Artificial Analysis benchmarks 22 providers serving gpt-oss-120B (high). Cerebras runs at 2,230 tokens per second. Fireworks 836. SambaNova 704. Cerebras also holds the lowest time-to-first-token on the board at 1.37 seconds against SambaNova’s 4.61. The speed advantage is documented. The institutional bet is that it survives the move to frontier-model production load.

Per Cerebras’s S-1 filing, MBZUAI accounted for 62% of 2025 revenue and G42 for 24%, totalling 86% from two Abu Dhabi-linked entities. The OpenAI contract diversifies the book without dissolving the concentration. A single sovereign customer pull-back is the entire bear case.

Three non-GPU inference plays have priced through three capital channels in twelve months: Cerebras via IPO, Fractile via venture (see below), Groq absorbed into NVIDIA at Christmas. Public-market capital just signalled inference compute is a separable market from training, and the second source for it does not have to be a GPU.

Bain, McKinsey, and Capgemini sell AI transformation engagements for a living. They have just signed up to sell OpenAI’s stack specifically. The conflict that exists when your "neutral consultant" is also a co-investor in the vendor’s distribution arm is the entire point of the structure.

A 17.5% guaranteed annual return over five years is not a venture term. It is a credit instrument with equity features. OpenAI has converted a slice of its growth optionality into something a private credit desk can underwrite.

TPG, Advent, Bain Capital, Brookfield, and Warburg Pincus collectively own thousands of portfolio companies. Each is now a Deployment Company prospect. Accenture took years to build a $3 billion AI practice. OpenAI committed more than that on day one.

The model is no longer the product. The integration is. Inference demand will increasingly flow through enterprise integration cycles, not consumer hype cycles. Harder to capture as a model vendor without engineers inside the customer.

Inference throughput gains come from software optimisation on top of the silicon, not from the silicon itself. Eigen AI optimises at the model layer. Clarifai optimises at the system layer. Nebius has now bought both. Token Factory is Nebius productising that optimisation as workload margin on top of the GPU rental box. Operators selling raw GPU hours compete on price for a workload that gets won on throughput per dollar.

NVIDIA owns equity in Nebius and just watched its portfolio company buy two inference-software stacks in eleven days. Software that wrings more tokens out of each GPU drives utilisation up, which drives more GPU purchases. Vertical integration with extra steps.

1.2GW broken ground in Missouri. 1.2GW more secured in Pennsylvania. Capacity at this scale does not get permitted, financed, and announced on speculation. Nebius is sizing for the inference workload it expects to win, not the GPU rental contracts it already has.

The unit of inference demand is shifting from chat turn to agent task. Chat fires one inference call per user message. A self-improving agent fires reflection, skill creation, and sub-agent calls on top of the primary task. Hermes hitting 224 billion daily tokens on OpenRouter alone is the public-facing evidence that the multiplier is live in production.

OpenRouter is the visible slice. Claude Code, Cursor, ChatGPT’s agentic modes, and every enterprise agent stack that does not route through public model marketplaces are invisible on this leaderboard. The total agent-driven inference workload across the market is a multiple of what the public rankings show.

The demand curve compounds. Self-improving agents take on more ambitious tasks as they accumulate skills, and more ambitious tasks generate more tokens per task. Throughput becomes the binding constraint on what users attempt next, which becomes the demand signal pricing every inference contract on the market.

Every other story in this issue prices against an inference workload class that did not exist eighteen months ago. Hermes is the public-facing proof that the workload class exists, is growing fast, and is not the hyperscalers’ captive market.

Capital is pre-paying for non-GPU silicon two years before it ships. The venture market is acting as though the inference architecture question is settled. The risk is not that the chip is bad. The risk is that someone else’s chip ships first and locks the workload before Fractile gets to market.

Pat Gelsinger ran Intel for four years and walked out with full visibility into where every wafer in the industry is sitting. He has now backed Fractile at the seed and the Series B. That is the most personal endorsement available in inference silicon right now, and it is not a passive cheque.

Fractile’s headline claim is 1,200 tokens per second against ~40 t/s on current GPUs. The chip doesn’t ship until 2027, so there is nothing to benchmark yet. A 30x throughput claim, if it holds, means a workload that takes a month on current GPUs takes a day. The capital is acting as though the if is settled.

AWS now sells three Claude tiers. Bedrock legacy, where AWS operates everything through its own Converse/InvokeModel API. Claude in Amazon Bedrock, where AWS still operates but with the Anthropic Messages API surface. Claude Platform on AWS, where Anthropic operates and AWS only bills. Each rung up gives the customer more Anthropic and less AWS. FedRAMP, IL4, IL5, and HIPAA-ready compliance only exist on the bottom two rungs.

Same-day feature shipping is a privilege AWS extends to exactly one model vendor. Meta, Mistral, Cohere, AI21, Stability. Every other model on Bedrock is on the previous release cadence. AWS just signalled which model vendor matters most, in writing, inside its own marketplace.

Anthropic manages rate limits, tier advancement, pricing, and customer support on the top rung. AWS keeps authentication, IAM, Marketplace billing, the compute commit drawdown, and the data centre colocation. AWS got the parts a hyperscaler monetises at scale. Anthropic got every part that touches the customer’s actual workload.

Azure and GCP now have a template. Third-party model vendors on hyperscaler marketplaces can ask for the same terms. The hyperscaler as inference operator narrows into a regulated-compliance niche. The hyperscaler as billing and identity layer is the new surface area.

Multi-gigawatt sites get built when a hyperscaler signs the offtake. The hyperscaler pipeline this quarter is going to named geographies with named partners. None of those geographies is West Texas. None of those partners is Fermi.

The headline loss is mostly non-cash. The cash burn is still material. The $785 million in Q1 equipment financing buys time, not customers. Heidrick & Struggles has the CEO search mandate. New CEOs do not typically run toward an active litigation file.

17GW of design capacity without a tenant is not the start of an asset story. It is the start of a balance sheet story. Power-first speculation without an anchor tenant is over as a financing pitch. Buyers want signed offtake before they sign capital commitments.