The Real AI War Isn't OpenAI vs Nvidia — It's Distributed vs Concentrated Power

By Ivana Tilca · September 1, 2026 · 7 min read

OpenAI revealed its own chip to stop depending on Nvidia — and Nvidia countered by reportedly buying Hugging Face, the home of open-source AI. Here's what's really being fought over: not two giants, but whether the power behind AI stays concentrated or spreads to all of us.

This week the AI industry quietly showed its hand, and the move underneath the headlines matters far more than the headlines themselves. OpenAI — the company that has bought millions upon millions of chips from Nvidia — revealed its own chip, built to stop depending on Nvidia. It's the corporate equivalent of saying, to your biggest supplier's face, "thanks, but I'll make it myself." And then Nvidia did something almost nobody saw coming. On the surface it looks like a rivalry between two giants. Underneath, it's a much bigger question: who ends up controlling artificial intelligence?

Let me walk you through it, because the chess here is genuinely fascinating.

OpenAI builds its own chip

OpenAI showed the first results of its custom chip — reportedly nicknamed "Jalapeño" and built with Broadcom — and the number is the kind that makes you sit up: up to 104x more performance on open-weight models. Not 10% better. Not "we improved things a bit." A different league entirely.

Here's the clever detail, the one that shows the move is smart and not just loud: they benchmarked it on purpose against public, open models. Why? Because that makes the claim verifiable. Anyone can take another chip and measure how it runs those same open models. Meanwhile, their own closed, frontier models run only inside OpenAI's cloud — so there, you have to take their word for it. And they imply that behind those closed doors the chip's advantage is even bigger. They show you the impressive number and hint that the real one is better still. Marketing-wise, beautifully played.

But there's a key point most people will skip over: this chip is for inference, not training. In plain terms, inference is what happens when you type something and the model answers — that's what runs here. Training a model from scratch is still done on Nvidia hardware, at least for now. Even so, the message is unmistakable: OpenAI has started making its own chips to depend less on Nvidia. Not long ago they partnered with Cerebras for ultra-fast inference; now they're essentially building their own version of that. Loosely translated: "thanks for the favor, we'll take it from here."

Nvidia's counter-move: bet everything on open source

Here's where it gets strange, in the best way. According to The Information — and to be clear, this is not officially confirmed by either Nvidia or Hugging Face as I write this — Nvidia reportedly agreed to acquire Hugging Face for around $12.9 billion.

If Hugging Face isn't on your radar, think of it as the GitHub of AI models. It's *the* place where companies publish open-weight models for anyone to use, and it also rents out GPUs to run them in the cloud. It is, quite literally, the beating heart of the open-model world.

So why do I find this move brilliant and a little unsettling at the same time? Follow the logic. OpenAI is making its own chips. Meta is making its own. Google has had its TPUs for years. Every giant is starting to build the whole house themselves and buy less and less from Nvidia. So Nvidia, instead of fighting that war head-on, does the opposite: it bets everything on open source. It buys the field where open models actually get run in the cloud, and turns itself into the owner of the stadium. If the world keeps shifting toward open models — and I'll show you in a second that it is — then everyone who migrates ends up paying for inference to... an Nvidia company. That's not a reaction. That's a ten-year chess bet, made two moves ahead.

The world really is going open

This isn't a hunch, there's hard data. On Vercel's AI gateway, the share of tokens going to open-weight models jumped from 28.4% to 62% in just two months. Two months — from less than a third to nearly two-thirds. That's not a ripple, it's a tidal wave.

And the models backing that shift are getting genuinely good. Strong open-weight releases keep landing — the GLM family, capable DeepSeek variants, the larger Qwen models — closing the gap with the closed frontier models that used to be untouchable. The distance that seemed unbridgeable a couple of years ago is now a hair's breadth.

The quiet revolution: AI on your own desk

While all of this plays out at the scale of billion-dollar deals, something just as important is happening at the scale of your own desk. Apple introduced new chips — the M6 and the M5 Ultra — built specifically for local AI compute: running models on your own machine, with no cloud involved.

The M5 Ultra is a little absurd, in the best sense: up to 4.5x the GPU AI power of the M3 Ultra, in configurations with as much as 512 GB of unified memory — most of which you can use as if it were VRAM to run models locally. The catch, of course, is price: a high-memory M5 Ultra lands near eleven thousand dollars, and that's not even the top configuration.

But the price isn't the real story. The real story is that more and more machines are becoming capable of running very good models at home. Models like the GLM family, some DeepSeek variants, the larger Qwen models — running on your desktop, without sending your data to anyone. As someone who works in this world and cares about its real impact on people, I think this is one of the biggest things happening right now, and almost nobody is paying attention to it.

Reading between the lines

Now the part I enjoy most: reading between the lines, because the official statement never tells you the whole movie.

Why would Nvidia make a move like this? Fear. Not "we go bankrupt tomorrow" fear, but the very real fear that its best customers — OpenAI, Meta, Google — are building their own houses and will buy less and less. So Nvidia diversifies and plants itself firmly on the side of the street that's actually growing: open source. When a company makes a strange move, don't listen to what it says — look at what it's hedging against. Here, it's hedging against being left out of the future.

And here's the detail that nags at me, the side not everyone tells. If this deal happens, Nvidia doesn't just sell you the chip — it also ends up owning the place where you run open models. Fewer and fewer hands controlling more and more of the chain. Personally, that concentration of power doesn't sit entirely right with me. It's wonderful that open AI is growing — truly, it makes me happy — but an "open" world that ends up with a single owner is a paradox worth watching out of the corner of your eye. Be careful about celebrating everything without reading the fine print.

The fight that actually matters

So let me set aside the corporate gossip and tell you what really matters about all of this.

The fight I actually care about isn't OpenAI versus Nvidia. It's distributed power versus concentrated power. Being able to run a good model in your own home — without sending your data to any cloud, without paying anyone for the privilege of thinking — is one of the most beautiful things this technology can offer us. And at the very same time, an "open" world that ends up with a single owner is exactly what we should be watching with narrowed eyes. Both things are true at once, and that's precisely why it's worth paying close attention to who ends up with what.

My prediction

Here's my prediction, and I'll sign my name to it: open source will keep taking ground from closed models, month after month. If you'd asked me three or four years ago, I'd have told you open models weren't remotely close to the closed ones. Look where we are now — neck and neck. For us, for ordinary people, that's the best news of all.

The real question underneath every chip announcement and every acquisition rumor is the same: will this power stay in a few hands, or spread into many? Keep your eye on that, and the rest of the noise gets a lot easier to read.

Sources

OpenAI's custom "Jalapeño" inference chip (CNBC): https://www.cnbc.com/2026/08/26/openai-jalapeno-ai-chip-nvidia.html

Nvidia's reported Hugging Face acquisition — original report by The Information (paywalled): https://www.theinformation.com/articles/nvidia-agrees-buy-open-source-model-repository-hugging-face-12-9-billion — free confirmation (TechCrunch): https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/

Open-weight token share 28.4% → 62% (Vercel AI Gateway): https://vercel.com/blog/ai-gateway-production-index

Apple M6 and M5 Ultra for local AI compute (Apple Newsroom): https://www.apple.com/newsroom/2026/08/apple-introduces-m6-and-m5-ultra-for-a-big-leap-in-performance-and-ai-compute/