September 18, 2026 — Billions for AI Factories, Data Lost to War

Host0:39

Thirty point nine billion dollars. That's the new price tag on a company called Crusoe, and if you haven't heard of it, that's kind of the point. Crusoe just closed the first part of a three point nine billion dollar funding round. The list of names writing checks reads like a who's who of people betting on AI infrastructure: Mubadala, Nvidia, the Qatar Investment Authority, Founders Fund. This is patient capital, sovereign wealth funds and strategic investors, not just venture money chasing a trend. Here's Crusoe's pitch, and it's a good one. Instead of just renting warehouse space and filling it with chips, Crusoe wants to own the whole chain. Power generation, the data center shell, the cooling, the chips, the cloud software on top. The company's CEO put it as controlling everything from electrons to tokens. Think of it less like a landlord and more like an oil company that also owns the gas stations. Crusoe says it's already sitting on more than one hundred forty billion dollars in contracted value, and more than six gigawatts of capacity either built or promised. That second number is doing a lot of work in this story, because right now, only about one gigawatt of that is actually delivered and running. And that's the open question. A lot of that huge contracted number is Crusoe's own accounting, not independently audited. Investors and rivals will want to know how much of it is firm, paid commitments, and how much is options and letters of intent that could evaporate. Building six gigawatts of power-backed data centers is not a software update, it's steel and turbines and years. Whether Crusoe can turn promises into buildings on that scale is the thing to watch, not the valuation.

2:30

Here's the flip side of building all this physical infrastructure. Sometimes the physical world hits back. Amazon Web Services just confirmed something that should worry anyone who thought the cloud was basically indestructible. Back in March, a drone strike hit AWS data centers in the United Arab Emirates. Another strike in April hit facilities in Bahrain. A third strike in July, this one attributed to Iran's Revolutionary Guard, further damaged what was left standing in Bahrain. AWS spent months saying it was working on recovery. On September fifteenth, it stopped saying that. The company posted an update saying it cannot restore access to resources and data across all three availability zones in its Bahrain region, and that one zone in its UAE region is gone for good. AWS's own explanation is worth hearing in its own words: "The damage to our infrastructure spanned multiple availability zones and exceeded what our regional and multi‑AZ services are designed to withstand." That's a striking admission. The entire selling point of spreading your data across multiple availability zones is that no single disaster takes out all of them. A war took out all of them anyway. Amazon already handed out one hundred fifty million dollars in customer credits after the first strike, and told customers back then to migrate out. Some clearly didn't, or couldn't in time. AWS says it's rebuilding and expects to update customers in early twenty twenty‑seven. The unanswered question is who actually lost what. Was this a handful of enterprise customers, or something touching healthcare records, financial data, government systems. Nobody outside Amazon knows yet, and Amazon isn't the neutral party here to ask. It's a strange contrast, sitting right next to Crusoe's story. One company is raising billions to build more physical AI infrastructure. Another is discovering, in real time, what happens when physical infrastructure sits in a war zone.

4:43

Now to something that sounds boring and is actually a pretty big deal: the SEC just said yes to trading stocks on the blockchain, sort of. The agency issued what it's calling an Innovation Exemption. In plain terms, certain trading venues can now let people trade tokenized versions of regular publicly listed stocks, without those venues being treated as full stock exchanges under the law. That legal distinction has been the roadblock holding this whole idea back for years. But read the fine print, because the SEC built in a lot of guardrails. The exemption runs five years, not forever. There are volume limits. Companies whose stock gets tokenized can object. And the venues doing this have to accept surveillance and reporting requirements, plus the SEC is asking for public comment on how it's going. Think of it like a learner's permit. You get to drive, but there's a chaperone in the car, a time limit on the permit, and the state can pull it whenever it wants. This matters because it opens a real, regulated path for tokenized trading in the U.S. after Congress stalled on broader crypto market legislation. It's a workaround from the regulator, not a new law. The pushback will come from a few directions. Traditional exchanges and broker-dealers who don't love a new kind of competitor operating under different rules. Some lawmakers who'll say the guardrails aren't tight enough. And it's worth being clear that this is not the SEC legalizing tokenized stock trading broadly. It's a conditional five year test with an expiration date. Whether it becomes permanent depends entirely on whether these venues can actually meet the surveillance and reporting bar the SEC set.

6:34

Fourteen thousand square meters. That's roughly two and a half football fields, and inside it, a Chinese company claims it can build a finished humanoid robot every ten minutes. The company is UBTech, the facility is in Liuzhou, in southern China, and the claimed output works out to about ten thousand robots a year. The plant makes two robot lines, called Walker S and Cruzr, and here's the detail that actually sells the story: the factory uses autonomous robots and driverless vehicles to build its own robots. It even has a fully automated warehouse that packs one hundred and twelve robots into a space of just sixty five square meters. A UBTech vice president credited this to what China has built up over the last ten to twenty years: engineers, supply chains, and manufacturing muscle that's hard to replicate quickly elsewhere. Here's why this matters beyond the neat factory tour. Robotics has spent years stuck at the lab-and-demo stage, one-off machines built by hand, expensive and slow to make. If this factory can genuinely hit ten minutes per unit at real quality, that's the difference between humanoid robots being a curiosity and them being a commodity, the way personal computers went from garage projects to assembly lines. That "if" is the whole story though. These are company claims about a company's own factory. Nobody outside UBTech has verified that every unit coming off that line is a fully functional, sellable robot versus a partially finished shell. Competitors and Western industrial buyers will be asking exactly that question: are these ten thousand robots a year, or ten thousand robot-shaped objects a year. The difference matters a lot if you're the one buying them.

8:28

Time to move fast. Five stories, sixty seconds each, here we go. OpenAI's lawyer bot. OpenAI launched Astra for Law, pairing its GPT‑6 Astra model with a legal search index covering more than two hundred thirty million web addresses and, according to OpenAI, ninety nine point nine percent of published U.S. case law. On OpenAI's own benchmark, it correctly answered fifty four percent of legal research questions, up from thirty eight point seven percent for the base model. Good for junior associates doing case research, less good if a bar association decides an AI citing case law without a human double-check is a malpractice problem waiting to happen. Anthropic's self-driving research. Anthropic says its Claude models now lead twenty six percent of the company's own research and development work. That's Anthropic's own internal number, reported by Bloomberg, with no independent way to check what "leads" actually means yet. Worth watching, because a company whose product is an AI model has every incentive to make that number sound bigger than it is. The power-sharing club. Nvidia, Google, and a startup called Emerald AI launched a consortium with eighteen other companies to build data centers that can throttle their own power use when the electrical grid is stressed. Google says it already disclosed more than one gigawatt of that flexibility. It's basically teaching AI data centers to be polite houseguests on the power grid instead of just taking whatever they want. Amazon's bargain server. AWS rolled out a new low-cost burstable EC2 instance family called T8i, which the company says delivers up to thirty percent better price performance than the previous T-series. Not exciting, but if you run a small website or a side project database, your cloud bill might just quietly get smaller. The DLP startup betting on chatty AI. Security startup Mind raised seventy two million dollars to stop employees from accidentally leaking sensitive company files into AI chatbots. The company says its revenue grew seventeen fold and its customer count grew eightfold over the past year. In other words: as fast as companies adopt AI tools, apparently they're also panicking about what those tools might leak, and paying good money for someone to build the guardrail. That's Fast Boot. Chatbots, chips, and one very expensive electricity bill.

11:11

Here's a number worth sitting with: eight hundred seventy five million dollars, over twelve years, to help fix America's air traffic control system with AI. The Federal Aviation Administration is calling the program SMART, short for Strategic Management of Airspace, Routes, and Trajectories. According to reporting from the Wall Street Journal and summarized in TechCrunch, the system is a cloud-based platform that chews through airline schedules, weather, airport capacity, and airspace conditions, trying to spot a traffic conflict before it becomes one. It's reported to be built by a firm called Air Space Intelligence, and it's starting in the Washington, D.C. area. Think of it as a forecasting model, but instead of predicting rain, it's predicting where two planes might get too close to each other, hours before it happens, so a human controller can act on the warning instead of reacting to an emergency. This lands at a moment when the FAA is short on controllers and running on aging equipment, which is exactly the environment where a genuinely useful prediction tool could help. It's also exactly the environment where a badly tested one could make things worse. The obvious question is how the FAA proves this thing is safe before it touches airspace with real passengers in it. Twelve years is a long runway, which suggests a slow, staged rollout rather than a flip of a switch. Expect controller unions, pilot groups, and independent safety researchers to push hard for transparent testing and a guarantee that a human, not the software, makes the final call. That's the story to watch as this moves out of Washington and toward the rest of the country's skies.

13:00

So that's today: Crusoe raised nearly four billion dollars to build AI infrastructure, AWS admitted customer data is permanently lost after strikes on its Middle East data centers, the SEC opened a five year test for trading tokenized stocks, UBTech switched on a factory that claims to build a humanoid robot every ten minutes, and the FAA is spending eight hundred seventy five million dollars to bring AI into air traffic control. Full sourcing on every story is in the show notes on our website. If you found this useful, follow The Daily Reboot wherever you're listening, and send it to one person who still thinks the cloud is unbreakable. We're back tomorrow. Take care of yourself out there.