Part 1: Trending Thursday 49: Chips, Cyber, and Capital

Transcript

SPEAKER_1: Trump flew Jensen Huang, Tim Cook, and Elon Musk to Beijing this week, and it unlocked something that hadn't moved in years: H200 chip clearances for Chinese tech giants. But the deal is stranger than the headline suggests. SPEAKER_2: Right, the US cleared roughly ten Chinese companies, including Alibaba, Tencent, ByteDance, and JD.com, to purchase Nvidia's H200 chips. Trump told reporters he brought Jensen along specifically to, quote, convince China to open up. Semafor reported Huang was effectively a bargaining chip in the negotiations. SPEAKER_1: And the roster on that flight was quite something. Tim Cook was there managing Apple's manufacturing concessions. Elon Musk showed up representing the Trump administration. And Dina Powell McCormick, Meta's new global emissary, joined the delegation and even met King Charles III on the way. SPEAKER_2: It was a who's who of American tech capital. But here's the catch: clearances were granted, yet deliveries are at zero. Reuters confirmed the US gave Nvidia licenses, but Beijing has halted actual deliveries, leaving the deal in limbo. SPEAKER_1: CNBC's Kristina Partsinevelos, who reported the story first, said the ball is in China's court after Chinese orders fell apart post-GTC. And the H200 is Nvidia's second-most-powerful chip. The H100 ban still holds. SPEAKER_2: The irony is sharp. China was already building around the chip ban. CNBC reported Chinese companies are aggressively ramping homegrown AI chips even as Nvidia tries to re-enter the market. Huawei, Alibaba's Tongyi lab, and a dozen stealth-stage startups are all shipping alternatives. SPEAKER_1: So the H200 opening may have arrived too late to matter for the most ambitious players. SPEAKER_2: Exactly. And there's a talent angle too. The Wall Street Journal reported that Chinese graduates returning home from abroad rose twelve percent year over year in 2025. Beijing is offering researchers six-figure signing packages, housing subsidies, and fast-tracked lab access. SPEAKER_1: While the US has been defunding research programs and cutting visa approvals. SPEAKER_2: Precisely. China isn't just waiting for Nvidia chips to arrive. It's building the talent pipeline to compete on the frontier regardless of whether those chips ever land. The US just opened the chip tap, but China's backup plan is already further along than anyone expected. SPEAKER_1: And if deliveries stay frozen, Beijing gets the diplomatic win without giving up anything. Jensen came home without a deal. The question founders should be asking: in a world where China has both the talent and eventually the chips, what does US AI leadership actually rest on? SPEAKER_2: That's the real question. And it connects to something else that came out this week that should be uncomfortable reading at every major AI lab, every cybersecurity team, and every government agency pretending to have this under control. SPEAKER_1: You're talking about the AISI report. Claude Mythos Preview became the first AI model ever to complete both of AISI's cyber ranges, the UK government's benchmarks for measuring AI offensive capability. SPEAKER_2: Right. Those ranges simulate real-world cyberattacks. GPT-5.5 completed only one. A newer Mythos checkpoint completed a thirty-two-step corporate network attack, estimated to take a human expert twenty hours, in six out of ten attempts. SPEAKER_1: And XBOW independently evaluated Mythos and found it substantially better than prior models at finding vulnerabilities. Anthropic has put the capabilities behind Project Glasswing and is working with defenders. But their own model is now the most capable autonomous offensive hacking tool ever independently benchmarked. SPEAKER_2: And the pace of improvement is alarming. AISI estimated in February 2026 that the length of cyber tasks AI can complete has been doubling every four point seven months since late 2024, down from an eight-month doubling time in November 2025. SPEAKER_1: And then Mythos and GPT-5.5 exceeded even that accelerated timeline. AISI's own candid note was: it is currently unclear whether this is a new trend or a one-off. SPEAKER_2: Matt Clifford's read on X was simply: there is no deceleration. Ethan Mollick flagged that both models appear to be token-limited rather than capability-limited. Throw more compute at them, and they go further. The ceiling isn't in sight yet. SPEAKER_1: And while all this is happening, Washington can't even figure out who's in charge. The Washington Post reported the White House's Office of the National Cyber Director and the Commerce Department's CAISI are actively fighting over which agency should lead AI model evaluations. SPEAKER_2: Neither has the budget or staffing to do it at scale. Both want the authority. So capabilities are doubling every four point seven months, and DC is stuck in a turf war. SPEAKER_1: Let's shift to markets. Cerebras priced its IPO at one hundred eighty-five dollars per share, raising five point five five billion at a fifty-six point four billion dollar valuation. The deal was twenty times oversubscribed. SPEAKER_2: Huge demand. But there's a concentration risk worth flagging: sixty-two percent of its revenue comes from one UAE government customer. That's a significant dependency for a company at that valuation. SPEAKER_1: On the enterprise side, Cisco posted record quarterly revenue of fifteen point eight four billion and raised its AI infrastructure order forecast to nine billion. And Foxconn reported sixty-seven billion in Q1 revenue, up twenty-nine percent. SPEAKER_2: Strong numbers across the board. The AI infrastructure buildout is clearly still in full swing, and those figures reflect real enterprise spending, not just hype. SPEAKER_1: And then there's the Musk versus Altman case. Closing arguments are in. Altman holds over two billion dollars in stakes at companies that did business with OpenAI. Microsoft testified it spent over one hundred billion dollars on the partnership. SPEAKER_2: Those are staggering figures on both sides. The conflict-of-interest questions around Altman and the sheer scale of Microsoft's commitment to OpenAI are both going to be central to how this case resolves. SPEAKER_1: Before we wrap, a quick note on Startup Intros. Their mission is to help tech founders navigate everything in the early stage, from compliance to capital. Two events coming up: Caffeine and Capital on May fifteenth at Corgi Cafe in FiDi, a morning coffee meetup for founders and investors. And on May nineteenth, Agentic AI Founders' Night-Out, focused on agent orchestration beyond copilots. Both are on Luma. SPEAKER_2: Great resources for early-stage founders trying to get connected and stay ahead of where the technology is actually going. SPEAKER_1: To recap this week: the US opened H200 chip access to China, but not a single chip has shipped, and China's homegrown alternatives and talent pipeline are already advancing. Claude Mythos set a new benchmark for AI offensive cyber capability, with doubling times accelerating and no agency in Washington clearly in charge. And markets showed strong AI infrastructure momentum, from Cerebras's oversubscribed IPO to Cisco's record revenue. Next time, we'll be going deeper into what all of this means for the next wave of AI infrastructure investment.