Analysis
Positron AI raised $875 million in a Series C financing, the company announced September 10, quadrupling its valuation in seven months. The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, Dylan Patel's SemiAnalysis Capital and Jim Clark, the Silicon Graphics and Netscape founder, Yahoo Finance reported.
- New round -- $875M Series C.
- New valuation -- $5B post-money.
- Prior mark, Feb 2026 -- $1B.
- Additional financial investors -- DFJ Growth, Qatar Investment Authority, Resilience Reserve, Arena Private Wealth, Natural Capital, Helena, 1517 Fund, Flume Ventures, Unless, Boardman Bay Capital Management, Fincadia Advisors, Banyan Ventures and U First Capital.
- Strategic investors -- VentureTech Alliance, Hudson River Trading, Cisco Investments and Naver Ventures, the last of which ties Positron to South Korea's Naver in a strategic-compute relationship rather than a purely financial one.
The technical bet: memory over raw compute
Positron designs inference-only chips built around memory capacity and bandwidth rather than the raw compute-per-chip metric Nvidia has optimized for. Its next-generation chip, Asimov, pairs Positron's compute architecture with between 288 gigabytes and 2,304 gigabytes of memory per chip using commodity LPDDR5X memory -- a far cheaper, more available memory type than the HBM4 that has become one of the tightest bottlenecks in AI hardware this year, the same shortage Pulse covered in this issue's Samsung-OpenAI chip story. The bet is that inference workloads, particularly for large-context and mixture-of-experts models, are increasingly memory-bound rather than compute-bound, making Positron's design cheaper per token served even if it can't match Nvidia's raw compute density.
Asimov is scheduled to tape out on TSMC's N3P process at the end of 2026, with production targeted for the second half of 2027 -- meaning Positron's $5 billion valuation prices a chip that won't ship commercially for another year, on top of the typical execution risk of any first-time chip design reaching volume production without yield or timeline slippage.
The competitive field
Positron enters a memory-centric inference-chip category alongside a widening field of specialized challengers to Nvidia's GPU dominance:
- Groq -- the LPU inference-chip pioneer whose technology and founding team Nvidia absorbed in a nearly $20 billion licensing-and-hire deal in December 2025, now under DOJ antitrust review; the remaining Groq entity has raised $650 million of its own capital to keep competing without its original leadership.
- Cerebras -- builds wafer-scale chips with roughly four trillion transistors per die, raised $1.1B in 2025 at an $8.1B valuation.
- SambaNova -- reconfigurable dataflow chips with tiered SRAM/HBM/DRAM memory, raised $1B in July 2026 alone.
Positron's $5 billion mark sits below Cerebras's $8.1 billion and well below what Nvidia paid for Groq's technology, but its quadrupling in seven months matches the general pace of 2026's AI-infrastructure re-ratings -- Harvey, Clay and Cognition have each re-rated at similarly fast multiples this same month on the application-software side of AI, while Positron represents the same dynamic playing out in inference silicon.
What the multiple assumes
None of Positron's investors have disclosed a revenue figure alongside this round, and Asimov's late-2026 tape-out means the company is being valued almost entirely on its architecture and roadmap rather than shipped, benchmarked silicon competing against Nvidia's current-generation Blackwell or Groq's already-deployed LPUs in production inference workloads. A memory-first architecture betting on the token-cost economics of ever-larger context windows is a real and increasingly validated thesis, but Positron's $5 billion valuation will ultimately be tested by Asimov's actual yields and per-token cost once it ships in the second half of 2027 -- eighteen months from now, not by this funding announcement.
