Deep-Sixing Technology and Innovation Hype
The “Manhattan Project” is a poor metaphor for innovation policy, but a great model for laundering money.
“Chinese startup DeepSeek shocked markets this week with powerful artificial intelligence models that might have been produced at a fraction of the cost of competitors’ technology, despite US restrictions on the most advanced semiconductors.” Thus intoned the Bloomberg Editorial Board on January 30, 2025.
I am not equipped to say anything about the capability of the DeepSeek AI platform, but the hype is that DeepSeek is as capable, or more capable, than existing platforms, and “the Chinese” managed to produce it at the cost of $6 million, an order of magnitude lower than the $billions that American AI competitors have thus far expended developing their own platforms. DeepSeek thus seems to have leapfrogged American rivals and may have situated itself to win “the AI race.”
The notion of a technology “race” may have a lot of intuitive appeal, but it is not representative of most R&D competition, at least to the extent it suggests that the “race” has a well defined finish line and winning the race amounts to dominating, if not even monopolizing, a nascent market. At most, the DeepSeek revelation may amount more to an important example of a “knowledge spillover”: DeepSeek may have revealed a more efficient way of deploying computing resources for AI development, and that more efficient way may be available free of charge to other developers. So, if the DeepSeek innovation is a real thing, then good for them. And good for the rest of us. DeepSeek will have done all of us a public service.
Now, there is a question of why DeepSeek would have revealed its secret sauce. Perhaps commercializing the technology required DeepSeek to reveal it. Perhaps the mere expression of the technology amounted to a revelation. Or perhaps the news is just a deep fake. I don’t know. But, much commentary on social media framed the DeepSeek revelation as a “Sputnik moment,” as an indication that American developers had been caught flat-footed and had allowed a virtually unknown competitor to leapfrog them and to secure an insurmountable lead in “the AI race.”
It is true that AI darling Nvidia did lose about one-seventh of its market capitalization overnight. It was trading around $140 per share on the NASDAQ exchange and is now trading around $130 per share. Spot the NVDA bloodbath on this 2-year stock price chart here:
https://www.barchart.com/stocks/quotes/NVDA/interactive-chart
One can also observe the broader AI bloodbath in the NASDAQ chart here:
https://www.barchart.com/stocks/quotes/$NASX/interactive-chart
Don’t worry if you can’t really discern a bloodbath. Because there isn’t one. There is a discernible effect, but it is hard to distinguish it from other discernible effects. It does not look exceptional. But, don’t tell that (oddly) to the professional America-haters and Jew-haters online. Many of them gleefully jumped on the DeepSeek news. (For example: https://x.com/thatdayin1992/status/1884214714995204467) Did the news make a mockery of President Trump’s announcement that various entities endeavor to invest something on the order of $half-a-trillion in the United States to develop AI and the energy sources required to run it?
In any case, market indices did not crash, and investments in AI proceed apace. But, that doesn’t mean that the notion an innovation “race” is useless. In the late 1980’s, for example, there was a race to develop certain superconducting materials. Someone had randomly discovered that ceramics blended with a little barium could superconduct electricity at much higher temperatures than had previously been achieved. Up to that point, available materials had to been subjected to a deep, deep freeze—we’re talking like -200 degrees Celsius—in order to conduct electricity with no resistance. The new materials could superconduct at some temperature higher than -200, and a race started to find materials that could superconduct at much higher temperatures. Imagine, for example, the Holy Grail of room-temperature superconductivity. Superconductivity would then likely become economical for a host of applications. The most common example would be magnets so powerful that they could levitate trains. Hence superfast trains gliding on air.
I had the privilege of interviewing an MIT engineer about the superconductivity hype. (He was very gracious in taking a cold call from me, an 18-year old.) He had been featured in the pages of magazines like Business Week, expressing concerns about Japanese consortia securing a lead in superconductivity and … about how that would be bad, because …?
Because such a lead, if any, would enable Japanese consortia to dominate or monopolize new markets? Japanese firms would definitively win the race, and American firms would thus get shut out, because, that’s what winning entails, right, shutting out the competition? So, therefore, the United States government should subsidize research on superconductivity.
This fellow had pressed for government subsidies in testimony on Capitol Hill, although I don’t think anything ultimately came of it. Which, is just as well. The superconductivity hype ultimately passed, and, these decades later, you still have to cool materials to extremely low temperatures to enable them to achieve superconductivity. It is thus not obvious that Japanese government subsidies achieved much, if anything.
My small bit of research into superconductivity constituted my first experience in thinking about how public policy can promote innovation. And, what I got out of it was more questions as well as a healthy dose of skepticism. And, yet, “innovation” likely remains a hot topic, in one guise or another, at the business schools or public policy programs. Hmm.
Spillovers and underinvestment: The economics rationale for R&D subsidies
If there is a policy-relevant question about subsidies, it relates to “knowledge spillovers” and underinvestment in research-and-development (R&D). Here the idea is that spillovers reflect value that an innovator cannot appropriate by commercializing a technology. But, that privately inappropriable value does reflect value that society as a whole gets to enjoy. So, economists will illuminate a potential problem of underinvestment in R&D: A given innovation might yield huge benefits to the whole of society, but, if a private party can’t expect to extract much private return to the costly R&D that would go into yielding that innovation, then that private party might decline to pursue R&D in the first place. So, economists would then be quick to identify a role for public policy: subsidize R&D in technologies that, were they to be realized, would (1) generate large social benefits but would yet (2) only generate small private benefits. Easy, right?
I don’t know that anyone has come up with a good scheme for systematically identifying technologies that satisfy both conditions, which is another way of saying that trying to centrally manage R&D processes may not really work. And, worse, trying to manage such processes may amount to little more than a game of “picking winners.” And, even worse, picking which projects to subsidize may degenerate into a variety of crony capitalism: the central authorities endeavor not so much to guess which R&D projects might turn out to be winners, but they direct subsidies to their friends. They end up anointing winners in the game for securing subsidies.
Over the last week, Elon Musk’s team over at DOGE (the “Department of Government Efficiency”) has demonstrated that government has been subsidizing a vast array of insane projects, although DOGE has not yet focused on R&D subsidies. People who have studied government R&D subsidies will note, however, that the government has been heavily subsidizing serious R&D since the immediate post-war era. A lot of public resources went into military programs, and observers will note that it was out of such programs in the 1950’s that we got things like transistors. Transistors constituted a huge advance over the vacuum tube technology of the time.
Now, there is a question of whether private parties would have developed innovations such as transistors. Basically, did public financing of certain innovation merely crowd out private efforts. Did public financing really help deliver innovations that would not have otherwise emerged from the private sector?
This is important question, because we can imagine private parties being very pleased to get subsidies to do work that they would have done with or without subsidies. In other words, it’s not obvious that any given body of subsidized innovation satisfies the second condition of “only generates small private benefits.” And we can’t expect reliable answers from private parties. Of course they need subsidies, they will say.
My own experience looking into government subsidies in R&D had focused on the Advanced Technology Program (ATP) managed by the National Institutes of Standards and Technology (NIST). This was not a big program by the standards of the last few years. But, it was something the Clinton Administration latched onto in the early 1990’s. The Administration took a program that put up grants for collaborative R&D projects totaling $tens-of-millions in a given year and ramped it up to $hundreds-of-millions in a given year. All well and good, but the managers of the ATP ended up having to face up to the criticism that it was not obvious that the program was funding projects that private parties were not eager to otherwise fund themselves. Indeed, ATP would advertise the fact that it had money to hand out. Applicants naturally showed up.
To the great credit of the ATP, it did make a big effort to come up with a better way of screening proposed projects. The ATP ended up consulting with a number of academics who do think very seriously about “knowledge spillovers” and “appropriability problems” in R&D—about the phenomena that induce underinvestment in R&D—and these people helped the people over at the ATP to appreciate the idea that ATP really should endeavor to identify projects that would likely (1) generate the largest social benefits while (2) yielding small private benefits, for, again, these would be the really good projects that private parties would nonetheless not have an incentive to pursue. But, again, how to identify such projects? Not easy.
So, it’s not obvious that the greatest thinkers about innovation processes could themselves design a good process for selecting good candidates for R&D subsidies. But, all is not lost. The conclusion may be negative, but there is a place for negative conclusions, because they do provide guidance about what the central authorities should do—or not do. And they shouldn’t get too ambitious about picking winners among specific R&D projects. Instead, the authorities might concentrate on funding much more “basic research” on technologies that are far from commercialization. Here the idea seems to be that we really don’t know where a given technology may go, so let’s give up on particular applications and just fund research that may prove to have very broad applications far into the future.
It is that kind of thinking that has motivated heavy funding of research of ostensibly small private parties. The Bayh-Dole Act of 1980 is central here. A private party could secure government support but would have to grant the government a royalty-free license to any technology that might ultimately be patented, and it was through this mechanism that much funding has gone toward university researchers.
I have not studied funding channeled through the Bayh-Dole process or through other channels to the universities, but one can wonder if the central authorities have concentrated resources on projects in politically favored research areas. And, have the authorities tended to favor certain conclusions in funded research? In this way, do we ultimately end up with bodies of Soviet, Lysenko-ized science? For example, I get notices about research positions, research conferences or research programs on “sustainability” all of the time. The idea here is that researchers can do any volume of research so long at it conforms to a certain orthodoxy. But, that’s not research. That’s not science. That’s Lysenko-ized science of establishing conclusions first and then building up a judiciously selected body of evidence that appears consistent with the conclusion.
* * *
The administration of Jimmy Carter (1977-1980) took some interest in innovation-relevant policy, and the Administration came to see that the federal government supported an ecosystem of programs for subsidizing R&D. A natural impulse was to understand that the ecosystem was more of a jungle than a clean, well-regulated place. There thus seemed to be some scope to impose some order on it. (A very nice book on this is Industrial Policy American Style, Richard Bingham 1998.)
The idea of imposing some order on a chaotic, largely decentralized system has some intuitive appeal. Centralization, rationalization—it sounds so sensible, but I’d suggest that chaotic, decentralized governance might be better adapted to harness chaotic, decentralized processes like the Schumpeterian gales of creative destruction—“Let a thousand flowers bloom!”—because one never knows where the good stuff is going to come from. Is a cadre of bureaucrats endeavoring to rationalize R&D processes really well equipped to know how to rationalize R&D processes?
But, wait. Didn’t the Manhattan Project demonstrate proof-of-concept? Did it not demonstrate that centralized management backed up by generous funding can generate fabulous advances in technology? Again, I suggest that a great hazard is that programs that throw money at R&D can become slush funds to Soviet science. I would further suggest, however, that “Manhattan Project” comprises a poor metaphor for how innovation unfolds. The Manhattan Project involved operationalizing a body of basic research. Others (Einstein, for example) had already done the basic research. The problem was to solve some (admittedly non-trivial) engineering problems to transform that basic knowledge into nuts-and-bolts know-how. This kind of late-stage R&D is the kind of thing most amenable to having money thrown at it. So, the authorities threw money at it and managed in just a few years to develop an operational technology.
Admittedly, there was some uncertainty that the Manhattan Project had to come to grips with. The main problem was to concentrate a mass of material in a deliverable package (a bomb) that could set off uncontrolled fission. Now, it’s one thing to get certain body of material (enriched Uranium) to support such an uncontrolled chain reaction, but, having started the chain reaction, could that chain reaction then extend to other matter? A concern was that setting off a fission process with enriched Uranium is one thing. But could that process set off uncontrolled fission of the atmosphere? That would be bad. So, perhaps a further bit of “basic research” might resolve that question—or, at least, impose more structure on that question?
We know the results. The difficult engineering problem involved inducing an uncontrolled chain reaction within a mass of enriched Uranium, and it didn’t set off a globe-spanning chain reaction in the atmosphere. Matter can be unstable (”radioactive”), but it also exhibits some degree of stability.
* * *
So, should the central government fund AI, or is our cadre of oligarchs already equipped to pour $billions into it? Is there really a problem of underinvestment, and, even if there were, would government subsidies really constitute a good remedy?
Further, would government subsidy programs not ultimately become slush funds that even a manager of USAID could envy?
Worth making the point that many are falsely comparing the marginal cost for deepseek's training run to the total cost for the US companies to produce a model. Which isn't to say that the narrative of doing good things at lower costs is completely wrong.
Also spot on with the sustainability and Lysenkoism analogy. The deluge of university slop with the magic word sustainability in it is enormous.