devdot
← All postsEngineering ·

Claude Ported a Model to AMD Chips in Three Days. Your "Too Expensive to Migrate" List Is Out of Date.

Anthropic says Claude Mythos Preview moved one of its models onto AMD MI355X GPUs in about three days, work that used to take dozens of engineers months. The rewrite is getting cheap. Proving it works is now the real cost of any migration.

Anthropic co-founder Tom Brown told IPO investors this week that Claude Mythos Preview got one of the company's models running on AMD MI355X GPUs in about three days. By his account, the model wrote or ported nearly all of the kernels with "basically no human intervention." The same job, Anthropic says, used to take dozens of engineers six to twelve months.

If that holds up, it matters well beyond GPUs. Most companies carry a mental list of migrations they've decided are too expensive to ever do. The old framework nobody wants to touch. The database you'd leave if it weren't for 400 stored procedures. The cloud provider you only stay with because its proprietary queue service is wired into everything. Those decisions were priced in engineer-months. That price is moving.

Read the footnote before the headline

It's worth being precise about what was actually claimed. According to the roadshow material, the three days covered prompt processing (prefill) for one model at "useful" performance. Not full inference. Not decode speed. There's no published comparison against Nvidia, nothing on numerical drift, and no test methodology. The S-1 filing itself only says engineering teams are using Claude to bring new chip platforms up faster, without a figure.

That's not a reason to dismiss it. It's a reason to notice where the hard part went.

Writing the kernels got cheap. Proving they're correct, fast enough and safe to run in production didn't. Anthropic could pull this off because it already had a reference implementation on other hardware to compare outputs against, and people who knew exactly what "useful" meant for that model.

Re-price your lock-in list

Here's what we'd do with this, even if you never touch a GPU.

Pull out the migrations you shelved on cost. Framework upgrades, ORM swaps, moving off a vendor SDK, porting a service from one language to another. Anything where the blocker was "it's six months of grunt work" rather than "it's risky" deserves a second look. Grunt work is exactly what agents are now good at.

Price the verification, not the rewrite. For each one, ask what it would take to know the new version behaves like the old one. If the honest answer is "we'd click around and hope," that's your real cost, and it's the part to fund first:

  • Parity tests that run both versions on the same inputs and diff the outputs
  • Recorded production traffic you can replay against the new system
  • Performance baselines with real numbers attached

Define "done" in writing. Anthropic's own claim has a scope: one model, prefill, useful performance. That's a good habit to steal. Before an agent starts a migration, write down which paths must match exactly, what latency is acceptable, and which edge cases are out of scope. Otherwise you'll get a confident "ported!" and spend the next month finding out what it meant.

Keep the old system runnable. The cheapest parity check you own is the original code. Don't delete it until the new version has carried real traffic for a while.

Switching costs are turning into testing costs

The interesting part of this story isn't that Nvidia's software moat got a bit shallower, though it probably did. It's that switching costs across the whole stack are starting to look less like labour and more like testing.

Teams with good tests, reference outputs and clear acceptance criteria can now move between vendors, chips and frameworks in weeks. Teams without them are still stuck, and an agent won't fix that for them. It'll just produce a lot of code nobody can sign off on.

So the next time someone says a migration is too big, ask a different question: how would we know it worked? If you can answer that, the rewrite might be the easy part now.

We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!

NEXT POST →Devin Bots Can Auto-Merge Low-Risk Changes. Decide What Counts as Low-Risk in Code, Not in a Chat.