What if AI chips stopped costing so much?
Training artificial intelligence systems today faces a bottleneck similar to cooking a banquet in a tiny kitchen: a lot of space is needed for the ingredients. In the hardware world, that "space" is high-speed HBM memory, an expensive component with limited supply. The startup Positron AI, recently turned unicorn, challenges this paradigm with a bold proposal: use conventional memory, but on a colossal scale. 🧠
The strategy: prioritize capacity over raw speed
Instead of the exclusive and fast Ferrari that HBM memory represents, Positron bets on a fleet of trucks. Its new accelerator, named Asimov, will integrate large amounts of LPDDR5x memory, the same efficient technology used by modern smartphones, but multiplied. The goal is to incorporate several terabytes of this memory, an enormous amount that far exceeds the usual. This trade-off approach exchanges some pure speed for massive capacity, lower cost, and more energy-efficient consumption for moving large volumes of data.
Key advantages of the Positron AI model:- Reduce costs: LPDDR5x memory is significantly cheaper than high-end HBM.
- Scale capacity: Allows integrating terabytes of memory, something very complex and costly with HBM.
- Improve efficiency: Manage large AI models in a more energy-sustainably way.
It's like financing the construction of a new type of airplane years before it flies.
A timeline that reflects long-term ambition
A surprising detail is the project's timeline. Positron AI has secured 230 million dollars in funding based on its promise, but the Asimov chip won't be ready for its first tests until the end of the first quarter of 2027. This deadline demonstrates the magnitude of the technical challenge and investors' faith in reinventing the basic infrastructure of data centers for AI, freeing it from traditional components that limit its growth.
Implications of this development:- Democratize AI: It could make training large models more accessible, not just for tech giants.
- Change the market: Offer a real alternative to the current dependence on scarce HBM memory.
- Accelerate innovation: By lowering the entry barrier, more companies could experiment with advanced AI.
The potential future of a simple idea
If Positron AI succeeds, we could witness a real democratization of high-level AI. The landscape could evolve from an environment where only a few companies have supercomputers to a model where AI power can be rented more affordably and scalably. In the future, your favorite AI assistant might run, in part, thanks to an architecture inspired by your phone's memory, but taken to a truly titanic scale. 🚀