Tech
OpenAI Filed a Chip Patent Before It Had a Chip Story
Application US18/903,427, published April 2, describes an inference chip designed for 20 HBM stacks where Nvidia's current data center GPU ships 8. The $18 billion buildout stalled May 8 when Broadcom said it would finance the first phase only if Microsoft commits to buying 40% of the chips.
OpenAI's $18 billion Broadcom chip buildout stalled May 8 when Broadcom made financing conditional on Microsoft buying roughly 40% of the chips.
Application US18/903,427, filed October 1, 2024 and published April 2, 2026, is assigned to OpenAI Opco LLC. The six named inventors are Clive Chan, Chian-Min Richard Ho, Christopher Leary, Ravi Narayanaswami, Devin Persaud, and Kaushik Vaidyanathan.
Chan joined OpenAI in January 2024 as its first chip designer, hired before the company had committed to custom silicon. Ho had built tensor processing units at Google before joining OpenAI. Chan and Ho are the only two inventors on this filing who had shipped a custom AI accelerator at scale.
Memory and Distance
The patent targets the shoreline problem: JEDEC standards restrict HBM chiplets to within 6mm of a compute chiplet, capping how many stacks a chip can reach. OpenAI's design inserts embedded logic bridges with active amplification circuitry, extending that distance to at least 16mm.
The architecture supports up to 20 HBM stacks per compute chiplet. Nvidia's GB300 Blackwell Ultra uses eight.
LLM token generation is memory-bandwidth-bound rather than compute-bound. Models or context windows too large for a single Nvidia GPU run on one Titan chiplet without distributing the load.
That count advantage narrows before Titan ships. Nvidia's Rubin Ultra, targeting H2 2027, reaches 16 HBM4e stacks for 1 TB of total capacity per GPU. HBM4e's higher bandwidth per stack means the 4-stack lead does not straightforwardly convert to an equivalent memory throughput edge in the same deployment window.
The company's public hardware story runs through io Products, the Jony Ive hardware company acquired for $6.5 billion, whose consumer device slipped to 2027. Chan's April 2026 X post put the program nine months from "the fastest and largest volume ramp of any first time chip," targeting 10 gigawatts.
The Microsoft Condition
The first phase covers 1.3 gigawatts of data center capacity and roughly $18 billion in cost. Broadcom agreed to co-design and deploy the chips, codenamed Titan, under a deal announced in October 2025.
Broadcom's condition is that Microsoft buy roughly 40% of Titan chips, install them in its own data centers, and lease capacity back to OpenAI. Microsoft has not signed a firm purchase commitment. The sticking point is data center design: OpenAI wants facilities built for Titan; Microsoft designs for fungibility, preserving the ability to serve multiple tenants and redeploy capacity if demand shifts.
Sachin Katti, OpenAI's head of compute, has called the commercial structure "financially unattractive."
The financing impasse reveals that OpenAI's route around Nvidia runs through Microsoft, whose multi-tenant data center preference converts the independence thesis into a conditional offer.
If Titan ships on schedule in Q1 2027, it arrives before Nvidia's Rubin Ultra, which targets H2 2027 with 16 HBM4e stacks. Miss the window and Nvidia closes half the stack-count gap while OpenAI is still negotiating who pays for the data center.