OpenAI Commits $20B+ to Cerebras Chips, Plus an Equity Stake

TL;DR

OpenAI has committed more than $20 billion to Cerebras over the next three years, roughly double the size of the $10 billion agreement the two companies signed in January. In return, OpenAI gets warrants for a minority stake in Cerebras that could grow to about 10% if spending climbs toward $30 billion, and it’s kicking in an extra $1 billion to fund Cerebras data centers directly. The Information broke the story today. The headline number is big, but the equity structure is the real signal: OpenAI is starting to own pieces of its compute suppliers.

What actually got announced today

The short version: OpenAI ripped up its own Cerebras contract and wrote a much bigger one.

Back in January, Cerebras announced that OpenAI had agreed to buy up to 750 megawatts of compute capacity over three years, in a deal worth more than $10 billion. At the time, that was a surprise. Cerebras is a distant second behind Nvidia in AI accelerators, and OpenAI had just signed a $38 billion contract with AWS and was deep in its $300 billion Oracle deal.

The new agreement, reported today by The Information and later picked up by Reuters, does three things the old one didn’t:

1. The dollar commitment roughly doubles. OpenAI is now on the hook for more than $20 billion with Cerebras over the same three-year window.
2. OpenAI gets warrants for equity. As OpenAI’s spending crosses certain thresholds, it earns rights to purchase Cerebras shares at a fixed price. If total spending climbs to around $30 billion, those warrants could convert to roughly a 10% stake.
3. OpenAI is funding the data centers. Separate from the chip spend, OpenAI has agreed to provide Cerebras about $1 billion to help build out the facilities that will host its own workloads.

That third part is easy to miss. OpenAI is paying Cerebras to buy chips from Cerebras, and also paying Cerebras to build the buildings that house those chips. It’s a tight circular loop.

Why Cerebras, specifically

Cerebras doesn’t make GPUs. It makes the WSE-3, a single chip carved out of almost an entire 300mm silicon wafer: 46,225 mm², about the size of a dinner plate, with 900,000 cores and 44GB of on-chip SRAM. Nvidia’s H100 is a fingernail-sized die by comparison.

The architectural bet behind wafer-scale is that by cramming everything onto one piece of silicon, you avoid the tax of moving data between chips. On-chip SRAM bandwidth on the WSE-3 is 21 PB/s. For context, a fully loaded GB200 system will get you a few tens of TB/s of HBM3e bandwidth per GPU. The numbers aren’t apples-to-apples (you use them differently), but the gap is why Cerebras pitches wafer-scale as better for workloads where data movement is the bottleneck.

Two workloads where that actually shows up in 2026:

• Long-context inference. Serving a model at 128K, 256K, or 1M-token context windows is a memory-bandwidth problem, not a compute problem. GPT-5.4 runs at 1M tokens. GPT-6, whenever Sam gets around to launching it, will push that further. Cerebras benchmarks on Llama 3.1-70B have reportedly posted 2,100+ tokens/sec per user — numbers that Nvidia hardware hits only with heavy optimization.
• Reasoning models. Thinking models generate long chains of tokens before answering. That turns inference latency into a user-facing problem, because the wait time is literal. Faster per-token generation directly translates into better UX for the next generation of agents.

OpenAI has been public about wanting frontier-scale inference that doesn’t bankrupt them. Cerebras gives them a second hardware stack tuned for exactly that.

Why the warrants change the calculus

Chip deals aren’t usually structured with equity. You sign a volume commitment, you get a discount, everyone goes home. The warrant structure here does something different.

It aligns OpenAI with Cerebras the way Microsoft got aligned with OpenAI. Microsoft pushed $13 billion into OpenAI and got a structural share of its upside. Nvidia committed up to $100 billion into OpenAI in September 2025, with OpenAI agreeing to take 10 gigawatts of Nvidia systems in return. Now OpenAI is running the same playbook in reverse: it’s the one taking the stake, and its supplier is the one getting the capital.

Here’s the pattern across OpenAI’s biggest compute deals:

The Cerebras row is the only one where OpenAI ends up on the owner side. That changes the relationship. If Cerebras goes public (it’s been trying to since 2024), OpenAI benefits from the upside directly rather than just paying invoices. If Nvidia aggressively discounts inference hardware to squeeze Cerebras out, OpenAI has a reason to defend Cerebras that goes beyond “they make good chips.”

Call it compute vertical integration without the ownership. OpenAI hasn’t bought Cerebras, and the warrant structure means it doesn’t need to.

What this does to Nvidia

In the short term, not much. OpenAI’s Nvidia commitments are still an order of magnitude larger than the Cerebras deal. The 10 GW of next-gen Vera Rubin systems OpenAI agreed to with Nvidia represents the single largest hardware commitment the company has made to anyone. Even Nvidia’s reported PC-maker acquisition talks don’t change that math — Nvidia is still the default.

But the directional signal is loud. Until this week, the plausible read on Cerebras was “interesting architecture, scrappy startup, waiting for its IPO.” As of today, Cerebras is a hyperscale supplier to the largest AI company in the world, with a warrant structure that guarantees the relationship deepens as spending rises. That changes how Meta, Anthropic, and Google think about their own diversification bets.

Nvidia’s market share in AI accelerators is sitting above 90% in 2026. A decade-long projection out to 2030 where it stays there looks very different after today than it did last week. The interesting number to watch is the inference/training split. Training workloads still heavily favor Nvidia’s maturity and software stack. Inference is where alternative architectures get a foothold, because the workload is more about memory bandwidth and latency than about the broader CUDA stack.

How this fits OpenAI’s $600B compute plan

OpenAI told investors in February it now plans to spend about $600 billion on compute through 2030. That’s down from the $1.4 trillion figure Sam Altman had floated late last year, and it’s the number that everything else gets measured against.

The Cerebras deal lands in that context like this:

flowchart TB
    A["OpenAI 2030 compute plan<br/>~$600B"] --> B["Oracle<br/>$300B"]
    A --> C["Microsoft Azure<br/>$250B"]
    A --> D["Nvidia hardware<br/>$100B+"]
    A --> E["AWS<br/>$38B"]
    A --> F["Cerebras<br/>$20B-$30B"]
    A --> G[Other/unallocated]

The Cerebras commitment is 3–5% of the total. Not dominant, but not trivial, and it’s the only line item in OpenAI’s supplier stack where the architecture isn’t a GPU. The rest of the chart is Nvidia chips running inside different cloud providers’ buildings. Cerebras is structurally different silicon.

If you want a one-sentence read: OpenAI is locking in a second chip architecture at meaningful scale, before GPT-6-class models make the question of “what do we run inference on” a pricing emergency.

What it means if you build on OpenAI APIs

Most developers reading this don’t care which brand of silicon runs their OpenAI API calls. Fair. But the infrastructure mix does bleed through to pricing and latency in ways worth tracking:

• Per-token prices keep dropping. OpenAI has cut inference prices roughly in half every 6–9 months since 2023. The Cerebras deal adds hardware optionality, which means OpenAI has more negotiating power against any one supplier and more room to keep pushing prices down on the long-context tiers.
• Latency on reasoning and long-context tiers should improve. GPT-5.4 Thinking is famously slow at high reasoning settings. Cerebras hardware is specifically good at shortening that tail. Don’t expect an overnight change. Expect a gradual drop over the next 12 months.
• Outage resilience gets better. Multiple chip architectures behind the API mean a single hardware supply issue (CoWoS packaging shortages, HBM allocation) bites less.

If you’re running serious inference volume, the practical takeaway is that you should assume long-context pricing keeps getting cheaper and architect for roughly half today’s per-million-token rate rather than the current sticker.

The part that should worry antitrust regulators

I’ll flag one thing. This deal, and the pattern it fits into, is the kind of arrangement the FTC and the EU have started calling “circular.” Nvidia invests in OpenAI; OpenAI buys Nvidia chips; OpenAI gets Microsoft’s cloud; Microsoft owns a chunk of OpenAI. Now OpenAI is buying Cerebras chips and owning a chunk of Cerebras.

Bloomberg has been mapping this graph for months under the label “AI circular deals.” At some point regulators are going to look at whether these interlocking stakes are actually competition or something that walks and quacks like coordination. Cerebras’s warrant structure is going to be Exhibit A in that conversation, because it’s the cleanest example yet of a hyperscaler taking a financial position in a hardware supplier rather than the other way around. Regulators and founders will read the same arrangement very differently.

FAQ

What is the OpenAI Cerebras deal?

OpenAI has committed more than $20 billion over three years to buy AI server capacity powered by Cerebras chips, up from a $10 billion commitment signed in January 2026. The new agreement also gives OpenAI warrants for a minority stake in Cerebras (up to roughly 10% if total spending hits $30 billion) and adds about $1 billion in separate funding for Cerebras data center construction.

Why is OpenAI buying Cerebras chips instead of Nvidia?

OpenAI is still buying a lot more from Nvidia. Its Nvidia commitments are well above $100 billion. The Cerebras deal is about hardware diversification and workload fit. Cerebras’s wafer-scale architecture has a clear bandwidth advantage for long-context inference and reasoning models, which is where frontier model usage is heading. Having a second chip architecture also reduces supply-chain risk.

What are Cerebras chips?

Cerebras makes the WSE-3, a single chip that occupies most of a 300mm silicon wafer, roughly 57 times the area of an Nvidia H100 GPU. It has 900,000 AI-optimized cores, 44GB of on-chip SRAM, and 21 PB/s of on-chip memory bandwidth. The design trades manufacturing complexity for massive single-chip memory bandwidth, which helps on workloads where data movement is the bottleneck.

Does OpenAI own part of Cerebras now?

Not yet. OpenAI has warrants (rights to buy Cerebras shares at a fixed price) that vest as OpenAI’s spending with Cerebras crosses thresholds. If OpenAI’s total commitment grows to around $30 billion, those warrants could convert into roughly a 10% ownership stake.

How much does OpenAI plan to spend on compute?

OpenAI told investors in February 2026 it plans to spend about $600 billion on compute infrastructure through 2030. That covers its cloud deals with Oracle ($300B), Microsoft Azure ($250B), AWS ($38B), Nvidia hardware ($100B+ across multiple arrangements), and now Cerebras ($20B+).

Bottom line

OpenAI just doubled its bet on the only non-Nvidia AI accelerator that ships at serious scale today, and structured the deal so that Cerebras’s success is partly OpenAI’s success. The dollar number, over $20 billion, is big enough to count on its own. The equity structure is the part antitrust reviewers and every other chip-building startup will be re-reading for weeks.

Nvidia still dominates. But the 2030 picture where Nvidia owns 90%+ of AI compute just got materially weaker.