Dylan Patel, founder of semiconductor research firm SemiAnalysis, has laid out a scenario in which Jensen Huang’s exit from NVIDIA could open the door for Google and AMD to take meaningful chunks of the company’s AI chip market, in an interview on the tech podcast TBPN.
Patel’s framing is built around a three-pronged theory of why NVIDIA has stayed so far ahead of every rival chipmaker. “It’s sort of a three-pronged strategy,” he said. “One is you have to have the technical progress, and that keeps you winning.” The second, according to Patel, is supply chain muscle, something he credits almost entirely to Huang personally. “Jensen’s the only executive in the semiconductor industry that is going out there and drinking and meeting and convincing all of these supply chain companies to invest like they’ve never done before, give him the allocations,” he said.
The Jensen Premium
Patel’s argument is essentially that a large part of NVIDIA’s lead isn’t really about silicon at all, it’s about Huang’s personal ability to extract allocation and commitment out of a supply chain that every other chip company is also fighting over. Strip that out, Patel suggests, and competitors that have been technically competent but commercially timid suddenly look a lot more dangerous.
He pointed directly at Google’s TPU program as the clearest example. “Now if all of a sudden there’s no Jensen, and Google’s like, actually, double my allocation,” Patel said. “The TPU, from our benchmarking, we test TPUs. We run benchmarks on them on open source models every night. They’re just as good as Nvidia.” That benchmarking is a reference to SemiAnalysis’ own work, which has repeatedly found Google’s chips closing the gap. The firm has reported that Google’s TPUv7, codenamed Ironwood, delivers up to 50 percent better performance per dollar than NVIDIA’s Blackwell Ultra on some workloads, and more recently that Google’s TPUv7 actually beat NVIDIA’s best chip on an agentic inference benchmark using just a handful of chips against a full NVIDIA GPU rack.
What’s held Google back from converting that silicon into share, in Patel’s telling, isn’t engineering. “But Google doesn’t have that DNA of wheeling and dealing with the supply chain, wheeling and dealing with customers,” he said. “Now if that DNA is gone, the thing that is Jensen is gone, does that now mean Google can just take a ton of share? It’s a real big possibility.” Google has already been chipping away at that reluctance to sell outward, most notably by open-sourcing its TPU Raiden inference library in what looked like a deliberate push to make TPUs easier for outside teams to adopt, and by locking in external customers like Anthropic, which agreed to use up to one million Google TPUs in a deal worth tens of billions of dollars.
AMD’s Problem Was Never The Chip
Patel applies the same logic to AMD, which has long had the awkward distinction of making GPUs that benchmark respectably against NVIDIA’s while still capturing a fraction of the revenue. “Same with AMD,” he said. “Technically, they’re awesome, but they’re just not able to wheel and deal like Jensen is able to. And that’s been a lot of the reason why they’ve been behind.”
The other half of AMD’s problem, Patel noted, is software, specifically NVIDIA’s CUDA ecosystem, which has functioned as a nearly two-decade moat by making it expensive and risky for developers to leave NVIDIA’s hardware. “Of course, the software as well,” he said. “One of Nvidia’s biggest moats being CUDA, and software is being rapidly displaced by the fact that AI models are so good at writing software.” That moat has already shown cracks: Anthropic recently said Claude got itself running on AMD chips in three days, a migration to AMD’s ROCm stack that used to take engineering teams months, exactly the kind of AI-assisted software work Patel is describing.
Where this leaves NVIDIA is still an open question. The company still controls roughly 92 percent of the GPU market by most estimates, a dominance built over years that won’t evaporate overnight just because of a leadership change. But Patel’s point is less about NVIDIA’s current numbers than about the fragility of what’s propping them up. Huang’s exit follows disclosures around his personal conduct and a board that was reportedly already working with an outside firm to line up his replacement even before he stepped down, with CFO Colette Kress stepping in as interim CEO and names like CUDA inventor Ian Buck and chief scientist Bill Dally floated as potential permanent successors. Whoever eventually takes over will inherit the technology and the factories. Whether they can also inherit the dealmaking is, in Patel’s view, the real question now hanging over NVIDIA’s market share.