Two of Nvidia's competitors used the Open Compute Project Global Summit on 12 October to make the same argument from different directions. AMD introduced Helion, an open rack blueprint for agentic AI whose CPU only configuration packs up to 32,768 sixth generation EPYC cores into a single rack.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source Intel formally launched the D270, an inference card formerly codenamed Crescent Island, built around up to 480GB of LPDDR5X rather than high bandwidth memory.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source Neither product tries to beat the leading GPU on peak performance. Both are bets that agent workloads reward something else.

Helion: the rack where the GPU is optional

Helion comes in two forms. The CPU only rack targets hyperscalers that need capacity for agent execution; the all in one rack adds AMD Instinct MI350P GPU nodes alongside EPYC CPU nodes and general purpose compute for enterprises.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source Both can be air or liquid cooled and include an integrated token router that sends agents' model calls out to frontier and open source models.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source AMD maps three Zen 6 Venice binnings to agent tiers: a dense 400 watt part for personal agents, a balanced 600 watt part for enterprise, and a high frequency 96 core part for technical agents.2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source

AMD's own framing, in the words of Dan McNamara, its head of compute and enterprise AI, is that agentic AI is not a single workload.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source The design follows from that. An agent spends much of its life not in a model's forward pass but in tool calls, code execution, browsing, retrieval and orchestration, all of which run on CPUs. The token router makes the point explicit: the rack is meant to host the agent and call out for the model, not necessarily to host the model.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source

A Futurum analyst called it the first rack scale AI product that treats GPUs as optional.2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source AMD says partner configurations from Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro and Wiwynn are planned for November 2026, and it will contribute the specification to OCP.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source Coverage disagrees on the ship window, with some reports pointing to the first half of 2027.2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source

D270: memory capacity over peak speed

Intel's card is an inference part on the Xe3P architecture rated at 1.3 petaflops of FP8, 1.5 terabytes per second of memory bandwidth and 350 watts in a standard PCIe 5.0 form factor with air cooling.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source Intel's own build ships with 160GB; the 480GB versions will come from ODM partners.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source It sits below Nvidia's Blackwell Ultra and AMD's MI455X on peak performance but is cheaper and denser on memory.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source Customer sampling is underway, with broader availability targeted for 2027.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source

The trade is deliberate. LPDDR5X is far slower than HBM but much cheaper and more available, and a large pool of it lets a single air cooled card hold a large model or many long context sessions. For inference serving where capacity per dollar matters more than raw throughput, that is a defensible position. It also lets the card go into existing air cooled racks, which most enterprise data centres still are.

Why both moves point the same way

We assess with moderate confidence that the industry is splitting AI infrastructure into two markets. Frontier training stays on dense, liquid cooled, HBM heavy GPU systems where Nvidia leads. Agent execution and mid scale inference move toward hardware optimised for cores, memory capacity and cost, deployable in ordinary facilities. AMD and Intel are both conceding the first market and contesting the second, which is where their existing CPU and enterprise relationships count.

What buyers should take from it

For a data centre buyer the practical question is cost per unit of agent work, not peak throughput. A CPU only Helion rack with an integrated token router lets an operator host many agents on general purpose cores and buy model inference as a service, which shifts spending from capital equipment to API bills.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source The D270 offers the opposite route for operators who want to host models themselves on cheap, large memory in air cooled racks.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source Both widen the menu beyond a single GPU vendor, and procurement teams negotiating GPU allocations now have credible alternatives to cite, even before either product ships at volume.2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source

Who gains and who loses

Gains: the OEMs and ODMs named as partners, which get differentiated rack products to sell.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source Enterprises with air cooled facilities, which gain inference options without a liquid cooling retrofit.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source Model providers on the other end of Helion's token router, whose APIs become the default target of CPU hosted agents.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source

Loses: GPU centric budgets for agent workloads, if buyers accept that most agent work does not need a GPU.2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source HBM suppliers, at the margin, in the inference tier.

The counter case

The counter case is that both companies are making a virtue of a gap. AMD and Intel trail Nvidia at the top end, and a narrative in which the top end matters less is convenient for them. The D270 comes after two earlier Intel data centre GPU cancellations, and its broad availability is a 2027 promise.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source Helion's ship date is already inconsistent across coverage.2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source Agent frameworks may also move more reasoning into the model itself, shifting work back toward accelerators.

Assessment

We assess with moderate confidence that at least one hyperscaler publicly commits to Helion CPU only racks within six months, because agent execution is already a large and growing CPU cost.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source We assess with low confidence that the D270 reaches meaningful volume in 2027, given Intel's record in this segment.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source

What to watch

First, the first named Helion customer and whether partner configurations actually ship in November 2026.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source 2 AI News Weekly 2026-10-12 Venice binnings: dense 400W, balanced 600W, high frequency 96 core; ship dates reported from November 2026 to 1H 2027; Futurum calls it the first rack scale AI product treating GPUs as optional. Open source

Second, independent inference benchmarks of the 480GB D270 against HBM based cards on cost per token by mid 2027.3 AI News Weekly 2026-10-13 D270 Xe3P: up to 480GB LPDDR5X via ODMs, 160GB Intel build; 1.3 PFLOPS FP8; 1.5 TB/s; 350W PCIe 5.0; sampling now, broader availability 2027; below Blackwell Ultra and MI455X on peak performance. Open source

Third, whether Nvidia answers with a CPU heavy agent rack of its own at its next major event.

Fourth, whether the Helion specification is formally accepted by OCP within the year.1 AMD Newsroom 2026-10-12 CPU only rack up to 32,768 EPYC cores; all in one rack adds MI350P GPU nodes; integrated token router; spec to be contributed to OCP; partners Cisco, Dell, Hon Hai, HPE, Lenovo, QCT, Supermicro, Wiwynn; November 2026 availability. Open source