Andrew Feldman on AI Chips, Data Centers and Compute Demand

Cerebras co-founder and CEO Andrew Feldman tells Harry Stebbings that he sees AI infrastructure being built behind current demand, unlike historical periods when infrastructure was built well ahead of use. He cites his company's reported backlog as evidence, while acknowledging the discussion is about his industry and business.

Feldman says data centers cannot be completed fast enough for buyers and cites a $25 billion backlog. Backlog is a company-reported measure of orders or commitments, not recognized revenue or proof that every order will be delivered or remain unchanged.

Stebbings asks whether permitting and construction delays may help prevent overbuilding. Feldman agrees that a metered rollout can smooth expansion, while praising customers that anticipate future demand and contract for capacity ahead of time. Neither view proves that future supply cannot exceed demand.

Discussing a reported compute purchase, Feldman says the equipment was an older generation and assesses the transaction as less than ideal for the buyer. This is his interpretation of a deal involving other companies, not independently established contract terms or performance.

He says rapid AI demand has put pressure on memory suppliers and manufacturing capacity. Feldman distinguishes Cerebras's use of on-chip SRAM from the high-bandwidth memory used in many GPU systems, and sees that difference as an advantage for his company.

Memory fabrication capacity is expensive and takes time to expand, he argues, so shortages could persist for several years if demand remains high. That is a conditional forecast; future demand, production and pricing may differ.

Feldman believes models became sufficiently useful around 2025 to drive a broader increase in inference use. He expects demand to grow if models continue improving and users find more applications; neither the pace of improvement nor consumption is guaranteed.

He argues that large cloud platforms provide security, credibility and integrated software that some enterprises value, while other customers may mainly want low-cost compute. His view of Nvidia's support for alternative cloud providers is his interpretation of its strategy, not a confirmed statement of intent.

Feldman says Cerebras's architecture and supply chain have benefited from avoiding certain memory and packaging constraints. As the CEO of a competing chip company, he has a commercial interest in these comparisons; costs and availability vary by product and contract.

He expects chip designs across the industry to improve output per dollar and per unit of power over the next several years. He also predicts Cerebras's performance gap will widen. Both are forecasts, not independently verified future results.

On Google's vertically integrated infrastructure, Feldman sees benefits in controlling the stack from data centers to models, but notes a potential constraint if a custom chip primarily serves the company's own demand. He calls the economics an open question, not a settled winner.

He says companies using their own hardware in their own facilities avoid paying the same intermediary margins as a provider buying chips and reselling capacity. The margin figures he cites are his estimates; actual cost depends on hardware, utilization, financing and the service delivered.

Feldman praises CoreWeave's financial and deployment innovation while distinguishing it from Cerebras's engineering approach. He does not establish a valuation for other providers or resolve how their margins will evolve.

Stebbings cites a Cerebras report that a Kimi model ran 6.7 times faster than a comparison GPU cloud. Feldman argues that speed is valuable, especially for long tasks. The multiplier is a company-reported comparison for a specific model and setup, not a universal hardware ranking.

He illustrates why latency can matter: if one workflow finishes sooner, a team may run more experiments within the same period. Whether speed translates into better economics depends on quality, workload, price and the rest of the system; the interview does not establish that every user needs the fastest option.

Asked whether Cerebras can serve additional major labs, Feldman says the company is focused on fulfilling a large OpenAI agreement. He responds to concentration concerns by saying that winning and serving one large customer helps build the capability to serve others. The cited contract is a company-reported multiyear agreement, not revenue already earned.

He describes a rapid change in the scale of data-center planning, from tens of megawatts to proposals for multiple gigawatts. Those examples illustrate industry ambition, not proof that every announced facility will be built or operate at full capacity.

Stebbings asks whether electricity will become the decisive bottleneck. Feldman says that is a reasonable view but is not certain: it also assumes models keep improving enough for additional energy use to make economic sense. He leaves that question open.

On data-center delays, Feldman says large construction projects routinely encounter equipment, contractor and permitting problems. The analogy does not establish that community opposition or environmental impacts are insignificant, but explains why he expects schedules to be difficult.

He says the industry has not engaged communities well enough and should be more transparent, pay its own infrastructure costs and provide local benefits. His examples of jobs and community projects are proposals or general claims; effects vary by site and are not independently evaluated in the interview.

Feldman says some data centers can reduce water demand through recycling and closed-loop designs, and argues operators should pay for grid upgrades. These are design possibilities and policy preferences, not evidence that every data center has low water use or no local costs.

Asked about layoffs, he expresses concern for affected families but believes many recent cuts attributed to AI reflect earlier hiring and other productivity changes. His percentage estimates and explanation are personal judgments, not a verified accounting of individual layoffs.

He expects more productive engineering tools to let Cerebras pursue more projects and says his company intends to hire engineers rather than fewer of them. This is a company intention, not a guarantee for its workforce or the industry.

Discussing enterprise spending on AI tools, Feldman compares it with historically expensive hardware-design software. He sketches a large potential market for tokens per engineer, but the user count and spending figures are illustrative assumptions, not a forecast of realized revenue.

He expects new roles around AI governance and says some information-heavy tasks in HR and management may change. The nature and number of jobs created or displaced are uncertain; his examples are predictions rather than established outcomes.

Stebbings asks whether data quality is the main barrier to enterprise AI. Feldman argues that legal and security approval often slows initial adoption, then says data organization becomes important once a company proceeds. This is his experience and interpretation, not evidence that oversight is unnecessary.

He says leaders must weigh productivity against risks that can be difficult to see in advance. He acknowledges that some risks are real; the discussion does not determine the right controls for a particular organization.

The guests discuss open models and licensing. Feldman praises several Chinese-developed models but says legal and security teams face difficult decisions about adoption. His performance comparisons and views on legal complexity are opinions, not a universal ranking or legal advice.

Asked whether US companies should sell advanced chips to China, Feldman says no. He believes the technology could be used for military and industrial purposes, while acknowledging arguments that sales might preserve influence or discourage a separate ecosystem. These are contested policy judgments, not facts established by the interview.

He favors rebuilding advanced manufacturing and packaging capacity in the United States and proposes a long period of streamlined local rules for fabs. This is his policy preference; the interview does not evaluate environmental, community or legal consequences of the proposal.

On Europe, Feldman worries that regulation and a lower tolerance for failure can inhibit technology development. He acknowledges counterexamples, and Stebbings points to successful European application companies. Their regional comparisons are opinions, not a comprehensive assessment.

Asked about the timing of Cerebras's public offering, Feldman says the company had tried to go public earlier and did not know how other listings or chip markets would unfold. He attributes the eventual timing to persistence and luck, rather than claiming perfect foresight.

He says public-company resources could create opportunities for partnerships, investments or acquisitions, but does not announce a specific transaction. He distinguishes the obligations and access of public investors from those of venture backers.

Feldman describes earlier regulatory review around the planned listing as an obstacle and says it was later resolved. His account of the process and the administration's role is his perspective; the interview does not independently establish all reasons for the review or resolution.

His advice to founders is to keep improving the business through conditions they cannot control. He recalls fundraising during the 2008 crisis and says adding customers and improving technology helped his teams remain ready when circumstances changed.

In the quick-fire round, Feldman says he was surprised by the number of advisers and vendors seeking business around an IPO. He also says he values the wealth created for employees; the number of employee millionaires he cites is a company-leader claim, not an independently verified count.

Asked about family life, he says a partner must understand the pressure and travel involved in leading a company. He describes the need to invest attention in a relationship rather than let work consume every interaction.

Feldman recalls an 18-month technical struggle during which the company was spending about $8 million a month and repeatedly failing to solve a manufacturing problem. He says an empathetic board and disciplined failure analysis helped the team persist. The financial figure and chronology are his account.

He says leaders do doubt themselves and that difficult technical work can take a toll at home. He credits his colleagues and his wife for supporting him through that period, and Stebbings closes the conversation with thanks.