Cerebras: Left on the Bench

OpenAI has found another way to sell speed. We are doubling our Cerebras short.

CORBANU · TRADE UPDATE · 29 SEPTEMBER 2026

Doubling Down

We propose taking our Cerebras short from 1.75% to 3.5% of capital. SemiAnalysis reports that OpenAI's GPT-6.1 Sol Ultrafast runs on Nvidia GPUs at low batch size, rather than Cerebras. OpenAI has passed over its speed specialist for the fast version of its new workhorse.

This is a consequential assignment. OpenAI pitches Sol 6.1 as approaching Astra's intelligence at one-fifth of its standard token prices, built for everyday use at scale. Sol Ultrafast is due in the next few days; Astra Ultrafast is available today. The SemiAnalysis report concerns Sol specifically. OpenAI launch · DevDay recap

If Cerebras is your strategic partner, your comrade in arms, why leave it off the frontline when speed is part of the sales pitch?

Why Nvidia?

Small batches mean serving fewer requests together. Think of a bus leaving before every seat is filled: passengers get moving sooner, but the operator gives up some efficiency. OpenAI is reportedly making that trade with its GPU fleet. We cannot infer the relative cost of the two systems from this report, but Nvidia evidently offers a usable alternative.

Perhaps OpenAI simply lacks enough available Cerebras capacity. Full machines could be good news for demand, though they are little help to a customer trying to launch a model. The announced 750 MW partnership comes online in stages through 2028. OpenAI partnership

Or OpenAI may prefer Nvidia for this job. Reliability, flexibility or economics could matter more than Cerebras's speed advantage. You can build an extraordinary chip and still lose the order.

The third possibility is that testing and integration on Cerebras missed the deadline. One late deployment would be understandable. If Cerebras struggles to keep up as OpenAI releases new models, the partnership becomes much less valuable. This launch gives us reason to worry about that, although one launch cannot establish a pattern.

These explanations could overlap. None requires a falling-out between the companies; any could leave Cerebras with a smaller role than investors expect.

A Very Expensive Stopgap

Cerebras already proved it could do useful work for OpenAI. In August, it powered the GPT-5.6 Sol Ultrafast preview at up to 750 output tokens per second, according to OpenAI. The preview was limited, but it was a working product. Having done that, Cerebras ought to have been well placed for the next Sol. OpenAI announcement

Our original short argued that the valuation required durable growth from a company whose biggest strategic customer was developing its own inference silicon, Jalapeño. At that September snapshot, the Street expected $7.83 billion of 2028 revenue and 56% gross margins. Our $118 base case allowed for substantial growth with less generous economics.

Sol running on Jalapeño would be a more direct blow. That is not what has been reported. The concern here is that Nvidia appears sufficient for this speed tier while OpenAI develops its own alternative. Cerebras could lose future work well before Jalapeño arrives.

We are more convinced that OpenAI regards Cerebras as a stopgap. Existing contracts can produce revenue without securing the follow-on work needed to justify the valuation.

We also flagged Anthropic's prospective IPO as competition for investors' AI allocations. Keeping OpenAI's business is the more pressing concern. If Cerebras cannot hold a central role with its marquee partner, why should less committed customers build around it?

The Bigger Short

At 3.5% short notional, a 50% rise in Cerebras would cost roughly 1.75% of portfolio capital, before funding and fees. Increasing the short makes this a directional bet; it no longer preserves the original CXMT pair's beta neutrality.

A broad Sol rollout on Cerebras would weaken this catalyst. An expanded OpenAI commitment or large, profitable independent customers would challenge our deeper thesis.

For now, the customer that gave Cerebras its best advertisement has reportedly chosen Nvidia for the next version. We are doubling the short because we think that choice says more about the relationship than the partnership announcement did.