Positron AI raised $875 million this week at a $5 billion valuation, roughly five times what it was worth in February. That is a stupefying jump in seven months, and the obvious question is whether it is madness or a real crack in Nvidia’s armour. The honest answer is: a bit of both, and the reason is worth understanding even if you never buy a chip.
Positron is not trying to out-Nvidia Nvidia at training the giant models. It is going after inference, the running of models once they are trained, which is where the bulk of AI’s actual day-to-day compute (and cost) now lives. Its next chip, Asimov, is built around memory rather than raw compute muscle, pairing its architecture with as much as 2,304 GB of commodity LPDDR5X memory per chip. The bet, in plain terms: most inference is bottlenecked by memory, not flops, and cheap commodity memory can undercut Nvidia’s expensive high-bandwidth stuff on the metric that matters, cost per token served.
The investment case, both ways
The bull view is straightforward. Inference demand is exploding, Nvidia’s margins on it are famously fat, and a challenger that delivers similar throughput at lower cost and power has an enormous addressable market to eat into. The investor list is not a joke either: NEA, the Qatar Investment Authority, Netscape’s Jim Clark, Hudson River Trading and Cisco all put money in. Smart money is treating “someone dents Nvidia’s inference lead” as a bet worth making.
The bear view is equally real. Asimov does not tape out until the end of 2026 and is not in production until the second half of 2027. That is a long time to promise performance you have not delivered at volume, in a field where Nvidia is not standing still and where the graveyard is full of “Nvidia killers” that never killed anything. A 5x valuation jump in seven months, on a chip that does not yet exist in customers’ racks, is exactly the kind of thing that looks visionary in a boom and reckless in a bust. (None of this is investment advice.)
The useful frame: Positron is a pure-play wager that the AI money is migrating from training to inference, and that Nvidia’s weakest flank is cost-per-inference rather than peak performance. If you believe that thesis, this is one of the cleaner ways to express it. If you think Nvidia’s software lock-in makes hardware price irrelevant, you will see an $875m bonfire waiting to happen.
Related: our even-handed look at whether Nvidia itself is a bubble or a bargain.