On June 12, 2024, Cerebras Systems CEO Andrew Feldman announced a $25 billion order backlog. I audited 50 ICO whitepapers in 2017. I know a pumped number when I see one.
This isn't an opinion swing at Feldman. It's a quantitative reality check. The market is frothy. AI compute demand is real. But $25 billion in backlog for a company that booked less than $500 million in revenue last year is not a signal of dominance. It's a signal of desperation for an IPO narrative.
Let me be precise. Cerebras builds wafer-scale engines (WSE-3) with 4 trillion transistors. The chip is a marvel. In training large language models, it benchmarks 2–3x faster than NVIDIA H100 on certain workloads. But that's a technical footnote, not a financial fact.
Context: The Architecture of the Claim
Cerebras's business has three revenue streams: direct chip sales (CS-3 systems), compute-as-a-service (CaaS) contracts, and government grants. In 2023, total revenue was roughly $500 million. The company has never disclosed a formal backlog figure. Now, a CEO goes on podcast and drops $25 billion.
The first question any quant asks: what is the contract status? Is it a non-binding letter of intent (LOI)? A framework agreement? A set of options that expire if milestones aren't met? Feldman said "backlog," not "signed contracts." In capital markets, backlog sometimes includes multi-year estimates of potential demand. It's a marketing metric, not a GAAP liability.
Consider the customer base. The largest disclosed deal: G42, an AI firm from Abu Dhabi, committed to deploy x number of Cerebras clusters. G42 itself is a private company backed by sovereign wealth. Their total AI compute budget is likely in the billions, but binding? Unclear. Other customers include the U.S. Department of Energy and a few research labs. None is a hyperscaler like Microsoft or Google.
I trade the ledger, not the hype cycle. A $25 billion backlog must be decomposed into three components: (1) confirmed purchase orders with cancellation penalties, (2) multi-year service contracts with opt-out clauses, (3) unfunded letters of intent. Cerebras hasn't filed an S-1 yet. Until that happens, we're in the dark.
Core: Order Flow Analysis – The Numbers Don't Lie
Let's build a bottom-up model. Assume the average selling price of a WSE-3 core is $100,000 (a conservative estimate given its 4T transistor die and wafer-scale packaging). $25 billion backlog implies 250,000 chips. But each chip consumes 15–25 kW. At 20 kW average, 250,000 chips would require 5 GW of power continuously. The world's largest AI cluster today – Meta's – uses about 150 MW. 5 GW is equivalent to five nuclear reactors.
Power constraints are a hard wall. Data centers take 18–36 months to build. The global high-voltage transformer supply is bottlenecked. 5 GW of incremental load would require tens of thousands of new cooling units, and grid interconnection queues already stretch years.
Now look at timing. Cerebras's manufacturing partner is TSMC, using advanced CoWoS packaging. TSMC's CoWoS capacity for 2024 was estimated at 250,000 units of H100 equivalents. Cerebras's WSE-3 die is much larger, consuming 3–4 times the area. If Cerebras wants 250,000 WSE-3 chips, that would consume TSMC's entire CoWoS output for years. Unrealistic.
What about CaaS? Feldman hinted that many orders are for compute-as-a-service, meaning Cerebras retains ownership and sells compute time. That reduces upfront revenue but increases future operating income. This is sensible: it lowers customer capital outlay and allows Cerebras to amortize chip costs. However, CaaS revenue is recognized over 12–36 months. A $25 billion CaaS backlog would take a decade to recognize. The net present value at 10% discount rate is roughly $15 billion. And that is if the customers stay online, which they don't always do.
In 2020, I ran a DeFi arbitrage team that exploited liquidity gaps between Uniswap and SushiSwap. We found that 80% of listed yield was accounted for by impermanent loss and gas costs. Similarly, 80% of Cerebras's backlog may be accounted for by optionality, non-binding intent, and aggressive time horizon expansions.
The hidden data point: Cerebras's 2024 run-rate is still under $1 billion. To reach $25 billion backlog, they would need to sign customers materially larger than G42. The only players at that scale are the hyperscalers: Amazon, Microsoft, Google, Meta. None has publicly deployed Cerebras silicon. If the backlog includes those names, we would have seen their names in press releases. We haven't.
Volatility is the tax on undiscerned capital. This statement is precisely for moments like this. The capital flowing into AI compute is undiscerning. It sees a headline and assumes it's a competitive threat to NVIDIA. The reality is far more nuanced.
Contrarian Angle: Retail Sees Competition, Smart Money Sees an IPO Pump
Retail narrative: Cerebras is the NVIDIA killer. $25 billion in orders proves customers are abandoning CUDA for wafer-scale efficiency. Smart money knows better.
Contrarian point one: NVIDIA's datacenter revenue in fiscal 2024 was $47.5 billion. Cerebras's claimed backlog is just over half of that – but it's all future revenue, spread over years. In any given quarter, NVIDIA's revenue is still 100x larger. Cerebras's market share gain is at the margin, if at all.
Contrarian point two: Feldman's timing. Cerebras has been rumored to file for an IPO in late 2024 or early 2025. The CEO is inflating the backlog to drive valuation. In 2017, I saw dozens of ICOs claim partnership with "top-tier blockchain projects" that turned out to be press releases. The same pattern emerges: a vague, huge number, no verification, just when external capital needs to be raised.
Contrarian point three: The real victim here is not NVIDIA, but GPU miners. If AI compute demand overheats and power costs rise, crypto miners suffer. In 2022, Terra's collapse taught me to watch energy markets as a leading indicator for hash rate crashes. If Cerebras's claims lead to more AI data center builds, the local grid constraints will push electricity prices up. PoW miners with fixed power purchase agreements (PPAs) may see their costs double.
Yield without protocol is just delayed loss. This applies to the AI compute market: compute without verified demand is just speculative buildout. If the $25 billion backlog never converts to revenue, the capital wasted in manufacturing and data center prep will be a deadweight loss for investors.
What is smart money doing? Look at the options market for NVIDIA. After the Cerebras announcement, NVDA implied volatility actually narrowed slightly. Institutional traders are not pricing in a competitive threat. Instead, they see a smaller player trying to grab attention. The smart move: short Cerebras-linked SPACs or pre-IPO secondary shares, and long NVDA on any irrational dip.
Takeaway: The Market Pays for Clarity, Not Complexity
The key signal to watch is the S-1 filing. Not the podcast. The SEC demands audited backlog definitions and contract terms. Until that document appears, Cerebras's $25 billion is a number without a shadow.
For traders: - If you're in crypto, monitor energy ETFs (like TAN or PBW). Any AI buildout that diverts power from mining will hit PoW coins. - If you're in equities, avoid IPOs from private chipmakers until their backlog is formally defined. - If you hold NVIDIA, don't panic. The monopoly is intact until I see a hyperscaler publicly switch to Cerebras at scale.
I trade the ledger, not the hype cycle. The ledger shows one company with $47B in infrastructure revenue, and another with a press release. The trade is simple: sell the hype, buy the fundamentals.
Speculation is noise; fundamentals are signal. Compute without verified demand is delayed loss. Watch the SEC filing, not the headline.