Why is Big Tech all making its own AI chips, and what does that mean for NVIDIA?
In June 2026, OpenAI announced Jalapeño, the first inference chip designed by itself, in collaboration with Broadcom. The detail that caught my attention the most was not the chip's launch, but the time: the design was completed in 9 months, with the help of OpenAI's own AI models. A pure software company, with no traditional hardware division, just completed a chip in less than a year. OpenAI is not the first. And that's what's worth discussing.
No one wants to depend on a single supplier
Look at the list: Apple has been making its own AI chips since 2017, Google since 2015, Tesla built Dojo since 2021, SpaceX operates Colossus for xAI since 2024, and now OpenAI with Jalapeño. If only one or two companies do this, it can be called a private strategy. But when nearly all the leading AI companies are doing the same thing, it's a signal of a systemic shift. The core reason is not complicated: NVIDIA holds about 85-92% market share in the AI accelerator hardware segment. In the model training segment, that number exceeds 90%. When one supplier controls nearly the entire supply of an essential resource, the biggest buyers soon realize the risks: high prices, long lead times, and more importantly, your product development roadmap is tied to someone else's.
Google has been making its own chips since 2015 Google is the earliest and also the most patient. Their first dedicated processor (Tensor Processing Unit — TPU) has been running secretly in Google's data centers since 2015, before being announced at Google I/O in 2016. The initial reason was not concern about NVIDIA, in fact NVIDIA at that time did not dominate AI the way they do today. The reason is economics: Google realized that with the scale of its operations, a chip designed for the right purpose would save significantly compared to a general-purpose GPU. In 2026, Google is operating the 8th generation of TPUs, with two separate variants for training and inference.
Apple also promoted its own chip development with the A11 Bionic integrated Neural Engine Apple goes in a different direction. Apple's target is not data centers but handheld devices. Since the A11 Bionic chip launched in 2017, the first iPhone with a dedicated Neural Engine processor, Apple has gradually integrated AI processing capabilities deeper into every chip they design themselves. To date, the entire M and A chip lines have their own Neural Engine, helping Apple Intelligence run right on the device without having to send data to the cloud, which Apple calls "on-device AI" and sees as a privacy advantage. Tesla started from a painful lesson. In 2016, Tesla used NVIDIA chips for its driving assistance system. By 2019, they launched their first self-designed chip for Autopilot and claimed performance many times better than commercial chips at a lower cost. Since then, Tesla has continued to invest in Dojo, a self-built supercomputer to train machine vision models for self-driving cars.
How difficult is it to make your own chips?
How will NVIDIA perceive this?
NVIDIA's 85-92% market share in the AI hardware segment sounds solid, but I'm paying more attention to the direction of movement: from a peak of about 87% in 2024, that number is forecast to decline to about 75% in 2026. That decline is not catastrophic, but it is happening in the context of the total AI hardware market exploding from $120 billion in 2025 to an expected more than 200 billion do. Simply put: NVIDIA lost some relative market share, but absolute revenue still increased strongly because the "pie" is growing very quickly.
It's not easy to make things difficult for NVIDIA, but they will be affected to a certain extent in the Inference segment