HPE announced a 256-CPU Vera rack and later access to Private Cloud AI. The 80 percent figure has no test setup in the accounts.
HPE's October 21 release announces a 256-CPU Vera rack and later API access. Mixed GPU clusters are available now. The 80 percent line has no test setup.
The short version
HPE's October 21, 2026 release announces a liquid-cooled rack with 256 NVIDIA Vera CPUs and says Private Cloud AI's developer APIs and MCP servers roll out in late 2026. StorageReview and an Investing.com account say mixed H200 and RTX PRO 6000 clusters are generally available now. The 80 percent performance line does not include a test setup.
A release from GTC Berlin
HPE's October 21, 2026 release, distributed on Business Wire, says the company and NVIDIA are extending HPE's AI factory portfolio. StorageReview and Investing.com both date the remarks to NVIDIA GTC Berlin. The hardware wire carried StorageReview. The release stacks several products. This page separates the ones with a date from the ones with an adjective. Asked directly, When do the APIs arrive? The accounts say Private Cloud AI's governed APIs and MCP servers for developers, software vendors, and agents are scheduled for late 2026. A self-service developer experience is planned for 2027. Scheduled is not shipped.FSR 4.1 reaches RX 6000.
A rack the release introduces
The new infrastructure line is the NVIDIA Vera CPU Rack by HPE: liquid cooled, 256 Vera CPUs, support for 22,500 sandboxes, for CPU-heavy agent work and reinforcement-learning evaluation next to GPU clusters. The release says this helps scale those jobs. It does not say a named customer has taken delivery this week. Introducing is the verb. A sandbox count is a capacity claim on the spec sheet. Asked directly, Is the rack in a customer's room? The verb in the release is introducing. The dated availability sentences in the secondary accounts attach to the mixed GPU clusters and to the late-2026 API rollout, not to a ship date for the Vera rack.ICEYE's Bandwagon-5 launch.
Eighty percent, without a setup
The same paragraph says the rack delivers 80 percent faster agentic CPU performance. StorageReview notes that HPE does not name the baseline, and that NVIDIA's page frames the figure against traditional CPU infrastructure without a test configuration. Investing.com repeats the figure as the company's. A percentage without a workload, a core count on the other side, and a date of the run is a marketing sentence. This page keeps it labeled that way.an evaluation a second lab can run.
What the accounts say is available now
StorageReview says mixed clusters of H200 and RTX PRO 6000 Blackwell Server Edition GPUs, plus KAI scheduling, are generally available. Investing.com says the same, and it says the developer API platform rolls out in late 2026, with self-service registration of applications and agents planned for 2027. Those two clocks are the useful part of the story. A mixed GPU pool is a configuration a buyer can ask for now. An MCP server in the console is a feature the release schedules.llm-d 1.0 routes a fleet.
Governed APIs are the late-2026 row
The release says developers and software vendors will reach inference, models, and knowledge services through production APIs, and that MCP servers for agents will be versioned and governed by role, tenant, and project, with audit logs. OpenAPI is described as the source for generated documentation. Air-gapped deployments are named as in scope. All of that is the shape of the feature. The calendar, in the secondary accounts, puts it in late 2026 rather than on October 21.what a process node measures.
Other names the release stacks
The release also says HPE will support NVIDIA's OpenShell reference architecture and a modular DSX POD design. Those are further announcements in the same document. They do not come with the availability sentence the GPU mix has. This page names them so a reader does not think the only object in the release is the Vera rack, and it does not promote them to shipped products. Asked directly, Where does the 80 percent come from? HPE's release says 80 percent faster agentic CPU performance. StorageReview says the release does not name a baseline, and that NVIDIA's framing is against traditional CPU infrastructure without a test configuration. This page does not supply one.the hardware desk.
Who is in scope
Enterprises that buy HPE's NVIDIA factory stack, including teams that want a mixed GPU cluster now and teams waiting on developer access. A lab that wants the 80 percent figure reproduced is in scope only as a warning: the sources do not give the experiment. Customers outside that factory portfolio are not handed a new SKU by this page. Asked directly, What does the Vera rack contain? The release says 256 NVIDIA Vera CPUs in a liquid-cooled rack, with support for 22,500 sandboxes, aimed at agent sandboxes and reinforcement-learning work that runs beside GPUs.Astra SR200 vendor speeds.
What stays on the record
October 21, 2026, GTC Berlin. A 256-CPU Vera rack announced, with 22,500 sandboxes and an 80 percent claim that lacks a published setup. Mixed H200 and RTX PRO 6000 clusters generally available, in the secondary accounts. Private Cloud AI APIs and MCP servers scheduled for late 2026. Self-service planned for 2027. Asked directly, What is available now? StorageReview and Investing.com say clusters that mix NVIDIA H200 and RTX PRO 6000 Blackwell Server Edition GPUs, and the KAI scheduler, are generally available. That is the row with a present-tense availability.Futureweb.
What would move the row
A ship date for the Vera rack, a baseline for the 80 percent, or the late-2026 API actually turning on would be a later document. The record today is an announcement with two clocks and one ungrounded percentage. The short answer this page will stand on is this: HPE's October 21, 2026 release announces a liquid-cooled rack with 256 NVIDIA Vera CPUs and says Private Cloud AI's developer APIs and MCP servers roll out in late 2026. StorageReview and an Investing.com account say mixed H200 and RTX PRO 6000 clusters are generally available now. The 80 percent performance line does not include a test setup.what nameplate capacity measures.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
What is available now?
StorageReview and Investing.com say clusters that mix NVIDIA H200 and RTX PRO 6000 Blackwell Server Edition GPUs, and the KAI scheduler, are generally available. That is the row with a present-tense availability.
When do the APIs arrive?
The accounts say Private Cloud AI's governed APIs and MCP servers for developers, software vendors, and agents are scheduled for late 2026. A self-service developer experience is planned for 2027. Scheduled is not shipped.
What does the Vera rack contain?
The release says 256 NVIDIA Vera CPUs in a liquid-cooled rack, with support for 22,500 sandboxes, aimed at agent sandboxes and reinforcement-learning work that runs beside GPUs.
Where does the 80 percent come from?
HPE's release says 80 percent faster agentic CPU performance. StorageReview says the release does not name a baseline, and that NVIDIA's framing is against traditional CPU infrastructure without a test configuration. This page does not supply one.
Is the rack in a customer's room?
The verb in the release is introducing. The dated availability sentences in the secondary accounts attach to the mixed GPU clusters and to the late-2026 API rollout, not to a ship date for the Vera rack.