The AI Cold War, Signalgate, CoreWeave IPO, Tariff Endgames, El Salvador Deportations

Gavin Baker argues Nvidia is going through the largest semiconductor product transition ever—from Hopper to Blackwell GPUs—akin to an iPhone upgrade where you must also rebuild the house’s electrical and cooling systems. Blackwell racks are heavier, draw twice the power, and require liquid cooling, making deployments slower and more complex. This complexity and recognition timing explain why Nvidia’s accounts receivable rose faster than sales; if AR remains elevated past the July quarter, then it becomes a real red flag, but for now it’s understandable rather than evidence of “round tripping.”

Baker contends Nvidia would have sold the same volume of GPUs without equity stakes in CoreWeave and similar players—they’d simply have gone to hyperscalers like Meta, AWS, and Microsoft. By funding smaller GPU clouds that adopt Nvidia’s reference server designs, Nvidia both accelerates GPU deployment and reduces dependence on a tight three-customer oligopoly. Chamath likens this to Intel’s “Intel Inside” strategy: building an ecosystem, fragmenting buyer power, and branding what’s inside the box as a competitive moat.

CoreWeave is widely criticized as a debt-laden, commodity GPU cloud with concentrated revenue from Microsoft. Baker pushes back, saying reliably orchestrating tens of thousands of melting-hot GPUs—handling cabling failures, synchronization, and training interruptions—is extremely non-trivial. He analogizes it to running 1,000 consistent retail stores nationwide: sounds simple, but few can execute. Historically, markets underestimated AWS and Azure for similar reasons; he believes CoreWeave’s operational know-how plus acquisitions like Weights & Biases meaningfully de-commoditize the business.

While Baker emphasizes that fully-realized agents will make compute the bottleneck and drive high ROI for Blackwell, Friedberg argues the bigger unlock is enabling small teams to execute projects that today require dozens of highly specialized experts. He cites examples like planning an underwater plant-breeding facility or efficiently managing mega-projects like California’s high-speed rail. With standardized protocols like Anthropic’s Model Context Protocol (MCP), agents can coordinate services like Stripe, turning two or three generalists into the productive equivalent of 200–300 specialists.

Export controls on advanced GPUs are adding friction, but Baker notes they also massively incentivize China to build its own semiconductor ecosystem and innovate algorithmically—DeepSeek being a prime example. He pegs China’s odds of matching or surpassing Nvidia-like hardware in the next five years at essentially zero, but over ten years he sees real uncertainty, especially because the CCP plans in decades and centuries while U.S. politics turns on election cycles. This frames current AI and hardware policy as a true ‘AI cold war’ where over-tightening controls could backfire.