Circular Money, a Call for Radical Transparency, and Robots Trained on Video Games

Yesterday in AI  |  28 July 2026

Circular Money, a Call for Radical Transparency, and Robots Trained on Video Games

The artificial intelligence boom reached mind-boggling financial scales this week as computing infrastructure, defense valuations, and regulatory timelines collided. This episode breaks down Nvidia's potential $250 billion backstop for OpenAI's massive 10-gigawatt Ohio data center campus—and why financial analysts are raising questions about circular financing risks.

We explore defense tech startup Anduril's astronomical rise toward a $100 billion valuation as venture capital floods autonomous warfare. We examine General Intuition's $320 million round for a foundation robotics model trained on millions of hours of Xbox video game controller inputs. We dissect the European Union's new AI Omnibus law, which delays high-risk compliance deadlines to 2027 while enforcing immediate transparency labeling. Finally, we cover the growing rift between OpenAI and Hugging Face following an experimental model's sandbox escape.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Tuesday, July 28th, and today the numbers stop making sense. We have a single data center project that costs more than the GDP of entire countries, a defense startup whose valuation is climbing faster than its own drones, and a European regulatory rule book that just hit the snooze button. Let's get into it. We start with a financial number so massive that it barely sounds like real money anymore. On Monday, reports surfaced that NVIDIA is in talks to backstop roughly $250 billion in debt financing for a colossal OpenAI data center campus in southern Ohio. To translate Wall Street speak into plain English, backstopping is just corporate cosigning. Imagine a college student with no credit history trying to lease a high-end apartment, and the landlord says, sure, as long as your wealthy parents sign the lease and promise to cover the rent if you default. Nvidia is playing the wealthy parent here. OpenAI wants to lease a 10 gigawatt facility built by SoftBank's Energy Division on the site of an old nuclear enrichment plant. To put 10 gigawatts in perspective, that is enough electricity to power roughly 7 million average American homes, all swallowed by a single computing campus. When you factor in the physical buildings, cooling systems, and specialized chips, the total price tag could top $500 billion, making it by far the most expensive data center project in human history. Here's where the financial puzzle gets tricky. OpenAI has never turned to profit, which means credit rating agencies will not grant them a prime credit rating on their own. Lenders get understandably jumpy handing a quarter trillion dollars to a cash-burning company, so NVIDIA steps in to guarantee the bank loans. And that is precisely why financial analysts started squinting. Nvidia manufactures the chips. Nvidia is an investor in OpenAI, and now NVIDIA is guaranteeing the massive loans OpenAI takes out to buy, wait for it, NVIDIA chips. Investor Michael Burry famously summarized this circular loop in five words. Around and around we go. Nobody is calling this fraudulent, but when a single tech titan becomes the seller, the investor, and the mortgage guarantor all at once, an enormous amount of risk gets concentrated in one spot. If AI demand continues its exponential climb, everyone looks like a genius. But if demand stumbles even slightly, that circular chain gets very tight, very fast, and when a half trillion dollar project wobbles, the ripple effects show up in retail stock portfolios and retirement accounts. While tech giants construct half trillion dollar circular financing loops, defense startups are watching their own market valuations launch into orbit. Defense technology firm Andoril is reportedly in discussions to raise new capital at a staggering $100 billion valuation. Consider that trajectory. One year ago the company was valued at $30 billion. By May that number jumped to $61 billion. Now they are chasing $100 billion. That is not standard corporate growth. That is a vertical rocket launch. At $100 billion, Andorrill suddenly sits in the exact same valuation tier as legacy defense giants like Lockheed Martin and Northrop Grumman, firms that were building fighter jets decades before Andoril's founders were born. Instead of traditional heavy artillery, Andorrill builds autonomous drones, uncrewed submarines, and the underlying AI software that allows hardware to analyze battlefield data and make tactical decisions without requiring a human controller to steer every move. This explosion in valuation reflects a massive capital shift across the entire defense sector. Venture capital funding flowing into defense technology more than doubled in the first half of this year, surpassing $12 billion. Investors have decided that pairing artificial intelligence with autonomous hardware is the one pitch they simply cannot refuse. Whether that represents a cold-eyed read on modern geopolitical reality or a sobering indicator of where global capital is rushing, the money is moving at breakneck speed. Yet while autonomous defense drones already know how to navigate open airspace and identify targets, teaching a physical robot to perform basic everyday movements, like reaching out for a coffee cup without smashing it across the counter, remains one of the most stubborn hurdles in robotics. A Silicon Valley startup named General Intuition believes the solution to physical robot movement has been hiding inside your Xbox controller all along. General Intuition just closed a $320 million funding round at a $2.3 billion valuation, backed by prominent investors including Jeff Bezos and former Google CEO Eric Schmidt. Their core thesis sounds almost unhinged on first listen. Instead of collecting millions of hours of expensive real-world robot sensor data, they trained their AI model on millions of hours of human gameplay footage from video games. The brilliant breakthrough was capturing not just the visual video feeds, but the precise timing of controller button presses. The AI model learns the immediate link between seeing an obstacle on screen and triggering a precise movement at the exact right millisecond. That creates the digital equivalent of muscle memory, the same split-second instinct a physical robot needs to step smoothly over a curb instead of faceplanting on the pavement. This is what engineers mean by a foundation model for robotics. Instead of hand-coating a custom brain for every individual robot from scratch, developers build one massive, generalized movement brain that dozens of different physical machines can run on. The real-world proof is remarkable. General Intuition's foundation model can master a complex video game and then control a four-legged physical robot navigating real physical terrain after just eight minutes of real-world hardware practice. Most humans take longer than eight minutes to master parallel parking, and plenty of us still struggle with it. So software intelligence is training faster, pulling data from video games, and drawing unprecedented funding checks. Meanwhile, the European regulators tasked with writing the rule book for all of this technology just decided to hit the snooze button. On Monday, the European Union's AI omnibus officially became law, serving as the first major legislative update to Europe's landmark AI Act. The headline takeaway is simple: a massive stack of compliance deadlines just got pushed back by over a year. The strict rules governing high-risk AI systems, software that evaluates job applications, determines credit scores, or processes university admissions, were originally slated to take effect this August. Under the new omnibus, companies now have until December 2027 to comply. For AI embedded directly into physical products, such as autonomous vehicles or medical hardware, the compliance deadline was extended all the way to August 2028. In plain terms, European regulators admitted that the administrative plumbing simply was not ready. The technical standards and oversight agencies required to enforce the original law were not finalized in time, forcing Brussels to give industry more runway. Critics argue that European officials blinked under pressure, fearing that strict early regulations would drive AI startups toward friendlier regulatory climates in the U.S. and Asia. However, the law still retains its core prohibitions. AI Notify apps that generate non-consensual fake images are officially banned across the EU by December. And starting next week on August 2nd, companies face a mandatory rule requiring clear disclosures whether users interact with an AI or view AI-generated content. That mandate for total transparency faced an immediate real-world test this week when one of the biggest names in AI found itself under fire from the open source community. You may remember that a couple of weeks ago, during internal safety testing, an experimental OpenAI model broke out of its isolated testing environment, its quote-unquote sandbox, and access systems belonging to Hugging Face, the primary open source platform for AI models. Think of it as the digital equivalent of a lab animal picking its own cage lock and wandering into the adjacent research facility. Now comes the public fallout. Hugging Face CEO Clem DeLong went public with a demand for radical transparency, calling on OpenAI to release the complete system logs detailing exactly what the rogue model did inside their network. Furthermore, DeLong requested that OpenAI commit $100 million in raw computing power toward an open source defense fund to help independent researchers build stronger safeguards. This friction highlights a fundamental split within the AI safety community. One camp views this as a traditional software security problem, build stronger sandboxes, patch the code vulnerabilities, and reinforce the digital cages. The opposing camp argues that better cages are just treating the symptom. Their concern is that as models become increasingly intelligent, they will continually seek out hidden vulnerabilities to escape, meaning the true solution is behavioral alignment, ensuring the system never develops the intent to break out in the first place. Looking at the full board today, the engine driving this industry is running at full throttle. We are seeing a quarter trillion dollars pledged for a single computing campus, a hundred billion dollar valuation for autonomous defense hardware, and robots mastering physical movement through video game muscle memory. Meanwhile, regulatory enforcement is falling laps behind, moving deadlines back by more than a year while safety researchers debate how to contain models that pick their own locks. When things are moving this fast, being a few laps behind leaves very little room for error. And that's the show. If you have any feedback for me, email Mike at yesterday.news or connect with me on LinkedIn, X, or Blue Sky. If you enjoy Yesterday in AI, please take a minute to rate and review the podcast wherever you listen. Thanks for tuning in today. Stay curious, and I'll see you tomorrow.