Here's the thing about scarcity: it reveals who actually matters. This week, the AI chip world cracked clean in two, and the fracture line runs through every startup that isn't Microsoft, Google, or OpenAI.
The winners
Nvidia closed at a record high, pushing past $5 trillion market cap for the first time since October. That's not news — that's inevitability. But the 4.3% jump on a day when Intel popped 24% (its best since 1987) tells a different story. The rally wasn't just Nvidia. It was the entire chip ecosystem remembering it exists.
AI startups are struggling to access Nvidia GPUs as Microsoft and other cloud providers divert supply to internal teams and large customers like OpenAI.
That's from The Information, and it's the kind of thing that sounds like a complaint but is actually a structural fact. The cloud providers aren't hiding this — they're not even pretending anymore. If you're an AI startup and you're not already in the queue for Nvidia's next allocation, you're already behind.
The deals getting done
Meta just signed a multibillion-dollar deal to rent hundreds of thousands of Amazon's Graviton chips. Not Nvidia. Amazon's homegrown silicon. That's a signal — when the biggest AI player in the world (by training compute) chooses Amazon's second-best option over fighting for Nvidia scraps, the market is telling you something.
And this is where it gets interesting. DeepSeek just dropped V4 Pro and V4 Flash, pricing them at roughly 60% cheaper than comparable models. V4 Pro costs $1.74 per million input tokens. That's half what OpenAI charges for GPT-4.5 — and the gap is shrinking.
What this means
The chip shortage isn't a temporary supply problem. It's a market correction. Every startup that built their business on "just use API tokens" is realizing those tokens have a hard cost floor that cloud providers control. The winners are the ones who either (a) have their own silicon, or (b) have the leverage to get allocated chips.
Everyone else is getting priced out — not because they can't afford the models, but because they can't guarantee the inference capacity to serve them.
This is the Great Chip Divorce: the divide between companies withchip access and companies without. It won't close. It'll widen.