One Chip. Four Buyers. No Backup Plan.

Two weeks ago, NVIDIA vice president Ian Buck strapped a server box into the back of a car and drove it across the Bay Area. His first stop: Anthropic's offices in San Francisco, where James Bradbury, Anthropic's head of compute, took delivery of the first production units of NVIDIA's new Vera CPU. Buck then drove south to Mission Bay, where OpenAI's infrastructure team unboxed their own Vera system on an outdoor balcony — San Francisco weather permitting. Then it was on to Palo Alto, where Elon Musk asked Buck about the cooling system and the memory layout. By Monday, Oracle's cloud team had theirs.

This is not a paper launch. Vera is real silicon, in real data centers, being evaluated by real customers at real scale. Oracle Cloud Infrastructure has committed to deploying hundreds of thousands of Vera CPUs beginning this year. That is not a pilot. That is an infrastructure bet.

NVIDIA announced Vera at GTC San Jose in March, positioning it as something genuinely new: a CPU designed from scratch for agentic AI workloads, not adapted from the x86 architecture that Intel and AMD have spent decades perfecting. The chip packs 88 custom Olympus cores, 1.2 terabytes per second of memory bandwidth via LPDDR5X, and a memory subsystem NVIDIA calls SOCAMM — field-replaceable modules that trade the serviceability of traditional DDR5 for efficiency and bandwidth density. For the specific work of running AI agents — sandbox execution, Python tool chains, code compilation, orchestration — NVIDIA claims 1.8 times the performance of leading x86 processors. Independent benchmarks from Phoronix, commissioned by NVIDIA, broadly corroborate those numbers. The counterpoint: Phoronix is not an unbiased third party.

What makes this worth covering is not the chip. It is the customer list.

Anthropic, OpenAI, SpaceX, and Oracle Cloud Infrastructure are the first buyers. Those are the four organizations most likely to have a credible alternative. Anthropic has its own ASIC research. OpenAI has been building its own silicon. SpaceX runs one of the most vertically integrated engineering operations on the planet. And Oracle runs a public cloud that could buy anything it wants from AMD, Intel, or the growing roster of custom silicon startups. All four chose Vera anyway.

That concentration is the monoculture story. For years, the conventional wisdom held that custom silicon from Google, Amazon, and AMD would eventually diversify the AI compute stack — that NVIDIA's dominance was a temporary artifact of the GPU lead in training, destined to erode as inference workloads proliferated and custom ASICs matured. AMD's 256-core Zen 6 Venice, due later this year, is a genuinely different chip aimed at general-purpose data center compute. Intel's Diamond Rapids follows a similar philosophy. Neither is trying to do what Vera does. The question is whether what Vera does — agentic workloads, the inner loop of AI agents running code, using tools, and evaluating results — is the workload that matters most.

The market reaction at Computex gave one answer. NVIDIA shares gained four percent on the Vera announcement. AMD fell between five and nine percent. Intel fell. Qualcomm fell. The semiconductor industry looked at the same numbers and decided NVIDIA had taken something.

Jensen Huang, speaking at Computex Taipei on June first, put it plainly: "AI agents will be the largest users of computing. Vera is the first CPU designed for that future." He was not modest about the scope. NVIDIA is positioning Vera as the beginning of a new infrastructure layer — one where the chip that runs the agent becomes as load-bearing as the GPU that trains the model.

There is a version of this story where the monoculture concern is overblown. Oracle's hundreds of thousands of units are a commitment, not a deployment — the gap between procurement and production can be years. The AI lab evaluations may conclude that existing infrastructure is sufficient, or that the CUDA ecosystem lock-in is the real cost, not the silicon. AMD and Intel have not exited the conversation; they are building toward different conversations.

But the delivery visits tell a different story. NVIDIA did not announce Vera and wait for orders. It built the chip, confirmed it worked, and then Ian Buck drove it across the Bay. That is what it looks like when a market crystallizes.

The question for the rest of the industry — for every company that wants to run AI agents at scale but is not Anthropic, OpenAI, or Oracle — is whether there will be an alternative when Vera defines what production AI infrastructure looks like.