Cerebras Founder Reflects on Chip Development Journey in Podcast
In the latest episode of his podcast Uncapped, Jack Altman invited Andrew Feldman, founder and CEO of wafer-scale chip company Cerebras, along with Benchmark general partner Eric Vishria, to discuss the complete journey of Cerebras from startup to scaled commercialization. The conversation covered various topics including the chip market landscape, the current state and future direction of the AI supply chain, and the role of investors in hardware companies.
Cerebras was founded by Andrew Feldman in 2015, who previously managed the business of server company SeaMicro. The company chose to directly challenge NVIDIA's dominance in the AI chip market by developing a giant AI accelerator chip known as the "Wafer Scale Engine" (WSE), with a single chip area of 46,225 square millimeters, nearly the size of a dinner plate. It employs an architecture that integrates massive computing units and on-chip SRAM storage on a single wafer, distinguishing it from the traditional GPU cluster solutions used by companies like NVIDIA.
The chip manufacturing process has continued to evolve with each generation: the second-generation product WSE-2, launched in 2021, used TSMC's 7-nanometer process, while the third-generation product WSE-3, set to launch in 2024, has upgraded to TSMC's 5-nanometer process. Official data shows its memory bandwidth reaches 21 petabytes per second (PB/s), claiming to be nearly a thousand times faster than NVIDIA's new generation Rubin chip in inference tasks.
Benchmark co-led Cerebras' $25 million Series A financing in 2016, the same year the company was founded, with Eric Vishria joining the board shortly thereafter, having been a general partner at Benchmark for only about a year and a half. It has been disclosed that Benchmark continued to invest through multiple rounds, with early shares costing about $18 million and subsequent investments totaling around $250 million, bringing the total investment to approximately $270 million.
This nearly nine-year bet will culminate on May 15, 2026, when Cerebras officially lists on NASDAQ under the ticker CBRS, with an initial valuation of $66 billion and raising approximately $5.55 billion. On its first day of trading, the stock price surged nearly 70%; at the closing price, Benchmark's shares were valued at over $5.3 billion, achieving a multiple return on its total investment. According to early prospectus disclosures, customer concentration was once a market concern, with a single customer, G42, accounting for 87% of the company's revenue. However, the company has since secured new clients such as Amazon Web Services, Meta, OpenAI, and Mistral AI, significantly diversifying its revenue structure.
From a funding flow perspective, the nearly 70% increase on Cerebras' first day of trading essentially reflects a strong scarcity premium expectation in the secondary market for "AI chip alternatives outside of NVIDIA": as leading cloud providers and model companies become its clients, the market bets that AI computing power demand has consistently exceeded the capacity that a single supplier can meet, thus willing to pay a premium for a second supplier with a differentiated architecture that has secured real large customer orders. The beneficiaries are the early investors in the hardware sector who have endured nearly nine years of high R&D investment cycles, as well as downstream cloud providers that locked in capacity and orders early, while the burden falls on traditional hardware investment institutions that previously remained cautious about Cerebras' commercialization prospects and missed the early entry window.
ABAB AI Insight
Andrew Feldman was responsible for the business of server company SeaMicro before founding Cerebras; after establishing Cerebras, the company first submitted its IPO prospectus in September 2024 but quietly withdrew the document in October 2025, delaying its listing plan, which was not completed until May 2026 after more than two years of adjustments. Eric Vishria led the early investment in Cerebras just about eighteen months after joining Benchmark, which had not engaged in hardware investments for a decade. From a funding pathway perspective, Benchmark phased its investment in Cerebras—initially acquiring early shares for about $18 million in 2016, followed by an additional investment of about $250 million, totaling approximately $270 million. It took nearly nine years to see the payoff from commercialization and the IPO, during which they continuously bore the risks associated with long hardware R&D cycles and high capital consumption; this pathway reveals the typical "long-cycle companion" characteristic of AI infrastructure investments, contrasting sharply with the rapid return models common in software investments. This journey can be likened to NVIDIA's own early growth from a fringe player to an industry standard-setter in the graphics chip field, as well as AMD's long-standing challenge to Intel in the CPU market, gradually gaining market share through a differentiated architecture. Cerebras is currently at a critical stage in the AI chip industry, "challenging the monopolist"—with its wafer-scale architecture and real large customer orders, it is beginning to carve out a niche in the NVIDIA-dominated market landscape. From a structural nature perspective, this event corresponds to a transfer of pricing power: the nearly 70% surge on Cerebras' first day of trading and the subsequent orders from leading cloud providers reflect that the pricing power in the AI computing supply chain is shifting from a single supplier to a more diversified supplier structure; the causal mechanism is that as downstream computing power demand continues to outpace the capacity expansion of a single chip manufacturer, buyers have a natural incentive to support second and third suppliers to gain bargaining power for supply chain security. This behavior pattern of "supporting alternative suppliers to hedge against concentrated bargaining power" is systematically reshaping the pricing power structure in the AI chip industry, which was originally highly concentrated in a single vendor. ABAB News · Cognitive Laws
Eight years of running alongside, finally leads to a payoff
Buyers support the second player just to regain bargaining power
Hardware's patience and software's speed each have their own moats