Positron AI Secures $875 Million to Build Memory-First Inference Chips

When Groq raised $640 million in August 2024 to accelerate artificial intelligence inference, it relied on a static random-access memory architecture that prioritized raw speed over capacity. Positron AI has now secured a larger $875 million Series C round by taking the exact opposite approach: designing an inference chip that runs large language models entirely on cheap, abundant consumer-grade memory. The funding, announced on September 10, 2026, catapults the hardware startup to a $5 billion valuation just seven months after its previous capital raise.[1][4]

The round was led by New Enterprise Associates, injecting capital into a market currently dominated by Nvidia's high-end graphics processing units. Nvidia's flagship H100 and B200 chips rely on High-Bandwidth Memory, a specialized component that costs roughly $15 to $20 per gigabyte and remains in chronic short supply globally. Positron's silicon architecture bypasses this bottleneck entirely, utilizing standard LPDDR5 memory—the same modules found in high-end smartphones and laptops—which prices out at less than $3 per gigabyte.[3][5]

That cost differential fundamentally alters the unit economics of running deployed AI models. While training a frontier model requires massive parallel processing power, "inference"—the act of generating responses for users—is primarily constrained by memory bandwidth and capacity. By pairing standard memory with a proprietary interconnect fabric, Positron claims its hardware can serve a 70-billion parameter model at one-fifth the operational cost of an equivalent Nvidia cluster.[5]

"The industry has spent the last three years brute-forcing inference by buying training chips to do a deployment job," said Positron AI chief executive officer Sarah Lin in the funding announcement. "We designed a processor specifically for the deployment phase, where memory capacity matters more than raw compute, allowing enterprises to scale their AI applications without waiting in line for HBM allocations."[2][5]

The $5 billion valuation represents a rapid ascent for the hardware developer, which emerged from stealth mode in late 2024. The Series C capital injection will fund the tape-out and mass production of its second-generation silicon, codenamed "Neutrino," scheduled for commercial delivery in the first quarter of 2027.[1][2][4]

Positron is not the only challenger attempting to break the inference bottleneck. Competitors like Cerebras Systems and SambaNova have also raised hundreds of millions to build specialized AI hardware. However, Positron's explicit focus on consumer-grade memory integration separates it from rivals that are largely attempting to build larger logic chips or wafer-scale processors.[3][4]

The shift away from high-bandwidth memory also insulates Positron from the supply chain constraints currently throttling the broader AI hardware market. Advanced packaging facilities, which assemble specialized memory modules onto logic chips, are booked out through late 2027. By utilizing standard packaging techniques and LPDDR5 memory sourced from suppliers like Samsung and Micron, Positron can theoretically manufacture its chips at a higher volume and lower defect rate.[5]

With $875 million in fresh capital, the immediate hurdle for Positron shifts from engineering to software ecosystem adoption. The company must convince developers to port their models away from Nvidia's entrenched CUDA software platform onto Positron's proprietary compiler. If the startup can demonstrate seamless model migration by its 2027 hardware launch, the cost savings of its memory architecture could force a structural pricing shift across the cloud computing sector.[3]