OpenAI unsatisfied with Nvidia chips seeks for alternatives

OpenAI is shopping for AI chips beyond Nvidia (NVDA).

After months of stalled talks and bruised egos, OpenAI has grown unhappy with parts of Nvidia’s chip lineup.

They are actively testing alternatives for the most important phase of AI going forward.

OpenAI and Nvidia have been circling a massive partnership that includes up to 10 gigawatts of compute and an investment as high as $100 billion. That number was later walked back as nonbinding by Nvidia CEO Jensen Huang, who also questioned OpenAI’s business discipline while pointing to competition from Google and Anthropic.

Huang now calls the reporting nonsense, but the talks have still dragged on for months. During that time, OpenAI has cut GPU supply deals with AMD and explored inference-focused chips from startups like Cerebras and Groq.

The disruption behind the news: The next phase of AI is all about inference.

Training gets headlines, but inference is where costs explode and users feel latency.

Every chatbot reply, every code suggestion, every agent action runs through inference.

At scale, milliseconds matter and memory access becomes the bottleneck.

Read more about the 4 AI bottlenecks here.

Nvidia still owns training. Its CUDA moat and H100 class accelerators are unmatched for brute-force model building. But inference is different. It rewards chips with massive on-chip memory, fast SRAM, and deterministic performance. That’s why OpenAI is sniffing around nonstandard architectures and why Nvidia itself rushed to lock up Groq in a reported $20 billion licensing deal.

OpenAI’s internal roadmap is now optimized for speed, not just intelligence. This surfaced first in Codex, where latency directly impacts developer productivity. Sam Altman said outright that coding models put a big premium on speed. If a response takes 800 milliseconds instead of 200, the product feels broken.

This changes the power dynamic. Nvidia’s pricing assumes customers have no alternative. Inference-focused chips threaten that assumption. Even partial substitution matters. If OpenAI can move 20 percent of its inference load off Nvidia, that’s billions a year in leverage. It also means Nvidia’s stock narrative shifts from absolute dominance to segmented dominance.

There’s also a deployment reality. Inference runs everywhere. Data centers, edge servers, consumer laptops. OpenAI just shipped a Mac app for Codex. That’s not an Nvidia environment. Apple silicon thrives on memory bandwidth and efficiency, not monster GPUs. That makes a single-company AI stack impossible.

What to watch next

First, watch pricing.

If Nvidia starts offering explicit inference discounts or bundled licensing, the pressure is real.

Second, watch product launches.

Any OpenAI announcement that highlights speed or latency numbers will telegraph which silicon won. Sub-300 millisecond response times at scale are the bar.

Third, watch capital flows.

If Nvidia’s eventual investment in OpenAI lands well below the rumored $100 billion ceiling, that’s loss of leverage.

Finally, watch startups.

If inference-first chip companies survive another 12 months, Nvidia’s grip has already loosened.

OpenAI unsatisfaction with Nvidia chips isn’t about betrayal or drama. Inference is where AI becomes a product. Nvidia can adapt, but it won’t be the only winner. Anyone betting otherwise is clinging to last year’s narrative in a market that’s already moved on.

Nvidia (NVDA) has a Disruption Score of 4. Click here to learn how we calculate the Disruption Score.  

Nvidia is also part of the Disruption Aristocrats, our quarterly list of the world’s top disruptive stocks.