Cerebras Systems is now eight trading days past its IPO. The chip-design company opened at $350 last Thursday against a $185 offer price — an 89% first-day pop that priced the company at roughly $34 billion, in the upper half of the pre-IPO bookbuild range Erik Torenberg circulated in April [1]. Through Friday's close the stock held above $300. The Memorial Day weekend put a comma in the price action; Tuesday will resume it.

The paper's Sunday read of Day Six past the IPO framed the position as the OpenAI concentration being simultaneously the floor and the ceiling. Two more trading days have not changed that. The 750-megawatt OpenAI build-out remains the named anchor in the S-1; it is also the reason the ceiling is structural rather than negotiable.

The mechanics are familiar to anyone who has read a 10-K: when one customer represents more than a third of forward bookings, the equity story becomes a derivative of that customer's roadmap. Cerebras's 750-megawatt commitment from OpenAI is the largest single named contract in any AI-infrastructure prospectus filed this year [2]. The valuation that allowed the IPO to print at $185 was a valuation that took that contract as fixed. The valuation the market gave the stock at $350 took it as fixed plus pipeline. The difference between those two views is the part the eight-day tape has been pricing.

OpenAI's own infrastructure posture is the upstream variable Cerebras cannot control. The Stargate program, the Microsoft training compute, the Anthropic-versus-OpenAI competitive frame that Anthropic's $300 million Stainless acquisition was structured to address [3] — each is a vector that determines whether OpenAI's 750-megawatt Cerebras commitment grows, holds, or moves to a different vendor on a different timetable. The closer the integration with OpenAI, the harder it is to grow with a non-OpenAI customer at scale.

This is the structural ceiling. A pure-play AI-infrastructure company priced as a derivative of a single AI lab cannot rerate above the rate at which the lab rerates without acquiring a second customer of similar scale. Cerebras's path to that second customer runs through the sovereign-compute deals it has been pursuing in the Gulf — the G42 partnership in the UAE, the inference clusters in Saudi Arabia — and through the federal-laboratory pipeline (Argonne, Lawrence Livermore) that has supplied much of the company's revenue history. None of those is a 750-megawatt commitment, and none of them is the kind of named anchor that an IPO can price against.

The Day 8 read is therefore not a price read but a calendar read. Tuesday morning's open will be the first session in which holders who took the 89% pop on Day 1 can settle into either trim-or-add discipline. The lock-up overhang on insider sales does not arrive for another 174 days; the analyst initiations from the IPO underwriters, by convention, will be published in the second half of the calendar week. The next two information events that matter are therefore initiations and any OpenAI capacity announcement that confirms or repaces the 750-megawatt commitment.

In the meantime, the stock holds above $300 on a structure that prices a single customer as the company. The wafer-scale chip — the WSE-3, larger than a dinner plate, capable of unified-memory inference at scales Nvidia's H100 cannot match in a single die — is real engineering [4]. The ceiling is not the engineering. The ceiling is the contract.

-- THEO KAPLAN, San Francisco