App Economy Insights — Cerebras: Demand Is Not the Problem

The record IPO, then reality

• Cerebras pulled off the biggest semiconductor IPO: the May 2026 deal priced at $185, surged to $386 intraday and closed at $311 — a 68% pop — then spent six weeks giving most of it back. Its first earnings report triggered a record two-day plunge that briefly pushed it below the IPO price.
• The pattern App Economy keeps flagging: the day-one pop is often short-lived, and the first earnings call is where the hype meets reality. The quarter itself was strong — revenue nearly doubled, a beat, and guidance raised above consensus.

What Cerebras actually builds — wafer-scale

• A typical AI processor is about the size of a postage stamp; Cerebras builds one the size of a dinner plate — an entire silicon wafer as a single chip. Its Wafer-Scale Engine packs ~4 trillion transistors vs ~200 billion on NVIDIA's Blackwell. It's sold inside a machine (the CS-3), not as a bare chip.
• The point is speed: keeping a whole model on one giant chip avoids the slow shuffle of data between thousands of smaller GPUs, letting Cerebras generate tokens roughly an order of magnitude faster — 1,000+ tokens/sec on frontier models. The pitch: fast AI is worth a premium because it gets more done.

Two revenue streams — and where gravity is shifting

• Hardware: CS-3 systems and clusters sold to customers who run them in their own data centers — $111M in Q1 FY26 (+59% Y/Y), the legacy side.
• Cloud & other services: customers rent the same inference horsepower by the token from Cerebras-operated data centers — $83M (+178% Y/Y), growing 3× faster than hardware and the strategic center of gravity going forward.

The moat — sidestepping the chokepoints

• Cerebras uses no high-bandwidth memory (HBM), no CoWoS advanced packaging and no bleeding-edge 3nm process — leaning instead on plentiful SRAM and mature 5nm. While the industry fights over scarce TSMC and HBM allocation, Cerebras isn't standing in that line.
• It's also the only AI-accelerator maker building its systems entirely in the US.

Customer concentration — UAE-heavy

• Cerebras was, until recently, almost entirely dependent on G42, an Abu Dhabi AI group; two Abu Dhabi entities (G42 and MBZUAI University) drove the bulk of 2025 revenue. Its risk disclosures now name four significant customers: OpenAI, G42, MBZUAI and AWS.
• The bull case rests on the first and last names — turning a UAE-heavy base into broader Western enterprise demand. But OpenAI is only starting to ramp (back-half-2026) and AWS revenue doesn't begin until 2027.

OpenAI wears three hats

• The deal: a multi-year agreement for more than $20B of Cerebras compute. The twist: to win it, Cerebras handed OpenAI warrants (the right to buy its stock at a discount).
• The accounting: those warrants count as a discount, so part of what OpenAI pays never shows up as revenue — the reported figure understates the true OpenAI business, and the gap widens as the deal ramps. OpenAI is also a $1B working-capital lender. The anchor customer is helping finance the company it buys from.

Q1 FY26 by the numbers

• Revenue +94% Y/Y to $193M ($12M beat); gross margin 45% (+3pp); operating margin −8% (+21pp); net loss narrowed to $14M, or $(0.22)/share (from $(0.46)). Adjusted EBITDA flipped positive to $13M (from −$15M) — a clean beat-and-raise.
• Cash flow tells the other story: OCF was +$12M (vs −$55M a year ago) but FCF was −$120M as capex rose to $132M. Balance sheet: $3.3B cash & investments, a $1.0B loan from OpenAI. FY26 guide: revenue $855–865M (+69% at midpoint), gross margin 38–41%, operating margin −28% to −32%.

The real bottleneck is real estate

• Feldman (CEO): "Demand is not the constraint. Supply is not the constraint. The constraint is data centers." That reframes Cerebras as an infrastructure-buildout company as much as a chip designer — leasing capacity back from a customer while racing to add data centers across the US, Europe and a new 120MW Bell Canada site.
• Cerebras has ~$2.3B of future data-center leases not yet on the balance sheet; it can sell compute faster than it can build the buildings to house it.

Margins are the problem, not pricing

• Komin (CFO): "Because there's tremendous demand, we've been able to see higher pricing from existing customers." Q1 core gross margin (47%) came in above plan; rivals' HBM costs lifted the price floor, so Cerebras is raising prices rather than discounting.
• The guided margin step-down is a capacity-cost problem (renting capacity back from a customer), not a pricing one. Management wants investors to look past a near-term trough toward a 60% gross-margin target; hardware margin should slide toward the low-30s as Q1's was inflated by one-time pricing.

What to watch next

• Backlog conversion: ~$25B in remaining performance obligations vs $865M of 2026 revenue — "the entire investment debate." Watch how much warrant contra-revenue suppresses reported OpenAI sales and how fast hardware revenue rotates into Cerebras Cloud.
• The buildout math: $132M quarterly capex, $2.3B of future leases, an $850M revolver drawn against a deficit balance sheet — can internal cash flow fund expansion, or is another raise (more dilution) coming? The concentration test: AWS and enterprise revenue showing up large enough to shrink the UAE share; the wildcard is selling "decode" inference to companies that already own GPUs.