Positron raises $875 million to build an inference chip that targets 16-trillion-parameter models
The Reno startup closed a two-tranche Series C at a $5 billion valuation to fund its Asimov tapeout and a production inference system it calls Titan. The bet: the largest models will need purpose-built silicon, not repurposed training hardware.
The inference bottleneck is real money now. On October 9, 2026, Reno-based chip startup Positron closed $875 million across two tranches — a $375 million Series C and a $500 million Series C-1 — at a post-money valuation of approximately $5 billion. NEA led the financing, with investor Jim Clark also participating. The size and structure of the round signal that at least one set of serious backers believes the inference market will not be served adequately by hardware built for training.
What the money is actually for
Positron is directing the capital at two concrete milestones. The first is the tapeout of its Asimov chip — the step where a chip design is finalized and sent to a foundry for production masks. Tapeouts are expensive, irreversible commitments; funding one is a statement that the architecture is locked. The second use of capital is a 2-megawatt engineering data center, the company's proving ground for validating Asimov at scale before Titan ships to customers.
Neither project is speculative in the abstract sense. Tapeout funding and a small-scale data center are exactly the infrastructure a chip company needs between design and production. The question is whether the timeline holds.
The Titan system and the 16-trillion-parameter target
The end product Positron is building toward is called Titan — an inference system designed to combine eight Asimov chips and serve models exceeding 16 trillion parameters. That parameter threshold is worth pausing on. The largest publicly discussed models today operate in the hundreds of billions to low trillions of parameters. A system architected around 16-trillion-parameter workloads is either anticipating a generation of models that does not yet exist at scale, or positioning for private deployments where model size has already outpaced what general-purpose accelerators handle cleanly.
Either way, it frames Positron's market thesis clearly: the company is not competing for today's inference workloads. It is betting that the frontier will keep moving and that operators running the largest models will pay a premium for silicon that was designed for that job from first principles — rather than adapting training-optimized hardware to inference tasks it was never tuned for.
Production of the Titan system is targeted for the second half of 2027. That gives Positron roughly a year after the tapeout phase to move through fabrication, bring-up, system integration, and validation — an aggressive but not implausible schedule if the Asimov design is clean and the foundry relationship is stable.
What this round says about the inference hardware market
The two-tranche structure — C and C-1 closing together — suggests the round was built to give investors staged exposure while letting Positron capture a larger pool of capital quickly. NEA leading is notable: the firm has a long record in semiconductor bets and the patience horizon those investments require. Jim Clark's participation adds a founder-operator signal alongside the institutional lead.
At $5 billion, Positron is priced as a company that has already cleared the credibility threshold — not a seed-stage architecture bet, but a funded execution story. The valuation implies that NEA and co-investors believe Positron has something defensible in the Asimov design, not just a slide deck about large-model inference.
The broader shift this reflects: inference hardware is separating from training hardware as a distinct investment category. For years, the same chips — broadly — handled both jobs, with inference treated as a secondary concern. As model sizes grow and inference costs become a line item that enterprise operators scrutinize, the calculus changes. Purpose-built inference silicon starts to look less like a niche and more like infrastructure. Positron is not alone in that thesis, but with $875 million committed and a production target in late 2027, it is now one of the better-capitalized companies making the argument with real silicon.