Cerebras Has a $25.4 Billion Backlog, and One OpenAI Agreement Is Behind Much of It

Cerebras reported Q2 adjusted revenue up 103% YoY to $209.9M and raised its full-year adjusted revenue outlook to $880M–$890M. The company ended June with $25.4B of remaining performance obligations, largely tied to a December OpenAI agreement (750MW committed, with an option for 1.25GW), but expects to recognize only ~22% of the backlog ($~5.6B) over the 24 months ending June 30, 2028. With shares around $215 (down ~44% from the $386.34 52-week high) and valuation near 58x expected 2026 adjusted revenue, the core positive is strong demand/operational execution, tempered by slow backlog-to-revenue conversion risk and customer concentration.

Analysis

The market should read this less as a growth print and more as a financing-and-execution test. The economic value of the contract book is heavily discounted by timing, concentration, and buildout risk: cash flow is gated by data-center ramp, not by customer intent, so the equity should not trade as if backlog were near-term revenue. That makes CBRS vulnerable to multiple compression if investors start underwriting the backlog at face value; the first catalyst is not the next headline contract, but evidence that recognized revenue is accelerating faster than the installation curve.

Second-order, this is a signal that inference capex is broadening beyond the incumbent GPU stack, which is incrementally negative for NVDA’s share of wallet in customized inference deployments but not necessarily for overall AI spend. The real external beneficiaries are the power, cooling, and electrical-grid vendors that sit upstream of capacity expansion; the risk is that any hiccup in financing or permitting converts those adjacent winners into delayed revenue as well. Over 6-18 months, the key question is whether this becomes a durable multi-customer platform or remains an extraordinary one-customer project.

Contrarian view: consensus is likely over-capitalizing the backlog because it treats contracted demand as if it were already monetized. The missing variable is counterparty optionality — if OpenAI changes architecture, timing, or mix, the valuation embeds too much certainty for a business still absorbing losses. The thesis is falsified if quarterly revenue from the anchor customer keeps stepping up and gross margin expansion continues without materially higher working-capital needs; otherwise, the stock remains priced for perfection.

Key Decisions for Investors

• Avoid chasing CBRS on backlog headlines; wait for 1-2 quarters of visible revenue conversion. If you must own it, size as a call-option on execution, not a core growth multiple.
• Relative-value idea: short CBRS / long NVDA into strength. Thesis: CBRS is pricing in too much of a single-customer buildout, while NVDA still captures the broader AI capex base. Cover if CBRS demonstrates sustained conversion above current run-rate and no increase in customer concentration.
• Buy a small starter short only on rallies above the post-print range, with a 1-3 month horizon. Risk/reward improves if the market starts asking for proof of backlog monetization rather than repeating the top-line story.
• Set a watch item on the next filing for the share of backlog tied to the anchor customer and any rise in working-capital/financing needs. A deterioration there would be the cleanest trigger to add to a CBRS short.
• If looking for cleaner AI exposure, prefer NVDA over CBRS on a 6-18 month basis: same secular theme, far less single-name execution risk and no dependence on a few mega-contracts to justify the multiple.