The Gigawatt War: SpaceX, Nuclear Power, and the New Cloud

Transcript

SPEAKER_1: Ok, so last time we established that Anthropic's rate limit changes on Claude 3.5 Sonnet weren't a product decision—they were the first visible output of industrial-scale compute activation. I want to go deeper on what that infrastructure actually looks like physically, because I think most people picture a server rack, not... what we're actually talking about here. SPEAKER_2: Right, and that's the mental model that needs updating. We're not talking about racks. Bloomberg, via Letsdatascience, reports that Anthropic's compute commitments now span roughly 3.5 gigawatts of contracted power. For reference, a typical nuclear reactor generates about one gigawatt. So Anthropic has essentially contracted the output of three and a half power plants—just for AI inference and training. SPEAKER_1: Three and a half gigawatts. How does that break down between Amazon and Google specifically? SPEAKER_2: The Google side is the more dramatic number. App Economy Insights reports that Alphabet is committing up to $40 billion in Anthropic—$10 billion upfront, $30 billion tied to milestones—and crucially, five gigawatts of computing capacity over five years. The Information confirmed on May 5, 2026 that Google will supply Anthropic with five gigawatts of server capacity beginning next year, tied to Anthropic's $200 billion spend commitment. Amazon's April 2026 move, per Bloomberg via Letsdatascience, was an additional $25 billion investment structured as AWS Trainium consumption. SPEAKER_1: So for someone like Anvesha, who's been tracking these deals, the question becomes: why does an AI company need to get into the energy procurement business at all? Why not just... buy more cloud credits? SPEAKER_2: Because cloud credits don't solve the physical bottleneck. Air Street Press reported in May 2026 that data center NIMBYism is now a first-order constraint on scaling AI compute—communities are blocking new builds. So FERC is fast-tracking transmission permits, and the DoE and DoD are coordinating on siting data centers near nuclear baseloads. The energy source isn't incidental. It's the strategic moat. SPEAKER_1: That's the part I want to push on—why nuclear specifically? What's the actual advantage over, say, solar or natural gas? SPEAKER_2: Baseload reliability. Solar and wind are intermittent; you can't run a training cluster at 95% utilization on intermittent power without massive battery buffers that don't yet exist at this scale. Nuclear delivers consistent gigawatt-hours around the clock. The DoE and DoD coordination Air Street Press described is essentially the government acknowledging that AI compute is now a national infrastructure priority on par with the grid itself. SPEAKER_1: And what are the risks there? Because that's not a trivial bet—nuclear has its own set of problems. SPEAKER_2: Permitting timelines are the biggest near-term risk. New nuclear in the US takes a decade or more to permit and build. So the near-term play is co-locating with existing plants, not building new ones. The longer-term risk is concentration—if your entire training stack depends on a handful of nuclear sites, a regulatory change or a plant shutdown creates a single point of failure that no amount of cloud redundancy can fix. SPEAKER_1: Ok, so that's the energy layer. Now where does SpaceX fit into this? Because in the last lecture we noted Anthropic is actually excluded from Pentagon AI deals—The Next Web confirmed Anthropic was ejected from Pentagon supply lines due to safety restrictions. So the SpaceX angle isn't military. What is it? SPEAKER_2: It's connectivity for distributed infrastructure. When you're running 3.5 gigawatts of compute across multiple geographies—because you can't put it all in one place due to the NIMBYism problem Air Street Press flagged—you need low-latency links between clusters. Undersea cables have fixed routes and chokepoints. Starlink's low-earth-orbit constellation can provide flexible, high-throughput links to data centers in locations where terrestrial fiber simply doesn't reach. SPEAKER_1: So it's less about speed and more about... geographic flexibility? SPEAKER_2: Exactly. The concept here is what infrastructure analysts call data gravity—compute needs to be close to where data is generated and stored, because moving petabytes across long distances is expensive and slow. Satellite connectivity lets you place a cluster near a nuclear baseload in a remote location and still integrate it into a global inference network without a decade-long fiber build. SPEAKER_1: That reframes the whole thing. It's not cloud computing in the traditional sense—it's something more vertical. SPEAKER_2: That's precisely the shift. Traditional cloud computing means you rent capacity from a provider who owns the stack. What Anthropic is building—through the Google TPU deal Devdiscourse reported, the Amazon Trainium commitment, and the energy siting strategy Air Street Press described—is a sovereign compute stack. They're contracting the silicon, the power, and the connectivity as integrated layers, not as separate vendor relationships. SPEAKER_1: And the TPU piece specifically—Devdiscourse reported that in April 2026, Anthropic collaborated with Google and Broadcom to access several gigawatts of tensor processing unit capacity anticipated to be operational by 2027. How significant is the Broadcom involvement? SPEAKER_2: Broadcom is the chip designer; Google manufactures TPUs using Broadcom's architecture. By going directly to Broadcom in that three-way agreement, Anthropic is essentially influencing chip design specifications for its own workloads. That's not a customer relationship—that's vertical integration into silicon. ResultSense confirmed this deal on May 6, 2026. SPEAKER_1: Meanwhile, Bloomberg reported Anthropic's valuation went from $380 billion at the Series G close in February 2026 to potentially $900 billion by late April—Letsdatascience noted that's a 2.4x jump in roughly eleven weeks. Does the infrastructure story explain that multiple? SPEAKER_2: It does, because investors are pricing in the sovereign compute thesis. A company that controls its energy, its silicon roadmap, and its connectivity layer is not a software company with cloud exposure—it's an infrastructure company with software margins. App Economy Insights framed the Google-Anthropic relationship as a paradox: Google is simultaneously Anthropic's investor, cloud provider, and chip supplier. That kind of structural lock-in commands a different valuation framework entirely. SPEAKER_1: One more thread—Air Street Press flagged that frontier cyber-offense capability is doubling every four months. How does that add urgency to all of this? SPEAKER_2: It means the compute race has a security dimension that compresses timelines. If offensive AI capability scales that fast, the window to establish infrastructure dominance is measured in quarters, not years. That's why the April 2026 deals moved so quickly and at such scale—everyone involved understands that the entity that controls gigawatt-scale inference in 2027 sets the terms for the entire decade. SPEAKER_1: So for our listener trying to make sense of all this—what's the single frame they should carry out of this lecture? SPEAKER_2: The alliance between Anthropic and the cloud giants has evolved beyond a vendor relationship into a vertical integration play where energy procurement and global satellite connectivity are now core components of AI model deployment. Whoever controls the gigawatts controls the models, and whoever controls the models sets the price of intelligence. That's the new cloud—and it runs on nuclear baseloads and low-earth-orbit satellites, not just server racks in suburban data centers.