🎧 Prefer to listen?

Anthropic is hiring chip designers — architecture, design, verification, the whole stack — to build custom silicon for Claude. Google has been doing this for a decade with TPUs. OpenAI, Meta, and Amazon all have their own silicon programs. When every lab that makes the models you use starts manufacturing the hardware they run on, one question lands in your lap that has nothing to do with semiconductors: does this change what you should pay, and when? I gave it 20 minutes instead of a shrug, and the answer changed two of my subscription decisions.

The facts first, because the details matter more than the headline. Anthropic’s job listings cover chip architecture, design, and verification, with the company stating its teams will co-design hardware and models side by side — the same vertical-integration play this column covered when Anthropic first confirmed the plan. What’s new in this round of reporting: the company says it will maintain a multi-chip strategy regardless, continuing to use Amazon, Google, Nvidia, and AMD silicon, and TechCrunch reported discussions about a Samsung manufacturing partnership. No timeline. No commitment to owning fabrication. Meanwhile the parallel plays are all running: Google’s TPUs, OpenAI’s custom accelerator work, Meta’s MTIA, Amazon’s Trainium — and software-side alternatives like ZML’s hardware-agnostic inference are attacking the same cost problem from below. The infrastructure era is consolidating, and whether the rules of that era get written by regulators or by labs, both are moving now — the same collision we covered in the EU AI Act labeling rules.

What I actually did

Here’s the logic chain that made this worth 20 minutes: custom chips exist to cut inference costs. Inference costs are a large share of what you pay for in a monthly AI subscription. Therefore the pricing of AI tools is about to become more volatile, not less — first as labs’ margins improve (pressure toward lower prices), then as competition forces those savings downstream (more pressure). Volatility cuts both ways, and volatility is exactly what annual contracts don’t handle well.

So I listed every AI subscription I pay for, marked when each renews, and sorted them into two buckets: core (I’d pay double before giving it up) and expendable (there’s a substitute I haven’t switched to yet because switching is annoying). The sorting took ten minutes and the results surprised me — two tools I treated as core are actually expendable, and one expendable tool is eating more per month than my core stack.

Do this yourself

Twenty minutes, no code:

  1. List your subscriptions with renewal dates. Every AI tool with a recurring charge. Note the monthly cost and the next renewal date for each.
  2. Sort core vs expendable. The test: if the price doubled tomorrow, would you pay it or shop around? Your honest answer sorts the list better than any feature comparison.
  3. Run the annual-vs-monthly break-even on core tools only. Annual pricing typically discounts 15–20%. If a tool survived the doubling test, annual is usually right — you’re locking in today’s price on something you’d pay more for anyway. For expendable tools, stay monthly; flexibility is the product.
  4. Check which subscriptions are inference-heavy. Tools that generate text or images on every use (chat assistants, image generators, agent platforms) benefit first from cheaper inference. Tools that mostly store or organize (note apps with light AI features) don’t. The first group’s pricing is where the chip wave will show up; the second group’s pricing follows the software market instead. This is the same distinction behind which AI subscriptions actually deserve your money.
  5. Set one review date 90 days out. Custom silicon won’t ship this quarter, but pricing moves ahead of hardware announcements. One calendar entry: re-check the list, see if anything dropped in price or added an annual discount.

If your spending problem is bigger than subscriptions — if AI coding tools are quietly eating your budget — that’s a different audit, covered in what solo builders actually pay for AI coding tools.

What failed

Two honest misses.

Annual locking backfired once. I locked in a year of a design tool that survived the doubling test — then its best competitor cut prices two months later. My “core” tool now costs more than the alternative for similar output. The lesson: annual pricing is a bet on stability, and in a market this competitive, stability is the thing you’re least likely to get. Lock in annual only where switching costs (templates, integrations, history) are real.

The inference-heavy sorting was harder than it sounded. A “note app with AI features” turned out to run summarization on every note I save — inference-heavy after all. The sorting isn’t about the product category; it’s about what the AI does per action you take. Check what triggers the AI, not what the company calls itself.

Who should skip this

If you pay for one AI subscription and you’re happy with it, this is noise — a single $20 decision doesn’t need a framework. Skip the whole exercise.

Skip it too if you’re on free tiers across the board. The chip wave will eventually reach free products, but nothing announced today changes your cost, which is zero. But if you pay for three or more AI tools, or if AI costs are a line item in your business budget, the hiring news is a leading indicator: the labs are positioning for a price war, and your renewal dates are the battlefield. The backdrop matters too — Google’s own AI reorganization and the custom-chip era generally all point the same direction: the cost of intelligence is the next competitive frontier.

The takeaway

Every major AI lab is now a hardware company, and none of them will ship custom silicon this quarter. But you don’t need the chips to arrive to make a better decision — you need to know which of your subscriptions are core, which are expendable, and when each renews. Twenty minutes of sorting turns pricing volatility from a threat into an opportunity: lock in the tools you’d pay more for, stay flexible on the rest, and let the labs fight over your marginal dollar.