The man who built Lambda's GPU empire - then left to make GPUs optional.
Spend eight years building one of the world's largest GPU clouds and you learn something few people ever will: where exactly the machine wastes your money. Mitesh Agrawal was Lambda's Chief Operating Officer. He watched GPU inference deployments up close. He saw idle silicon, runaway power bills, and memory bottlenecks that no firmware patch would fix. So he left - not for a competitor, but for the only company that was attacking the problem at the hardware layer.
Positron AI was founded in spring 2023 with a deceptively simple thesis: the GPU was never designed for inference. It is a tool optimized for training runs, for the long bruising grind of teaching a model to think. Once you need that model to actually respond - in milliseconds, at scale, twenty billion times a day - you are running Ferrari V12 engines to power a delivery truck. Agrawal saw this from Lambda's operations desk. He also saw that no one had built the right truck yet.
He joined Positron as CEO in February 2025, arriving at a company that had already done something remarkable: shipped a production AI inference system - Atlas - with fifteen people and under $12.5 million. Most semiconductor startups spend that getting their first tape-out design review approved. Positron got Cloudflare running on their hardware. That is a different kind of proof.
Atlas delivers three times the performance of Nvidia's H100 while consuming a third of the power. The numbers read like a benchmark sheet that got loose from marketing - except Cloudflare's engineers are the ones validating them. When a customer with Cloudflare's operational discipline puts your chips into production, the conversation about benchmarks is over.
By February 2026, Positron had closed a $230 million Series B at a $1 billion-plus valuation - 34 months after founding. The round was co-led by ARENA Private Wealth, Jump Trading, and Unless, with strategic participation from the Qatar Investment Authority, Arm Holdings, and Helena. When Arm - the architecture underlying most of the world's chips - bets on you, the message is not subtle.
The next target is Asimov. Named after Agrawal's favorite author (he counts Isaac Asimov's Foundation series among his regulars), the next-generation custom silicon is on track for tape-out by the end of 2026, with production beginning in early 2027. Asimov will ship with 2,304 GB of RAM per device - against 384 GB for Nvidia's forthcoming Rubin GPU. That is not a spec-sheet quirk. Memory is the actual bottleneck in large language model inference, the number that determines how many parameters you can load and how fast tokens flow to the user. Positron is building the chip that stops being the bottleneck.
The path to Positron runs through three universities and several career pivots that most chip executives would find baffling. Agrawal graduated from Georgia Tech with degrees in biomedical engineering and biology. He went to Stanford for a Master's in chemical engineering. Then Berkeley Haas for the MBA. Then Cornerstone Research as a financial analyst. Then VMware as a software consultant. Then he co-founded Aerigo Water Technologies. Then - almost as an afterthought that lasted eight years - Lambda Labs, where the COO job turned out to be a front-row education in every constraint that would later define Positron.
That trajectory is not chaos. It is a man who moves toward hard problems, regardless of whether his resume says he belongs there. Chemical engineering teaches you to think in systems under constraints. Finance teaches you capital efficiency. Operations teaches you what actually breaks at scale. All three disciplines show up in how Positron was built: first production system under $12.5 million, customer validation before hardware scaling, and an architecture designed around the specific failure modes of real-world inference deployments rather than around impressive benchmark numbers.
Outside the office, Agrawal is a tennis player who admits the court rarely sees him these days. He is a science fiction reader - the Foundation series informed the chip name, and it is not the first piece of his work to carry a literary fingerprint. He became a father recently, navigating new parenthood alongside Series B negotiations and chip roadmap decisions. He says, simply, that he is grateful for an industry that aligns with what he actually cares about.
That alignment matters. Positron is making a bet that the next decade of AI infrastructure is not a Nvidia story - it is a differentiation story. The companies that win the inference market will do it on efficiency, cost-per-token, and memory per dollar, not on training pedigree. Agrawal built his career reading the inefficiencies in systems others considered optimized. He is doing it again. This time with custom silicon, $305 million in capital, and a chip named after the man who imagined robots before anyone had built one.
"AI hardware looks glamorous from the outside, but it's brutally hard... stay close to customer pain."- Mitesh Agrawal, CEO, Positron AI
Performance benchmarks from public Positron AI disclosures
Lower power consumption is better. Performance measured on transformer inference workloads.
Biomedical / Medical Engineering & Biology
Chemical Engineering
Business Administration
Energy availability has emerged as a key bottleneck for AI deployment.
A customer becoming an investor is one of the strongest validations we can receive. It signals both technical conviction and real-world demand.
Memory is the other giant bottleneck in inference, and our next generation Asimov custom silicon will ship with over 2304 GB of RAM per device next year, versus just 384 GB for Rubin.
Competing with Nvidia means matching their shipping frequency. Development speed is an essential competitive advantage.
GPUs are incredible for training, but once you move to deployment, most of their silicon sits idle.
We didn't want to optimize around benchmarks or theoretical FLOPs. We wanted to design silicon that solved the real-world pain of deploying AI at scale.
"Positron is the company I wished existed when I was at Lambda."- Mitesh Agrawal