The most expensive part of many AI chips is not the processor but the memory next to it. Positron is betting it can do without that memory, called HBM. The startup announced on September 10 a funding round of up to $875 million that values it at $5 billion.
HBM, short for high-bandwidth memory, is stacked chips packed tightly beside the processor. It is fast, but supply is tight, and Reuters reported price rises of 20 to 50 percent for Chinese chipmakers the same day. Positron's next chip, Asimov, uses LPDDR5X instead, a common type of memory that is easier to buy and lets one chip hold much more.
Built for inference
Positron makes hardware for inference, the work of running an AI model after it has been trained. Each Asimov chip is to carry between 288 gigabytes and 2,304 gigabytes of memory, depending on the setup. A Titan system with four to eight chips is designed for models above 16 trillion parameters and inputs of more than 10 million tokens. A token is a piece of a word.
These are design targets. Asimov will be made by TSMC on its N3P process, with tapeout, the point when the design goes to the factory, planned for the end of 2026. Production is targeted for the second half of 2027.
How the money is split
The round has two parts. A $375 million Series C is firm, at a $3.5 billion valuation before the money. A second part of up to $500 million, led by NEA and Netscape founder Jim Clark, is not fully committed. In February Positron was valued at about $1 billion.
The company already sells an earlier chip, Atlas, which runs in more than 50 racks at Oracle's cloud. Asimov has no published performance figures yet.
