Positron AI has raised $875 million to develop Asimov, an inference chip designed to tackle one of AI’s least glamorous but most expensive problems: moving enough memory quickly enough to keep models fed. The startup says its approach could make large-scale AI serving more efficient, but the headline performance claims remain unverified.
Watch Desk analysis
What happened
According to SDxCentral, Positron secured the money across two tranches, including a $375 million round co-led by Atreides Management and Valor Equity Partners and a $500 million Series C-1 led by New Enterprise Associates and Jim Clark. The funding values the company at about $5 billion.
Positron plans to use low-power LPDDR5X memory rather than the high-bandwidth memory more commonly associated with AI accelerators. Its Asimov chips are intended to sit inside Titan inference servers, with configurations ranging from 277GB to 2.3TB of memory per chip. The company says Asimov is designed for tapeout on TSMC’s 3nm process by the end of 2026, with production expected in the second half of 2027.
The company claims Asimov can use more than 90% of available memory bandwidth, compared with just under 30% for GPUs running the same models. Those are Positron’s figures, not an independent benchmark, and the chip is still on the road to production rather than available for customers today.
Why it matters
AI infrastructure is often discussed as a contest between ever larger accelerators. Positron is betting that the more stubborn constraint is memory: the data has to reach the computing hardware, and quickly, or an expensive processor spends its time waiting. A very costly waiting room is still a waiting room.
If the design works as advertised, the prize is more than a new chip. It could mean inference servers that handle larger models or longer contexts with less power and lower memory pressure. But tapeout, manufacturing and real customer workloads are still ahead, and the company must secure enough LPDDR5X supply to make the strategy practical at scale.
Our read
This is a serious hardware bet, not merely another funding round with an AI label attached. The interesting idea is the architectural choice to make memory capacity and bandwidth central to inference economics. The important caveat is equally plain: the strongest numbers are company claims, while the product remains pre-production.
What to watch
- Whether Asimov reaches its planned 2026 tapeout.
- Independent performance and efficiency results on real inference workloads.
- Whether Positron can secure LPDDR5X capacity at the volumes it needs.
- Customer deployments of Titan servers and evidence of lower cost per useful inference.
Discussion spark: If AI inference is increasingly limited by memory rather than raw compute, should the industry put more faith in specialised low-power chips, or will general-purpose GPUs remain the safer bet once software and supply chains are counted?
Sources and evidence
- Former Lambda, Groq alums secure $875M to build ‘memory-first’ AI inference chips – SDxCentral (16 September 2026, 08:40 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.