Positron AI raised $875M total to build memory-focused inference chips and is targeting hyperscalers with plans to tape out this year

Positron AI

Positron AI has raised $875M in total to build memory-focused inference chips, with a $230M Series B closed in February 2025 that Mitesh Agrawal says remains largely untouched because the company has been generating revenue this year. The raise isn't about near-term burn — it's about signaling to TSMC and hyperscalers that Positron has the balance sheet to scale. Even $875M in equity, Agrawal notes, isn't enough to fund a 200-megawatt deployment. The real financing for that scale comes from infrastructure capital, the way Lambda structured its deals.

What the chip actually does

Positron's architecture is built around commodity DRAM — specifically LPDDR5x — rather than the HBM that sits at the center of NVIDIA's supply chain constraints and CoWoS packaging bottlenecks. The bet is that commodity memory sidesteps those chokepoints, making it easier to scale supply. The tradeoff is raw bandwidth: commodity memory is slower, and the technical innovation Agrawal describes is in the architecture's ability to compensate for that. He's candid that there's no free lunch in silicon.

The first-generation product is FPGA-based — pre-taped-out silicon, used to validate the architecture before committing to a full tape-out. 50 Atlas racks are deployed at Oracle today. A full chip tape-out is targeted for end of 2025, with production ramp in H2 2027.

“One of the reasons we raised, you know, $875,000,000 is not like we need $875,000,000 to spend tomorrow or even in the next six months. I mean, look, we raised $230,000,000 in series b in February of this year. Untouched... we are taping out this year production kinda ramp up in second half of twenty twenty seven... if you really wanna get to, like, the frontier labs and hyperscalers, you're really talking about hundreds of megawatts.”

Demand dynamics

Agrawal frames the customer conversation around two questions hyperscalers and frontier labs actually ask: can you fabricate at scale, and can you fit in our data centers without requiring liquid cooling infrastructure already allocated to GPUs and TPUs. Performance and TCO matter, but those are table stakes — the real filter is supply chain credibility and deployment compatibility.

The scale threshold for hyperscalers is stark. Agrawal says the rooms he walks into want proposals framed in gigawatts. For 2028, he says Positron needs a credible path to hundreds of megawatts — meaning 300 to 500-plus megawatts — as a stepping stone. A gigawatt of AI compute, even on cheaper ASIC economics, represents tens of billions of dollars of infrastructure spend.

Alongside hyperscalers, Positron is selling into inference-as-a-service providers, sovereign AI clouds, and quantitative finance. Agrawal mentions Jump Trading and i3d.net as existing customers. Those channels can carry the company to $200M–$2B in revenue on their own trajectory, but landing a frontier lab or hyperscaler is what makes Positron a long-term viable company at scale.

Software integration

Positron's software stack is being built around open-source inference frameworks — primarily vLLM and SGLang — because that's where the broad market operates. Agrawal says the speed of model bring-up is improving fast: when Gemma dropped recently, the team had it running on Atlas within hours. But raw bring-up speed isn't the same as optimization, and when a customer is large enough to matter, they come in and co-optimize regardless. Anthropic, OpenAI, and hyperscalers bring their own inference stacks and will drive that process themselves.

Team and timeline

Positron crossed 100 employees this week, with more than 50 hired in the last three months. The company shipped its first-generation product to a customer in 15 months from founding, with a team of fewer than 20 people. The founders — Thomas Summers and Edward Kemet — came out of Groq. Agrawal joined from Lambda Labs in early 2025, where he had been part of the founding team.

Investor Gavin Baker framed the opportunity in a board meeting as a 1% market share equals $100B enterprise value problem. With NVIDIA approaching a $5T market cap, the math holds without requiring Positron to displace anyone — just to be credible enough to capture a sliver of an expanding market. That framing also captures the operational reality: Agrawal says no one should be billing themselves as an NVIDIA rival until they're at 10% of NVIDIA's revenue.