Positron, a U.S. startup building chips dedicated to AI inference, is in talks to raise around $750 million in new financing split across two rounds, according to a report on July 8, 2026 — part of a wider push by companies to compete with Nvidia.
July 8, 2026 · Positron
Positron Chases $750M to Challenge Nvidia in AI Inference
The U.S. inference-chip startup is in talks for a two-tranche raise — at valuations near $3.5B rising to about $5B — betting energy-efficient, high-memory hardware can undercut Nvidia's dominant GPUs.
~$750M
New financing sought across two rounds
$5B
Target valuation on second tranche
3×
Claimed perf-per-watt vs H100/H200
Inference speed, in a leaner power budget
Company claim: tokens/sec per user on Llama 3.1 8B (BF16). Atlas hits more throughput while drawing a fraction of the power.
280
Positron Atlas
tokens/s · ~2000W
180
8-way DGX H200
tokens/s · ~5900W
More throughput on roughly one-third the power draw — the core of Positron's efficiency pitch.
Shipping now
Atlas
FPGA-based inference appliance
~256 GB memory in some configs
System draw ~2000W
Up to 500B-parameter Transformers
Planned 2027
Asimov
Purpose-built inference ASIC
2 TB+ memory per chip
>4 TB/s, under 400W peak
8 TB+ "Superintelligence-in-a-Box"
Why it's attractive
Air-cooled, higher rack density, no major overhaul
Drop-in: Hugging Face + OpenAI-compatible API
Live use: Cloudflare, Parasail, CDNs, trading
U.S. design, fabrication and assembly
Open questions
Few independent third-party benchmarks
Claims come mainly from company/investors
Unverified real-world scaling at hyperscale
Nvidia's mature software ecosystem lead
A crowded inference field — Groq, Cerebras, Etched, FuriosaAI — makes this raise a real test. Founded 2023 in Reno, Positron reached unicorn status via a $230M Series B and has now pushed total funding above $300M .
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…