The companies that trained AI on Nvidia's chips are now designing the hardware to run it, and that's not a diversification play—it's a declaration of independence.
The Summary
- OpenAI announced Jalapeño, a custom inference chip built with Broadcom, joining Google, Apple, and SpaceX in designing proprietary silicon
- The move signals the end of total dependence on Nvidia for AI infrastructure, as Big Tech shifts from single-supplier risk to vertical integration
- Custom inference chips optimize for deployment cost and speed, not training horsepower, which is where the real money gets spent at scale
The Signal
OpenAI naming its chip Jalapeño is either the best or worst branding decision of 2026, but the strategy underneath is deadly serious. The company partnered with Broadcom to design custom silicon specifically for inference, the process of running trained models to generate responses. This isn't about training the next GPT model. It's about serving billions of ChatGPT queries without paying Nvidia's premium on every single one.
Inference and training are different beasts. Training requires massive parallel compute, the kind Nvidia's H100s and H200s excel at. Inference needs efficiency, low latency, and cost optimization at enormous scale. When you're serving millions of requests per day, shaving pennies per query compounds into millions in savings. That's the math driving this wave of custom chip development.
"The era of total dependence on Nvidia might be ending."
Google, Apple, and SpaceX are already down this path. Google's been running TPUs (Tensor Processing Units) since 2016. Apple's Neural Engine powers on-device AI in iPhones. SpaceX needs custom chips for Starlink's edge computing. Each company has specific workloads that generic GPUs handle inefficiently. OpenAI's inference demands are massive and growing. ChatGPT alone processes more queries in a day than most companies see in a year. Buying Nvidia chips for that is like hiring a Formula 1 car to deliver pizzas.
The Broadcom partnership is the tell. Broadcom doesn't make chips for hobbyists. They design custom ASICs (Application-Specific Integrated Circuits) for hyperscalers who know exactly what they need and are ordering in volume. This isn't OpenAI hedging. It's OpenAI committing to infrastructure independence at a scale that justifies the engineering investment, the fabrication costs, and the multi-year roadmap.
Key implications of the custom chip shift:
- Nvidia still wins on training, but loses recurring inference revenue
- Companies with enough scale can now optimize their entire stack, hardware included
- The barrier to entry for AI infrastructure just went up for everyone without chip design teams
The Implication
If you're building AI products, watch where the hyperscalers spend their chip budgets. Training will stay on Nvidia and AMD for now, but inference is fragmenting fast. That means lower costs for serving models, which means cheaper API calls, which means more applications become economically viable. The agent economy runs on inference, not training. Every autonomous system, every real-time AI interaction, every background process is an inference call. Cheaper inference means more agents doing more work.
For Nvidia, this is the beginning of margin compression on the inference side, but they're not dead. They still own training, and training is where the frontier models get built. The question is whether custom inference chips become good enough, fast enough, that companies stop buying Nvidia's all-in-one solutions and start designing hybrid stacks. OpenAI just made that future more likely.