"The success of DeepSeek, the Chinese rival to American goliaths with radically more cost-effective artificial intelligence, reveals the futility of U.S. sanctions policies.
Under the Biden administration, the American government was captured by some of the world's most ham-handed national-security socialists, while the Chinese private sector under Xi Jinping commands some of the world's most nimble capitalists.
The entrepreneur behind DeepSeek's apparent breakthrough is Liang Wenfeng, who founded the High-Flyer hedge fund in 2015. Since DeepSeek's launch less than two years ago, the venture has received no further outside funding.
China has roughly nine times as many engineers as the U.S. and perhaps 15 times as many science and technology graduates.
That means Mr. Liang had a cornucopia of technical talent at his disposal, all galvanized by the challenge of doing AI without violating U.S. restrictions on the memory bandwidth of their Nvidia graphics processing units. These chips, like the leading GPUs in U.S. AI data centers, are nearly all fabricated by Taiwan Semiconductor Manufacturing Co.
"Do more with less" is the Chinese entrepreneurial answer to American "Stargate" program socialism, mobilizing a half-trillion dollars to do more with more, as governments and politicians usually try to do.
By discrediting U.S. sanctions and subsidies, again, Chinese capitalists are performing a service for U.S. capitalism.
American entrepreneurs are hamstrung by a putative $6 trillion in global climate-change mandates and subsidies for obsolete technologies, such as windmills and solar panels, specified by zero-sum Green New Dealers.
The U.S. has been dissipating the bonanzas conferred in recent decades on our economy by Chinese manufacturing prodigies from Foxconn in Shenzhen and other Chinese fabricators. Chinese factories have been crucial to enabling American companies to command as much as 70% of global equity market capitalization, compared with 10% at best for China.
DeepSeek, by using microchips more efficiently, is similarly favorable to the U.S. economy. As my chip-guru colleague John Schroeter wrote in his newsletter -- and both Nvidia's Jensen Huang and Microsoft's Satya Nadella have said -- semiconductors are an example of the Jevons Paradox. William Stanley Jevons, a 19th-century British economist, discovered that when a resource is rendered more efficient, we use more of it, often so much more that total spending on the resource rises. When people used only fire for lighting, the world was a very dark place. Nobel laureate William Nordhaus has pointed out that as we progressed from candles to oil lamps to incandescent lights and now LEDs, the cost of lighting dropped by 99.97%, yet we buy more of it than ever.
Advancing at an even faster pace, the number of transistors a dollar buys has increased by several million percent in 70 years. At the same time, annual global spending on semiconductors has grown from less than a few hundred million dollars to nearly $700 billion. The cheaper computing became, the more it was demanded.
Today, the key breakthrough in technology isn't some ingenious trope in AI software but the emergence of an era altogether beyond microchips. Called wafer-scale integration, it obviates the usual data-center welter of chips and "chiplets" in plastic packages backed by snarls of wire and racks of computer servers. Instead, the new regime banishes chips and integrates the essence of an entire data center on a single 12-inch wafer. A wafer is a silicon slice that serves as the target for semiconductor lithography usually inscribing the design of thousands of separate chips. In wafer scale, by contrast, it is just one integral system.
Pioneering this breakthrough are U.S. companies such as Cerebras and Tesla. Cerebras, an AI computer innovator beyond chips, has demonstrated wafer-scale computing on about four trillion interconnected transistors. With finance from G42, a tech company in the United Arab Emirates, Cerebras had planned an initial public offering until it ran into resistance from the U.S. government based on possible links between China and the U.A.E.
The most advanced wafer-scale project is Tesla's Dojo system for AI training. It is based on the vast accumulation of video data from the cameras on Tesla's automobiles. This system is based not on chips or internet data, but on real sensory inputs and "training tiles," which are interconnected across entire wafers. Since large language models such as DeepSeek and ChatGPT use unreliable internet data, they are inherently less likely to achieve intelligence in the real world than the pixel processors on Tesla's Dojo tiles.
Working with Taiwan Semiconductor Manufacturing Co. to overthrow the existing data-center era, these ventures promise processing economies of a scale millions of times greater than anything contemplated at DeepSeek or other AI companies.
As outlined in a January 2024 article in the journal Nature, a team from Georgia Tech led by a Dutchman, Walter de Heer, achieved a further wafer-scale breakthrough using a layer of graphene atop a silicon carbide wafer. Because graphene, a two-dimensional carbon sheet, switches 1,000 times faster than silicon, Mr. de Heer's technology, the fruit of roughly 20 years of research, foreshadows a new epoch in the materials science behind information technology.
The chief obstacle to the success of such ventures is the U.S. national-security apparatus, which somehow imagines that by inflicting sanctions on China, it can help Americans. Beyond the huge challenges of replacing the existing paradigm of semiconductor fabrication, Mr. de Heer's main obstacle is his previous links with Tianjin University in China and his Chinese students at Georgia Tech. He is under investigation by a congressional committee on China for alleged links between his research and the Chinese military. Mr. de Heer said several of his students are back in China, collecting about $350 million in investments for a wafer-scale project.
Technology is the key adventure of human progress, and it is intrinsically global. The key test of the Trump administration will be whether it can come to terms with this fact of life and enterprise.
---
Mr. Gilder is author of "Gaming AI: Why AI Can't Think but Can Transform Jobs" and "Life After Capitalism: The Information Theory of Economics."" [1]
1. DeepSeek AI Is the Competition America Needs. Gilder, George. Wall Street Journal, Eastern edition; New York, N.Y.. 30 Jan 2025: A17.
