Positron's $875M Bet: Inference's Real Fight Is Memory, Not FLOPS

Positron, an inference-chip startup that did not exist until 2023, raised $875 million this week at a $5 billion valuation to take its next-generation silicon to market. That marks a roughly fivefold jump from February, when it was worth about $1 billion after its Series B. The money is not aimed at out-computing Nvidia. It is a wager on a narrower, more interesting idea: at inference, the constraint that decides how much money an AI model makes is no longer compute — it is memory.

The difference between training and inference is the whole argument. Training is the one-time job of building a model, weeks of heavy math on thousands of chips. Inference is every prompt your chatbot gets, every token it generates — and it runs forever. As models roll out to real users, the industry's spending is following the steady revenue. Deloitte projects inference will be roughly two-thirds of AI compute by 2026, in a market near $50 billion. That is where the fight for the next phase of this cycle will be decided, and it is not primarily a fight over raw FLOPS.

Inference is bounded by the memory wall. A GPU can compute far faster than it can feed data into its arithmetic units, so for cost-sensitive inference the expensive compute sits idle while the memory bus limits throughput. Every token has to be streamed out of memory, so the practical economics — cost per token, power per token — are set by how much memory bandwidth a chip actually puts to use, not by its headline FLOPS. And here is the twist in Positron's approach: it does not try to beat Nvidia at raw speed. It builds for how inference actually runs.

That is where commodity memory enters. Nvidia's flagship parts rely on high-bandwidth memory (HBM), the fastest tier of DRAM, packaged tightly onto the chip. HBM is fast but scarce, expensive, and supply-constrained, which is why the whole market sees an HBM crunch. Positron instead attaches ordinary LPDDR — the same kind of memory in high-end phones and laptops — to an organic substrate. Slower in raw terms, but abundant and far cheaper, and not rationed by the HBM supply chain.

The loaded question is whether that cheap memory can actually do the job, and here Positron's claim is the crux of its bet. It reports memory bandwidth utilization above 90% — as high as 93% — while best-in-class HBM inference workloads typically use only 20% to 50% of their bandwidth. In practical terms that means Nvidia's Rubin may advertise a raw ~16 TB/s of memory bandwidth, but only an estimated 2.5–8 TB/s is usable; Positron's roughly 11 TB/s total, at its claimed utilization, delivers about 10 TB/s of effective bandwidth. Same fight, won differently: more of what you bought actually gets used.

The power story is what makes the economics compound. Positron's shipping Atlas chip claims about three times the compute per watt of an H100, drawing around 2,000 watts for a workload an H200-based system runs at roughly 5,900. It is air-cooled and fits inside a normal rack, so it can be dropped into the installed base of older data centers — the 15–30 kW, air-cooled racks that cannot run liquid-cooled Blackwell or Rubin systems without a rebuild. That is a real customer. Positron is deploying more than 50 racks of Atlas into Oracle Cloud for mixture-of-experts inference, deals its chief executive has described as tens of millions of dollars, with Jump Trading as an investor-customer.

Now the part of the bet that is not yet delivered. The efficient-chip claims I just used to make the story work belong to Atlas, which is shipping. The reason Positron needed $875 million is Asimov, its second-generation chip: a TSMC-built part targeting tape-out late this year and production in the second half of 2027, with a stated goal of five times the tokens per watt of Nvidia's Rubin, and more than two terabytes of directly attached memory per accelerator against Rubin's 384 gigabytes. Five times is a roadmap target, not a result — it has not reached any financial statement.

That gap between claim and delivery is the discipline this story needs. As of late 2025 Positron had roughly $5.4 million in revenue. The company's own chief executive frames the ambition modestly, saying Positron would be "lucky" to capture a tenth of a percent of that ~$50 billion inference market. Read that against a $5 billion valuation and you understand the price is not paying for what the company has sold; it is paying for whether Asimov can be manufactured on schedule at the claimed efficiency. There is friction already: the company could not secure allocation at TSMC's Arizona fab — Apple and Nvidia get priority — so initial production is in Taiwan, and U.S. manufacturing is not expected before 2028.

None of this is an argument to buy or avoid Nvidia on its own. It is an argument about where the inference cycle's economics are heading, and on that the leader and the challenger largely agree. Nvidia has reportedly considered cutting memory on its Rubin Ultra chips amid the HBM shortage — a sign that even the incumbent sees inference as a memory-and-power problem and is moving to make its parts more memory-efficient. That is both validation that Positron is aiming at a real constraint and the deepest threat to it: if memory efficiency is where inference goes, the market leader with the software lock and the scale is already heading there too.

For an investor the useful takeaway is not a new ticker — Positron is private and not something a retail account can own. It is the mechanism. The memory wall is why inference cost, HBM supply, and power have started to move the AI trade more than FLOPS do, and it is why the contested ground of this cycle sits on whoever turns the least memory and power into the most usable throughput. Positron's $875 million is a sharp, coherent bet on that direction. Whether it earns the $5 billion price is a separate question, one that will only have an answer when Asimov's five-times claim lands as revenue — or does not.