Latest / UpNext AI / SpaceX’s $2.8B AI Power Bet, OpenAI’s Math Breakthrough, and the Rise of Medical AI Agents | UpNext AI – May 21, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Thursday, May 21st, 2026, and here's what matters in AI today.\n\nFirst up, the infrastructure story.\n\nWIRED reports that SpaceX has committed more than $2.8 billion in recent months to buy gas turbines to power AI data centers tied to Elon Musk’s AI unit. The article says this comes as Musk’s AI operation faces complaints, a lawsuit, and regulatory scrutiny over emissions tied to those turbines.\n\nThe bigger point here is that power is becoming one of the hardest constraints in AI, not just chips. According to WIRED, SpaceX and xAI operate data center sites in Memphis, Tennessee, and Southaven, Mississippi, and as of March had enough servers across those sites to use about 1 gigawatt of power, roughly the electricity demand of a large U.S. city. Portable gas turbines are being used as a fast workaround while longer-term energy sources catch up.\n\nThe filing cited by WIRED also sketches the scale of the build-out: an $805 million turbine agreement in March, a separate $2 billion deal in late April that is still pending, and more than $14 billion in construction in progress. WIRED also reports that SpaceX is leasing some server access at those data centers to Anthropic and that Musk said additional deals are planned.\n\nSo the takeaway is simple: the AI race is now as much about securing electricity as securing accelerators. And that has real consequences for where data centers get built, how fast capacity comes online, and what environmental tradeoffs companies are willing to make.\n\nNext, a strong second act from the research frontier into reasoning itself.\n\nTechCrunch reports that OpenAI says one of its reasoning models produced an original proof that disproves a geometry conjecture first posed in 1946. That would make it an 80-year-old open math problem. And crucially, this time the report says mathematicians including people who criticized OpenAI’s earlier claims are backing the result.\n\nThat context matters. TechCrunch notes that months ago OpenAI had made earlier claims around unsolved Erdős problems that did not hold up as originally presented. In this new case, the company also published supporting remarks from mathematicians including Thomas Bloom, who had previously called that earlier episode a dramatic misrepresentation.\n\nIf this result stands, the significance is not just that AI got a math proof. It’s that a general-purpose reasoning model, rather than a system built only for theorem proving, may be starting to sustain longer chains of thought across a hard open-ended problem. TechCrunch says OpenAI frames this as the first time AI has autonomously solved a prominent open problem central to a field of mathematics.\n\nThe practical implication is broader than geometry. If systems can genuinely reason through unfamiliar, multi-step problems, that matters for science and engineering too. But for now, the cleanest version is: OpenAI is making a much stronger math claim than last time, and this one appears to have more credible outside support.\n\nFor research today, we’re staying practical.\n\nA paper in Eye asks whether AI agents could be a game changer in ophthalmology, or eye care. The core setup is easy to understand: the paper says a single clinician often has to combine patient history, examination findings, and diagnostic imaging in a high-volume setting. That workflow is getting strained by workload, referrals, and documentation.\n\nThe paper argues that agent-style systems could help because they are designed to connect multiple models and multiple inputs inside one working system, rather than handling one narrow task at a time. In plain English, this is less about an AI replacing a specialist, and more about software helping assemble the case: what the patient said, what the exam showed, what the images suggest, and what needs attention next.\n\nWhat we do not have in the supplied summary are hard results like sample size, accuracy numbers, or clinical outcomes. So this is better read as a framing paper than a definitive proof point.\n\nBottom line: one of the clearest near-term uses for AI in medicine may be coordination, helping clinicians synthesize a messy case faster, rather than trying to hand diagnosis over to a fully autonomous system.\n\n...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ...\n\nTechCrunch reports Anthropic will pay xAI $1.25 billion per month for compute through May 2029, in a deal that could be worth more than $40 billion in revenue to xAI. The details came from SpaceX’s S-1 filing, according to the report. Whatever the final economics look like, it is a vivid sign of how expensive frontier AI infrastructure has become, and how rivals can also become suppliers.\n\nBloomberg Opinion has a column arguing that the AI boom is rewarding some of Stanford’s worst instincts. This is commentary, not straight reporting, but the piece uses student anxiety, recruiting pressure, and the orbit around frontier labs to argue that AI is reshaping elite computer science culture in ways that are not entirely healthy.\n\nForbes makes a useful enterprise point in a piece on what it calls tokenomics. The argument is that the important metric is not the price of a token, but the cost of getting actual work done. In other words, cheaper model calls do not automatically mean lower AI bills if a workflow still needs lots of retrieval, validation, tool use, and human review.\n\nAnd The Decoder reports Stability AI has launched Stable Audio 3.0, with open weights and support for tracks up to six minutes long. Even with limited detail in the supplied summary, the headline is clear: open audio generation is still moving forward, and longer-form output is becoming part of the competition.\n\nBefore we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers.\n\nIf you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!