Claude moved to AMD chips in 3 days. The footnote is the story

Rows of server racks with small status lights inside a data centre

A chip switch used to cost Anthropic dozens of engineers for 6 to 12 months. Anthropic told its IPO investors that its Claude Mythos Preview model just did that porting work in about 3 days, the software work that gets a model running on AMD MI355X chips. That is a money story before it is a technology story. A lab that can change hardware in days negotiates with every chip supplier from a stronger seat, and its capacity stops waiting in 1 vendor’s queue.

The direct answer: The 3 days cover prompt processing for 1 Anthropic model on AMD MI355X GPUs, at what the company calls a useful level of performance. Mythos ported or wrote almost 100% of the kernels, the low-level code that talks to the chip. The amended prospectus, filed November 10, 2026, makes the same point in drier words: the company can onboard new chip platforms in a fraction of the time it once took. No performance figure against NVIDIA hardware was published, and the claim says nothing about the reply-generating step that dominates serving cost.

  • What the claim covers, line by line
  • Why a Canadian buyer should care about a chip port

What the claim covers, line by line

Read the scope before the headline. Prompt processing is the reading step, where a model takes in the request. The reply step, called decode, carries most of the cost when a model serves millions of answers, and the roadshow claim does not address it. The baseline is also the company’s own: 6 to 12 months from Anthropic’s historical experience, not an independent measurement. None of that makes the claim empty. A buyer-side company, heading into a listing, put the narrow version in writing in a filing. Narrow and filed beats broad and marketed.

Advertisement

The wider record supports the direction. Anthropic runs its models as 1 pool across NVIDIA and AMD GPUs, Google TPUs and AWS Trainium, and its AMD agreement covers up to 2 GW of capacity. Porting speed makes a pool like that real instead of theoretical. Jensen Huang resigned. Nvidia now has an avatar problem covered the trust question at the top of the largest chip supplier last week. This week the question moved down a level: how fast can a big buyer leave a chip behind when it wants a better deal.

Why a Canadian buyer should care about a chip port

Your firm does not buy MI355X chips. It buys tokens, seats and finished work, and the price of each one carries the lab’s compute cost inside it. A Canadian firm pays that bill in CAD, so a lab’s compute saving reaches it through 2 filters: the vendor’s price list and the exchange rate on the day the invoice lands. Gemini 4 Carbon is out. The token price is the small number made the point that cost per finished task decides a bill, not the headline token price. Labs with cheaper, more flexible compute have room to cut that task cost. Labs locked to 1 supplier’s queue do not.

Routing is where the reader’s own leverage sits. Qwen4-Max output costs $6.42 CAD. Route the easy steps, keep client data home set out the practice of sending easy steps to a cheaper model. The Mythos claim is the same practice at the hardware layer: the buyer with options pays less than the buyer with a single door. And when a billing model changes under you, as Fable leaves your plan tonight. Credits start at $57.03 CAD showed this week, the firms that already know their alternatives move first.

The recap is short. The 3-day figure is real, narrow and filed: prompt processing, 1 model, no performance comparison published.

Ask your main AI vendor 1 question at the next renewal, and get the answer in writing: which chips serve our workload today, and what happens to our price if you move it. A vendor that can move has options to share. A vendor that cannot move has a queue to feed, and your bill is part of the feed.

That buyer leverage is the demand side of the quarter just reported: Nvidia sold $115.8B in 1 quarter. Canada’s bottleneck is power and permits shows demand still rising, while Canadian builds wait on permits and power.

Total
0
Shares
Prev
An agent clicks through the clinic software now. The review step is yours to set
A computer workstation on a desk in an empty clinic office

An agent clicks through the clinic software now. The review step is yours to set

OpenAI’s launch post of November 16, 2026 puts a number first: reviewers

Next
Count accepted comments, not posted ones, before you renew your AI review tool
A close-up of programming code on a computer screen

Count accepted comments, not posted ones, before you renew your AI review tool

Open 1 pull request queue from last week and count 2 numbers by hand

You May Also Like