Cerebras Systems filed an updated S-1 prospectus late Monday that prices the company at $115 to $125 a share for 28 million shares of Class A common, raising about $3.5 billion at the top of the band and pricing the company at roughly $26.5 billion at the midpoint. [1] Order indications already past $10 billion are intact going into Tuesday's bookbuild day. [2] Pricing is set for May 13. The shares list on the Nasdaq Global Select Market under the ticker "CBRS" on May 14. [3] The headline number that anchored Monday's papers — including this one — is now off the page. The syndicate halved it.

This paper led Monday's coverage with a $40 billion target at roughly seventy-eight times trailing sales and called it the AI-capex referendum meeting Berkshire's $397.4 billion in cash. The S-1 filed Monday night cuts that target by about a third while leaving the demand book at more than $10 billion. The bookbuild made an institutional choice with the cut: clear into the room — the room being Greg Abel's "no AI for AI's sake" Berkshire Q&A and Palantir's after-the-bell print — instead of pricing through it. The decision is the news. The decision was not made by demand. It was made by the syndicate's read of the tape demand will land in.

The tape is doing two things at once. Palantir reported Monday after the close: revenue $1.63 billion against $1.54 billion consensus, adjusted earnings per share of $0.33 against $0.28, U.S. commercial revenue up 133 percent year-over-year, government up 84 percent, full-year guide raised to $7.65 to $7.66 billion against an LSEG consensus of $7.27 billion — a 71 percent year-over-year growth bar. [4] Net income quadrupled to $870.5 million. The eleventh consecutive earnings beat. Karp's commentary on AIP and Department of Defense ramp turned the after-hours tape into a counter-receipt: the AI commercial backlog is real. [5] Berkshire's Saturday print and the Sunday Q&A produced the opposite document: $397.4 billion in cash, up $24 billion from the prior quarter; $11.35 billion in operating earnings against the $11.56 billion FactSet consensus — Abel's first miss as CEO; $10.1 billion of net income, double the prior year on $5.8 billion of investment gains; insurance profit up 28 percent; Geico off 34 percent; the Apple position trimmed to roughly $80 billion; March-only buybacks of $234 million, the first since May 2024. [6] Sherwood News on Sunday called Berkshire "the ultimate anti-AI stock." [7]

Two simultaneous receipts: Palantir says the AI commercial backlog has revenue this quarter; Berkshire says it does not have a multiple worth paying. The Cerebras syndicate had to price into both. It chose the band that lets the order book absorb the cut without breaking demand on opening day.

The math the cut implies. At $26.5 billion at the midpoint, against the company's most recent disclosed revenue base, Cerebras prints at roughly fifty times trailing sales — still a generous multiple by any pre-2024 benchmark, but no longer the seventy-eight-times number that read as a referendum. The IPO retreats from referendum and toward "premium AI compute pure-play" framing. The $40 billion target was an aspiration; the $26.5 billion midpoint is a product. The syndicate priced the product. The $10 billion demand book is the buy-side telling the syndicate that the product clears, perhaps clears with a pop, and that the syndicate left money on the table to make sure of it. Secondary-market quotes pre-IPO have closed the gap with the band into the $26 to $28 billion range — so the public-market clearing price is approximately the band's ceiling, not the bookbuild's IOI-implied premium. [8]

The Cerebras float is $3.5 billion of new capital. Twenty-eight million shares is a meaningful supply for a deal of this size; the syndicate could have priced fewer shares at a higher band and reached the same proceeds. The choice to issue volume at a moderate price is the syndicate's lockup-period bet: a deal that prices at $26.5 billion and trades up has a different post-listing arc than a deal that prices at $40 billion and trades flat. The first creates retail demand for the secondary; the second creates a holding pattern. The May 14 listing day will be a microstructure read on whether the cut produced clearing or under-pricing. Either is acceptable to the syndicate. The unacceptable outcome was a $40 billion print that broke book on day one. That outcome is the one the cut prevented.

The reference set the prospectus invokes is now very short: Nvidia at one trading multiple; Palantir at another; AMD's Tuesday-after-the-close print at a third. AMD is expected to report Q1 revenue of about $9.84 to $9.87 billion (up roughly 32 percent year-over-year) with data-center revenue near $5.56 billion (up about 51.5 percent), the MI355 ramping into the first half of 2026 and the MI450 inflection landing in the third quarter. [9] Cathie Wood's ARK funds sold roughly $80 million of AMD into the print — the bear-tape signal. The OpenAI 6 GW and Meta 6 GW H2 2026 deals are AMD's bull case; Lisa Su's "anxious customers" line on MI450 is the bear case. The Cerebras prospectus sits next to AMD's Tuesday print on every IPO investor's desk. The cut to $26.5 billion gives Cerebras headroom against a Tuesday after-close that may surprise either way.

The institutional signal in the cut. The May 4 paper read the $40 billion target as the syndicate's first answer to whether the AI-capex regime — Microsoft's $190 billion calendar-2026 capex, Google's $185 to $190 billion, Amazon's $200 billion, Meta's $135 to $145 billion — could float a pure-play infrastructure provider at premium multiples through a roadshow window dominated by Berkshire. The Monday-night S-1 is the second answer: the syndicate could have priced through the room; it chose to price into it. This is not a vote against AI capex. It is a vote against pricing as if the regime were uncontested. The Cerebras deal will list, will likely trade up, and will set the multiple the next AI-pure-play files behind. The next deal can price at $40 billion. This one will price at $26.5 billion.

The footnote that runs through the SpaceX prospectus due to file publicly between May 15 and May 22: the Cerebras cut is now a reference comparable in the SpaceX risk-factor set. [10] Cerebras at $26.5 billion is about 1.5 percent of SpaceX's Polymarket-implied valuation midpoint. SpaceX is not an AI infrastructure deal; it is a launch-and-Starlink deal that has AI-adjacent contracts. But it will price into a tape where the headline AI infrastructure bellwether had its target halved. The institutional read is whether the SpaceX bookbuild interprets that signal as discipline — pure-play AI compute had to come down to clear, but a launch-and-satellite operator with Department of Defense backlog may not — or as precedent: the public-market multiples for any deal floated against the AI-capex backdrop have a ceiling, and that ceiling is no longer assumed to be where the private rounds priced.

Berkshire's Tuesday-morning analyst notes are the institutional anchor running through the same tape. [11] $397.4 billion in cash is a vote with a number on it, and the number is bigger than the public-market value of any of Cerebras, Palantir, or AMD individually. Abel's "no AI for AI's sake" was not Saturday-Q&A boilerplate; it was a position that has now produced a buyback line item — $234 million, March only — and a market read on Berkshire as the dry-powder reservoir against $710 billion of Mag-Five capex. [12] Sherwood News and Tker.co are running the "ultimate anti-AI stock" framing because the framing has institutional resonance. [13] The Cerebras syndicate priced into that resonance on Monday night.

The one number worth tracking through Wednesday morning. May 13 pricing, May 14 listing, $115 to $125 band, 28 million shares, $26.5 billion midpoint, $10 billion order book, fifty-times trailing sales. The deal that lists at any of those numbers is a different deal from the one that did not list at $40 billion. The paper's position from yesterday — that the AI-capex referendum was inside this roadshow — still holds. The referendum's first question was answered Monday night. The answer was: cut the headline.

-- THEO KAPLAN, San Francisco