Positron Raises $875M Series C At $5B To Chase Nvidia On Inference

NEA leads an $875M Series C for Positron at $5B, with Atreides, Valor Equity, Andra, SemiAnalysis Capital and Jim Clark backing the memory-first Asimov inference chip.

Positron Raises $875M Series C At $5B To Chase Nvidia On Inference

Positron just quadrupled its price tag. The Reno-based inference-chip maker last valued at ~$1B in early 2026 said on September 10 it has raised $875M in a Series C led by NEA at a $5B post-money valuation, adding Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital and Netscape co-founder Jim Clark to the cap table.

Memory-first Asimov chip, Titan servers, real customers

Positron's pitch is that inference is now a distinct compute market from training. Its second-generation Asimov processor packs up to 2.3 TB of on-package memory per chip and slots into the Titan server, which the company has already delivered to Oracle, Jump Trading and Parasail after the earlier Atlas system. That customer list — and the pricing pressure on Nvidia's H200/B200 inventory — is the crux of the Series C thesis: cheaper tokens-per-second at production scale, without CUDA.

AI inference chip architecture close-up

Six months, 5x the valuation

The step-up is aggressive even by 2026 standards. Positron traded at roughly $1B in the $230M Series B from Valor and NEA earlier this year; today's $5B mark is a 5x jump in about six months, funded by an inference boom that has pushed Anthropic to lock in 14.8 GW of compute and prompted DOJ scrutiny of NVIDIA's latest capacity plays.

Where the money goes

Positron plans to expand production of the Titan server, finish qualifying the next Asimov revision at a US foundry partner, and open a second design center in Silicon Valley. The company is also hiring aggressively on the compilers side — the piece of the stack that decides whether developers see Positron as an Nvidia alternative or a bespoke ASIC that requires re-writes.

The specialized-inference race

Positron now sits alongside Groq, Cerebras, Etched, Rebellions and SambaNova in a widening pool of Nvidia challengers focused specifically on inference. Two forces are pushing capital into this niche: HBM shortages that make each Nvidia GPU more expensive, and enterprise buyers who want lower serving costs on models they already trained elsewhere. If Positron can hold its Oracle deployment and add a hyperscaler design win in 2027, the $5B mark will look conservative.

Reporting based on coverage from The Wall Street Journal, TechCrunch and SemiAnalysis.

Category: Funding & Investments

Tags: AI Models Nvidia AI Chips GPU Compute

Related Articles