Anthropic's not asking for export controls on Alibaba. It's going after the NYSE listing — a weapon we haven't seen a lab pick up before. This is the AI Daily Briefing. It's Friday. Today: Anthropic's securities play against Alibaba, and whether R$2.7 billion in Brazil tells us anything real about who's actually building. Plus, a step back on what those big data-center headline numbers even measure. Cassidy — start with the cloning letter. If today's show was useful, follow us wherever you're listening — the next one will be waiting. From Ars Technica:
Anthropic accused Alibaba of “brazenly” racing to make a copycat Claude, seemingly unfazed by Trump’s threats to crack down on foreign efforts to copy US frontier models despite depending on US investors. “Alibaba is listed on the New York Stock Exchange, maintains business operations in the United States, and is accountable to US investors and regulators,” Anthropic’s letter noted, “yet this activity unfolded in the weeks after” Trump’s memo warned that cloning attempts were “unacceptable.”
Anthropic just sent a letter accusing Alibaba of the largest Claude cloning attack to date — and the new part is the target. Instead of leaning on export controls or a safety panic, Anthropic points to Alibaba's New York Stock Exchange listing, and to its accountability to US investors and regulators. Right. That's a securities lever, more than a policy plea. I want to know whether being cross-listed is an actual handle someone can grab, or just a clever sentence in a letter. And the timeline is the knife. Anthropic says this unfolded in the weeks after Trump's memo called cloning attempts 'unacceptable.' So a publicly traded, US-regulated company read the memo and kept going. Which tells you the memo had roughly zero deterrent effect on a player sophisticated enough not to care. If you're already suing the administration that same week — and Alibaba is, over that military blacklisting — you're not exactly trembling at a memo. Mechanically, a cloning attack means somebody ran Claude at scale to distill it. So Anthropic's cost per token becomes part of the attack surface. The cheaper you make inference, the cheaper you make yourself to copy. That's the uncomfortable math. And Anthropic's own framing — that this helps China hit Mythos Preview-level capabilities sooner — basically admits distillation works. The output is the leak. Valor Econômico, with Daniela Braun:
Data center infrastructure company Ada Infrastructure, headquartered in the United States, has begun construction of three data centers in Brazil and plans to invest R$2.7 billion in the project in Franco da Rocha, São Paulo, by 2030.
Ada Infrastructure, R$2.7 billion, three facilities in Franco da Rocha outside São Paulo — and now we've got the fourth geography in three days: West Texas, Japan, Stockholm, then greater São Paulo. So I'll stop calling the dispersal a thesis. It's a documented pattern now. What I want answered is why Brazil is on the map: cheap power, lower latency into Latin American markets, or a spot outside US and EU regulatory reach? Hold on, though — R$2.7 billion is a capital commitment through 2030. They didn't disclose the investment or the capacity for that first Franco da Rocha building. We've got one named facility in greater Rio at 50 megawatts and a 2027 date, and that's it. So before anyone calls this contracted demand, I want signed customers or contracted power. Construction started three months ago. Steel in the ground isn't occupancy. And the tell is the language — they call Franco da Rocha the leading hub in Latin America because hyperscalers like it for connectivity and power availability. That sounds aimed at future tenants; it doesn't name current ones. When a company announces a '$30 billion data center buildout' or a '20-megawatt compute deal,' what are they actually measuring — and how do we tell real infrastructure from a press release dressed up in big numbers? Great question. The honest answer is that headline figures can be measuring a few different things, and people mush them together. A dollar figure is usually a capital commitment — money pledged, not yet spent. A megawatt number is power capacity, and that's the real constraint. Deloitte's 2025 AI Infrastructure Survey puts the stakes in sharp relief: U.S. power demand from AI data centers could grow more than thirtyfold by 2035, from about 4 gigawatts today to 123 gigawatts. JLL Research estimates it would take roughly $3 trillion in investment to bring 100 gigawatts of new supply online by 2030. So a single $30 billion announcement matters, but it's still a small slice of what's needed. The catch is the gap between announcing and building. Per reporting tracked by Sightline Climate's data center pipeline tracker — cited by writer Aren Deu in Bricks & Bytes — big tech committed around $400 billion to AI infrastructure in 2025, with another $650 billion signaled for 2026. But only about 5 of the 16 gigawatts in the 2026 pipeline are actually under construction right now. And Janus Henderson analyst Ian McDonald's research adds another filter: even among projects supposed to come online by 2030, his 'capacity ledger' — which accounts for permitting timelines, scarce resources, and grid bottlenecks — suggests scarcely more than half of announced energy-generation capacity will be completed on schedule. So if you want the cleanest read on durable infrastructure versus bubble-era overbuild, look at megawatts currently under construction, with permits in hand. So if under-delivery is actually the bigger risk than overbuild, does that flip the standard 'bubble' framing on its head? Largely, yes — at least on the physical infrastructure side. Janus Henderson's McDonald makes that argument directly: the key risk is announced capacity arriving too slowly to meet demand, because permitting and grid interconnection move far slower than capital commitments. KKR's November 2025 analysis echoes this, noting that while there is genuine froth in parts of the AI ecosystem, long-term data center demand looks likely to justify current investment levels — and historical overbuilds in railroads and fiber actually seeded durable economic change. So watch the construction-start rate and grid approval timelines, instead of the size of the next splashy announcement. If AI Daily Briefing has you watching the infrastructure behind the models, try The Data Center Daily. It's a daily briefing on AI compute, hyperscaler capex, power grids, semiconductor supply, and the energy markets being reshaped by intelligence at scale. Find it wherever you listen to podcasts.
One thing we’ll keep an eye on: Ada expects its first São Paulo data center to begin operations between December 2027 and June 2028.
You’ll find links to every story we covered today in the show notes, so if one caught your ear, you can dig in there. That’s AI Daily Briefing for today. This is a Lantern Podcast.