Transcript
SPEAKER_1: Anthropic just shipped its most powerful public model with the brakes welded on. Claude Fable 5 debuted as a so-called safe Mythos-class model that Anthropic says can't be used for cyberattacks, and this came just days after the company warned that AI is getting too dangerous. SPEAKER_2: And the quality is genuinely not in dispute. Tom's Hardware calls it state-of-the-art on nearly all tested benchmarks. Ethan Mollick built an interactive isochrone map and a data analysis tool in nine and a half hours. Theo Browne's verdict was simply that Fable is Mythos, and it is really good. SPEAKER_1: And the pricing is notable too, right? It's less than half of Mythos Preview. SPEAKER_2: Exactly. Ten dollars per one million input tokens and fifty dollars per one million output tokens. But here's the catch: Anthropic says Fable 5 carries invisible safeguards using prompt modification, steering vectors, or PEFT to limit its usefulness for building frontier LLMs. And there are conservative safety classifiers that silently fall back to Opus four point eight in under five percent of sessions. SPEAKER_1: So the complaints aren't really about the model itself. They're about the restrictions. SPEAKER_2: Right. Cybersecurity researchers complained the guardrails reject innocuous tasks like reading blog posts or reviewing code. Dean Ball argued the sabotage strengthens the case that labs hype safety to justify monopolistic behavior. Microsoft is limiting Fable 5 internally, citing Anthropic's thirty-day data retention rules. SPEAKER_1: And power users are reporting serious token burn too. SPEAKER_2: Theo says he churned through over one thousand dollars in tokens in a single day on a two hundred dollar plan. By Wednesday night, Anthropic backtracked on the quiet part: restricted requests will now visibly fall back to Opus four point eight instead of silently degrading. SPEAKER_1: And Dario Amodei was busy on the policy front that same week too. SPEAKER_2: He published policy proposals on AI's exponential progress and called for mandatory third-party testing of frontier models for cyber, bio, and autonomy risks. Anthropic also put one hundred fifty million dollars into Claude Corps, a fellowship placing one thousand paid fellows with US nonprofits. SPEAKER_1: And the competition is reading all of this very differently. SPEAKER_2: The Wall Street Journal reports OpenAI is considering drastic token price cuts in anticipation of similar cuts from Anthropic. The translation here is that the best model on the market is also the most deliberately restricted one, and that's not a contradiction. It's Anthropic's enterprise pitch. SPEAKER_1: The big takeaway being: if your product depends on a frontier model, this week proved the capabilities underneath you can change invisibly. SPEAKER_2: Exactly. What's your fallback when your model provider decides your use case is the dangerous one? SPEAKER_1: Now let's talk SpaceX. The largest IPO in history just closed its books roughly four times oversubscribed. SPEAKER_2: SpaceX drew more than two hundred fifty billion dollars in investor demand against the seventy-five billion it aims to raise, with a Nasdaq listing planned for this week. Retail investors alone placed over seventy billion dollars in orders and are expected to get at least twenty percent of available shares, which is an unusually large retail cut for an offering this size. SPEAKER_1: And the buyer list itself is quite a story. SPEAKER_2: BlackRock placed an order for at least five billion dollars in shares. A single family office requested over one billion on its own. Saudi Arabia's PIF and Kuwait's KIA each placed orders worth one billion to five billion dollars. And Chinese investors are buying tokenized stocks with stablecoins like USDT to circumvent capital controls and simulate bets on the IPO. SPEAKER_1: When investors are inventing synthetic exposure through offshore stablecoin rails, demand has officially outrun the order book. SPEAKER_2: And the risk factors aren't hiding either. Axios reports Elon Musk spent IPO week stoking far-right culture wars across the West, while India froze Starlink's commercial approvals over concerns its terminals were used in Iran. SPEAKER_1: And the proceeds are funding some very ambitious moonshots. SPEAKER_2: SpaceX wants to launch orbital AI computing tests by late twenty twenty-seven and has asked regulators for up to one million data-center satellites, while also prepping to absorb Cursor at four billion dollars in annualized revenue. The translation: this is the purest key-man risk ever sold to the public. Every late-stage private valuation in the space economy just got marked up with it. SPEAKER_1: Now, moving on to what may be the most jaw-dropping funding story of the week. Jeff Bezos's Prometheus raised twelve billion dollars at a forty-one billion dollar valuation. SPEAKER_2: That's the largest Series B on record, and it came just seven months after launching with a six point two billion dollar Series A. The company is building what Bezos calls an artificial general engineer: AI that compresses the path from design to manufacturing for physical objects, from bridges to chips. SPEAKER_1: And Prometheus has roughly one hundred fifty employees. That puts funding at over one hundred million dollars per head. SPEAKER_2: Before most of the world has even seen a product. The Series B drew JPMorgan, BlackRock, Goldman Sachs, DST Global, and Arch Venture Partners. And Bezos is clear: the company is not building robots. It's building the AI that designs what robots and factories make. SPEAKER_1: And physical AI checks kept landing all week beyond just Prometheus. SPEAKER_2: Neura Robotics, the German humanoid maker, raised one point four billion dollars from Tether, Qualcomm, Amazon, and Nvidia at around a seven billion dollar valuation. Standard Bots raised two hundred million dollars led by General Catalyst at a one billion dollar valuation to make AI-powered robotic arms in the US. SPEAKER_1: And there were European deals too. SPEAKER_2: Theker, the Barcelona industrial-robotics startup, raised eighty-five million dollars led by CRV, the largest robotics Series A in European history, with Samsung and LVMH making their first Spanish startup bets. Alta Ares raised a fifty million euro Series A for AI-powered air defense against drones and missiles. SPEAKER_1: And one name keeps appearing across all of these deals. SPEAKER_2: Nvidia. It backed Neura, signed a multiyear next-gen memory pact with SK Hynix, expanded an AI factory partnership with LG, and is powering Naver's gigawatt-scale AI factories for physical AI. SemiAnalysis also argues China's Unitree is running the BYD playbook to dominate the global humanoid market, so the competitive threat is real. SPEAKER_1: The translation being: software ate the world, and now the money wants AI that ships atoms. Nvidia is quietly becoming the index fund of the entire category. SPEAKER_2: And a twelve billion dollar Series B for one hundred fifty people resets every comp in physical AI. If you're building in robotics, the question is whether you raise into this euphoria or against it. SPEAKER_1: Let's talk about the AI buildout's bill, because Morgan Stanley just put a number on it: nearly five hundred seventy billion dollars in AI-tied debt issuance in twenty twenty-six. SPEAKER_2: That's more than double this year's total. Hyperscaler unsecured bond supply hit one hundred fifty-five billion dollars year-to-date by May, over forty-five percent more than all of twenty twenty-five. DoubleLine and Oaktree are openly bracing for AI credit pain. Amazon followed fourteen billion Canadian dollars in bonds with a seventeen point five billion dollar loan led by Citigroup. SPEAKER_1: And there are some very large new commitments too. SPEAKER_2: KKR launched Helix Digital with over ten billion dollars committed from Nvidia and the Kuwait Investment Authority. And OpenAI is in advanced talks to lease a ten gigawatt, over five hundred billion dollar campus in Ohio. SPEAKER_1: But the first stalls are showing up too. SPEAKER_2: Crusoe was pushed aside at its one point eight gigawatt Wyoming campus after failing to win customers including Google, amid cost and timetable concerns. SoftBank's attempt to raise over six billion dollars in margin loans against its OpenAI stake stalled, weeks after it already cut the target from ten billion. When lenders hesitate on OpenAI-backed collateral, the easy-money phase is over. SPEAKER_1: And communities aren't waiting for Washington to act either. SPEAKER_2: Seattle's City Council voted nine to zero for a one-year moratorium on new large data centers. Texas Governor Abbott released recommendations to make data centers shoulder the costs of their own growth. Ireland now requires new data centers to bring their own power plants or nearby energy contracts. And Amazon disclosed for the first time that its data centers used around two point five billion gallons of water in twenty twenty-five. SPEAKER_1: While China is running the opposite play entirely. SPEAKER_2: A state-drafted plan of around two hundred ninety-five billion dollars to fund a nationwide AI buildout sourcing over eighty percent from local suppliers. The translation: the buildout's constraint is no longer GPUs. It's credit spreads and city councils. SPEAKER_1: Now let's run through the product launches. OpenAI and Visa are letting AI agents make purchases after users grant permission, with enterprise AI-payment applications next. SPEAKER_2: And DoorDash shipped Ask DoorDash the same week: an AI chatbot that orders food, builds grocery carts, and books reservations from photos and prompts. Chatbot carts run thirty-five percent or more higher in value. SPEAKER_1: Waymo launched a Premier membership: thirty dollars a month with ten percent cashback, five free cancellations, and waitlist skipping in SF, LA, and Phoenix. SPEAKER_2: The same week Bloomberg counted Tesla's robotaxi fleet at just fifty-nine vehicles across three Texas cities after almost a year, nowhere near Musk's promises. SPEAKER_1: Citigroup launched a blockchain platform letting wealthy and institutional clients trade tokenized private-company shares. And OpenAI acquired Ona, a cloud platform for AI agents, to join the Codex team. SPEAKER_2: Google AI Plus dropped to four ninety-nine a month, down from seven ninety-nine, with storage doubled to four hundred gigabytes. The AI subscription price war is on. Gemini three point five Live Translate delivers near real-time speech-to-speech translation in over seventy languages. SPEAKER_1: Google also released DiffusionGemma, an experimental twenty-six billion parameter open model using text diffusion for faster generation than autoregressive models. And NotebookLM is going agentic, running on Gemini three point five and Antigravity. SPEAKER_2: Deezer launched an AI music detector that works across twenty services including Spotify and Apple Music, after no company licensed the tech. Xiaomi's MiMo-V2.5-Pro-UltraSpeed claims one thousand tokens per second from a one trillion parameter model on an eight-GPU commodity node. And Instagram's Your Algorithm feature is hitting the main feed, letting users steer feed topics directly. SPEAKER_1: Now for the personnel moves. Xbox CEO Asha Sharma is warning of a one hundred-day reset, citing annual revenue down nearly five hundred million dollars over five years, with major layoffs reportedly planned for next month. SPEAKER_2: Tata's chairman says AI will replace half of TCS jobs. Asia's largest outsourcer plans to slow hiring as it adopts AI. Shopee cut around eight percent of its developers globally during its pivot to AI. Opendoor shut down India operations, laying off nearly two hundred fifty people and replacing them with smaller AI-enabled teams in the US. SPEAKER_1: Alibaba swapped DingTalk's CEO after an internal debate over the enterprise app's role in Alibaba's AI strategy. And Sriram Krishnan is leaving the White House at the end of June, with sources saying he plans a pro-Trump AI policy institution. SPEAKER_2: A federal judge blocked Trump's one hundred thousand dollar H-1B fee, ruling it violated the Administrative Procedure Act and the Constitution. And FISA Section seven-oh-two expired after the House rejected a last-minute extension one hundred ninety-eight to two hundred eighteen, the program's first lapse since two thousand eight, though existing certifications run into twenty twenty-seven. SPEAKER_1: So let's bring it all together. What's the pattern underneath all of this? SPEAKER_2: It's concentration. Capital is piling into a handful of names at the extremes: SpaceX, Prometheus, the hyperscaler debt complex, while the rest of the market fights over what's left. And at the same time, the constraints are getting local and physical: city councils, credit desks, water disclosures, and model providers deciding what you're allowed to build. SPEAKER_1: And there are two specific things founders are watching coming out of this week. SPEAKER_2: First, whether SpaceX's debut marks up every late-stage valuation in hard tech, or just proves the public buys stories at any price. Second, whether trust-tier models like Fable 5 become the enterprise default, making model guardrails the new platform risk. SPEAKER_1: To recap: Anthropic deliberately limited its best model and then backtracked on the silent part. SpaceX drew two hundred fifty billion dollars in demand for a seventy-five billion dollar raise. Bezos closed a twelve billion dollar Series B for one hundred fifty people. AI debt is heading toward five hundred seventy billion dollars next year. Agents got payment rails. Waymo started charging for loyalty. And the buildout's real constraints are now credit spreads and city councils. Big weeks make big markets, and this one made several.