Custom chips: OpenAI, Google, and SpaceX challenge Nvidia's AI dominance

The answer is brutal: Nvidia sells the chips and sets the prices. If you're OpenAI or Google training billion-parameter models, depending on one company means endless wait times, skyrocketing costs, and zero control over the technology roadmap. The real issue isn't Nvidia's quality, it's their power: whoever supplies the chip controls the game.

OpenAI just launched Jalapeño, developed with Broadcom, for inference (running already-trained models, not training from scratch). Google has had Tensor Processing Units for years, Apple has Neural Engines in its chips, SpaceX is building its own processors. Each has different reasons, but the pattern is identical: no single-supplier dependency.

It's not an all-out war yet. Each company builds specialized chips for its own needs, not to compete directly in the general market. OpenAI targets inference with Jalapeño, Google focuses on its data centers, Apple integrates everything into phones. Nvidia still dominates model training, where raw power still matters most.

The real shift is strategic choice: yesterday, Nvidia dependency was accepted (what else would you do?), today it's a risk. Building a custom chip is expensive, but savings across millions of daily requests pay for it. Plus, owning the design means optimizing for your own software without compromises.

Nvidia won't disappear tomorrow, that's clear. But the monopoly crumbles piece by piece. As OpenAI, Google, Apple, and SpaceX build their own chips, Nvidia's orders drop, growth slows, prices might finally ease up. For regular users, this means AI could get cheaper and faster in the medium term.

The real winner? Broadcom and other component suppliers helping design these custom chips. The real loser? Anyone locked into Nvidia's platform with no alternatives. For the market overall, more competition is always good news.