KorvaJoin the waitlist
Tech

David Sacks Says America's AI Chip Lead Comes From Speed, Not One Chip

Sacks argues new generations arrive yearly and are several times more powerful, and he warns rivals gain ground when that cycle stalls.

AI image
All-In with Chamath, Jason, Sacks & Friedberg · Jensen Ousted as Nvidia CEO: "Spy" Scandal, Export Restrictions, Market Impact

David Sacks, a commentator on Nvidia and United States chip policy, argues that America does not lead in artificial intelligence chips because it owns one unbeatable chip. He says it leads because it replaces that chip about every year with one that is several times more powerful.

He says people in Washington who are outside the chip industry see technology in static terms, focused on one particular chip. The real edge, in his view, is momentum and the ability to keep producing stronger designs.

In practice, Nvidia has released successive processor lines that include Ampere, Hopper, Blackwell and Vera Rubin, each name marking a new generation of its AI processors. To tape out means to finalize a chip design and send it to manufacturing, and Sacks says the advantage is that yearly tape-out cycle, not any one design.

China's Huawei shows why Sacks focuses on pace rather than a single model. Its individual Ascend chips remain less powerful and use more energy, Sacks says, but Huawei links 384 of them together in a system called Cloud Matrix.

Networking means connecting many chips so they work as one larger computer. Nvidia links 72 chips in its NVL-72 rack, and Sacks says that smaller system is still better, but Huawei is getting closer by using more chips and extra electricity to make up the difference.

Sacks uses the word tempo for this release schedule. By losing a tempo or two, he means losing a step in that cycle, with new generations normally arriving roughly every year to year and a half.

Phones are different, he argues, pointing to Apple under chief executive Tim Cook. Apple has largely sold incremental versions of the iPhone, while chips face faster turnover and sharper competition.

That difference shapes his warning about a slowdown. Nvidia slowing down would not by itself make rival chips better, a point Chamath Palihapitiya, another participant in the discussion of Nvidia's future, made, but Sacks says Nvidia would no longer advance at the same pace. If it stalls for two or three years, he says, rivals could overtake it and Huawei could catch up much faster.

Questions people ask

What is a rack system in a data center?

A rack is a tall metal frame in a data center that holds many processors, memory, and networking gear in one unit. Engineers measure AI systems at this scale because no single chip handles the largest models alone.

Why does connecting many weaker chips sometimes match fewer stronger ones?

Linking chips lets them split up the math behind training and running AI models and share results very quickly. The trade-off is that more chips need more electricity and faster connections, which is why Huawei's approach uses extra power.

What gets better in each new chip generation?

Each generation usually improves the design, memory connections, and energy efficiency, so it can do more AI calculations per second. Software improvements also matter, because AI labs must adapt their programs to use the new hardware.

More in Tech