Three chip companies. Three different answers. Same question: does agentic AI need its own CPU?
Arm dropped the AGI CPU on March 24, 2026 — 136 Neoverse V3 cores, TSMC 3nm N3P, 300W TDP, one thread per core, no SMT, and a pair of 128-bit vector units for SIMD workloads. Meta is the lead partner, co-developing the silicon and deploying it alongside its own MTIA accelerators at gigawatt scale. Arm CEO Rene Haas told CNBC the chip alone would generate $15 billion in annual revenue by 2031. Arm stock jumped 16% the day after the announcement. The company's market cap sits at $166.8 billion against roughly $4 billion in 2024 revenue — a multiple that prices in a future this chip has to deliver.
One week later, the man who used to run Arm's infrastructure business told a different story.
Kevork Kechichian joined Intel as executive vice president and general manager of the Data Center Group in September 2025, leaving behind a more than 30-year stint at Arm Intel's leadership announcement confirms the move. He spent much of that time shaping Arm's data center roadmap. And now he thinks the AGI CPU is solving a problem that doesn't exist. "The new chip is solving a problem that doesn't exist," Kechichian told The Register, "and Nvidia shipped its own CPU a week earlier, because apparently the rules don't apply."
The rules, apparently, being: if you're going to make a purpose-built chip for agentic workloads, you need to explain what that actually means architecturally. Arm says the answer is no SMT, minimal SIMD, high memory bandwidth, deterministic scaling. Kechichian's answer is that Intel already makes exactly that chip — and the people actually buying it are using it for packet processing, not AI agents.
The chip in question is Intel's Xeon 6+ series, codename Clearwater Forest. It has up to 288 Darkmont E-cores, fabricated on Intel's 18A process, with 12 DDR5 channels and no SMT DataCenterDynamics. The company unveiled it at Mobile World Congress in Barcelona in March 2026 and expects to launch in the first half of this year. Kechichian's point: the density is there, the core count is there, the architectural profile is there — and Intel isn't marketing it as an agentic chip because its customers aren't asking for one. "We are told the chip is most popular in networking applications like packet processing," The Register reported.
The irony is architectural. Arm and Intel have arrived at almost the same place by opposite roads. Arm's AGI CPU opts out of simultaneous multithreading because, as Arm EVP for Cloud AI Mohamed Awad put it, "one thread per core allows for more deterministic performance scaling." Boost modes on legacy architectures "are not sustainable across long periods of time. They're not sustainable across a chip." When you increase frequency, you increase power — and for sustained agentic workloads, that's the wrong trade-off. Intel's Clearwater Forest also has no SMT. Kechichian noted that for orchestration workloads — "mostly traditional data movement types of things" — not having heavy SIMD engines "is a good thing."
Both chips keep SIMD minimal. Arm calls this the future of agentic computing. Intel ships it as a networking part. The difference is not the silicon. It's the story.
The third voice in the room belongs to Nvidia, which announced its Vera CPU at GTC 2026 in March — the same week Arm was preparing its launch. Nvidia's Vera CPU has 88 Olympus ARM cores with spatial multithreading and 1.2 terabytes per second of memory bandwidth per chip, Nvidia's press release confirms. Jensen Huang's framing at GTC: "The CPU is no longer simply supporting the model; it is driving it." A rack of 256 liquid-cooled Vera CPUs sustains more than 22,500 concurrent CPU environments, Nvidia said. Nvidia put SMT in. Arm didn't. Arm says that's a feature. Kechichian says it's an architectural constraint Arm has no choice but to call a philosophy.
Kechichian, speaking to The Register, drew the sharpest contrast between the two approaches: "If they had the option, they would have put it in. They don't have the option, and none of the cores have SMT at Arm." Whether that constraint is a deliberate design choice or a competitive disadvantage is the question neither company will answer directly.
The market will. Arm's AGI CPU ships at the end of 2026, with material financial impact expected from 2028 onward — EE Times reports it is Arm's first production silicon in its 35-year history, a significant shift for a company that built its business licensing reference designs. The company claims more than double the performance per rack versus x86 and up to $10 billion in capex savings per gigawatt of AI data center capacity, Arm's newsroom blog states — though independent benchmarks have not been published. Arm projects AI data centers will require more than four times the current CPU capacity per gigawatt as agentic AI scales. The launch partner list includes Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom, with systems available to order from ASRock Rack, Lenovo, and Supermicro, Arm's newsroom confirms.
OpenAI's presence on that list is the detail worth sitting with. The company most associated with pushing the frontier of what large language models can do is also a launch partner for a chip whose own branding acknowledges it won't run AI models — that is a job for GPUs and AI ASICs, as The Register noted in its initial coverage. The Arm AGI CPU is for orchestrating agents, not running them.
Kechichian's skepticism isn't vendetta. It's calibration. Intel's Clearwater Forest has the cores, the memory bandwidth, the same SMT-free profile — and its actual customers are using it for packet processing. If the agentic workload is real, if it's distinct from general orchestration, if it requires a purpose-built chip rather than a well-provisioned general-purpose one, the market hasn't proven it yet. The deployments will. Arm needs those launch partners to actually deploy at scale, not just co-develop and reference. Haas needs the revenue materializing from 2028. And the industry needs to decide whether "agentic workload" is a real architectural category or a rebranding of things CPUs have been doing for decades.
TheRegister ran the defection story first, on March 31, 2026.
Our read: Kechichian is the most credible voice in the room because he built the chip he's now dismissing. That doesn't make him right. But it makes his skepticism worth taking seriously. The architecture tells one story. The marketing tells another. The deployments — when they arrive, at the end of 2026 and beyond — will tell us which one mattered.
Sources: The Register (Kechichian quotes, Awad quotes); Arm newsroom (blade/rack specs, launch partners); Arm newsroom — launch (AGI CPU announcement, Haas quote); CNBC (Haas revenue forecast); HPCwire (AGI CPU specs, stock jump); EE Times (ship date, financial impact timeline); DataCenterDynamics (Intel Xeon 6+ specs); Nvidia news (Vera CPU specs); Nvidia blog — GTC (Jensen quote); Intel leadership announcement (Kechichian bio)
