In a recent episode of the a16z podcast, famed investor Gavin Baker and a16z partner David George engaged in a deep discussion, with both agreeing that NVIDIA is "in a very, very favorable position."
Baker believes Jensen Huang has built an exceptionally deep moat around the company. Through vertical integration of its supply chain, locking up wafer capacity at TSMC and other critical global component supplies, and constructing a financeable data center ecosystem, NVIDIA has become the "central bank" of the AI supply chain, a moat that is incredibly hard to replicate. Host David George also remarked: "The past 26 years have taught me one thing — never bet against Jensen Huang."
Nine Chips, One Unified Ecosystem
Baker described NVIDIA's current product matrix: nine chips, including multiple acceleration chips, CPUs, Ethernet switches, two types of GPUs, and InfiniBand networking. This is not a simple product line expansion. Baker stated that NVIDIA's strategy is "vertically integrated but horizontally open," meaning that even if a truly excellent competitive chip emerges, as long as it can plug into NVIDIA's ecosystem, it will almost certainly perform better.
This leaves competitors facing a dilemma: confront NVIDIA head-on, or integrate into its ecosystem. Baker's advice to all semiconductor CEOs is succinct: "The only words you need to say are 'thank you Jensen, thank you for creating this opportunity, how can we cooperate with you.'" He added that each 1% of market share is currently worth roughly $100 billion, so "find a niche and take that 1%."
Supply Chain Lockdown: The Hardest Barrier to Replicate
NVIDIA's moat is not just about chip design capability; it lies in its control over the supply chain. According to Baker's podcast insights, NVIDIA has secured 70% to 80% of global critical supplies, including TSMC's wafer capacity, DRAM capacity, NAND capacity, laser capacity, capacitor capacity, and everything needed to build out a full rack system.
"Over the past 15 years, he has upgraded bets of hundreds of millions of dollars every two to three years into bets of hundreds of billions of dollars, all while bringing the entire supply chain and financing system along with him," Baker said. This scale of supply chain integration means that even if a competitor produces a chip with comparable performance, they face the overwhelming constraint of having no access to manufacturing capacity. Baker pointed out directly: "Hardware is hard, the physical world is hard. And Jensen, at his current scale and speed, keeping the entire supply chain and financing system with him, that is truly hard to replicate."
Residual Value Guarantees: Turning Financing into a Moat
One of the most overlooked aspects of NVIDIA's moat is its financing structure. Baker detailed this mechanism in the conversation: assume a NVIDIA data center costs $50 billion. The buyer only needs $15 billion in equity; the remaining $35 billion can be financed. Institutions like Blackstone, KKR, Apollo, Goldman Sachs, and JPMorgan are willing to participate because NVIDIA provides a residual value guarantee. The key is that as long as the residual value guarantee is lower than the gross profit NVIDIA earns from selling chips into that data center, NVIDIA carries almost zero risk while also generating income from revenue-sharing agreements.
In comparison, Baker noted that the TPU might be the second most financeable option, "but it would require at least double the equity commitment, with higher interest rates on financing." "The cost of capital is a huge advantage, and it's why you simply want to be part of his ecosystem," Baker said. George added that NVIDIA's residual value guarantees help smaller players compete with Anthropic and OpenAI, "just like when he supported NeoClouds — this essentially democratizes compute, which is beneficial for the world."
Open Source Is a Tailwind, Not a Threat
There is a prevailing concern that the rise of open-source models will compress NVIDIA's profit margins. Baker's assessment is the exact opposite. "Some people think this is a huge risk to his business — the logic is completely inverted," Baker said. "Open source means that the margin on tokens produced on NVIDIA GPUs might drop from 90% to 40%, but it also means more tokens will be consumed, requiring more compute. In a supply-constrained world, that's a huge tailwind for him."
Baker also pointed out that NVIDIA's incentive structure naturally aligns with AI fragmentation, model diversification, and the decentralization of compute, "which is entirely consistent with U.S. national interests." Jensen is the biggest proponent of open source, and it strengthens his business rather than weakening it.
The Challenger's Dilemma: Don't Poke Michael Jordan
For those attempting to challenge NVIDIA, Baker used a recurring metaphor. "Sometimes you see someone talking smack in front of him, like seeing someone talking trash to Michael Jordan when he's in his prime — game 50 of the regular season, he's a little bored, and some cocky young player decides he wants to poke the bear, and then... that's my favorite moment to watch."
He cited the TPU team as an example, suggesting they "pulled on Superman's cape," with less than ideal results. Regarding Apple's self-developed AI chip, "Jalapeno," Baker offered some praise — "it's the first truly competitive internal ASIC I've seen, built in a relatively short amount of time, and credit should be given where it's due" — but added, "Jalapeno is pulling on Superman's cape; let's see what happens."
Baker also highlighted a structural reason explaining why general-purpose GPUs are difficult to replace with specialized chips: DeepSeek, Kimmy, and Qwen, three mainstream Chinese open-source models, are evolving in very different ways. "They all run fine on general-purpose GPUs, but if you want to specialize, you need general-purpose chips to handle that evolution and uncertainty."
Deal Structures Reveal True Preferences
Baker also offered a unique lens for gauging real market preferences: examining the deal structures that chip companies sign with their customers. He outlined four tiers of deal structures, ranked from best to worst: direct equity investments by chip companies into customers (such as Google and Amazon's TPU and Tranium deals with Anthropic); residual value guarantee structures (with Blackstone and KKR involved in financing); warrants tied to fixed token prices; and finally, plain warrants ("which could be negative NPV").
"Through these four tiers of structures, you can infer true customer preferences. The deals NVIDIA does are usually very good, and smart people are participating in his deals — that says a lot by itself," Baker concluded.
Comments