Finland just became a very expensive answer to the question of who gets to control the next layer of AI. This is AI Daily Briefing. Google, a debt-funded Nvidia cloud in Southeast Asia, and China’s homegrown GPU push all raise the same uncomfortable question: who owns the stack when the racks go live? And the security fight just joined the build-out fight. Let’s start in Finland. From Kirsi Heikel at Financial Post:
Alphabet Inc.’s Google is planning its biggest investment in Europe, an artificial intelligence infrastructure build out worth at least 13 billion euros (US$15.1 billion) in Finland. The investments over the next two years will include the construction of at least three new data centres as well as the expansion of its existing data centre in the southeastern city of Hamina, the company said in a statement on Wednesday.
Google’s €13 billion Finland build has the detail most AI-infrastructure announcements conveniently skip: Fortum is selling it half the output from the Loviisa nuclear plant. There’s an actual power plan behind it, not just a glossy rendering. The build includes three new data centers and an expansion in Hamina, all over two years—and the electricity is already spoken for. That makes the capacity number believable: somebody did the boring work before the GPUs arrive. And Google is putting its biggest European investment into a country where 96% of power is carbon-free and server heat can warm homes. Turns out inference likes cheap, reliable electrons more than launch-event adjectives. From Reuters:
Sept 9 (Reuters) - AI platform Zankore said on Wednesday it has signed an up to $3.1 billion senior term loan facility to fund the rollout of Nvidia-powered GPU and cloud-computing infrastructure, as it expands AI capacity in Indonesia and across Southeast Asia.
We covered Zankore’s Nokia-and-Nvidia launch last edition. Now Reuters has the financing: up to $3.1 billion. It’s senior debt, which is a lot more serious than a flashy capacity announcement. Right. A signed term-loan facility means lenders signed off on repayment from an Nvidia GPU cloud in Indonesia and Southeast Asia. They’re betting those racks will have paying workloads, not just very expensive launch photos. Compare it with the Finland build: Google can pair infrastructure money with power commitments because it owns the whole machine. Zankore’s borrowing to assemble an Nvidia-powered regional layer inside somebody else’s stack. $3.1 billion buys a lot of GPUs; it does not buy low latency, clean data pipelines, or customers whose workloads survive the invoice. But debt has a useful feature: eventually, it demands the utilization number. Here's Ashley Belanger at Ars Technica:
The United States has now named six Chinese AI firms accused of waging industrial-scale attacks distilling US frontier AI model capabilities and perhaps sparing billions in Chinese development costs. In a joint release Tuesday, the National Security Agency (NSA), Cybersecurity and Infrastructure Security Agency (CISA), and Federal Bureau of Investigation (FBI) alleged that DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have been attacking US models since at least late 2024.
The NSA, CISA, and FBI named DeepSeek, Alibaba, and four other Chinese firms in one advisory. That’s an unusually specific federal accusation, and it makes closed-model access a security perimeter. Since late 2024, they’ve allegedly pulled capabilities from Claude, GPT, Gemini, and Grok. If you run a frontier-model API, query anomaly detection just became a board-level engineering problem. Finland and Zankore make this harder to dismiss as a side issue. The U.S. agencies are saying access to the models themselves can shave billions off a rival’s training bill. The release says “likely” and alleges Chinese government awareness, so keep that allegation label attached. Still, six firms and four major U.S. model families over nearly two years is a long way from somebody benchmarking an API for a weekend. Hacker News, weighing in:
Every US frontier lab trained on the entire internet without asking or attributing Reddit, books, GitHub, Wikipedia and they called it fair use. OpenAI, Anthropic, Google all did this. They unethically ingested the output of every human who ever typed into a browser, paid nobody (oh did I forget about art and art styles being stolen as well without attribution?) and built trillion dollar companies. But Now it's "industrial-scale theft" requiring a joint NSA/FBI/CISA advisory. The hypocrisy.
The hypocrisy charge is fair on training-data fights. Frontier labs have made expansive claims around web data while demanding tight control over their own outputs. But industrial-scale capability extraction can still do real harm. Labs may owe creators answers, and API operators still need to defend the product they’re selling. Digitimes’ Jingyue Hsiao is tracking it. JD Cloud says it’ll build a 100,000-card Moore Threads cluster. Great—now show us the utilization curve. Show us the interconnect failures, and what a real training job looks like on card 87,000. Digitimes says JD is calling this the first 100,000-card core cluster at a leading provider using domestic Chinese GPUs. China is trying to own the silicon layer instead of relying on Nvidia workarounds. Set that beside the NSA, CISA, and FBI accusations, and the strategy looks pretty plain: protect access to frontier capability however you can while building hardware you control. A planned cluster still has to prove it performs. At 100,000 cards, though, this is well past the demo stage. Outside China, Zankore is borrowing $3.1 billion to buy into an Nvidia-powered cloud. Inside China, JD Cloud is placing a domestic-hardware bet at the same scale. The inference map is starting to split along the supply chain. If you’re enjoying AI Daily Briefing, take a moment to subscribe or leave a review wherever you’re listening. It helps other people find the show—and helps us make it better.
Links to every story are in the show notes if you want to dig deeper. That’s AI Daily Briefing for today. Thanks for listening—we’ll be back tomorrow. This is a Lantern Podcast.