Positron AI raises $875M to build cheaper inference hardware

Positron AI has secured $875 million in new financing to expand its hardware business. The company is valued at $5 billion following this Series C round, which was co-led by NEA, Atreides Management, and Valor Equity Partners. The capital is intended to scale the production of systems designed specifically for running AI models efficiently.

The core of Positron’s strategy is a shift away from raw computing power toward memory capacity. As noted by GN technics/hardware (en-US), the company believes that memory bandwidth, not just processing speed, is the primary constraint for modern AI workloads. This approach aims to make inference cheaper and less energy-intensive for large enterprises.

Memory capacity drives the new design

Traditional AI hardware often relies on high-bandwidth memory, which is expensive and in short supply. Positron instead uses standard LPDDR5X memory, a component that is more widely available. This choice reduces dependency on constrained supply chains and allows for higher memory volumes without the premium cost of specialized chips.

The trade-off for this accessible approach is a focus on architectural efficiency. The systems are engineered to utilize over 90% of available memory bandwidth. This means that while individual components may not be the fastest on the market, the overall system delivers a higher volume of processed data for a lower energy cost.

Current deployments face scaling challenges

Positron’s first-generation Atlas system is already in use by several major customers, including Oracle Cloud Infrastructure and Jump Trading. Over 50 racks are currently deployed in production environments. These early systems have provided critical data on how memory-heavy architectures perform in real-world data centers.

However, scaling this technology presents significant engineering hurdles. The company must balance the need for massive memory capacity with the physical limits of heat dissipation. While the hardware supports both air and liquid cooling, integrating these systems into existing data center infrastructure requires careful planning to avoid power bottlenecks.

Next generation chips arrive in 2027

The new funding will accelerate the development of Asimov, a chip scheduled for production in late 2027. Designed on TSMC’s 3-nanometer process, Asimov will support up to 2.3 terabytes of memory per chip. This capacity is targeted at AI models with trillions of parameters, which require vast amounts of data to operate.

The risk lies in the execution of this massive technical leap. Positron must successfully transition from a startup to a large-scale manufacturer while simultaneously developing complex new silicon. If the next-generation systems fail to meet performance targets, the high valuation may prove difficult to sustain.