Why we chose Cerebras in 2023, and what the market saw in 2026

Why we chose Cerebras in 2023, and what the market saw in 2026

On May 13, 2026, Cerebras Systems priced its initial public offering at $185 a share, above a marketed range of $150 to $160, raising $5.55 billion (Cerebras press release, 2026-05-13). CNBC described it as the largest US tech IPO since Uber (CNBC, 2026-05-13). Shares opened on Nasdaq under the ticker CBRS at more than double the offer price (TechCrunch, 2026-05-14).

CrossWork became a Cerebras shareholder in 2023. This essay is about why, and about what the three years in between taught us. The method is the more useful part.

The thesis: compute, not models, was the bottleneck

In 2023 the market's attention was on models. ChatGPT had just reset public expectations, and capital was flooding into application companies and model labs. Our view, shaped heavily by conversations with former operators who had run infrastructure at scale, was that the binding constraint on the whole ecosystem would be compute, and specifically the cost and speed of inference once models moved into production.

That framing mattered because it pointed away from the crowded trade. If compute was the bottleneck, then the companies with differentiated silicon and system architecture had leverage over everyone downstream. Nvidia was the obvious beneficiary and was already priced accordingly. The question was whether there was a credible architectural alternative still available privately.

Cerebras had made an unusual bet: a wafer-scale processor, a single chip the size of a dinner plate, designed to keep an entire model's working memory on-chip rather than shuttling data between many smaller processors. In 2023 that bet looked eccentric to many generalist investors. To the operators we spoke with, people who had watched memory bandwidth, not raw arithmetic, throttle real workloads, it looked like the right problem to be solving.

The operator signal

The most useful information we had was not in a deck. It came from people who had bought, deployed or evaluated large compute systems and could describe, concretely, where the pain was. Operators do not tell you whether a company will IPO. They tell you whether a product works in production, whether a customer would renew, and whether an engineering team is shipping or explaining.

In Cerebras's case, the operator read was that inference speed was going to become a product feature in its own right, not just a cost line. That view has aged well. By the time of the IPO, the company was positioning itself explicitly around fast inference for production AI workloads (CNBC, 2026-05-13).

Access: the secondary route

Primary rounds in companies like Cerebras are allocated to a small set of institutions. Our route in, as is often the case for a firm of our size, ran through structured secondary access: buying existing shares from holders who wanted liquidity. Secondary access has trade-offs. Pricing is negotiated rather than set by a lead investor, information rights are often thinner, and transfers require company consent. But in 2023, when AI infrastructure sentiment was mixed and some early holders were ready to sell, it was the practical door.

The part that nearly went wrong

A candid case study has to include the difficult middle. Cerebras first filed publicly to go public in September 2024. The filing disclosed that a single customer, the Abu Dhabi group G42, accounted for 87 percent of revenue in the first half of 2024, and that a G42 investment was under review by the Committee on Foreign Investment in the United States (Cerebras Form S-1, SEC, 2024-09-30). Reuters reported in October 2024 that the IPO was likely to be postponed because of the CFIUS review (Reuters via KFGO, 2024-10-09). A year later, in October 2025, the company formally withdrew its registration, shortly after raising a $1.1 billion Series G at an $8.1 billion post-money valuation (CNBC, 2025-10-03; Cerebras press release, 2025-09-30).

For a private investor, that period was a test of conviction. Concentration risk was real. Regulatory risk was real. The public narrative in late 2024 was that the listing had stalled. What had not changed was the underlying thesis: demand for fast, efficient inference was growing, and the company was broadening its customer base and building its own data center capacity.

The lesson we draw is not that we were right to hold. It is that the variables we had underwritten, the architecture and the inference demand curve, were distinct from the variables that caused the delay, customer concentration and regulatory timing. Knowing which risks you are actually underwriting is most of the discipline.

What the market saw in 2026

When Cerebras returned in 2026, three things had changed. The customer base had diversified. Inference had become the dominant growth driver in AI compute. And public investors had developed an appetite for AI infrastructure that simply did not exist in 2023. The offering was priced above range and implied a market capitalization of roughly $67 billion (SiliconANGLE, 2026-05-13). Early trading was volatile, as it often is after large listings: shares surged and then gave back part of the gain in subsequent sessions (Yahoo Finance, 2026-05). Commentators have also cautioned that the scale of AI IPOs this year deserves scrutiny, not just enthusiasm (Forbes, 2026-05-21). We agree. Public markets will judge the company from here.

The method, generalised

Stripped of the specifics, the Cerebras position followed a pattern we look for:

1. A bottleneck thesis. Identify the constraint the whole ecosystem will hit, and invest where the leverage over that constraint sits.
2. An operator read. Test the thesis with people who have lived the problem, not people who have modelled it.
3. A practical access route. Accept the friction of secondaries when that is the door that is open.
4. Separate the risks. Know which risks you are underwriting and which you are merely enduring.
5. Patience measured in years. From our entry to listing was roughly three years, including one withdrawn filing.

The categories that will define the next cycle are forming now, and the bottlenecks are again not where most attention is. CrossWork Midas II is built to apply the same method to them, through the same advisory network of former operators. Access is invitation-only.

---

Sources

• Cerebras Systems, "Cerebras Announces IPO Pricing at $185 per Share" (2026-05-13).
• CNBC, "Cerebras prices IPO above expected range" (2026-05-13).
• TechCrunch, "Cerebras raises $5.5B, then stock pops 108%" (2026-05-14).
• SiliconANGLE, "Cerebras prices its shares at $185" (2026-05-13).
• Yahoo Finance, "Cerebras stock slides after near-70% surge" (2026-05).
• Forbes, "The PermaBull's Blind Spot: What the Cerebras IPO Tells Us" (2026-05-21).
• Cerebras Systems, Form S-1, SEC EDGAR (filed 2024-09-30).
• Reuters via KFGO, "Cerebras likely to postpone IPO due to CFIUS review delay on G42 deal" (2024-10-09).
• CNBC, "AI chipmaker Cerebras withdraws IPO" (2025-10-03).
• Cerebras Systems, "Cerebras Raises $1.1 Billion at $8.1 Billion Valuation" (2025-09-30).

Past performance is not indicative of future results. This is not an offer; offering documents govern. Detailed performance is available to verified investors in the CrossWork portal. This is not investment advice.