint report, September 12. Available at: https://greyhoundresearch.com/backlog-isnt-bank-balance-a-reality-check-on-oracles-ai-story/

Greyhound Research (2026a) Panic Has A Shelf Life: Why Greyhound Research Remains Bullish On IBM After July 14. Research note, July 17, including findings from Greyhound CIO Pulse 2026. Available at: https://greyhoundresearch.com/panic-has-a-shelf-life-why-greyhound-research-remains-bullish-on-ibm-after-july-14/

Greyhound Research (2026b) Everyone Is Building AI Agents. Oracle Is Moving The Contest Inside The Runtime. Research report, July 20. Available at: https://greyhoundresearch.com/everyone-is-building-ai-agents-oracle-is-moving-the-contest-inside-the-runtime/

Greyhound Research (2026c) Oracle Fusion Claw – Autonomy Arrives With A Receipt. Research report, October 14. Available at: https://greyhoundresearch.com/oracle-fusion-claw-autonomy-arrives-with-a-receipt/

Greyhound Research (2026d) Oracle Agentic Implementation – When Agents Draft The Setup, Review Becomes The Control. Research note, October 29. Available at: https://greyhoundresearch.com/oracle-agentic-implementation-when-agents-draft-the-setup-review-becomes-the-control/

Greyhound Research (2026e) Oracle AI Security – When AI Finds The Flaws Faster, Fixing Becomes The Control. Research note, October 29. Available at: https://greyhoundresearch.com/oracle-ai-security-when-ai-finds-the-flaws-faster-fixing-becomes-the-control/

Greyhound Research (2026f) Oracle Fusion Claw By Function – When Agents Reach Every Function, Ownership Becomes The Control. Research note, October 31. Available at: https://greyhoundresearch.com/oracle-fusion-claw-by-function-when-agents-reach-every-function-ownership-becomes-the-control/

↑ Back to the section guide

Disclosures: Independence, Financial, Intended Use, Methodology, Authoring, And Copyright

The evidence in this research note is current through October 31, 2026. Any written response received from an organization named in it after that date 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. Greyhound Research did not offer any named organization a pre-publication factual review of this research note. Editorial control remained entirely with Greyhound Research throughout. 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. Oracle is a client of Greyhound Research (part of The House Of Greyhound). The engagement covers only thought-leadership content published under the Greyhound Consilium brand and did not fund this research note. Greyhound Research holds no other financial interest in Oracle or its subsidiaries.

Intended Use

This research note is written for chief information officers, chief technology officers and heads of infrastructure, chief financial officers and procurement leaders, heads of vendor management, chief information security officers and data protection officers, chief data officers, public sector technology leaders and authorizing officials, and the press. It reads Oracle’s infrastructure and multicloud announcements at Oracle AI World 2026, and the documents beside them, to inform capacity planning, contracting, support, placement, commitment, and exit decisions. 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 authorization and contract terms is to be confirmed with counsel and the relevant authorizing officials. Its assessments rest on the documents listed in Sources And References as they stood on October 31, 2026, and can change as Oracle, its partners, and others publish more. Product availability, regions, features, and terms are described as Oracle and its partners published them through October 31, 2026, so confirm them with your Oracle account team before acting on them.

Research Methodology

Greyhound Research’s analysis rests on source evidence verified against the document that supports each claim. The evidence comprises the forty-four sources listed in Sources And References, in seven classes, from Oracle’s announcements, blogs, documentation, and investor day to partner announcements, named media, and the research this note builds on. A finding that something is not stated or not described is bounded to the documents the text names, as they stood on October 31, 2026. Statuses are given in Oracle’s own words, and where two Oracle documents word the same status differently, both are quoted.

Statements from Oracle’s investor day are cited for product and customer facts only, and no financial figure from it is used. The Greyhound CIO Pulse 2026 figure is cited as published in Greyhound Research’s research note of July 17, 2026, and the Greyhound CIO Pulse 2025 figure as published in its commentary of May 27, 2025, each labeled with its edition. The five terms, the readings in Tables 2 to 5, the incident path, exit paths, risks, and signals in Exhibits 1 to 12, and every term proposed for an order are Greyhound Research’s analysis, not statements by Oracle or its partners. No Greyhound Fieldnotes are cited, and no factual claim in this research note rests on unpublished material. 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.

Copyright Policy

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