Mr. President, Here Are Some More People to Fire (and Elon Knows Where They Live)
We would not for a moment disparage the good work being done by the President and his excellent Mr. Musk in such bold undertakings as canceling U.S. taxpayer funding of transgender drama productions in the upper reaches of lower Estonia.
We merely note that there are some even more important targets, and possibly no one in the world understands them better than the immigrant from South Africa. Among his many other accomplishments, Mr. Musk, over at Tesla (NASDAQ: TSLA), is offering the semiconductor industry a way out of the cul-de-sac in which it finds itself.
Meanwhile, Mr. Trump’s own government, courtesy of the CHIPS Act, resembles nothing so much as that Drum Major in Animal House leading the band’s assault on that brick wall.
At issue is how the industry reacts to the end of Moore’s Law. Initially indicating the periodic doubling of the number of transistors that could be fit on a single microchip, for the past decade or so it has evolved into a measure of the periodic doubling of computing power at a given price.
Even by that broadened standard, Moore’s Law has been repealed. As our brilliant colleague Dr. Robert Castellano shows in an analysis to be published this Thursday in the Gilder Technology Report, the semiconductor industry no longer offers us more for less, though, as of now, it still offers more for more.
Robert starts from the observation that although revenues from semiconductor sales continue to rise, semiconductor unit shipments are nearly flat and, for many categories, are sinking. In 2024, semiconductor unit shipments rose by just 2.8%. Meanwhile, revenue jumped by 19.3%. That’s because average selling prices (ASPs) have risen as the buyers—especially data centers—have opted for more technologically advanced chips.
This is big news. The industry has always advanced by selling more for less. Less for more is a revolution. It is also unsustainable. The semiconductor industry is perhaps the most dramatic example ever of Say’s Law: supply creates its own demand. The more abundant—and affordable—a resource, the more purposes will be found for its use.
To some extent, current pricing of elite chips is no more than a traditional supply problem readily fixable by expanding production. Some of the high price of Nvidia GPUs comes down to the producers—especially Taiwan Semiconductor (NYSE: TSM)—not being able to make enough. TSM is adding capacity and supply should meet demand by 2026.
Some of the pricing problem, however, is less easily solved. As Dr. Castellano points out, high-performance chips—both logic and memory—are built on advanced process “nodes” with smaller circuits and higher manufacturing costs. “With each new technology node, the cost per wafer and per transistor rises, leading to higher prices.”
This is happy news for the chipmakers—for now. In the end, though, prices are not set by cost but by value to the customer. If (when) the cost of making elite semiconductors exceeds their value, the industry, accelerating now for more than 70 years will hit a wall, hard.
The way forward will not come from attempts to revive Moore’s Law by shrinking circuits at ever higher cost, nor from complex “3D” architectures that shift rising costs towards creative but ever more expensive chip packaging.
The way forward requires a new semiconductor paradigm and, as it happens, Elon Musk is its leading prophet. The new paradigm does not save the chip; it leaves the chip behind. No longer will we “chip” wafers into hundreds of pieces, the wafer itself will become the computer.
The most ambitious version of this new paradigm so far is the Cerebras Wafer Scale Engine with its four trillion transistors yielding 126 petaflops of compute, i.e., 126 quadrillion operations per second. As many as 2,048 of these can be linked together, bringing total computational power to 256 exaflops (quintillion operations per second).
So far, however, the Cerebras engine is being used for only the most extraordinary, custom applications. AstraZeneca uses a first-generation version to analyze thousands of scientific papers at a pace seven times faster than AI running on more conventional hardware.
Closer to the ground is Tesla’s Dojo supercomputer, driven not quite by wafers but by “tiles.” In their first iteration, DOJO tiles were built up from 25 palm-sized (645mm2) chips, each bearing 50 billion transistors. Each tile can run at nine petaflops. Dojo is built up by “exapods” consisting of 120 tiles.
Dojo is designed to train fully self-driving cars, but following generations will be designed for general purpose AI computation.
The new paradigm will not necessarily displace the current leaders of the industry. TSM is manufacturing both the Cerebras engines and the Tesla tiles. Nevertheless, the semiconductor industry and its current biggest customers, the hyperscale data centers, are on their way to an almost unrecognizable future.
Yet you can scan the website of the National Institute of Standards and Technology, which hands out CHIPS Act money, and find nary a mention of the impending revolution, just $50 billion or so of subsidies for aging technologies.
Mr. President, ask Elon to spend five minutes scanning that site. After he stops laughing, tell him to go ahead and do the right thing.