Compute deals keep getting bigger—and AI has wandered into a risk category where you don’t get to ship a hotfix. This is AI Daily Briefing. A cloud bill, Amazon’s private AGI build, and a lab result that should make everyone sit up. First up: who’s paying for all this compute—and what happens when the output can’t be recalled? Tap follow so the next episode finds you. From Rebecca Bellan at TechCrunch:
The deal is worth upward of $100 million, Mirendil’s co-founder and CEO, Behnam Neyshabur, told TechCrunch. That’s roughly half of what Mirendil raised in seed funding at a $1 billion valuation in late June. The deal gives the startup access to both Google’s TPUs and Nvidia GPUs, as well as managed training clusters with which Mirendil will work on its self-improving AI.
Mirendil raised about $200 million in June, and now it’s committing upward of $100 million to Google Cloud. Half that seed round is effectively already spoken for by one vendor, before there’s a production system to price. And Google gets to sell it TPUs, Nvidia GPUs, and managed training clusters—the whole stack. At that point, the cloud provider isn’t just selling capacity; it’s part of the lab’s operating model. Self-improving AI gets expensive fast when every iteration burns rented compute. I’d love to see the unit economics for a system that’s supposed to improve itself while its cloud bill starts north of a hundred million dollars. Mirendil’s founders came from Anthropic, so the ambition is familiar: automate pieces of frontier research. For now, though, Google Cloud has a very large claim on a seed-stage lab’s balance sheet. The Economic Times writes:
Microsoft launched its largest India data center in Hyderabad on Thursday. This new facility brings Microsoft's cloud regions in India to four. Adani Group and HDFC Bank are among the early users of this new center. Microsoft has committed $20.5 billion to expand its India operations. Rivals Alphabet and Amazon are also investing in India's data center capacity.
Microsoft’s Hyderabad region is live: three availability zones, plus local residency for workloads banks can’t toss overseas. Early users include Adani Group, Bajaj Finance, and HDFC Bank—regulated customers with very unglamorous reasons to care. That’s a better sign than another capacity rendering. HDFC Bank using a local region for continuity and compliance may not make a killer demo, but it’s how a $20.5 billion commitment starts finding a revenue path. Four Indian cloud regions, plus two data centers with Jio, gives Microsoft a serious local footprint. What matters is who can keep the data, the fine-tuning, and the serving stack inside the customer’s compliance boundary. Sure—but I want the next update to be utilization, not another map pin. Adani and HDFC are named customers; now show us what they’re running, what it costs, and whether those three Hyderabad zones stay busy. Eugene Kim, writing in Business Insider:
Amazon is redesigning part of a massive AI data center campus in rural Indiana into a sprawling cluster of powerful computers to build its next frontier AI models, Business Insider has exclusively learned. Internal planning documents reviewed by Business Insider describe an effort to consolidate multiple data centers and deploy thousands of Trainium-powered AI servers. According to people familiar with the matter, the effort is part of a broader initiative called "AGI Pivot" supporting the company's AGI organization and its future in-house AI models.
Amazon is redesigning a rural Indiana campus mid-build around thousands of Trainium servers for an internal program literally called AGI Pivot. AWS clearly doesn’t want its frontier-model future priced by Nvidia. Business Insider’s internal documents show Amazon consolidating data centers for its own AGI organization. It’s taking control of what runs, where it runs, and whose silicon it runs on. Trainium can be a cost play and a supply-chain hedge at once. Consolidating multiple facilities says Amazon thinks the scheduling and networking gains outweigh the pain of redesigning a giant campus. That’s a serious architectural bet. Mirendil just committed upward of $100 million to Google Cloud. Amazon’s response: fine, we’ll be our own cloud customer. Different balance sheet, same fight for control of the inference stack. Scientific American, with Adam Kovac:
Scientists used artificial intelligence to generate new viruses. The landmark first could lead to new antimicrobial drugs—and may pose immense danger to human health. Scientists at Stanford University and the Arc Institute, a nonprofit AI and biology research organization, made the viruses using the actual genome of a bacteriophage—a type of virus that can kill bacteria—called ΦX174 (pronounced FYE-ex-174) as a template.
Stanford and Arc used models trained on more than two million bacteriophage genomes, then produced viruses that kill antibiotic-resistant bacteria. Once those viruses are produced, there’s no hotfix to pull them back. And ΦX174 was the template here—a bacteriophage, not a human virus—but Scientific American is right to put the misuse concern in the headline. Evo 2 can suggest entirely new genome designs. The therapeutic upside is real; antibiotic-resistant infections need better tools. But in software, we ask what happens when something works in the wrong context. Here, the paper’s successful lab result is the warning label. Congress already bundles model safety, misinformation, and biosecurity into one convenient fog bank. Add AI-generated pathogens to that hearing-room soup, and watch them demand rules for three different problems with one slogan. Here's Cha Min-ju at The Herald Business:
NHN Cloud is pushing into large-scale AI infrastructure with NHN Factory X Seoul, an AI data center in Yeongdeungpo-gu equipped with liquid cooling technology. As AI adoption has made heat management in high-performance GPUs a central challenge, the company is betting its proven operational capabilities will help it win market share.
Seven thousand, six hundred fifty-six B200s in Seoul, with direct liquid cooling holding each server to 12.3 kilowatts. Good—now we’re talking about the constraint that decides whether all those GPU promises actually run flat-out. NHN is running its own stack, not renting someone else’s cloud logo. Factory X is 5.7 times larger than its Gwangju site; this is a regional operator making a very specific bid for control. Water cooling isn’t glamorous, but a B200 cluster that can sustain performance without turning the room into a jet engine is a better enterprise story than another glossy model demo. The private-market push still has to earn its keep—but this is actual hardware, actually installed. For more on AI policy and defense, try Anthropic Pentagon Watch, a daily briefing on Anthropic’s fight with the DoD over Claude, military AI use, autonomous weapons, and AI procurement blacklisting. Find it wherever you listen to podcasts.
Links to every story are in the show notes. If one caught your attention, go read more when you have a minute. That’s AI Daily Briefing for today. This is a Lantern Podcast.