Cerebras Fixed Its Concentration Problem. It Replaced 86% UAE Dependency With 86% OpenAI Dependency. Now OpenAI Is Also Its Lender.

Cerebras Systems spent 18 months unable to go public. When it finally did — Nasdaq debut May 14, 2026, priced at $185, opening at $350, closing at $311.07 on first-day volume — the cause of that delay appeared to be resolved. The underlying condition had not changed.

The IPO delay traced to Cerebras's September 2024 S-1 filing, which was withdrawn after CFIUS scrutinized the $335 million equity stake held by G42, an Abu Dhabi-based technology investment company with historical ties to Huawei and Chinese entities. G42 resolved the review by divesting its Chinese holdings and converting its Cerebras stake to non-voting shares, completed March 2025. The refiled S-1 appeared in April 2026. The CFIUS concern was addressed. The revenue story embedded in that S-1 was not.

The concentration that didn't move.

Cerebras's FY2025 revenue: $510 million. Customer breakdown: Mohamed bin Zayed University of AI (MBZUAI, Abu Dhabi) — 62%. G42 (Abu Dhabi) — 24%. Combined UAE sovereign-linked entities: 86%.

MBZUAI purchases under the same G42 Framework Agreement. The shift from G42 to MBZUAI across fiscal years is a legal-entity change, not a reduction in counterparty exposure. The 86% concentration figure held steady from FY2023 through FY2025.

OpenAI revenue in FY2025: zero. The Master Relationship Agreement was signed December 24, 2025. Revenue recognition began February 2026. The $25 billion RPO that dominated Cerebras's IPO narrative was entirely forward-looking backlog at the time of the offering.

In Q1 2026 — the first quarter of actual OpenAI revenue — Cerebras reported $193.4 million GAAP revenue (+94% year-over-year, beating consensus of approximately $181 million). OpenAI contributed approximately $121.8 million — roughly 63% of quarterly revenue. Concentration rotated from UAE sovereign entities to a single US technology company. The magnitude remained functionally identical.

Patrick Moorhead of Moor Insights characterized it precisely: "Concentration rotated, it didn't go away."

OpenAI as customer, lender, and future equity holder simultaneously.

The OpenAI relationship extends well beyond revenue. Disclosed in the S-1 and 10-Q: OpenAI loaned Cerebras approximately $983 million total at 6% annual interest, secured by warrants for 33.4 million Cerebras shares at near-zero exercise price. The loan is current ($621 million) and non-current ($362 million) on the balance sheet. The termination clause is material: if the Master Relationship Agreement terminates for any reason other than OpenAI's own uncured breach, OpenAI can demand immediate full repayment and seize collateral.

OpenAI also committed to a potential ~11% equity stake at IPO. The company is therefore simultaneously Cerebras's largest revenue source, its largest creditor with callable debt, and a potential major shareholder — three positions that, taken together, give OpenAI structural leverage over Cerebras's business continuity without requiring any formal acquisition.

The $25 billion Remaining Performance Obligation was "overwhelmingly" attributed to OpenAI in SEC filings. The RPO delivery schedule: approximately 15% (~$3.75 billion) recognizable within 24 months; approximately 43% in months 25–48; approximately 42% after 2029. The long-dated back-loading means Cerebras's near-term financial performance depends heavily on OpenAI's willingness to take delivery and maintain the relationship — while OpenAI holds a callable $983 million loan secured by Cerebras equity.

What the Q1 earnings actually said.

Cerebras reported Q1 2026 earnings on June 23, 2026. Revenue beat. EPS missed: -$0.22 actual vs. -$0.16 consensus. The stock fell from approximately $226 pre-earnings to an absolute trough of $161.26 on June 25 — breaking below its $185 IPO offer price. The -58% decline from the $386.34 all-time high (IPO-day intraday) is not the "47%" that appeared in most financial coverage; that figure reflected the landing zone around $205, not the absolute low.

The primary catalyst for the selloff: gross margin guidance. Q1 actual gross margin was 46.5%. Q2 guidance: 36-38%. Full-year FY2026 guidance: 38-41%. Approximately 1,000 basis points of compression guided in one quarter.

CEO Andrew Feldman told CNBC the market "misunderstood" the guidance, attributing the compression to a structural arrangement in which Cerebras must rent back capacity from a large customer while building owned data center infrastructure. This explanation raised more questions than it answered: if the capacity-renting arrangement is structural — required at each new data center buildout — then the gross margin compression may recur at each expansion cycle. Analysts maintained Buy ratings across the board (10 of 10 covering analysts; price targets $273–$340, average approximately $299, implying roughly 71% upside from the current ~$175 price). The disconnect between unanimous analyst Buy ratings and a stock trading below IPO price reflects uncertainty about the gross margin profile's duration.

The S-1 disclosed three material weaknesses: revenue recognition controls, inventory accounting, and IT segregation-of-duties controls. All three were flagged at the time of a $5.55 billion IPO that raised money based substantially on a $25 billion RPO from a single customer. The controls for recognizing the OpenAI contract were themselves identified as material weaknesses.

Three things coverage got wrong.

"Largest US tech IPO of 2026": At the time of Cerebras's May 14 debut, the claim was accurate. SpaceX priced its IPO on June 11, 2026 and raised $75 billion — the largest IPO in history by any measure. Cerebras's $5.55 billion in proceeds is approximately one-thirteenth of SpaceX's raise. Coverage that called Cerebras "the largest US tech IPO of 2026" has not issued corrections.

"$10B+ OpenAI deal": The $10 billion figure appeared at announcement (January 14–15, 2026) as attributed sourcing in TechCrunch, CNBC, and Reuters — "a source familiar with the details." It did not appear in either company's official statement or press release. The S-1 language is "up to more than $20 billion" — explicitly an option ceiling, not a committed minimum. The base commitment is 750 megawatts of CS-3 inference capacity staged through 2028. Neither $10 billion nor $20 billion should be reported as a committed contract value without qualification.

"57x larger than NVIDIA GPU die": Mathematically accurate against the NVIDIA H100/H200 GH100 die (814 mm² → 46,225 mm² WSE-3 = 56.8x). Against NVIDIA's B200, which functions as a module of approximately 1,600 mm², the ratio is approximately 29x. Cerebras's marketing chose the more favorable comparison; coverage repeated it without noting the contemporaneous alternative.

Why Cerebras has a real technical moat anyway.

The "57x" debate is a distraction from a genuine architectural differentiation. LLM inference — the process of generating tokens — is memory-bandwidth-bound, not compute-bound. Each token requires loading model weights from memory. The WSE-3 delivers approximately 21 petabytes per second of on-chip SRAM bandwidth; an NVIDIA H100 delivers approximately 3.35 terabytes per second of HBM bandwidth. The ratio is roughly 6,000-to-1 in bandwidth.

Weights never leave the WSE-3 during inference. The memory wall that limits GPU cluster inference performance is eliminated by architecture, not by engineering optimization. Cerebras benchmarks claim approximately 6x faster token throughput than Groq on identical models (Llama 3.3 70B, Llama 4 Maverick). The architecture also supports training, unlike Groq's inference-only LPU design.

The constraint is the flip side of the architecture: the WSE-3's 44GB on-chip SRAM is the maximum model size that fits. Models larger than 44GB require external MemoryX storage, which reintroduces off-chip bandwidth bottlenecks that degrade the core throughput advantage. The chip is also the system — there is no "add half a WSE-3." Scale-out requires multiple CS-3 units connected via external fabric, which reintroduces the inter-chip bandwidth problem that wafer-scale was designed to eliminate within a single die.

The last independent NVIDIA alternative.

NVIDIA acquired Groq in December 2025 for approximately $20 billion — NVIDIA's largest acquisition — taking a non-exclusive perpetual license to the LPU architecture and approximately 80% of Groq's engineering staff including founder Jonathan Ross. Groq nominally continues as a legal entity operating GroqCloud. Choosing Groq for AI inference now means running on hardware that NVIDIA licenses and manufactures.

Cerebras is the only significant remaining independent alternative to NVIDIA for high-throughput AI inference at scale. The $95 billion day-one market capitalization — even after the subsequent 54% decline from that peak — reflects, in part, a market premium on independence. The IPO raised $5.55 billion in cash. That cash is the buffer for a company whose entire revenue-generating infrastructure depends on a single customer that also holds a callable loan.

The OpenAI deal is a lifeline and a leash simultaneously. If it performs — if the $25 billion RPO recognizes on schedule, if AWS Bedrock physical deployment in H2 2026 adds a second revenue stream, if FY2026 gross margins recover to the guided 38–41% — Cerebras has a viable independent path. If OpenAI terminates the MRA for any reason that allows loan acceleration, the $983 million callable obligation lands on a company that raised $5.55 billion and has yet to prove it can sustain gross margins above 40%.