BREAKING: Cerebras Systems CEO Andrew Feldman claims a $25B backlog. But the numbers don’t add up. This isn’t a signal of dominance—it’s a carefully orchestrated IPO prelude.
Context: Why Now? Cerebras, the AI chip startup known for its wafer-scale WSE-3 processor, has historically been a niche player—~$300M revenue in 2022, ~$500M in 2023, and perhaps $1B in 2024. The $25B figure is 20–30 years of revenue at current run rate. The company is rumored to be prepping for an IPO in 2025, and this announcement is a classic valuation inflator. The market context: NVIDIA dominates AI compute, but demand is so intense that even second-tier alternatives are being hoarded. Cerebras targets large-scale training clusters, claiming WSE-3 is 2–3x faster than H100 for specific workloads. But speed without precision is just noise; the market pays for timing.
Core: The Numbers Don't Lie—But They Can Be Bent I’ve spent years auditing smart contracts and liquidity pools. When I see a $25B backlog claim from a company that hasn’t delivered $10B cumulatively, I dig into the assumptions.
1. Financial Reality Check: - If WSE-3 costs ~$100K per chip (my estimate from public pricing), $25B implies 250,000 chips. Cerebras' annual production is ~100–200 systems, each containing one chip. Even with scaling, hitting 250,000 chips would take decades. - The industry standard for “backlog” includes non-binding letters of intent (LOIs) and multi-year framework agreements. These are not purchase orders. Based on my experience with the 2017 Parity multi-sig vulnerability, I know that unverified claims can cause real damage when investors act on them.
2. Customer Concentration Risk: - G42, an Abu Dhabi AI firm, is Cerebras' biggest known customer. Their contracts are in the $100M range. To get to $25B, you’d need 250 G42s. Even with sovereign wealth funds, this is implausible. - The U.S. Department of Energy has a few small contracts. No hyperscaler (Microsoft, Google, Amazon) has publicly committed. That’s a red flag.
3. Competitive Comparison: - NVIDIA’s data center revenue was $47.5B in FY2024. Cerebras claiming half of that in a single backlog is akin to a DeFi protocol claiming $50B TVL without a working frontend. The 2020 Yearn surge taught me that growth metrics must be verified against on-chain data. Here, the only data is a CEO’s word.
4. Infrastructure Constraints: - 250,000 WSE-3 chips would require ~375MW of power (15kW/chip), plus cooling and networking—total ~500MW. That’s a nuclear reactor’s worth. No data center of that scale is planned for Cerebras. - TSMC’s wafer-scale packaging capacity is limited. Even if orders are real, delivery will stretch into the 2030s.
Contrarian: The Unreported Angle—This Is an IPO Pump, Not a Pipeline The market is bullish on AI, and Cerebras is riding that wave. But the $25B figure is structurally flawed: - Double Counting: Some of that backlog may include “commitments” from investors like G42, who also hold equity. That’s circular capital, not real demand. - YoY Decline in Conversion: Many LOIs expire. In 2021, I analyzed BAYC floor price liquidity and saw how whale commitments vanished when prices dropped. Same dynamic here: if NVIDIA launches B100/B200 with better efficiency, these “commitments” evaporate. - Trust Deficit: Feldman’s history includes aggressive claims. The 17 reveals the true cost of trust. In crypto, we call this a “fake APY”—promising yields that can’t be sustained.
The BAYC crash wasn’t an art correction; it was a liquidity trap. Cerebras investors should watch for the same pattern: a sudden flood of supply (IPO shares) and no real buyers.
Takeaway: What to Watch Next Ignore the $25B headline. Watch these signals: - SEC Filing: If Cerebras files an S-1, the backlog will be broken down by binding vs. non-binding. That’s the truth. - Customer Names: If Microsoft or OpenAI emerges as a client, the thesis changes. Until then, treat it as a marketing stunt. - On-Chain Reality: Compare Cerebras’ actual chip deliveries to their public cloud offerings. If they can’t fill existing orders, the backlog is hot air.
Speed kills. Precision saves capital. This article is your first-person audit: I’ve seen the same playbook in 2017 with Parity, in 2020 with Yearn, and in 2022 with Terra. The only difference is the asset class. The question remains: Will you trust the narrative or the data?