Five gigawatts in France, a gigawatt in Korea, nearly two billion pounds in the UK — and somewhere in there, somebody forgot to tell us what any of this is actually going to run. It's Monday, this is the AI Daily Briefing, and today's rundown reads like a power grid wishlist with funding rounds stapled on. And Bill's in full infrastructure-fatigue mode, so good luck to every press release that says the word 'sovereign' today. Oh, we're getting to 'sovereign.' Let's start with SoftBank and France, because five gigawatts is a number you say, not a thing you've built. Right. Pair that with Naver and NVIDIA's gigawatt factory in Korea — the announcement leans hard on 'strengthening sovereign AI.' It reads like a procurement story dressed up as national security. That's the question hanging over the whole batch. Is sovereign AI a real technical requirement, or is it the line every data center vendor has learned customers will sign for? Because Nebius's £1.7 billion UK buildout is selling the exact same comfort. Data residency is real for regulated workloads. But 'sovereign' has quietly expanded to mean 'please give us the GPU contract' — and nobody's correcting them. And here's where it actually gets interesting — the one paper in today's rundown I got excited about. DeepProve, out of IACR: verifiable end-to-end LLM inference. Cryptographically proving the inference ran the way the provider claims. Exactly. That's the untrusted-cloud-provider problem Nebius and every GPU cloud is implicitly selling against. If you can prove the model did what the provider said it did, you don't have to trust the data center — you verify it. It's academic today, but the threat model is real. Funny enough, that's the more honest version of 'sovereign' — trust the math, not the flag on the building. Then, on the other end of the credibility scale — Generalist AI. Four hundred million dollars at a $2 billion valuation to build, quote, 'general intelligence for robotics.' Pre-revenue, vague noun, very large number. You've got that 2021 SaaS eyebrow up, don't you. Both eyebrows. 'General intelligence' in a manipulation context — I just want to see the error rate at step seven of picking up a coffee cup. Physical workflows don't paper over failures; they drop the cup on the floor. And there's no technical report attached, so until there is, it's a $400 million demo reel. Two billion dollars says trust me. DeepProve says prove it. Guess which one I'm rooting for. From SoftBank Group Corp.:
Paris, France — May 30, 2026 — SoftBank Group Corp. (TSE: 9984, “SoftBank Group”) today announced its commitment to develop and operate 5 GW of AI data center capacity in France, representing an investment of up to €75 billion.
Five gigawatts, €75 billion, announced at Macron's Choose France summit. Before we even get to the compute, the procurement pitch is wrapped in the flag — 'European technological sovereignty' right there in the subhead. And the first phase is the only number with actual edges on it — €45 billion for 3.1 gigawatts in Hauts-de-France. The other 1.9 gigawatts is a press-release rounding up to a headline figure. What I actually want is the Dunkirk piece. Anchoring data center manufacturing with Schneider Electric — that's the part that survives contact with reality, because someone has to build the gear and pour the power. This is Son's largest AI infrastructure bet in Europe, and 'sovereign' is carrying the politics. Whether France controls the inference stack or just hosts SoftBank's racks — that's a very different question than the gigawatts suggest. Right. Owning the building doesn't mean you own the model. You can own the substation and still rent the intelligence by the token. From The AI Journal:
LONDON–(BUSINESS WIRE)–Nebius (Nasdaq: NBIS), the AI cloud company, today announced it is investing approximately £1.7 billion to build out capacity in the UK with three new deployments of NVIDIA infrastructure, as the company continues to expand its commercial and AI R&D hub in London.
Okay, £1.7 billion, four sites, and when you read past the headline, it's 65 megawatts fully ramped in 2027. After the SoftBank 5-gigawatt number we just hit, 65 megawatts is almost refreshingly... legible. And look at who's standing next to them — the UK AI Minister, quoting the AI Opportunities Action Plan. The sale is being framed as national strategy. Right, and the customer list is the part I actually buy — financial services, healthcare, life sciences. Those are real latency and data-residency buyers, not a demo reel. Blackwell Ultra in November, three more NVIDIA deployments now. Nebius is basically reselling Jensen's roadmap with a London postcode attached. Kim Jinyeong, writing in The Asia Business Daily:
NAVER will use its data center 'GAK Sejong' as a forward base. Starting with 55MW operation in the first half of 2027, the company plans to expand its infrastructure to 100MW within the same year and to 200MW overseas by 2028. Through these efforts, NAVER aims to establish GW-scale infrastructure.
So that's SoftBank's 5 gigawatts in France, Nebius in the UK, and now Naver and Nvidia talking gigawatt-scale in Korea — all in one rundown. At some point, I want a tally of how much of this is actually running inference versus announced in a press briefing. And listen for the key word here: sovereign. Raj Mirpuri from Nvidia says they're expanding sovereign infrastructure to the GW level through the DSX platform — that's the framing every one of these deals is reaching for now. Right, but is 'sovereign AI' a technical requirement here, or just the procurement wrapper? Because the common denominator in all three of these is Nvidia silicon. You can host it in Seoul, Paris, or Manchester — the dependency doesn't move. It's sovereign right up until the GPU allocation meeting. And the detail that matters: they're starting at 55 megawatts, not the headline gigawatt. The gigawatt number is four times Naver's existing Gak Sejong data center, so that's the ambition, not Tuesday. 55 megawatts I believe. The gigawatt I'll believe when it's drawing power. And Naver's sharing the risk as a global partner — which is the polite way of saying they're on the hook if demand discovery comes up short. Here's Nicolas Gailly at IACR Cryptology ePrint Archive:
In this work, we present DeepProve, the first system to enable efficient end-to-end verification of full LLM inference (i.e., for all generated tokens of a prompt) on untrusted cloud servers using zero-knowledge proofs (ZKPs).
Okay, after a whole rundown of people pouring gigawatts into untrusted GPU clouds — Nebius, SoftBank, all of it — here's the paper that actually asks the uncomfortable question. DeepProve, IACR 2026/1112: how do you prove the cloud server actually ran the model it said it ran? Right, and that's the gap nobody on the data-center side wants to dwell on. You rent the inference, you get tokens back, and you're just trusting that's the model you paid for. What grabs me is the scope. Prior work — zkGPT at USENIX, zkLLM at CCS — could verify one token or one component like Softmax. DeepProve claims all generated tokens, end-to-end, with zero-knowledge proofs. That's the difference between a demo and something a builder can actually reason about. And notice they didn't brute-force it by stuffing the whole inference in-circuit — that goes quadratic in sequence length. They certify the output sequence instead. Smarter threat model, not just bigger hardware. Here's Kyt Dotson at SiliconANGLE:
Artificial intelligence startup Generalist AI Inc., a startup building embodied robotics intelligence, said today it has raised $400 million in new funding, bringing the company’s valuation to $2 billion. Radical Ventures led the round.
$400 million at a $2 billion valuation for, and I'm reading directly here, 'general intelligence for robotics.' Pre-revenue, as far as I can tell. And look, the pedigree's real — Pete Florence built RT-2 and PaLM-E at DeepMind. That's serious. But 'general intelligence' is the squishy part of that sentence. What stands out to me is the cap table. Nvidia and Bezos Expeditions were already in, and they doubled down — Nvidia keeps planting flags in anything that sells more GPUs into physical AI. Right, Nvidia's the one calling robotics the next trillion-dollar industry — and they're an investor. Convenient framing when you sell the picks and shovels. So what's the actual product timeline? Radical leading, 8VC and Union Square in — that's a serious round, but I haven't seen a benchmark, a demo with a technical report, anything I'd treat as more than a deck. And that's exactly my question. In a language model you can eat a hallucination at step seven. In physical manipulation, step seven of a ten-step chain is the robot dropping the thing or putting a hand through a wall. Show me that error rate, then we'll talk about 'general.' If AI Daily Briefing helps you keep up, subscribe wherever you're listening. And if you have a moment, a quick review really helps other people find the show. Thanks for being part of it.
You'll find links to every story we covered today in the show notes, so if something caught your ear, you can dig into the original reporting there.
That's AI Daily Briefing for today. Thanks for listening. This is a Lantern Podcast.