Nvidia is the central bank of the AI supply chain — How deep is the moat built by Jensen Huang?

Recently, in a podcast released by a16z, renowned investor Gavin Baker and a16z partner David George had an in-depth conversation, with both agreeing that Nvidia is "in a very, very advantageous position."

Baker believes that Jensen Huang has built a deep moat. Through vertical integration of its supply chain, securing TSMC wafer capacity and key component supplies worldwide, and building a financeable data center ecosystem, Nvidia has become the "central bank" of the AI supply chain—a moat that is extremely difficult to replicate. Host David George also commented: "The past 26 years have taught me one thing — never bet against Jensen Huang."

Nine Types of Chips, One Ecosystem

Gavin described Nvidia’s current product matrix: Nine types of chips — various accelerator chips, CPUs, Ethernet switches, two types of GPUs, plus InfiniBand.

This is not a mere expansion of the product line. Gavin stated that Nvidia’s strategy is “vertically integrated but horizontally open”—even if a truly excellent competing chip appears, as long as it can plug into Nvidia’s ecosystem, it will almost certainly perform better.

This means competitors face a dilemma: direct confrontation, or integration. Gavin’s advice to all semiconductor CEOs is only one sentence: "All you need to say is 'Thank you, Jensen Huang, thank you for creating this opportunity, how can we work with you.'"

He added that each 1% market share is worth about $100 billion today: "Find a niche market; getting 1% is enough."

Supply Chain Lock-In: The Most Difficult Barrier to Replicate

Nvidia’s moat is not just in chip design capability, but even more in its control over the supply chain.

According to Gavin in the podcast, Nvidia has locked in 70% to 80% of global critical supplies, including TSMC wafer capacity, DRAM capacity, NAND capacity, laser capacity, capacitor capacity, and everything needed to build racks.

"In the past 15 years, he turned bets of several billion dollars every two to three years into bets of hundreds of billions, bringing along the entire supply chain and financing system," Gavin said.

Such integration of this scale means that even if a competitor comes up with a chip with comparable performance, they still face the practical constraint of not being able to access production capacity. Gavin pointed out directly: "Hardware is hard, the real world is hard. And Jensen Huang, at this scale and speed, brings the whole supply chain and financing system with him—it’s really not easy to copy."

Residual Value Guarantee: Turning Financing into a Moat

One of the most overlooked parts of Nvidia's moat is its financing structure.

Gavin detailed this mechanism in the conversation: suppose 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, J.P. Morgan are willing to participate in financing, mainly because Nvidia provides a residual value guarantee.

The key is, as long as the residual value guarantee is less than Nvidia’s gross profit from selling chips to the data center, Nvidia is almost risk-free, and can also get a share of the income.

In comparison, Gavin noted, TPUs may be the second most financeable option, "but probably require at least twice the equity investment, and at higher financing rates."

"Cost of capital is a huge advantage. This is why you simply want to be part of their ecosystem," Gavin said.

David further added, Nvidia helps small and mid-sized players compete with Anthropic and OpenAI through residual value guarantees: "Just like when he supported NeoClouds—it’s essentially enabling inclusive computing power, which is good for the world."

Open Source Benefits, Not a Threat

There is a concern in the market that the rise of open-source models will squeeze Nvidia’s profit margins. Gavin’s judgment is exactly the opposite.

"Some even think this is a huge risk to his business—the logic is completely flawed." Gavin said, "Open source means tokens produced on Nvidia GPUs may see profit margins drop from 90% to 40%, but this means more tokens will be consumed, requiring more computing power. In a supply-constrained world, this is a big positive for him."

Gavin also pointed out that Nvidia’s incentive mechanism naturally favors AI fragmentation, model diversity, and distributed computing power, "which is entirely consistent with U.S. national interests." He is the biggest advocate of open source—and this only serves to strengthen his business, not weaken it.

Dilemma of Challengers: Don’t Provoke Michael Jordan

For competitors trying to challenge Nvidia, Gavin used a recurring analogy.

"Sometimes you’ll see someone trash-talking in front of him, like someone talking smack to Michael Jordan when he’s in the zone—Regular Season Game 50, he’s a little bored, and then some self-confident young player decides to provoke him, and then...that’s my favorite moment to watch."

He cited the TPU team as an example, saying it once "tried to wear Superman’s cape," and the outcome wasn’t ideal. As for Apple’s in-house AI chip Jalapeno, Gavin gave some credit—"the first truly competitive internal ASIC I’ve ever seen, done in a relatively short time, and it deserves recognition"—but at the same time, he said, "Jalapeno is trying to wear Superman’s cape, we’ll see."

Gavin also pointed out a structural reason why general-purpose GPUs are hard to replace by specialized chips: DeepSeek, Kimmy, Qwen—three mainstream Chinese open-source models—evolve in very different ways. "They all run on general GPUs, but if you want to specialize, you need general chips to cope with the uncertainty of such evolution."

Chip Deal Structure Reveals True Preferences

Gavin also offered a unique perspective for judging the market’s real preferences: examine the deal structures chip companies sign with their clients.

He categorized the structures into four tiers, from best to worst: chip company directly invests in the client (such as Google and Amazon’s TPU and Tranium deals with Anthropic); residual value guarantee structures (with Blackstone, KKR involved in financing); warrants tied to a fixed token price; and simply giving warrants ("which may be negative NPV").

"Through these four deal structures, you can infer true client preferences. Nvidia’s deals are usually pretty good, and smart people are all eager to participate—which tells you something," Gavin said.