The Chips Under Your AI Features Just Got Political

You ship AI features. You do not think much about the silicon underneath them. That layer used to feel like plumbing, boring and reliable. It is not boring anymore. The companies that make your models run are now fighting over chips, poaching each other's people, and getting pulled into trade fights between governments. Let me catch you up on what changed and why it lands on your desk.

Everyone wants off Nvidia

The model makers are done relying on one supplier. Anthropic is now talking to Samsung about a custom chip of its own, even though it hasn't decided what the chip will do or how powerful it will be. It's still leaning on Google, Amazon, and Nvidia for now. OpenAI already teamed up with Broadcom on a custom inference chip called "Jalapeño." Amazon and Google both sell their own TPUs.

The reason is plain. Nvidia is the undisputed leader, and being stuck with one vendor at these prices is a bad spot to be in. So the biggest AI buyers are spending real money to build their own hardware and buy themselves some independence.

What this means for you: the compute powering your features is about to come from more places, on more timelines. Bake that into your vendor plans.

Money is chasing anyone who speeds up inference

Inference, the part that happens after a user hits send, is the biggest cost center in AI right now. Whoever makes it cheaper and faster gets funded fast. Etched hit a $5 billion valuation and booked $1 billion in orders for systems built to run frontier models cheaper and with less power.

Here's the part worth remembering. In 2023, Etched pitched investors with a 30-page memo arguing AI would need specialized chips, and every major investor passed. The company was running month to month, close to out of cash. Now it has $800 million raised and Andrej Karpathy and Geoffrey Hinton on the cap table.

The lesson for your roadmap: your inference bill is not fixed. New hardware aimed straight at that cost is coming, and it could reprice the thing you spend the most on.

Talent is the tell

Watch where the senior people go. Nvidia just hired Nick Parker away from Microsoft, a 26-year Microsoft veteran who ran worldwide commercial sales, with a package worth more than $40 million. He now runs global sales at Nvidia, reporting to Jensen Huang.

That's two close partners pulling talent across the table. When a company pays $40 million to lock down the person who sells chips to the world's largest buyers, it tells you the seller thinks this market stays hot and stays contested for years.

For you, this is a signal, not a to-do. The people closest to the money are betting demand holds. Plan capacity like they're right.

The whole thing runs through two companies and one embargo

Underneath all of it sit two firms. ASML makes the lithography machines, and TSMC makes the chips. ASML builds about 90% of all chip-lithography tools worldwide, and its newest machine costs $400 million each. If you make advanced chips, you go through ASML. There is no way around it.

That chokepoint is now a political weapon. The US pressured the Dutch to block ASML from selling top machines to China. A proposed MATCH Act would go further and cut off older machines too, which is why the Dutch trade minister flew to Washington to push back in Congress. China accounts for 19% of ASML's system sales, so this is not a small fight.

The deep cut

The supply of chips behind your features is now shaped by two things you don't control: a handful of vendors racing to build custom silicon, and governments deciding who can buy what. IBM's new stacked-transistor design could add 10 to 15 years to the chip roadmap and cut data center energy use by up to 70%, but none of that reaches you on a clean schedule. So stop treating compute as a fixed line item. Treat it like a supply chain with real risk: know who supplies your inference, know what happens if that supplier gets squeezed by price, capacity, or policy, and have a second option before you need it.

Three questions for your team

1. Who actually supplies the compute behind our top AI features, and what's our fallback if that supplier's price or availability shifts in the next year?
2. How much of our cost is inference, and would a cheaper chip from a new player like Etched change what we can afford to ship?
3. If a trade rule or capacity crunch cut our access to certain chips, which features break first, and how fast could we route around it?