OpenAI taps Broadcom and TSMC for its first in-house chip

Sam Altman's multi-trillion-dollar dream of building a global network of chip factories has officially hit the brakes. Instead, OpenAI is taking a far more realistic route to break free from its hardware bottleneck.

According to reports from Reuters, OpenAI is developing its first proprietary AI chip in partnership with Broadcom and TSMC, while simultaneously shelving its ambitious plan to build its own foundries due to massive costs and timelines.

Focusing on inference with Broadcom and TSMC

Rather than burning years and billions trying to build fabrication facilities from scratch, OpenAI is adopting the playbook used by Big Tech: keep chip design in-house and outsource manufacturing.

The company has been developing a custom chip with Broadcom that TSMC is expected to manufacture by 2026. The chip will focus specifically on AI inference—the stage where a trained model actually responds to user prompts and generates outputs.

Running models at scale for hundreds of millions of users has become an existential expense for OpenAI, which expects to lose $5 billion this year on $3.7 billion in revenue. Custom inference silicon can drastically cut energy use and lower cost per query compared to general-purpose GPUs.

To pull this off, OpenAI brought in former Google TPU engineers Thomas Norrie and Richard Ho. Broadcom, which has long helped Google design its custom TPU chips, is assisting OpenAI with manufacturing optimization and inter-chip data transfer.

Diversifying suppliers without alienating Nvidia

OpenAI isn't waiting until 2026 to ease its hardware squeeze. The company is already deploying AMD's MI300X processors through Microsoft Azure alongside its existing clusters.

Following the news, Broadcom shares rose 4.5% and AMD gained 3.7%.

However, OpenAI is walking a tightrope. Nvidia still commands over 80% of the AI chip market, and OpenAI remains deeply dependent on landing allocations of Nvidia's next-generation Blackwell GPUs for future model training. Adding AMD and building custom silicon is about hedging bets, not cutting ties.

Why it matters

OpenAI dropping its foundry fantasy to focus on custom inference silicon is a major reality check, but it is easily the smartest move the company could make right now.

For developers and everyday users, custom chips won't make models magically smarter, but they could dramatically lower API costs and response latency once production ramps up in 2026. Sustainable pricing for massive AI services cannot survive indefinitely on subsidized GPU rentals.

For the semiconductor industry, this further cements the trend toward in-house silicon. Google has TPUs, Amazon has Trainium, and Meta has MTIA. While Nvidia's moat in training remains practically untouchable for now, the recurring inference market is up for grabs, and its biggest customers are increasingly designing their own alternatives.