Cerebras IPO: Wafer-Scale Performance, Wafer-Scale Pricing

Cerebras is described as having a $24.6bn RPO backlog and a marquee OpenAI partnership, but the stock trades at 18x FY28E sales, implying very strong execution is already priced in. Customer concentration is a key risk: UAE entities made up 86% of FY25 revenue, and while OpenAI could diversify revenue, it also creates a new dependency if it shifts inference in-house. The note frames the OpenAI relationship as both a strategic moat and a binary downside risk.

Analysis

This is a classic “good asset, bad price” setup: the technology can be real and still be a poor equity if the market has already discounted near-perfect scaling. The backlog is useful only insofar as it converts into repeatable, diversified revenue; right now the mix suggests the business is still being underwritten by a handful of strategic counterparties, which makes the equity more like a venture-style binary than a durable infrastructure comp. That matters because at this valuation, even modest execution slippage can compress multiples sharply before any fundamental deterioration shows up in the P&L.

The more interesting second-order effect is competitive: if the marquee AI customer relationship works, it validates the category and likely pulls larger hyperscalers and chip vendors deeper into custom silicon/inference optimization. That is a double-edged sword for the company because it expands the market, but it also accelerates competitor learning curves and increases the odds that major customers internalize the capability over a 12-24 month horizon. In other words, the partnership is simultaneously a demand signal and a roadmap for replacement.

The main catalyst structure is asymmetrical. Near term, sentiment can stay elevated on backlog conversion, design wins, and AI capex enthusiasm; over months, the market will start pricing concentration and customer control risk if revenue doesn’t broaden materially. The biggest tail risk is not a single lost order, but a shift in inference economics: if large buyers conclude they can replicate sufficient performance in-house or on a more standardized stack, the strategic moat narrows quickly and the multiple could de-rate well before growth does.

The contrarian take is that the market may be underestimating how hard it is to actually operationalize custom AI silicon at scale, which can keep the company relevant longer than skeptics expect. But that argument supports a trade in the private-markets/venture sense, not an 18x FY28E public-equity multiple. The setup looks more attractive as a volatility short or relative-value short than as an outright directional short, because the stock can remain expensive until a single customer relationship changes state.

Key Decisions for Investors

• Avoid initiating a long at current levels; the implied downside from multiple compression is larger than the upside from incremental backlog conversion over the next 6-12 months.
• If borrow/liquidity allows, short the name on any post-news strength or IPO-style pop; target a 20-30% drawdown if customer concentration remains unresolved into the next reporting cycle.
• Use call spreads to express a capped upside view for the next 3-6 months rather than naked long equity; the thesis is event-driven, not compounding-driven.
• Pair trade idea: long a diversified AI infrastructure beneficiary and short this name as the high-concentration, high-multiple leg; the spread should benefit if the market rotates from narrative to durability.
• Set a catalyst watch for any evidence of customer diversification or in-house inference moves over the next 1-2 quarters; a confirmed loss of strategic customer share would be a strong add-to-short signal.