Andrew Feldman on AI Chips, Inference and Competing With Nvidia

Cerebras co-founder and CEO Andrew Feldman joins Harry Stebbings to discuss AI chips, inference, model development and the contest with Nvidia. Cerebras sells competing AI hardware, so Feldman's technical comparisons and market forecasts carry a direct commercial interest. The interview does not independently verify its benchmark, cost or financial claims.

Feldman says the founders started Cerebras in 2015 after seeing AI emerge as a new computing workload. They expected it to place unusual pressure on memory bandwidth and communication, creating an opportunity for a machine designed for that workload. He says he underestimated how large the market would become.

He describes chips as calculating, moving and storing data. In his account, individual AI calculations can be relatively simple, but moving inputs and intermediate results among compute and memory is difficult. Cerebras set out to reduce that movement, aiming for more speed and lower power use; those advantages depend on the workload and system.

Training from scratch and fine-tuning involve approximately similar computational work, Feldman says, while generative inference has distinct communication demands. He gives an illustrative 70-billion-parameter model with 16-bit weights: moving its roughly 140 GB of weights for each generated word would require substantial memory bandwidth. Actual movement and performance depend on implementation and model configuration.

Feldman contrasts high-bandwidth memory, which offers capacity, with fast on-chip SRAM, which has less capacity per conventional chip. Cerebras uses a wafer-scale processor to place more SRAM on a single device and reduce data movement. His comparison simplifies architectural tradeoffs; the interview does not supply a matched independent cost, energy or performance test.

Asked why GPUs still dominate, he says their memory design served graphics well but creates a constraint for inference, and argues Cerebras has an advantage. He acknowledges Nvidia's architecture has been commercially successful. As a competitor selling a different architecture, his broad characterization should not be treated as a result for every GPU or workload.

Feldman says Cerebras had been fastest across a set of models tested by Artificial Analysis and others since launching inference in August. This is his report of selected benchmarks, not evidence that Cerebras leads every model, metric or deployment. He says power savings come partly from keeping more data on chip, while wafer-scale production creates yield challenges.

He explains yield using a silicon wafer and defects: a larger conventional chip is more likely to contain a flaw, potentially forcing a manufacturer to discard or downgrade it. Cerebras, he says, built the processor from many repeated tiles and uses redundancy to route around defective ones. He claims the resulting yield is as good as or better than smaller-chip approaches; no comparable yield data is supplied.

Stebbings asks whether speed, efficiency or cost matters most. Feldman says the answer depends on the task: a batch job generating training data may tolerate delay and prioritize price, whereas an interactive search or chat product can suffer from latency. His medical diagnosis example is a hypothetical about accuracy and delay, not evidence that any product can safely diagnose cancer.

He compares faster inference to broadband enabling new products that were impractical when video arrived by mail. His framework for inference demand multiplies the number of users, their frequency of use and compute per use. He believes all three were growing at the time of the interview; it is a conceptual model rather than a measured market forecast.

Feldman argues AI became more useful in everyday workflows around late 2024, after an earlier period he characterizes as novelty. He predicts the market will be more than 100 times larger in five years, without defining a precise market metric. Both the timing and growth estimate are his judgment, not established outcomes.

He acknowledges AI's heavy power consumption and says the industry must deliver substantial social value to justify it. He argues the US has power in places that do not always coincide with desired data-center locations and criticizes local barriers to new infrastructure. These are policy views, not a complete assessment of grid capacity, costs or environmental effects.

Stebbings raises concerns that some new data centers are poorly provisioned. Feldman says quality remains to be seen, but points to experienced operators, including former cryptocurrency mining businesses, that located facilities near inexpensive power. He acknowledges some less experienced entrants while arguing many builders are sophisticated.

He breaks inference costs into data-center power and space, computer costs and algorithm efficiency. He claims GPUs are often only 5% to 7% utilized during inference, and expects cheaper, faster hardware, better facilities and improved algorithms to reduce costs. The utilization claim lacks a specified workload, measurement method or comparable benchmark and should not be generalized.

Feldman believes substantial algorithmic improvement remains and cites OpenAI's o1 as evidence that additional inference compute can improve answers. He points to mixture-of-experts and sparsity as ways to avoid activating all parts of a model for every input. His account is a view of available approaches, not proof that scaling yields the same returns indefinitely.

He expects models to become faster, more accurate and efficient, and challenges the suggestion that compute, data and algorithms are already mature. Asked about training data five years out, he predicts it will be almost entirely synthetic. That is a forecast, and the interview does not establish that synthetic data is universally equivalent to human data.

His analogy is a flight simulator: training can repeatedly present rare, difficult events such as an engine failure instead of spending most time on uneventful flight. Likewise, he says synthetic data might supply hard-to-collect driving situations such as an unprotected turn in snow. These are examples of potential use, not validation of safety or real-world coverage.

Feldman argues that cheaper and faster computation makes new applications feasible and can increase demand. He predicts the industry will depend less on transformer architectures in three to five years, although he does not know what will replace them. He identifies attention costs and other limitations as reasons to seek alternatives; the timeline is speculative.

He praises DeepSeek as focused engineering and says its result shows a smaller team can make substantial progress without the resources of the largest labs. His statement about the team's size and available hardware is an estimate, not a verified accounting. He also praises the technical impact of releasing a model openly.

Asked about distillation, Feldman says he does not consider it wrong and calls for consistency when discussing use of copyrighted data as well. This is his opinion, not a legal determination about any model or dataset. Stebbings notes that an open release also allows others to learn from its innovations.

On defensibility, Feldman argues a brief model lead is less valuable than sustained performance near the top over years. In a mature market, ease of use and a good-enough product can matter even without the strongest technology. He considers hardware difficult and capital intensive, and says chip and systems companies can build lasting value.

Feldman says there is no CUDA lock-in for inference when a customer can switch hosted services with a small code change. He also acknowledges Nvidia's advantages as the default market leader and the difficulty of compilers. His claim concerns the examples he describes; it does not establish that every migration has negligible engineering, cost or operational friction.

He predicts Nvidia will hold roughly 50% to 60% of the market in five years, while retaining meaningful training and inference businesses and competing seriously. He expects Cerebras and others to gain shares as the market grows. The market definition is not specified, and the percentages are the forecast of a direct competitor.

He predicts chip providers will have larger enterprise values than model providers over five years, while acknowledging that short-term valuations can reflect considerable uncertainty. He argues mature businesses will eventually be assessed on profitability and future growth. This is a market opinion, not an investment recommendation.

Stebbings describes Cerebras as cash-flow positive. Feldman attributes its ability to maintain margins to technical differentiation, contrasting this with negative gross-margin businesses. The interview does not specify the period or cash-flow measure, and gross margin is not the same as net income or cash flow.

Asked about revenue concentration in G42, Feldman calls the relationship both a strength and a weakness: building one large strategic partnership can teach the organization how to serve others. He describes deployment, software hardening and manufacturing growth as lessons from the relationship. These are the CEO's claims, not an independent assessment of customer diversification.

On the decision to file for a public listing, Feldman says public-company readiness requires forecasting, processes and accountability. He believes some large enterprise customers prefer working with public companies. The filing and his reasons do not establish that an IPO occurred or guarantee access to customers.

In discussion of G42, Stebbings cites 87% of revenue, while Feldman says some estimates of the announced deal exceeded $1 billion. The interview does not specify the accounting period for the revenue percentage or equate estimated deal value with recognized revenue. Feldman says scaling hardware requires extensive preparation with supply-chain partners.

Stebbings asks whether delayed Nvidia shipments could leave customers unhappy. Feldman says that potential dissatisfaction offers an opening for competitors because customers unable to obtain equipment may test alternatives. It is a competitor's commercial assessment, not evidence that particular customers received obsolete chips.

On export controls, Feldman distinguishes physical hardware, which can be located and inspected, from software that is easier to diffuse. He says restrictions can have unintended effects, describing investment in Chinese electronic-design tools as an example, and considers slowing another nation's technical development difficult. His claims and policy analysis are not independently verified in the interview.

He considers the current US administration better for AI than its predecessor, while acknowledging the technology itself has changed. Asked why Cerebras chose not to sell to China, he says he believed a proposed deal could be used for purposes he found unacceptable, mentioning surveillance and military use. This is his account of the decision and concerns, not a finding about a particular end user.

Feldman argues China should not be underestimated, citing infrastructure, engineering talent and industrial policy. He praises economic zones such as Shenzhen and asks what the US can learn about building infrastructure, including rail. These are his geopolitical and policy judgments, not a comprehensive comparison of the two countries.

In quick fire, he expresses a hope that peace in the Middle East may be closer than many think, pointing to changes in regional business-focused states. He also predicts AI penetration could approach that of mobile phones within one or two years. Both are personal predictions, not established outcomes; G42 is a Cerebras partner in the region.

He recalls opposing co-founder JP's 2016 proposal for water cooling and later realizing he was wrong. Making many decisions requires accepting mistakes, he says, including misjudgments about people. He contrasts the venture-capital model, where a few large successes can dominate a portfolio, with a CEO's need to make many sound operating decisions.

Feldman argues founder experience matters for a company with manufacturing, supply chains and large engineering teams, while acknowledging that young founders may understand consumer markets particularly well. He also identifies tiny, low-power inference chips near sensors as an underappreciated opportunity, although he prefers building large data-center systems.

Looking three to five years ahead, he hopes Cerebras technology will help find a therapy for a condition affecting more than a million people a year, power new applications and be used indirectly by a meaningful portion of people in the US and Europe. These are aspirations, not evidence that a therapy or those adoption levels have been achieved.