Cerebras’s
FY2025 GAAP profit was mostly a one-time accounting gain, not operating
leverage — and its S-1 disclosed that 86% of revenue depends on two related UAE
customers.
Cerebras Systems reported net income of
$237.8 million in fiscal 2025 — its first profitable year on paper, arriving
just before a May 2026 IPO that valued the company at roughly $95 billion on
its first day of trading. Strip out one line item, though, and the picture
flips: excluding a $363 million non-cash gain from revaluing a pre-IPO
financing instrument, Cerebras actually lost $145.9 million at the operating
level that year — a wider loss than the year before. The headline number and
the underlying business tell two different stories, and the gap between them is
the most important thing to understand here.
Breaking the “Reticle
Limit”
Cerebras was founded in 2016 by five
engineers who had worked together at SeaMicro, a microserver startup Andrew
Feldman co-founded and sold to AMD for $334 million in 2012. Feldman (CEO),
Gary Lauterbach, Sean Lie, Jean-Philippe Fricker, and Michael James set out to
solve a problem they believed GPUs were never built for: training large neural
networks means moving enormous data between thousands of small chips, and that
inter-chip communication — not raw compute — was becoming the real bottleneck.
Their bet was that breaking through the semiconductor industry’s “reticle
limit” — the maximum size a chip can be manufactured at using standard
lithography — could put an entire large-model workload on one giant piece of
silicon and avoid that data-movement tax entirely. It took roughly three years
of relative stealth, building a chip 56 times larger than a typical GPU die
while engineering around defects that would normally ruin a wafer that large.
Argonne National Laboratory took delivery of the first CS-1 system in 2019,
giving Cerebras a marquee reference customer before it had any commercial AI
customers at all.
A Moat
Built From a Decade and Hundreds of Millions of Dollars
What makes wafer-scale computing hard
to copy is two stacked barriers. First, defect-tolerance: a single flaw
anywhere on a standard wafer would ordinarily ruin the whole chip, so Cerebras
had to design in redundancy across the entire surface. Second, none of that
hardware is useful without a from-scratch compiler and software stack that maps
AI workloads onto an architecture nothing like a GPU. Both took roughly a
decade and hundreds of millions of dollars to reach a shipping product, part of
why no pure-play wafer-scale competitor has emerged since 2016. The bet paid
off in speed: on some benchmarks, Cerebras runs faster than rivals — 1,800
tokens per second versus SambaNova’s 1,084 and Groq’s 750.
The
Customer Concentration That Paused an IPO
Here’s the risk that shaped Cerebras’s
entire path to the public markets: 86% of fiscal 2025 revenue came from two
related UAE entities — Mohamed bin Zayed University of Artificial Intelligence
at 62.0% and Group 42 (G42) at 24.0%. MBZUAI alone represented 77.9% of
year-end accounts receivable. This wasn’t boilerplate risk-factor language; it
was disclosed directly in the S-1 as a single point of failure, and it had real
consequences. Cerebras’s first IPO attempt, filed in September 2024, was paused
by a CFIUS national-security review over whether its technology could reach
China through G42’s broader business relationships. G42 is also simultaneously
a top customer and an investor, having put $335 million directly into the
company — a dynamic that has already drawn regulatory scrutiny once. What
changed the calculus was a single new customer: in January 2026, OpenAI signed
a compute deal worth over $10 billion, covering roughly 750 megawatts of
capacity. That relationship did more to de-risk Cerebras’s customer base than
additional G42 revenue could have, and the company refiled its S-1 in April
2026, IPO’ing the following month.
Reading Past the
Headline Number
That non-cash gain is the clearest
lesson in the whole dossier. The company’s own fiscal 2026 guidance — a core
operating margin of negative 28% to negative 32% — is the more honest read of
where things stand: a loss-making, capital-intensive hardware company, full
stop. Quarterly cloud gross margins told a similarly unsettled story through
2025, swinging from 68% down to 16% and back to 21%, a sign the company is
still learning to price newly built data-center capacity. And the trailing P/E
of roughly 140x is computed off an EPS figure still carrying the tail of that
one-time gain — anyone pricing the stock off trailing earnings is pricing off a
number that doesn’t reflect the underlying economics.
The stock itself has moved just as fast:
it priced at $185 a share on May 14, 2026, popped as much as 108% intraday, and
closed up 68% at a roughly $95 billion market cap — before round-tripping to
roughly $199 by late July 2026. A securities investigation was announced by
Kaplan Fox & Kilsheimer on July 17, 2026; the dossier treats that as
pending and unresolved. The competitive field also narrowed in December 2025,
when Nvidia struck a roughly $20 billion deal to license Groq’s core
technology, collapsing the independent inference-accelerator field from three
credible players to two, and confirming the dominant incumbent is willing to
buy its way past architecture-level threats.
Watch the Full Breakdown
Cerebras is a genuine engineering
achievement — a decade-long bet against industry orthodoxy that shipped a
working alternative to the GPU. Whether it becomes a durable business depends
on diversifying past two customers and proving out real operating margins, not
one-time gains. For the full financial charts and how Cerebras stacks up
against Nvidia, AMD, and SambaNova, watch the complete deep-dive video on the
Company Narratives YouTube channel.
Comments
Post a Comment