U.S. Opens a Disruptive Window for Chinese Semiconductors

Disruptive innovation was the late Clay Christensen’s term for what happens when an originally inferior technology becomes “good enough” for customers who had been buying elite products at premium prices, but no longer need the latest, greatest boost in performance if they can get good enough at a much lower price.

When the steamship replaced the clipper ship, it was still slower, but it was good enough and cheaper. The clipper sank beneath the waves of history.

Disruption can be dramatic, with the elite product disappearing almost overnight and its maker in bankruptcy just as quickly.

Most importantly, Clay argued that the elite incumbent company usually has no defense from a disruptive competitor. That’s because it was the incumbent’s excellent management that created the danger. Good managers are always pushing “up and to the right” catering to their best customers with their highest margin products. The better the company is the more likely it will eventually offer more performance than its customers need or are willing to pay for–if they can get by with a cheaper, good enough, alternative.

U.S. sanctions aimed at keeping America ahead of China in semiconductors may be setting the domestic semiconductor industry up for a drastic disruption from good enough alternatives.

DeepSeek made a lot of headlines for creating an AI model at a fraction of the cost of, e.g., Open AI’s models such as ChatGPT. Most of the cost savings came because DeepSeek’s model used less expensive hardware compared to Nvidia’s latest and greatest GPUs. These can cost from $30,000 to $70,000 for the floor model and are needed by the thousands in hyper-scale data centers where AI models are trained.

The price pressure alone was always going to lure lower-cost AIs into the field. DeepSeek, contrary to the hype, was not the first, nor the last. Our friend and eminent economist Gale Pooley, the co-inventor of time-prices, notes in the latest Gilder Technology Report that a new model competitive with DeepSeek was created on two dollars and 62 cents worth of rented hardware.

DeepSeek and other Chinese entrepreneurs, however, had an extra incentive beyond saving some bucks. The hardware Open AI, and Microsoft, and Google, and everyone uses is not legally available at any price in China because of U.S. export bans.

With price pressure providing so much incentive anyway, it may be a stretch to “blame” sanctions for DeepSeek. Not so for other looming disruptions.

The most punitive of all the U.S. sanctions is on a Dutch product, ASML’s extreme ultraviolet (EUV) photolithography machines, which are required for state-of-the-art chipmaking using the so-called “7nm”, “5nm”, “3nm” and now “2nm” processes.  The Chinese have none of these machines and are forbidden to acquire them. That puts Chinese semiconductor makers at least six years behind the United States.

The largest and best Chinese semiconductor manufacturer, Semiconductor Manufacturing International, or SMIC, has found work-arounds to make chips at the “7nm” level by using earlier generation ASML machines. So far, SMIC has been getting disappointing yields of about 50% of usable chips per wafer, making its “7nm” chips less economical. But SMIC’s yields will improve. That’s what the experience curve does.

And it is from these three-generations-behind 7nm chips that the potential disruption is coming.

7nm is not (apparently) good enough for Nvidia’s latest AI chips, or even AMD’s. But now comes a revolution in the industry that may make 7nm more than good enough for any AI model.

This is the wafer-scale revolution, the movement by companies such as Cerebras and Tesla to make integrated circuits comprising an entire 12-inch diameter silicon wafer, normally used to make 300-400 chips.

Wafer scale makes a single integrated device into a supercomputer. The Apple M1 Ultra chip bears more than 100 billion transistors, the most of any single logic chip. The second-generation Cerebras device “Wafer Scale Engine 2” carries 2.6 trillion transistors. And it was manufactured on TSM’s “7nm” process.

Wafer Scale Engine 3, manufactured on TSM’s 5nm process integrates 4 billion transistors, but Engine 2’s 2.6 trillion is still more powerful than any single chip will ever be and plenty good enough to train big AI models.

Getting to wafer scale will not be easy for the Chinese. We know of no manufacturer other than TSM capable of producing anything comparable to the Cerebras Devices. But it very likely will be easier to get to wafer scale than to duplicate one of ASML’s extreme machines, probably the most complicated machines on earth ever. In any event, wafer-scale is a second path to “good enough” Chinese chips to take on the world.

Absent sanctions, the Chinese using ASML’s best machines, would likely have followed America’s best chip companies pushing up and to the right, for ever-higher performing, higher margin devices. Instead, we are forcing them along the path of good enough, and good enough may be enough to set the U.S. chipmakers on their ears.

We put the nm measures in quotation marks because these have become mere marketing labels for succeeding generations of chips. No part of a “7nm” chip actually measures 7nm.