Europe's Sovereign AI Factory — The Nvidia Reality

Eighty-seven percent, for nothing

In March 2025 Nvidia sold a graphics card for $8,565. Today the same card lists at $16,000. No new silicon, no new memory, no new architecture. Eighty-seven percent in sixteen months.

That card is one of the ways a company stops sending its data to somebody else’s server, so this week I priced every option I could actually buy. The list is shorter than you’d think. These are published figures rather than my own bench, and I’ll say plainly what I could test and what I could not.

There’s a good version of this question and a bad one. The bad one is what can I run for $10,000? — a hobbyist question, answered better elsewhere by people with hardware I don’t own. The good one is the question a company asks: what does it cost to stop sending our data to an American or Chinese API, and are we allowed to do it?

Most of the answer is in that second half.

Three places to buy, one that will sell to you

China is the shortest answer. The United States has issued a worldwide prohibition on Huawei’s Ascend accelerators — the 910B, 910C and 910D. Not restricted, not difficult to source: using them is stated to violate export controls wherever you are. For a European or American buyer, that column is closed.

Europe surprised me, and I got it wrong when I started. I assumed Europe’s answer was a camera chip. It isn’t. Axelera, based in Eindhoven, makes a processor called Europa: 64GB of memory on a single chip, up to 256GB on a four-chip card, designed for language models.

So why isn’t it the answer? Two hundred gigabytes per second of memory bandwidth, against roughly 1,200 on a Mac Studio M5 Ultra. The capacity is there and the speed sits below the cheapest American option. Then the thing that actually decides it: Europa runs Axelera’s own toolchain, not the American stacks that every open-weight model ships against on day one. When Alibaba released a model this morning, the people running it were running it on American hardware, because that’s what the release was built to load into.

Europe has a serious AI processor and it is a generation of software behind. China’s is prohibited. Every option I could buy and run today, at every price point, is American.

And here is the awkward part. The hardware you can buy is American. The best open models are Chinese. Chinese silicon is prohibited and Chinese weights are state of the art, and those two facts sit inside the same procurement decision. Worth holding on to while the phrase sovereign AI is being sold this hard.

Three price points

Five thousand dollars buys one machine for one person — the solo attorney, the therapist, the accountant, anyone for whom don’t paste it into ChatGPT is a rule. Two options, both American: a Mac Studio with an M5 Max and 128GB of unified memory, or AMD’s Ryzen AI Max+ 395, the chip they call Strix Halo, in a mini PC. At the consumer show, AMD’s Lisa Su called it the smallest AI development system in the world.

Memory holds the model; bandwidth decides speed. Eleven tokens per second on the AMD box is usable, and slow enough that you’ll watch it think. Gemma 4, Mistral Small 4, and as of this morning Alibaba’s Qwen 3.8 Flash Next — 125 billion parameters, 262,000 tokens of context, and at 4-bit it fits inside 128GB. The hosted version costs 16 cents per million tokens. The cloud option got cheaper this morning; the sovereign option got 87% more expensive this year.

Fifteen thousand buys a machine for a team of fifteen to fifty. A regional law practice, a clinic group, a European company that read Article 50 and asked where its data goes. A Mac Studio with an M5 Ultra and 256GB lands at $14,299. The Nvidia card from the opening, the RTX Pro 6000 at 96GB, costs $16,000 — it blows the budget, gives you a third less memory, and runs the same memory bandwidth as a 5090. This tier runs DeepSeek V4 Flash at a million tokens of context, Qwen at server sizes, GLM 5.2.

Forty thousand buys a cluster: four Mac Studios linked over Thunderbolt 5, pulling one and a half to two terabytes. Published figures from people who have built this — four machines running Kimi K2 Thinking under 500 watts for the whole rack, another build reporting 28 to 30 tokens per second. I have not built one; I’m citing their numbers. The ceiling above it is Kimi K3, whose published weights run to 2.8 trillion parameters, need roughly 1.6TB of memory, and come with a commercial licence review attached.

The licence decides it, not the hardware

Llama 4 has the widest tooling ecosystem, and its community licence carries multimodal rights that exclude developers based in the European Union. Underneath that sits a second trap: licences attach to checkpoints, not to model families.

I spent 25 years in cybersecurity watching organisations buy the appliance and skip the contract, then discover during an incident that the contract was the thing deciding their exposure. Same shape here. Open weight is not open source — these models publish weights, not training data.

And the sovereignty story has a wrinkle at the infrastructure layer too. IT4LIA, the Italian AI factory with data centres in Bologna and Naples, is part of the EU sovereign AI factory initiative. It is an Nvidia machine with a European wing. Axelera is inside Europe’s sovereign AI factory as an option; the primary is Nvidia. Running a European model in the EU on EU hardware is harder than it sounds.

Where I’ve landed

I work on a Mac Studio M2 Max with 96GB, running a minimal local model in production for my own workflow — mostly testing. I’m about to order the M5 Ultra with 256GB, which runs Qwen 3.8 Flash Next comfortably, and move part of my workflow off the API.

I landed on 256GB because I don’t believe 512GB buys me much more. I could run DeepSeek V4 Pro locally, and Apple hasn’t announced the price for the 512GB version — I’d guess around $5,000 more, which to me isn’t worth it.

Tell me if I’m wrong. What does 512GB let you do that 256GB doesn’t?