The inference stack just picked up a passport — and by the time you get to the fine print, it's already in three countries. This is AI Daily Briefing. Today: a chip company brokering capacity between labs in Europe, a Bitcoin miner reborn as a 20-year landlord, and a study asking whether any of this compute does actual work. Cerebras and the map, first. Tap follow so the next episode finds you. This one's from National Post:
The new 1.7 MW facility marks DeepInfra’s ninth data center location and its first outside the United States. It will host over 1,000 NVIDIA Blackwell B300 GPUs, significantly expanding the company’s global inference capacity. The deployment strengthens DeepInfra’s infrastructure strategy, enhancing its ability to deliver low-latency, cost-efficient AI inference services across North American and international markets.
So DeepInfra opens its ninth site, first one outside the U.S. — Toronto, 1.7 megawatts, over a thousand Blackwell B300s. A thousand B300s is a serious capital commitment. What the release doesn't give me is a single utilization number. And they're leaning on that McKinsey line — inference hitting 40% of data-center demand by 2030, growing 35% a year. That's the number every inference shop is quoting this quarter to justify the buildout. Right, the projection's the pitch deck. Selling tokens only works as a real model if those boxes stay warm — and 'we opened a facility' tells me nothing about fill rate. That's the second time in a week I've watched capacity get announced with no occupancy attached. It's the same inference-in-a-new-geography move, though — the stack quietly shifting north. Toronto here, and later in the show you'll hear the European version with a much bigger name on the customer line. Yeah, and that's the tell. DeepInfra keeps adding halls and won't publish a warm-box rate. At some point, the silence on utilization becomes the data point. Here's Antoine Buteau at Antoine Buteau:
A new paper, “Mapping AI into Production: A Field Experiment on Firm Performance,” by Hyunjin Kim, Dahyeon Kim, and Rembrand Koning, offers an empirical answer. The bottleneck is not access to models, budget shortages, or poor prompting, but the organizational mapping problem.
Okay, finally — a number this week that isn't a megawatt or a lease dollar. Five hundred fifteen startups, controlled field experiment, and the finding is pretty blunt: giving people the tool does basically nothing by itself. The gains only show up when firms actually reorganize the work around it. Which is the thing I've been saying for months — most companies are buying a faster horse and calling it a new vehicle. And notice the timing. Kim, Kim, and Koning out of INSEAD, and it lands the same morning we're talking Toronto data centers and 20-year leases. You can sign for the hall and fill it with compute, and this paper says you'll still get nothing if nobody maps where the AI goes in the workflow. Right — and their point on why firms fail is sharp. Teams default to the obvious stuff: support chatbot, writing assistant. The high-value use cases sit there undiscovered because they'd require actually restructuring how work flows. It's a quiet rebuke to every announcement in the rundown. A developer can be fifty percent faster without the company becoming fifty percent faster, because the bottleneck's the org. Which is the least fundable sentence in AI. Nobody's writing a check for 'you need to rethink your org chart.' Markus Kasanmascheff, writing in Winbuzzer:
Digital infrastructure company and former Bitcoin miner TeraWulf has signed a 20-year lease with Anthropic for a purpose-built AI infrastructure campus at its Justified Data site in Hawesville, Kentucky. Anthropic is the public benefit corporation behind the Claude AI assistant and model family.
So the TeraWulf story lands again, but the Winbuzzer framing finally makes the whole thing legible: this is a Bitcoin miner that's basically become a 20-year AI landlord. $19 billion over the term, 401 megawatts in Hawesville, Kentucky. And the part I keep coming back to — Anthropic is spreading the hardware risk around. GPU-heavy campus in Kentucky, and later today you'll hear them buying into a very different chip stack in Europe. They're making two bets. Here's my problem, Cassidy — it's a 20-year fixed-rate lease on a facility that doesn't exist yet. The TL;DR literally says 'planned.' Capacity starts arriving late 2027. So Anthropic's $19 billion clock is running against a campus that hasn't racked a single server. Financing, construction, cooling design, power controls — all still open. That's a lot of promise stacked up before year one. And it was built to mine Bitcoin, not host enterprise IT. Those power assets were designed to burn watts, not run a fine-tuning cluster. The conversion is the actual product here. Right, and that's the bet I can't shake — if inference cost curves keep falling the way they've been falling, a fixed rate signed in 2026 looks awfully expensive in year fifteen. Here's Crypto Briefing:
Cerebras Systems is planting a massive flag in European soil. The AI chip company announced plans to deploy 200 megawatts of AI compute capacity across Europe by the end of 2027, with initial capacity expected to come online by late 2026.
Okay, this is the one I've been waiting for all week. Cerebras, 200 megawatts in Europe by the end of 2027, France plus Norway and Finland — and a chunk of it is running OpenAI workloads. Remember the DeepInfra Toronto piece we just hit, and Korea before that? The seat was empty — nobody named the tenant. Here, the tenant has a name, and the name is OpenAI. And this traces back to the January partnership — 750 megawatts, over $20 billion through 2028. So a chip company is basically subletting capacity to a lab outside the U.S. The inference-stack-control question is hitting the silicon layer, in Europe. The Nordic siting is a tell, too — Norway and Finland for the hydropower. Feldman gets to say 'responsible' and 'sovereign' in the same breath at a Paris summit. Convenient framing, but the electrons are real. What I like is that it's a hardware answer, not a slide. Cerebras wafers, not a rack of somebody else's GPUs. If OpenAI's willing to run production traffic on that architecture in Helsinki, that's a real bet on the economics, not a demo. If you follow AI’s daily moves, check out The Data Center Daily — a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and the energy markets being reshaped by intelligence at scale. Find it wherever you listen to podcasts.
What we’re watching next: Cerebras expects its first European AI compute capacity to come online by late 2026, with 200 megawatts planned by the end of 2027. TeraWulf says Anthropic capacity delivery should begin in phases in late 2027 and wrap in early 2028.
You’ll find links to every story we covered in the show notes, so you can dig into the pieces you want to read more closely. That’s AI Daily Briefing for today. This is a Lantern Podcast.