Transcript
SPEAKER_1: Let's start with the compute race. OpenAI has been moving fast on infrastructure. What's the headline there? SPEAKER_2: OpenAI signed contracts for ten gigawatts of US AI compute capacity. They added more than three gigawatts in just ninety days, and they hit a goal they once thought they wouldn't reach until twenty twenty nine. That's a massive acceleration. SPEAKER_1: And there's a structural shift happening too, right? Away from the Stargate joint venture? SPEAKER_2: Yes. Sources say OpenAI has abandoned the Stargate JV in favor of bilateral deals. And Amazon expanded its partnership with OpenAI, making models and Codex available on Amazon Bedrock. SPEAKER_1: Now there's also the ongoing legal drama with Musk. What's the latest there? SPEAKER_2: It's gotten messy for both sides. Musk admitted to distilling OpenAI's technology, and there's an alleged covert liaison inside the company. The judge won't let him argue about existential risk in court. SPEAKER_1: And OpenAI isn't coming out clean either. SPEAKER_2: Not at all. OpenAI missed its growth targets, and its own employees are raising alarms about safety failures. When the plaintiff admits to the behavior he's suing over and the defendant's employees are sounding alarms, it's hard to know who the jury is supposed to root for. SPEAKER_1: Let's move to chips. Cerebras is going public. What's the scale of this IPO? SPEAKER_2: Cerebras is targeting four billion dollars in its IPO at roughly a forty billion dollar valuation. They build wafer-scale processors designed for training large models. It would be the largest AI chip IPO since Arm's fifty four point five billion dollar debut in twenty twenty three. SPEAKER_1: So public markets are signaling they're ready to price AI infrastructure at scale. SPEAKER_2: Exactly. And there's more coming. Nigerian mobile payments service OPay is preparing a US IPO at a four billion dollar valuation, with Citigroup, Deutsche Bank, and JPMorgan advising. SPEAKER_1: What about Intel? There's been a remarkable turnaround story there. SPEAKER_2: Intel's stock jumped a hundred and fourteen percent in April, hitting a record on April twenty fourth and pushing its market cap past four hundred and seventy billion dollars. It was Intel's best month on record. After years of losing ground to TSMC and Nvidia, the market is repricing Intel as a foundry play. SPEAKER_1: And TSMC made a move around Arm shares too. SPEAKER_2: Right. TSMC sold its remaining Arm shares for two hundred and thirty one million dollars. They had originally invested around a hundred million at Arm's twenty twenty three IPO. SPEAKER_1: SoftBank is also making some longer-term infrastructure bets. SPEAKER_2: SoftBank is planning to make lithium and cobalt free data center batteries in Japan as soon as fiscal year twenty twenty seven. Japan is trying to cut its reliance on Chinese metals. Japan's twenty three billion dollar data center market is set to grow about fifty percent by twenty thirty, with ninety percent of sites concentrated in densely populated regions. SPEAKER_1: Now let's talk about the China gap. Huawei is making a big push in AI chips. SPEAKER_2: Huawei expects AI chip revenue to hit around twelve billion dollars in twenty twenty six, up more than sixty percent from seven point five billion in twenty twenty five. Orders for its Ascend nine fifty PR chip are surging as Nvidia stalls in China. SPEAKER_1: But Chinese AI models are still trailing, right? SPEAKER_2: According to NIST's CAISI evaluation, DeepSeek V four Pro lags behind leading US models by about eight months. It's the most capable Chinese AI model to date, but it's still trailing. SPEAKER_1: And Nvidia's supply chain is shifting significantly. SPEAKER_2: Nvidia's Asian suppliers now account for about ninety percent of its production costs, up from sixty five percent in twenty twenty five. And Nvidia's B three hundred servers have nearly doubled to around one million dollars each in China, as a crackdown on chip smuggling dries up black market supply. SPEAKER_1: That's pushing companies to look for alternatives. What's happening on the diversification front? SPEAKER_2: Anthropic is in early talks to buy AI inference chips from UK-based Fractile when they become available in twenty twenty seven. Nebius agreed to acquire Eigen AI for around six hundred and forty three million dollars. Eigen optimizes chip inference performance. SPEAKER_1: So the supply chain is diversifying away from Nvidia, but slowly. SPEAKER_2: That's the key word. Slowly. When Nvidia's B three hundred servers cost a million dollars each in China and everyone is scrambling for alternatives, the question is whether the AI chip shortage is creating the next generation of winners or just inflating the incumbents. SPEAKER_1: Now let's get into military AI. There's been a major classified expansion. What deals were struck? SPEAKER_2: The Department of Defense struck deals with AWS, Microsoft, Nvidia, Oracle, and Reflection AI to use their AI tools on classified military networks for lawful operational use. This is the broadest classified AI deployment in US military history. SPEAKER_1: And this came right after the Pentagon's AI chief warned against over-relying on one vendor. SPEAKER_2: Just days after, yes. And the White House is preparing a policy memo telling national security agencies to avoid single-vendor reliance. Meanwhile, the NSA is testing Anthropic's Mythos to find vulnerabilities in Microsoft products. SPEAKER_1: Anthropic's situation is particularly contradictory here. SPEAKER_2: Very much so. Anthropic is simultaneously blacklisted by the Pentagon's supply chain review and deployed by the NSA for offensive security work. Both things are happening at the same time. SPEAKER_1: There are also some significant operational contracts on the ground. SPEAKER_2: The US Navy awarded Domino Data Lab a contract worth up to a hundred million dollars for AI software that teaches underwater drones to identify new mines in the Strait of Hormuz. Scout AI raised a hundred million dollars to build an AI model that operates and commands autonomous military ATVs. SPEAKER_1: And there's a space component too. SPEAKER_2: True Anomaly raised six hundred and fifty million dollars at a two point two billion dollar valuation for autonomous spacecraft and Space Force software. Combined, that's one billion dollars in total funding across those three deals. SPEAKER_1: There's also a cultural dimension to all of this. Hollywood is drawing a line. SPEAKER_2: The Academy of Motion Picture Arts and Sciences issued new rules saying acting and writing must be performed by humans to be eligible for Oscars. And OpenAI, Palantir, and a sixteen z-backed Leading the Future is paying influencers to promote US AI and oppose Chinese AI. SPEAKER_1: The line between Silicon Valley and the Pentagon really is thinning. SPEAKER_2: It's never been thinner. And that brings us to the guardrails question. The Five Eyes nations, the US, UK, Australia, Canada, and New Zealand, published joint guidance on agentic AI systems. SPEAKER_1: What's the core warning in that guidance? SPEAKER_2: They say many organizations give AI more access than can be safely monitored. The guidance specifically flags autonomous agent deployment as a risk vector that most organizations underestimate. SPEAKER_1: So to recap: OpenAI hit its compute goals years early and is shifting to bilateral deals. Cerebras is going public at forty billion. Intel had its best month ever. Huawei's chip revenue is surging but DeepSeek still lags the US by eight months. The Pentagon is deploying AI across classified networks at historic scale. And Five Eyes is warning that agentic AI is getting more access than anyone can safely watch. Next time, we'll be wrapping up the final lecture in this series of Weekly Download forty eight.