Why OpenAI, Google and SpaceX Are Suddenly Building Their Own AI Chips
The biggest companies in tech just decided Nvidia isn't enough. Here's what's actually going on.
For the last few years, if you wanted to run serious AI, you bought Nvidia GPUs. That was the deal. Nvidia built the best hardware, charged whatever it wanted, and everyone paid up because there was no real alternative. OpenAI, Google, Meta, Amazon — they all lined up and waited their turn.
That arrangement is now falling apart. And the pace at which it's happening is almost hard to keep up with.
In just the last few months, OpenAI unveiled its first custom chip, Google announced its eighth generation of in-house processors — this time splitting the design into two separate chips for the first time — and SpaceX is actively building a chip fab as a joint venture with Tesla. These aren't side projects. These are multi-billion dollar bets that the AI industry's dependence on a single chip supplier is a structural problem that needs to be solved.
OpenAI's First Chip Has a Name — and It's Called Jalapeño
On June 24th, OpenAI and Broadcom officially unveiled Jalapeño, OpenAI's first custom-built AI chip. The name is memorable, but what's more interesting is the context around it.
OpenAI president Greg Brockman has said bluntly that the company "cannot get compute fast enough." That's not a complaint — it's a structural reality. ChatGPT alone runs at a scale most people don't think about. Every conversation, every image generated, every code snippet reviewed costs compute. At the volume OpenAI operates, even small inefficiencies in how chips handle those tasks add up to enormous sums of money.
Jalapeño is built specifically for inference — the process of running a trained AI model in response to actual user requests. This is different from training, which is the expensive, weeks-long process of teaching a model in the first place. Inference is the everyday work: someone types a message, the model responds. Do that billions of times a day and suddenly the cost of each individual response matters enormously.
The chip went from initial design to manufacturing tape-out in nine months, which by the standards of semiconductor development is exceptionally fast. OpenAI also used its own AI models to speed up parts of the chip design process — which is either a fascinating recursive loop or a great piece of marketing, depending on how you look at it.
Initial deployment is planned for late 2026, with gigawatt-scale data centers being built alongside Microsoft and other partners. Brockman has framed this as a broader shift: OpenAI now wants to own more of its own stack — the chips, the infrastructure, the networking — not just the models running on top of it.
Google Has Been at This Longer Than Anyone
While OpenAI's chip is brand new, Google has been building its own AI silicon since 2015. They call them TPUs — Tensor Processing Units — and the eighth generation was announced in April 2026.
What's notable about TPU 8 is that for the first time, Google split the design into two distinct chips. The TPU 8t is built for training: massive, power-hungry model development runs that go on for weeks. The TPU 8i is built for inference: low-latency, cost-efficient serving of models at enormous scale. Before this, one chip was expected to do both jobs reasonably well. Google decided to stop making that compromise.
The numbers are difficult to put into plain language. A single TPU 8t superpod packs 9,600 chips and can scale to over a million chips across multiple data centers in a single training cluster. Google projects shipping 4.3 million TPU chips in 2026 alone, growing to over 35 million by 2028.
For ordinary people, the most tangible outcome is that Google Cloud customers get cheaper, faster AI infrastructure — which eventually filters down to better products and lower prices on the services built on top of it.
SpaceX Is Building a Chip Factory
SpaceX's angle is different, and considerably stranger.
The company is developing AI1 — an orbital data center satellite with a 70-meter deployed wingspan, designed to run AI workloads from space. The compute payload peaks at 150 kilowatts, roughly equivalent to a single Nvidia GB300 rack on the ground. SpaceX has filed to launch up to a million of these satellites, and has already signed compute deals — including a $920 million per month agreement with Google.
To supply chips for this program, SpaceX is building Terafab, a chip fabrication facility running as a joint venture with Tesla. The catch is that they don't have enough chips yet — SpaceX's own IPO filing warned that current chip supplies are insufficient. Terafab is the long-term answer, but it takes years to stand up a working fab.
In the meantime, SpaceX's orbital satellites use an interchangeable chip design, meaning they're not locked into any single supplier. Whoever makes the most competitive AI silicon can plug in. It's an unusual hedge, but it reflects the same underlying anxiety driving OpenAI and Google: if you depend on one supplier for a critical resource, you're at their mercy.
So Why Is This All Happening at Once?
The short answer is money. Running AI at scale is extraordinarily expensive, and the cost lives in the chips. Nvidia's GPUs are the best general-purpose option available, but they're designed to be good at everything — which means they're not perfectly optimized for any one thing. A chip built specifically for running large language models during inference can, in theory, do that job more efficiently and cheaply than a GPU designed to also handle gaming, scientific simulation, and video rendering.
Apple proved this model works. When Apple ditched Intel processors and built its own M-series chips, the performance and efficiency gains were significant and immediate. Every major AI company has watched that story unfold and drawn the same conclusion.
There's also the supply chain problem. During the AI boom of the last few years, Nvidia couldn't manufacture chips fast enough. Companies were waiting months for delivery. When your entire product depends on a component you can't reliably get, building your own starts to look less like ambition and more like common sense.
Broadcom CEO Hock Tan said recently that compute demand from his company's six biggest customers is "simply insatiable" — and that the same elevated demand is expected through 2028. That kind of prolonged scarcity is exactly the environment where custom silicon stops being a luxury and becomes a strategic necessity.
What This Means for Nvidia
Nvidia isn't going anywhere. Jensen Huang, Nvidia's CEO, has pushed back on the custom chip trend, arguing that general-purpose GPUs will always have an advantage because AI workloads keep evolving. A chip optimized for today's models may not be the right chip for tomorrow's.
That's a fair point. But it's also the kind of argument a dominant supplier makes when they see the market shifting around them. Nvidia's data center revenue hit $170 billion in fiscal 2026, but growth projections are already decelerating as more hyperscaler spending moves toward custom silicon.
The more likely outcome isn't that Nvidia gets replaced — it's that the AI chip market stops being a monopoly and becomes an ecosystem. Companies like OpenAI and Google will run custom chips for the workloads they understand deeply, and reach for Nvidia when they need flexibility or raw power for something new.
For the rest of us, the competition is probably good news. More chip options means more pressure on pricing and performance across the board. The cost of running AI doesn't just matter to the companies building it — it shapes what products exist, what they cost to use, and how accessible they end up being.
The Nvidia era isn't over. But it's getting a lot more crowded.
