Cerebras: Real AI Growth, Priced Like It's Already Proven

Baron Global Opportunity Fund just made a revealing pair of trades. It sold the last of its Wix stock, and it bought a new position in Cerebras Systems (CBRS), the AI inference-chip maker that went public in May.

Both moves came from the same discipline. In its second-quarter letter, the fund said it exited Wix because core growth was decelerating: bookings and new-user cohorts had softened, and the company's newer AI offering was not big enough to offset the slowdown, which undermined the durable subscription revenue and expanding free cash flow at the heart of the thesis. It redeployed the money toward what it called "clearer growth trajectories." Cerebras is the sharpest test of whether that filter holds on the buy side too.

The fund had a standout quarter from which to make these calls, with institutional shares up 26.7% in the second quarter versus a 14.9% gain for its MSCI ACWI benchmark. And its buying spree tells you where it sees the next growth cycle: alongside Cerebras it added Alphabet, Samsung Electronics on an AI memory bet, plus power and industrial-component names — a rotation from consumer internet into the AI-infrastructure supply chain.

The $20 billion contract is the growth — and the risk

Cerebras is not an AI pitch with no numbers behind it. It reported first-quarter core revenue of about $191 million, up 92% from a year earlier, and it guides to roughly $860 million of core revenue for 2026, about 69% growth. Its differentiator is the wafer-scale engine, a single giant chip purpose-built for fast inference rather than a rack of linked GPUs, which it markets as far faster than leading GPU alternatives.

The demand is anchored by one enormous deal: a multi-year agreement with OpenAI to deliver 750 megawatts of low-latency inference capacity, worth more than $20 billion and running through 2028. That single contract is about 23 times the company's full-year 2026 revenue guidance. OpenAI even lent Cerebras $1 billion to help fund the build-out. As an anchor agreement goes, it is hard to imagine a stronger one.

But it is also, in effect, the whole story.

$46 billion of market value on $860 million of revenue

The stock has been a volatile ride. Cerebras priced its IPO at $185 in May and closed the first day at $311, a 68% pop, before giving much of it back. By mid-September it traded near $195 — roughly 50% off its 52-week high — with a market value around $46 billion. Even after that pullback, it sits at about 68 times trailing sales. Nvidia trades at about 17 times sales, Broadcom about 19, Micron about 12.

That gap is the crux. A multiple well north of 50 times forward sales is what you pay for hypergrowth that is already proven and whose economics are secure. Cerebras' economics are doing the opposite this year: gross margin is guided down from 47% in the first quarter into the high 30s as it races to stand up hyperscale capacity, and management guides to a full-year core operating margin of roughly negative 30%. In plain terms, the larger the company gets this year, the more it is expected to lose.

Then there is concentration. A year ago, UAE-affiliated buyers including G42 generated about 86% of revenue, and that overhang helped stall an earlier IPO attempt amid a national-security review. The OpenAI deal was framed as diversification — but the backlog is now about $24.6 billion, largely a single customer. In the words of one analyst, the concentration did not go away; it rotated. If the OpenAI relationship wobbles, there is nothing behind the multiple. Insider unlocks around the reporting window add share supply on top of a thin float.

This is not a broken business. It is a genuinely fast one with a real contract. But measured by the very standard that sent Wix out the door — growth must convert into durable, reasonably priced cash — Cerebras is a good company whose stock is a full-price bet on flawless execution of a single deal. Until the valuation resets closer to the operating reality, or the contracted megawatts and the margin actually arrive at the pace the price assumes, "too early" is the more honest call.