"Cerebras Deep: How Valuable is the Ultra-Large Chip?"

Cerebras CEO Andrew Feldman often holds something in his hand during speeches: a "chip" the size of a dinner plate, which has an area 56 times that of the NVIDIA H100.

In the past, the chip industry did the same thing: cutting, splitting a wafer into dozens or hundreds of small chips, then packaging, wiring, and clustering them. Now, Cerebras has inverted this: do not cut—turn the entire wafer into a single chip, with all the memory soldered onto the chip itself.

Cerebras is betting on a concrete issue: the bottleneck of large model inference lies in the movement between chips.

Many people recognize this. OpenAI signed a $20 billion contract with them and also lent them $1 billion to purchase wafers from TSMC.

Whether or not one agrees with this route, Cerebras will occupy a central position in the AI industry chain for a long time, and it will be one of the stocks attracting attention on the US stock market over the next few months. Today, let’s explain this company in plain language.

*This is the 20th article in the "AI Map" series, with a total of 11,000 words

I. What Kind of Company is Cerebras

1. A Unique Business Model

Cerebras provides AI inference computing power, but sells it in a completely different way than NVIDIA—it designs chips, assembles complete systems, builds data centers, and operates cloud services, selling computing power directly to customers. Customers pay for what they use or can directly secure a portion of capacity, similar to a subscription model.

In other words, it is a chip company + system integrator + data center operator + inference cloud service provider, and it does all four things.

This "do everything ourselves" model is the starting point for understanding all the numbers associated with this company. There is currently no other company doing this.

2. The Founder and an Old Case from 2007

CEO Andrew Feldman is a serial entrepreneur; his previous company SeaMicro was sold to AMD for $334-$357 million, proving himself commercially. However, there is an old case worth noting from the S-1/A.

From 2000 to 2003, Feldman was the Vice President of Marketing at Riverstone Networks. This company went public in 2001 and was sued by the SEC in 2006 for its former management allegedly falsely inflating revenue—by recognizing revenue prematurely through contingent sales.

Feldman was one of the defendants.

The outcome of the case in 2007 (from federal court records):

- Civil—signed a permanent injunction, neither admitting nor denying the allegations, and must pay a disgorgement of $289,500

- Criminal—Feldman admitted to one federal felony, violating internal accounting controls of public companies (Exchange Act Section 13(b)(5) and 32)

- Sentencing—3 years probation + $5,000 fine

This is disclosed in compliance within the S-1/A; the probation has long since ended, and there is no ongoing legal impact. However, it still holds reference value today for three reasons:

First, the GAAP profit conversion operation of Cerebras has some degree of opacity. Out of the projected $237.8 million GAAP net profit for 2025, $363.3 million comes from the non-cash accounting gain of G42's cancellation of future contract liabilities. This is not fraud but a compliant GAAP treatment. However, it is not simple bookkeeping; it heavily relies on management's judgment: when to recognize, how much to recognize, and whether to recognize as a one-time item, all of which directly affect the profit for that year.

Second, the essence of the Riverstone case is precisely "revenue inflation driven by accounting judgment." The accusations at that time were about pushing the boundaries of accounting judgment to the limit.

Third, Cerebras's financial statements have a particular abundance of certain elements. How much “invisible discount” should be deducted from revenue for the equity granted by OpenAI, how the $400 million debt from G42 turns into accounting gains, and how the costs of data centers are accounted for as both revenues and costs—all of these are not mechanistic bookkeeping; they rely entirely on management’s judgment. A CEO who has previously violated in accounting judgment running a company that heavily depends on accounting judgment for valuation might raise concerns.

3. Four Continuous Years of Internal Control Deficiencies

While one could say the old case of the CEO from 19 years ago is "outdated," there is an even more noteworthy issue in the S-1/A: Cerebras identified significant internal control deficiencies (Material Weaknesses) in its financial statements for the years 2022 to 2025. This is the most serious level under this framework, and according to the SEC, it means there is a reasonable possibility of financial statement misstatement that cannot be timely detected and corrected.

The deficiencies disclosed in the prospectus focus on two categories:

- Insufficient accounting and financial resources that affect revenue recognition, inventory management, cost accounting, and accounting for data center assets

- Inadequate general IT controls, insufficient segregation of duties, and weak management of system access permissions

For more of the full text, please visit:

https://mp.weixin.qq.com/s/NXpkoWgtS7SJGXPTYk-Tog

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。