Pricing Non-Nvidia Inference Against Nvidia Before You Sign

Price non-Nvidia inference on four numbers: cost per token, MW-backed capacity, customer concentration, and consolidation risk. The trade-off is lower token cost and MW-backed capacity against vendor concentration and consolidation risk. Cerebras' multi-year OpenAI inference agreement worth more than $20 billion covers 750 megawatts. Divide the contract value by the megawatts and the floor is about $27 million per megawatt before tokens, latency, and exit terms. If the non-Nvidia stack does not beat the incumbent on all four, keep the contract with the incumbent.

The spread is the entry fee

Start with the token, not the chip. Ask each inference provider for tokens per second per dollar at your target model, context length, and latency class, then convert that number into monthly cost for your expected traffic. If the non-Nvidia provider cannot show a lower $/token at the same latency, the chip story does not become a serving story. You are buying inference, and the chip is only one input.

Nvidia's Q1 FY27 revenue was $81.61 billion, up 85.2% year over year, and its Data Center revenue was $75.25 billion, up 92%. That scale gives the incumbent stack pricing power, but it does not prove it is cheaper for your workload. The incumbent option carries a cost: dependence on one supplier and the assumption that its roadmap will keep your token price down.

Cerebras' third-generation Wafer-Scale Engine includes 900,000 AI-optimized cores and 44GB of on-chip SRAM. Its Q1 GAAP revenue was $193.4 million, up 94% year over year, and its cloud services revenue grew 178%. Those numbers show a serving business that is scaling, but they do not replace a $/token quote. The smaller-stack cost is a bet that the same token economics will hold.

MW capacity is the real pricing unit

Token price ignores power. Inference contracts are often priced by capacity, and capacity is measured in megawatts. Use the capacity floor from the large inference deal as a sanity check when a provider quotes you for a smaller slice. Get the power plan, the data center location, and the delivery date before you accept a token price.

If your data center plan has a small slice of headroom, multiply it by that floor. The result is the capacity commitment before tokens, software, and operations. The test is whether the provider can fill that capacity with tokens your users will actually buy. If it cannot, the MW number becomes a liability in your capex model.

Concentration is the exit tax

Now price the risk. Two UAE customers accounted for 86% of Cerebras' 2025 revenue. That is the clearest concentration number in the stack. If one of those customers pauses, the provider's pricing power changes. You need to know whether your contract is tied to the same large demand that is propping up the provider's revenue. Put the top-customer share and the renewal terms in the same email.

Cerebras priced its IPO at $185 per share and closed its May 14 debut at $311.07, a gain of about 68%. That closing price put its basic market value near $66.95 billion, with fully diluted estimates briefly above $100 billion. A public market cap is not a guarantee, but it changes the provider's ability to fund capacity and survive a demand shock.

A public listing gives you a price signal, but it does not remove the need for contract protections. The stack's valuation already prices in growth, which can show up in renewal terms.

Cerebras has direct inference relationships with Mistral and Perplexity. It has also disclosed commercial ties to Meta, OpenAI, and Amazon Web Services. Those relationships show demand, but they also show where the provider's attention goes. If your workload is not a large customer, your service level may be lower. Buy non-Nvidia inference only when the $/token quote is lower, the MW capacity is real, customer concentration is low, and consolidation risk is priced into the contract.