Cerebras Compute Leasing Bet Trap: 10-Billion Orders Hard to Fulfill, Contracts Lock Up Future Cash Flow

Who Gets Targeted

The victims are mostly founders of SMEs rushing into the large-model track, digitalization heads at traditional companies, and early-stage AI startup teams. They generally lack experience in procuring compute infrastructure and the ability to conduct independent technical benchmarks, making them easy to be dazzled by scarcity narratives such as “wafer-scale AI chips” and “the world’s largest chip” and by endorsements from a U.S.-listed company. Their psychological weaknesses center on three points: first, fear of missing out in the large-model race—fear of not being able to buy compute, fear of being unable to afford the wait in queue; second, equating “expensive” with “good,” mistakenly believing that a high-unit-price contract comes with resource priority; third, insufficient contract risk auditing, misreading “revenue sharing for compute” as capital cooperation rather than high-cost debt, only discovering after fulfillment is blocked that they have been locked into future cash flow.

骗局怎么运作

• Concept hype: packaging scarcity with a “wafer-scale chip” identity. Cerebras repeatedly emphasizes to the outside world “the world’s largest AI chip” and “a single wafer is a chip,” creating a compute myth that surpasses NVIDIA. In technical pitches, the sales team interchanges different units such as “peak compute” and “FP8-equivalent compute,” misleading customers into thinking every system can run their current models at full speed, while deliberately downplaying the shortcoming that actual performance depends on specific computational graph structures.
• Benchmarking against NVIDIA to acquire customers with low prices. Sales often use “cheaper than NVIDIA” and “specialized in large-model training” as selling points, quoting a “starting rental price” far below an H100 cluster to attract SMEs to sign quickly. But the quote does not clearly list supporting racks, operations and maintenance, data center cooling, and network resources. After customers sign supplementary agreements, total procurement costs far exceed expectations, creating a clear disconnect between marketing rhetoric and fulfillment.
• Embedding “revenue sharing”-type bet clauses in contracts. To lock in major customers, contracts may include non-standard structures such as “compute investment in exchange for equity or revenue sharing” and “future revenue offsetting rental fees.” On the surface, such clauses reduce customers’ upfront funding pressure, but in reality they turn customers’ future cash flow into collateral. Once model commercialization falls short of expectations, customers will face multiple pressures at once: compute rent, penalties, and revenue-sharing obligations.
• Differentiated priority for delivered resources. Public financial reports and media coverage show highly concentrated customers and revenue, with top customers accounting for the bulk of revenue. After SMEs sign, their compute scheduling priority is pushed to the back, and actual available compute is seriously inconsistent with advertised “exclusive full-rack” and “dedicated compute.” When customers complain, sales fobs them off with rhetoric such as “public cloud resource fluctuations” and “peak-period queuing,” dodging contractual commitments.
• Blocking lease termination with “hardware has been delivered” clauses. Contracts often characterize compute leasing as a “transfer of equipment use rights” or a “customized, irrevocable” service. When customers seek to terminate on the grounds that actual performance fails to meet standards, the counterparty cites the hidden interpretation that “compute underperformance does not equal contract breach” to refuse refunds, leaving companies in a passive situation of paying long term while receiving insufficient resources.
• The combination of “paper growth” in financial reports and collapsing gross margins adds pressure. The 2026 interim report shows revenue doubled, but gross margin guidance plunged, and the market publicly questioned how the $25.4 billion RPO would be fulfilled. If new customers are dazzled by big-order news and rush to buy queuing slots, they can easily overlook the risk that delivery timelines will be indefinitely delayed, becoming bagholders in the next round of fulfillment cracks.
• Industry sentiment shifts abruptly, but sales still reassures customers with “market misreading.” After NVIDIA halted the “revenue sharing for financing” model, market doubts about similar structures escalated, but the sales team still downplays risk with “we are different” and “it’s just a financial reporting metric issue,” urging customers to sign quickly to lock in supposedly scarce capacity—in effect treating external regulatory warnings as a window to close deals.

红旗信号（看到这些快跑）

• 🚩 Only talking about peak compute, not sustained available compute; claiming “near 100% utilization” verbally but refusing to specify concrete metrics in the contract.
• 🚩 Non-standard payment structures such as “revenue sharing,” “compute for equity,” and “future revenue offsetting rental fees” appear, accompanied by complex default clauses.
• 🚩 The key customer list is highly concentrated, with top customers accounting for the bulk of revenue, yet SMEs are promised “equal priority compute.”
• 🚩 Actual support shortcomings for mainstream models such as Transformer are avoided, wrapped only in rhetoric like “our chip architecture is more advanced.”
• 🚩 No reproducible third-party benchmark report is provided; only vendor-developed white papers or demo videos are given as performance evidence.
• 🚩 After financial reports show gross margin guidance plunging or the stock price crashing in a single day, sales still urges customers to sign quickly, citing “market misreading” and “one-off factors.”

真实案例

• In August 2026, Cerebras released its first earnings report after going public. Its GAAP results were described by the market as “confusing,” gross margin guidance plunged, and the stock fell 17% in a single day. Several brokerages issued risk warnings on the actual profitability and fulfillment capacity of compute orders. Longbridge Securities conducted a special analysis contrasting “nearly doubled growth yet triggering a 17% plunge.”
• In August 2026, Phoenix Finance reported under the headline “10-Billion Compute Mega-Order Pressed on ‘How to Fulfill,’” noting that the market continued to question the fulfillment path for Cerebras’s RPO of up to $25.4 billion (remaining performance obligations). It pointed out that many “10-billion-level compute mega-orders” lack clear delivery schedules and acceptance standards, and that the so-called orders are more like paper commitments. (Source: https://longbridge.com/zh-CN/news/296000955)
• On August 28, 2026, Sina Finance reported that NVIDIA officially halted the “revenue sharing for financing” model, making clear that such structures, which tie compute leasing to customers’ future revenue, are widely questioned by the capital markets. 36Kr reported in the same period that Cerebras’s $5 billion-level “redemption contract”-style agreements with OpenAI and others are the same type of high-risk structure, with contracts containing complex arrangements for revenue sharing and exchanging resources for equity. (Source: https://www.36kr.com/p/3808516069891585)

Official Stance

• On August 28, 2026, NVIDIA officially announced the halt of the “revenue sharing for financing” model, noting that business models tying compute leasing to customers’ future revenue are widely questioned by the market; the industry views this as a clear warning against similar compute sales structures. (Source: Sina Finance)
• In August 2026, after Cerebras released its first earnings report following its IPO, several brokerages issued risk warnings, noting that the company’s GAAP results were “confusing,” gross margin guidance plunged, and the stock fell 17% in a single day. They reminded investors and customers to view the profitability and fulfillment capacity of compute orders prudently. (Source: Longbridge Securities)
• In August 2026, Phoenix Finance publicly questioned Cerebras’s 10-billion compute mega-order under the title “How to Fulfill,” pointing out that the recognition and delivery boundaries of the $25.4 billion RPO are vague, and warning SMEs not to treat “paper orders” as supply guarantees. (Source: Phoenix Finance)

How to Protect Yourself

• ✅ Before signing, complete pilot validation on the compute leasing platform using your own business data, require the counterparty to provide measured throughput, latency, and stability reports covering more than 7 consecutive days, and refuse to rely only on vendor demos or white papers.
• ✅ Write all key metrics such as “peak compute,” “sustained compute,” “availability SLA,” and “queuing priority” into the contract annex, and agree on pro-rata daily compensation clauses for failure to meet standards, so that oral promises are not left unfulfilled.
• ✅ Be vigilant about non-standard payment structures such as “revenue sharing,” “compute for equity,” and “future revenue offsetting rental fees”; they must be separately submitted for joint legal and financial review, and the cash-flow lockup under worst-case scenarios must be calculated.
• ✅ Verify the compute operator’s public financial reports, customer concentration, and deferred revenue structure; avoid procuring long-term compute from suppliers highly dependent on a few major customers, reducing the risk of being squeezed by “priority supply.”
• ✅ Adopt a phased procurement strategy: first use small, short-term leases to verify performance and service levels, then decide whether to scale up; tie the payment schedule to acceptance milestones, and strictly prohibit one-time prepayment of the full-year fee.