← Back to feed

Jensen Huang used a Dwarkesh Patel interview this week to deliver two messages: one about his own moat, one about what happens if China fills the gap Nvidia isn’t allowed to close.

On ASICs: “Look at the Number That Have Been Cancelled”

Huang was direct when asked about the threat from custom AI silicon — the category of application-specific chips being developed by Google (TPUs), Meta (MTIA), Amazon (Trainium/Inferentia), Microsoft (Maia), and a long tail of startups.

“Look at the number of ASICs that have been canceled. Just because you’re going to build an ASIC… you still have to build something better than Nvidia. It’s not that easy building something better than Nvidia. It’s not sensible, actually. Nvidia’s got to be missing something, seriously.”

The sarcasm is pointed. Huang is not wrong that the ASIC graveyard is real — several well-funded attempts at custom training silicon over the past five years failed to achieve meaningful market share against H100 and A100 clusters. But the framing conveniently elides the cases that are working: Google’s TPU v5p powers the world’s largest Gemini training runs, and Meta’s MTIA 400 is now in production at 6 petaflops serving GenAI inference at scale. Neither of those is cancelled.

His ecosystem argument is stronger. Huang invoked x86 and ARM as analogies:

“Computing is not like that. There’s a reason why the x86 deal exists. There’s a reason why ARM is so sticky. These ecosystems are hard to replace.”

CUDA, the programming model that locks model training to Nvidia hardware, has 15 years of accumulated tooling, libraries, and institutional expertise behind it. Retraining that ecosystem — frameworks, optimisers, profilers, the engineering talent that knows how to debug on it — is a multi-year project, not a chip procurement decision. That stickiness is arguably Nvidia’s real moat, and Huang is right to lead with it.

On DeepSeek V4 and Huawei: “A Horrible Outcome”

The second and geopolitically sharper moment came when Huang was asked about DeepSeek preparing to launch its V4 foundation model optimised for Huawei’s Ascend 950 chips — China’s domestic answer to H100 that has become the primary training platform for Chinese AI labs operating under US export controls.

Huang called it “a horrible outcome” for America.

The specific concern: a model as capable as DeepSeek V4 — which leads on code generation and cost-efficiency benchmarks in the current generation — running natively on Chinese hardware creates an optimised CUDA-equivalent stack for Ascend. That stack, once built, lowers the barrier for every subsequent Chinese AI lab to build and run models without Nvidia. Export controls intended to slow Chinese AI capability could accelerate the development of a competing hardware ecosystem if they push Chinese labs to optimise aggressively for what they have.

Huang has separately argued against chip export restrictions on exactly this basis — contending that cutting off China does not stop Chinese AI development, it redirects it toward Huawei Ascend, which ultimately produces an Nvidia-less AI hardware ecosystem that is harder to constrain.

The Tension

Huang’s two messages are in tension with each other. If ASICs are a dead end — if ecosystems are too sticky to replace — then DeepSeek optimising for Ascend should not produce a “horrible outcome.” Either custom silicon can reach parity with Nvidia (in which case the ASIC threat is real) or it cannot (in which case Ascend-optimised DeepSeek hits a ceiling).

The more likely reading: Huang believes ASICs cannot beat Nvidia in an open market with full access to CUDA tooling, but that in a controlled environment — where export controls force Chinese labs to invest years of optimisation effort into Ascend — something close enough to CUDA could emerge. The danger is not that Ascend beats H100 globally. The danger is that Ascend becomes good enough for Chinese AI development and exports, permanently splitting the global AI compute stack into two incompatible ecosystems.

That split, not chip performance parity, is the “horrible outcome.”