Cerebras CEO Andrew Feldman on AI Chips, Demand and Infrastructure Risk

Harry Stebbings interviews Cerebras co-founder and CEO Andrew Feldman about AI chips, infrastructure, financing and competition. Feldman speaks as the executive of a company competing with Nvidia, so his performance comparisons and commercial judgments are interested claims, not independent benchmarks.

Feldman discusses a roughly $1.1 billion late-stage round described in the episode information as Series G at an $8.1 billion valuation. He praises Fidelity and Atreides as leaders of the round and names other investors. The financing and valuation are not revenue or realized investment returns.

He says the capital will help Cerebras expand manufacturing, add data centers and pursue new ideas. The company added five US data centers that year, he says. He argues that small, incremental changes will not address all the challenges in AI, a judgment reflecting his own product strategy.

Asked why the company raised privately instead of going public, Feldman says Cerebras still intends to go public. He presents the pre-IPO round as a way to raise money quickly without distracting from opportunities. His intention is not a confirmed listing date.

Feldman cautions readers to examine the time horizon and “up to” language in large AI deals. A headline maximum over five years may not be committed spending or near-term revenue. He sees the announcements as signs of strong but uncertain demand rather than precise forecasts.

He says some customers give very wide ranges for possible future queries, illustrating how hard demand is to predict six to twelve months ahead. The example is a company anecdote, not an independently measured market forecast.

Cerebras must make multi-year data-center and supply-chain commitments despite the uncertainty. Feldman favors more frequent planning and buying options on future capacity: pay for flexibility, accepting that some of the cost will be lost if the capacity is not used.

Asked whether he is still underestimating demand, he says yes and points to how rapidly valuations and demand have changed. He also believes AI may substantially raise productivity and reshape the economy over five years. These are forecasts, not guarantees of growth or returns.

Stebbings asks whether Nvidia is unassailable. Feldman credits its achievements but argues that using a balance sheet and investment deals can indicate a large company is concerned about future growth. That interpretation comes from a competitor and does not establish Nvidia’s motives.

On chip depreciation, Feldman says H100 and A100 hardware continues to be useful after the shortest proposed replacement cycles. His point is that economic life depends on the next generation’s actual performance and power efficiency: even a paid-off chip may be replaced if a new system generates more value in the same power-constrained facility.

He estimates that comparable generation-to-generation gains may be roughly two to two-and-a-half times rather than the largest marketing claims, and stresses comparing the same numerical precision and whole-system performance. Memory bandwidth can become the bottleneck even when theoretical compute increases. His estimate is not a universal benchmark.

Stebbings relays a concern that fast on-chip SRAM cannot hold enough for large models. Feldman agrees with the capacity constraint on conventional chips and explains Cerebras’s wafer-scale design: use a much larger piece of silicon to fit more fast memory alongside compute. This is the company’s explanation of its architecture, not proof that it wins every workload.

Stebbings asks why a larger chip was not an obvious solution. Feldman says the concept was easier than the engineering and describes earlier industry attempts that failed. His account emphasizes manufacturing and yield as obstacles; the interview does not independently test his historical comparisons.

The host suggests Nvidia may be better for training while Cerebras is better for inference. Feldman disputes that split and claims Cerebras is faster at both, while saying training has a harder software migration problem because model recipes are often developed on GPUs. He says inference users can change an API with relatively little work; these are vendor claims requiring workload-specific verification.

Feldman models inference demand as the number of users multiplied by their usage frequency and the compute per use. He sees all three rising and says this makes planning difficult. The framework identifies potential drivers, but it cannot by itself establish future demand or profitability.

He uses the history of electricity and computing to argue that productivity gains may arrive after organizations redesign work around a new technology. In his view, replacing a search query with an AI query may yield modest improvements, while new workflows could create more. The analogy and future gains remain predictions.

Stebbings asks whether AI’s energy requirements are feasible. Feldman says yes, while distinguishing feasibility from whether the use is socially desirable. He argues that US power resources and data centers, people and fiber networks are often in different locations; his broad claim is an assessment rather than a regional grid study.

He says an industry consuming large amounts of power has a responsibility to produce social benefits, such as better health care. Stebbings questions whether energy use on entertainment is valuable. Feldman replies that markets fund many experiments and it can be hard to know in advance which technologies will yield useful results.

Feldman argues that public funds, tax incentives and permitting support should favor projects with social value. Asked about the Trump administration’s AI policy, he gives a generally positive assessment while later criticizing immigration restrictions. These are political judgments from a business executive, not neutral policy evaluations.

He does not think nuclear power is the only possible solution, noting that countries have different natural resources and constraints. He sees nuclear as a possible long-term choice for some countries. Costs, construction timelines and grid impacts depend on local conditions.

On the concentration of value in the largest technology companies, Feldman says risk becomes dangerous when investors do not recognize it. A broad index may expose holders to a small number of companies and a common sector even if they believe they are diversified. He does not establish a particular investor’s holdings or risk tolerance.

Asked about Nvidia’s market valuation, he calls it an extraordinary company and says he cannot determine whether a specific multitrillion-dollar price is correct. His comments on a competitor’s business value are not a stock recommendation.

Feldman names expertise as a bottleneck: he says the US needs more engineers, AI practitioners and people who understand data pipelines. He argues that immigration helps attract talent and that domestic education must improve if that route is restricted. These are policy preferences, not measured causal estimates.

Stebbings asks whether extraordinary AI compensation has become excessive. Feldman says unusually talented people can contribute things larger teams cannot, while acknowledging he does not know the right comparison with athletes. His assertion that exceptional employees cannot be overpaid is a management opinion, not a financial rule.

He also cites limited semiconductor fabrication and data-center capacity. He says large fabs can cost tens of billions of dollars and take time to build, constraining supply for many chipmakers. His figures and construction timelines are estimates, not costings for a specific project.

Stebbings asks whether data centers are attracting enough investment. Feldman says their property-and-lease structure is familiar to capital markets, but expects some builders to lose money. Low-cost power, permits, construction discipline and reliable tenants all matter; a long lease alone does not eliminate these risks.

Asked whether AI companies must be vertically integrated, Feldman says the answer is unclear. He cites OpenAI and Anthropic’s use of outside cloud infrastructure as examples of approaches that have worked, while acknowledging that future decisions could differ. Neither model guarantees future success.

On model companies designing their own chips, Feldman says the result is an open question. Chip development takes years and costly design errors can cause long delays, he argues; he points to acquisitions and long-running efforts by large companies as examples. His company sells competing hardware, which is relevant to his assessment.

Asked whether the chip industry will become a single monopoly in ten years, he distinguishes AI accelerators from the broader silicon market. He argues that different segments have historically had different leaders and predicts that the whole industry will not accrue to one or two companies. That is a forecast.

Feldman says Cerebras had margins that helped it raise financing, contrasting this with other companies he says had negative margins. He does not give Cerebras’s margin figure or accounting basis in this exchange. Stebbings asks about Nvidia, and Feldman cites high gross-margin figures and attributes them to pricing power; those figures and motives are his claims, not audited comparisons here.

He discusses Mistral’s European positioning and says it uses Cerebras hardware for inference. He praises its strategy and claims their combined offering is fast. This is a commercial partner’s perspective, and no comparative test is supplied in the interview.

On US-China competition, Feldman says an arms race benefits neither side and favors peaceful engagement, while expressing concerns about military uses of technology. He says Cerebras passed on a potential China deal in 2019 because of those concerns. His description of China’s industrial policies and technological strengths is his assessment.

He argues that fragmented power infrastructure and local permitting can slow US projects, citing a Samsung fab as an example. He also wants the country to attract international talent and give universities more compute. These recommendations represent his policy priorities; the interview does not independently establish the example’s causes or costs.

Stebbings raises concerns about abuse of H-1B visas and alternatives. Feldman acknowledges that abuse can occur but disputes that it is unusually prevalent and argues for a path from US education to work and citizenship. The interview does not provide comparative data to settle the abuse question.

In the quick-fire round, Feldman predicts peace in the Middle East within his lifetime, arguing that moderation and economic growth can reinforce each other. Stebbings asks about his business interests in the region. This is a political forecast, not an assessment of the views of all parties to the conflicts.

Feldman recalls that an S-1 covering approximately the first half of 2024 showed roughly 75% to 80% of revenue associated with the UAE. He says later figures had not been published in the interview. The concentration is historical, approximate and attributed to him; it should not be treated as a current revenue mix.

Stebbings asks whether the revenue concentration influences his positive comments about the region. Feldman says his favorable impressions preceded the business and acknowledges his visits color his perspective. He says large orders consumed much of Cerebras’s manufacturing capacity, including business with G42; his commercial exposure is relevant to that assessment.

Asked about an unsuccessful bet, Feldman recalls roughly 15 months from 2017 to early 2019 when Cerebras could not produce a working wafer-scale chip while spending about $6 million to $7 million a month. He describes repeated failure analysis and the moment the team saw the device work. The timing and spending are his recollection.

He argues that chip startups benefit from experienced teams and relationships with fabrication, design and intellectual-property suppliers. He contrasts that with consumer or student-facing products where young founders may understand customers especially well. His views about founder age are generalizations, not rules for individuals.

Feldman believes data cleaning and pipelines are often undervalued despite their importance to AI projects. Asked about data-labeling providers, he says he does not know how durable the market will be if machines eventually do the work as well as humans. Current importance does not establish the future size of the market.

Stebbings cites a prediction that AI will cause major labor shortages within five years. Feldman disagrees with that timeframe, saying economic adjustment takes time, though he allows a longer horizon could differ. He uses AlphaFold as an example of the lag between scientific breakthroughs and practical benefits; his specific claims about downstream outcomes are not independently established in the interview.

He imagines AI helping education by detecting patterns in students’ mistakes and suggesting appropriate exercises. He also predicts changes to entry-level analytical work. These are potential applications and forecasts, not demonstrated improvements in learning or employment outcomes.

Asked what he would do if he could not fail, Feldman instead describes the satisfaction of competing every day against a much larger incumbent. He sees Cerebras as an underdog and credits the small group of investors and colleagues who backed it. This is his account of his company’s position.

In closing, Feldman praises Stebbings’s writing about entrepreneurship and argues that building an exceptional company requires intense, sustained work, while acknowledging costs. That is his personal view of entrepreneurship, not evidence that a particular number of work hours guarantees success.