Andrew Feldman

Andrew Feldman is the co-founder, chief executive officer, president, and chairman of Cerebras. He has served as CEO, president, and a board member since April 2016. His work at Cerebras centers on building purpose-built AI computing systems: in his public explanations, moving data, distributing training workloads, and reducing inference latency are central constraints on what researchers and developers can accomplish.[2][3][4][5]

In a February 2020 interview with Patrick Kennedy, Feldman described Cerebras as a reunion of colleagues from SeaMicro. He credited Gary Lauterbach with SeaMicro’s original idea and tentatively with the germ of Cerebras’ idea. Sean Lie, Michael James, and Jean-Philippe Fricker, he said, each contributed a different part of the vision for the system.[3]

Feldman recalled that Lauterbach began by questioning why a processor tuned for graphics should also be well suited to deep learning. The team studied the workload’s memory requirements and communication patterns. Their conclusion was that “the hard part of this work was moving data”—between memory and cores, and between cores themselves. That analysis led them toward a very large processor with substantial on-chip memory and fast internal communication, rather than assembling the same resources from many separate chips.[3]

The wafer-scale approach also required work beyond chip architecture. Feldman described packaging, power delivery, cooling, and software as reasons Cerebras became a complete-system builder. Ordinary manufacturing equipment could not even pick up a chip of that size. Concentrating computation also concentrated power and heat, requiring custom packaging and cooling. As he put it, “that is why we are a system builder.”[3]

Feldman remembered the first working chip with the words “It took my breath away.” He also recalled stretches when the team made no progress on a technical problem, and investors who supported continued work on fundamental solutions. One sent cookies to the entire company when they did not know how else to help. His account emphasized repeated experiments, changing one variable at a time, and documenting results so the team could compare failures and improvements.[3]

In his September 2022 explanation of weight streaming, Feldman argued that distributed computation had become a major barrier to training large language models: “Ironically, what’s hard about big AI isn’t the AI. It’s the distributed compute.” He presented Cerebras’ architecture as a way to reduce the need to divide individual model layers across devices and coordinate multiple forms of parallelism.[4]

The mechanism separates parameter storage from computation. MemoryX stores model parameters and optimizer states, streams weights to CS-2 systems, receives their computed gradients, and calculates weight updates. Activations remain on the CS-2. SwarmX broadcasts the weights to multiple CS-2 systems and reduces their returned gradients into a single stream for MemoryX. Each CS-2 processes a different portion of the training data with the same configuration. Feldman’s article reported perfect linear scaling specifically for GPT-3 XL across up to 16 CS-2 systems.[4]

By January 2026, Feldman’s public argument emphasized inference speed as a constraint on practical AI capabilities. He argued that faster execution allows models to do more reasoning within the same time budget, crediting Sean Lie and James Wang’s “The Cerebras Scaling Law” with explaining that relationship. He assessed wafer-scale computing as the best architecture for large-model inference; that was his architectural judgment.[5]

In a January 14, 2026 announcement, Feldman described a multi-year agreement between OpenAI and Cerebras to deploy 750 megawatts of wafer-scale systems for OpenAI customers, with deployment planned in stages beginning in 2026. He said the teams had met frequently since 2017 and framed speed as a fundamental driver of technology adoption. The 750-megawatt figure describes the announced deployment plan, rather than completed operating capacity.[6]

Feldman has also connected infrastructure choices to the applications he wants to enable. In 2020, he said cancer’s effects on his family informed Cerebras’ early work with Argonne National Laboratory and the team’s decision to allocate time toward cancer models. He envisioned wider access to diagnostic expertise through AI, including remote assessment of skin lesions; these were aspirations, rather than demonstrated clinical outcomes.[3]

Before Cerebras, Feldman worked in networking and dense microservers, and SeaMicro supplied several of the collaborators who later joined him at Cerebras. He held marketing and corporate-development responsibilities at Riverstone Networks from March 2000 to August 2003, followed by marketing and product-management leadership at Force10 Networks from August 2003 to December 2006.[2][3]

Feldman co-founded SeaMicro with Lauterbach and served as its CEO from November 2007 to February 2012. AMD completed the acquisition on March 23, 2012, for approximately $334 million, net of cash assumed. Feldman served as an AMD corporate vice president and general manager from February 2012 to June 2014. He holds a Stanford MBA and a Stanford bachelor’s degree in economics and political science.[3][2][7]