Thomas Sohmers

Thomas Sohmers is a co-founder of Positron AI, where his work centers on making transformer inference more economical through specialized hardware. He has emphasized memory-system design and early customer deployment over a long initial commitment to custom silicon. Positron lists him as chairman as of September 2026; in his September 2024 Cerebral Valley interview, he was its founder and CEO.[6][4]

Sohmers chose field-programmable gate arrays (FPGAs) for Positron’s first-generation hardware: chips whose logic can be reconfigured rather than fixed in a custom fabrication run. In 2024, he explained that this reduced non-recurring engineering costs and let the team deliver hardware sooner, revise designs, and learn from customer use. He acknowledged the tradeoff: FPGAs might be less efficient or performant than application-specific integrated circuits (ASICs). His priority was to avoid spending years building a chip that could be outdated at launch.[6]

Positron’s first Atlas server reached a customer in August 2024. Sohmers explained the shipping decision plainly: “We decided to start shipping because we hit the minimum feature set and stability that both we and our early customers agreed on.” At the time, speculative decoding and paged attention remained planned additions. The decision put a usable system into customers’ hands while leaving further capabilities to subsequent updates.[4][6]

FPGAs were a starting point rather than a permanent constraint. In September 2024, Sohmers said the next-generation product would use an ASIC informed by customer experience with the deployed system. His presentation on rapid iteration states the criterion: “Once the architecture proves stable and performant, harden it into custom silicon for mass scale and dramatically lower unit costs.”[6][5]

Sohmers identified the memory system as Positron’s principal architectural differentiator. His argument was that transformer inference could be constrained by memory bandwidth, so increasing floating-point compute alone would not solve the problem. He described combining directly attached high-bandwidth memory with DDR5 to balance bandwidth and capacity. He framed those choices around serving longer contexts and more concurrent users, with performance per dollar and performance per watt guiding purchase decisions.[6]

That economic focus also shaped his commercial strategy. In 2024, Sohmers said Positron intended to partner with existing cloud providers and data centers, selling them hardware rather than competing with their services. He described two early customer groups: organizations keeping proprietary models or data on premises, and cloud providers seeking cheaper inference with better throughput per user.[6]

Sohmers has linked inference efficiency to broader concerns about energy use and access to AI. In the same interview, he argued that continued expansion of expensive Nvidia-based infrastructure was environmentally unsustainable and risked concentrating compute among large technology companies. He wanted smaller companies to retain the ability to develop and run machine-learning systems. Internally, he described an engineering culture in which “no job is above or below anyone,” including leaders who should be willing to rack servers, scrub floors, or work through detailed technical problems.[6]

Before founding Positron in spring 2023, Sohmers worked on processor design, GPU-cloud infrastructure, and AI-chip strategy. Those earlier roles supply the technical and operational context for his inference-hardware work.[6]

Sohmers worked at MIT’s Institute for Soldier Nanotechnologies from February 2010 to May 2013, prototyping augmented reality, gestural interfaces, and laser communications for possible military use. His historical résumé describes a growing interest in embedded and high-performance computing during that period. In 2013, he received a Thiel Fellowship and left the Advanced Math and Science Academy Charter School to accept it.[3]

At REX Computing, Sohmers designed and built processors for mobile base stations and high-performance computing, beginning at age 17, according to his 2024 account. He subsequently worked on cryptocurrency ASICs, became Lambda’s principal hardware architect, and helped build its GPU cloud—an experience he described as his first deep involvement in AI compute using Nvidia GPUs. He then joined Groq as Director of Technology Strategy. He said the growth of generative AI and transformer workloads led him to start Positron, seeing a need for specialized compute.[6]