1. What is a DUV immersion Lithography Machine?

2. What is the strategic importance of DUV immersion Lithography Machines to China?

3. How much are U.S. export controls to blame for China’s pursuit of “self-sufficiency” in DUV immersion Lithography?

4. How likely is it that reports of Chinese progress in immersion DUV lithography are accurate?

5. How many immersion DUV machines can China make? How close does that bring China to its goal of self-sufficiency?

6. Are China’s immersion DUV machines competitive with those of ASML?

7. Can Chinese chipmakers produce advanced AI chips using ASML’s immersion DUV machines?

8. Can Chinese chipmakers produce advanced AI chips using locally-manufactured immersion DUV machines?

9. What is the U.S. government’s policy on China buying vs. locally producing immersion DUV lithography machines?

10. What Would the MATCH Act Do About This?

11. Conclusion: How big is China’s reported DUV semiconductor lithography breakthrough?

1. What is a DUV immersion lithography machine?

A modern chip factory (aka “fab”) requires many different kinds of extremely precise and complicated machines. The ultimate goal in every chip fab is to turn blank silicon wafer discs into processed wafers with many finished chips printed on top of each wafer. Lithography machines use light to “print” the circuit design onto the physical wafer. The features on modern circuit designs are measured in billionths of a meter (nanometers), and thus the machines must be extraordinarily precise.

The most advanced types of semiconductor lithography machines, called Extreme Ultraviolet (EUV) lithography machines, are almost unfathomably precise and have a credible case for being called the most advanced machines that humanity has ever made.

DUV immersion lithography machines are the generation of lithography machines that immediately preceded EUV machines. While there are different kinds of DUV machines, immersion DUV lithography machines are most accurately called Argon Fluoride immersion or “ArFi” lithography machines because of the argon fluoride chemical they use to generate the desired wavelength of light.

“Immersion” means a layer of ultrapure water sits between the final lens element and the wafer, raising the numerical aperture past the hard pre-immersion ceiling of 1.0. Explaining the meaning of numerical aperture is beyond the scope of this article, but suffice it to say that having a small pool of water touching both the lens and the portion of the wafer being printed upon meaningfully shrinks the size of the smallest features that the DUV lithography machines can print (improved resolution). The engineering enabling this is not merely a matter of “just add water.” Going from “dry” ArF to ArF immersion meant redesigning nearly everything inside the lithography machine.

Immersion DUV lithography was began testing in customer fabs in 2004. China likely imported its first immersion DUV lithography machine in 2008, likely from ASML. Despite being older and lower resolution than EUV machines, immersion DUV machines remain the workhorse of the industry and the best commercially available option for a wide range of chips — especially mature-node production, but also many non-critical layers on leading-edge chips. Immersion DUV machines remain commercially competitive and relevant for three major reasons:

• Immersion DUV machines are far cheaper than EUV machines (~$60 million vs. $200-400 million);

• EUV levels of resolution are unnecessary for making many types of chips that are still central to many commercial and industrial applications; and

• Even on leading edge chips, not all layers require EUV levels of resolution. On layers that DUV can pattern in a single exposure, DUV offers higher throughput and lower cost per wafer — though on leading-edge layers requiring multi-patterning, that advantage reverses. Common practice for production on leading edge chips still uses EUV on only a minority of the layers.

The dominant global supplier of all types of semiconductor lithography systems is the Dutch company ASML, which is the monopoly supplier of EUV lithography machines, the near-monopoly supplier of immersion DUV lithography machines, and leading provider of nearly all the major older lithography machine types.

The different generations of lithography technology from the past 40 years are summarized in the table below. Even the older models of machines that ASML no longer manufactures continue to be used commercially as part of a vibrant resale and refurbishment market. ASML estimates that more than 90% of even its old models of I-line lithography machines are still in operational use somewhere in the world.

The semiconductor industry groups chips by manufacturing technology node, which names a certain era of manufacturing process and characterizes its performance and technical sophistication. The below timeline maps the logic chip nodes (using the TSMC naming conventions) since 90nm based on the type of lithography they used and the iPhones whose core application processor chips used those manufacturing processes. Note: Apple used Samsung as its iPhone chip manufacturer exclusively until 2014, and Samsung’s naming convention for each node generation differs somewhat from TSMC’s.

The key takeaway from this chart is that immersion DUV lithography technology was industry-leading from 2006 to 2018 and remains not only commercially relevant but essential even at the leading edge today.

2. What is the strategic importance of DUV Immersion Lithography Machines to China?

Semiconductors are a strategic industry and especially so for China. The analogy that “chips are the new oil” has its limits, but it is helpful to illustrate just how deeply woven semiconductors are into every aspect of the economy and how painful it is to lose reliable access to semiconductors. As one recent real-world example: the massive COVID-19 semiconductor supply chain disruption placed the economic importance of semiconductors in a spotlight reminiscent of an oil embargo. Analysis by the U.S. Department of Commerce found that “the [chip] shortage shaved an estimated $240 billion off U.S. GDP in 2021.” In other words, U.S. GDP was a full one percent lower than it would have otherwise been had the semiconductor shortage never occurred.

China is arguably the country most economically exposed to a loss of reliable access to semiconductors. By dollar value, China imports more semiconductors annually than it does oil, mostly as inputs for manufacturing products that are re-exported. China has long sought to increase its “self-reliance” in semiconductors by eliminating the need for foreign chip imports. This extends not just to making the chips that China’s economy needs, but also to making the machines that make those chips.

China’s official five-year plan adopted in October 2020 declared the semiconductor industry to be a top technology priority. The official summary of the CCP’s Central Committee meeting adopting the five-year plan was blunt: it stated that China would “make technological self-sufficiency a strategic pillar of national development.” At the time, Chinese policy recognized that the Chinese semiconductor equipment industry was not sufficiently technologically advanced to end China’s reliance on imports. But, China viewed equipment imports as a temporary waypoint on the path to technological self-sufficiency.

3. How much are U.S. export controls to blame for China’s pursuit of “self-sufficiency” in DUV immersion lithography?

While it would be foolish to say that U.S. and allied semiconductor and semiconductor equipment export controls have had no impact on Chinese strategic thinking and planning, China’s goals for self-sufficiency long predate the modern era of U.S. semiconductor export controls that began in 2018 during the first Trump administration.

Three years earlier, in 2015, China announced its goals for self-sufficiency in semiconductors and semiconductor manufacturing equipment with its Made in China 2025 strategy. These goals specifically targeted:

• Domestic production of immersion DUV lithography machines before 2025

• Domestic production of EUV lithography equipment before 2030.

The roadmap document stated that

“meeting domestic market demand, improving the self-sufficiency rate of integrated circuit products, meeting national security needs, and occupying the strategic product market have always been the greatest demand and driving force for the development of the integrated circuit industry.” (emphasis added)

Even earlier policies, though less well resourced, sought to dramatically reduce use of foreign semiconductors and semiconductor manufacturing equipment. For example, the 2006 “Medium Long Range Plan for the Development of Science and Technology” explicitly called for self-sufficiency in semiconductor technologies and initiated so-called “mega projects” to drive toward that goal.

An earlier analysis by Decision Tree Research evaluated data on Chinese government semiconductor subsidies and equity investments and found no compelling evidence that U.S. and allied export controls caused China to increase its already sky-high level of government investment in semiconductor self-sufficiency.

4. How likely is it that reports of Chinese progress in immersion DUV lithography are accurate?

In general, China’s government and Chinese media like to boast and even exaggerate progress towards semiconductor self-sufficiency. It is therefore a bit odd that news of this technical achievement entered the public discourse via anonymous sources talking to journalists as opposed to an announcement from the company itself or Chinese government-run media. The Global Times, a state-run newspaper, published an editorial on July 28^th that gloated about the decline in ASML and related stock prices, but still described the report of local Chinese immersion DUV production as “unconfirmed.”

The official silence is especially odd against the backdrop of a documented pattern of Chinese actors overstating semiconductor progress to make U.S. export controls look futile. Announcements are often timed for signaling value, as when Huawei’s Mate 60 Pro launch coincided with then-Commerce Secretary Gina Raimondo’s visit to Beijing in August of 2023. And indeed the Global Times did take the opportunity to attack export controls, writing,

“That an unconfirmed report could trigger such violent market volatility illustrates current paranoia of some forces in the West regarding China’s technological rise.

The evaporation of massive market value shows that Western investors’ blind faith in the effectiveness of “blockades” has been shaken, and the US-led “decoupling” strategy has lost much of its momentum. Though they may not admit it openly, they know in their hearts that a Chinese breakthrough in “chokepoint” technologies is merely a matter of time […] Even Western tech companies like ASML understand that once China achieves a genuine breakthrough, their market share will face a direct hit.”

However, given The Information’s exemplary track record in its previous reporting of chip smuggling and other stories China was not keen to make public, it’s unlikely that this is disinformation. But if the reports are accurate, why the muted announcement?

The answer likely lies in this key excerpt from The Information article: “DUV machines have traditionally relied on a global network of specialized suppliers. The Chinese DUV uses mostly domestic components, though some key parts come from Japan.”

China most likely did not publicize this story because it also highlights the country’s continued reliance on foreign inputs for DUV machines, meaning that rather than building a fully “Chinese” immersion DUV machine, Aishengna seems to be taking critical and, at least for now, irreplaceable Japanese components and assembling them into a mostly “Chinese” DUV machine. Taken against the backdrop of self-sufficiency to which China aspires, the key question is whether China is truly developing a fully Chinese machine from Chinese parts, or simply cutting one foreign company (ASML) out of the equipment supply chain while remaining reliant on critical component suppliers in Japan (and perhaps elsewhere).

No high-quality reports have thus far emerged regarding what share of the machines are made locally, but Chinese online industry commentators have claimed that the predecessor machine that reportedly began trials at SMIC last year had between 70% and 85% localization of parts. Once identified, any foreign suppliers to Aishengna would be obvious targets for export control restrictions and enforcement. That’s plenty of reason for China to try to downplay such progress or keep it a secret.

5. How many immersion DUV machines can China make? How close does that bring China to its goal of self-sufficiency?

Two considerations are important here: quantity and quality.

On the quantity side, The Information article reported that Aishengna will make 5 immersion DUV machines in 2026 and expects to make 20 in 2027. These output metrics are consistent with the pace at which ASML scaled its own immersion DUV production during the 2003–2005 validation phase, during which ASML deployed 13 immersion machines for commercial validation. ASML shipped 131 ArFi systems in 2025 alone, and the overall count of ASML’s deployed immersion DUV systems worldwide is likely more than 1,400, based on ASML’s annual reports. From 2023 to 2025, Chinese companies purchased roughly 230 immersion DUV lithography systems from ASML. So while Aishengna’s current capabilities align with ASML’s early production scaling, they are still nowhere near the overall capabilities of the incumbent.

The House China Select Committee published a report that included data on what share of ASML’s immersion lithography machines went to China for the years 2021-2024. These are summarized in the below table:

While Aishengna’s expected sales next year of 20 units are small relative to ASML’s 2025 production and tiny relative to ASML’s global install base of machines, they do show a potential path toward meaningfully increased DUV lithography self-sufficiency for China. To begin, a significant share of China’s recent purchases of immersion DUV machines are likely future demand pulled forward in expectation of future export controls. In other words, Chinese firms have been buying as many machines as they can during the window of availability that is feared to be closing. By 2027 China’s annual demand for immersion DUV lithography machines may be considerably below the 90 machines purchased in each of the years 2024 and 2025, perhaps closer to the 21 machines purchased in 2022.

But ASML also shows how fast production can ramp: once production of the ArF immersion machines got underway, ASML was making 93 units annually only five years later.

2005: 13
2006: 23
2007: 38
2008: 56
2009: 31
2010: 93

Moreover, dramatically increasing China’s stock of immersion DUV machines is likely of less strategic importance to China than ensuring that Chinese firms can keep their existing fleet of machines operating and producing chips. This requires replacing parts that wear out or otherwise fail. For example, only two companies are known to produce ArF immersion light sources with sufficient power for the 20nm node or better: Gigaphoton of Japan and Cymer, an ASML subsidiary based in the United States. The consumable modules in these light sources degrade with usage and require replacement every year or so.

Neither The Information nor the Reuters article included information regarding whether or not the Aishengna machines are copies of ASML machines. However, industry participants have said that Chinese semiconductor manufacturing equipment designs are typically near-copies of foreign machines. If true in this case, Aishengna’s ability to successfully manufacture integrated machines could imply that it is also capable of manufacturing a meaningful share of the spare parts required to keep the hundreds of immersion DUV machines already in China running. Two exceptions would apply: (1) instances where Aishengna is still dependent upon foreign subsystem and component suppliers even for its own machine and (2) instances where Aishengna’s equivalent components cannot match the performance or reliability of those used in ASML machines.

6. Are China’s immersion DUV machines competitive with those of ASML?

This brings us to the quality part of the story. A machine that frequently produces defects cannot make very many functional chips, much less be economically competitive.

Performance characteristics like yield, throughput, and downtime determine the true return on investment for lithography machines and their ability to be productive. JP Morgan analysts correctly observed on July 28 that “producing a handful of immersion DUV tools is not the same as producing tools that can be used for high-volume manufacturing, where yield, overlay, throughput and reliability over thousands of wafer runs are what matter.”

Though immersion DUV lithography is a 20-year-old technology, ASML has been refining and improving its designs for that entire period. The reports from The Information and Reuters did not clarify whether Aishengna machines approach the quality of ASML’s best DUV machines, or whether they are more closely comparable to what ASML was producing 15-20 years ago. If the latter is true, it will undoubtedly not take Aishengna 15-20 years to catch up or come close—replicating is almost always easier than innovating. However, China could still be many years away from meaningfully reducing its true dependence on ASML in both new machines and spare parts.

There is little reporting thus far on the performance of the Aishengna machine, but the corporate history of Aishengna is intimately tied up with that of three other Chinese firms: Shanghai Micro Electronics Equipment (SMEE), Shanghai Yuliangsheng, and SiCarrier [a Huawei-linked firm]. According to reporting by Asia Times,

“Yuliangsheng, SMEE and Aishengna are actually sharing the same team of engineers, who are developing a machine similar to SMEE’s SSA800, which has not yet been officially launched.

Yuliangsheng was founded in 2022 and is jointly owned by SiCarrier and the state-backed Chuangkewei (Shanghai) Technology. Last year, SMEE underwent a restructuring that moved its immersion DUV lithography technologies and engineers into Yuliangsheng, while it kept its extreme ultraviolet (EUV) lithography project.

Yuliangsheng’s team is now housed within Aishengna and continues developing the SSA800, though the machine could be renamed to avoid ownership disputes if Yuliangsheng or Aishengna eventually goes public.”

This strongly suggests that the Aishengna machine is actually the direct successor to SMEE/Yuliangsheng’s SSA800, about which there was some noteworthy reporting last year when a Yuliangsheng (which Asia Times claims took over the SMEE immersion DUV project) tool reportedly began trials at the Chinese chipmaker SMIC. The Financial Times described it as a “28 nanometre Chinese-made DUV lithography machine” capable of “utilizing so-called multi-patterning techniques to produce 7nm chips.”

However, Asia Times reported that the SMEE SSA800 immersion DUV machine relied on imports to source components, “including a 193 nm excimer laser mirror, Zeiss lenses, a vacuum chamber, synchronization control algorithm software and argon fluoride (ArF) immersion light sources.” That suggests that local Chinese producers are not yet capable of making many of the key components that would be needed for China to achieve independence in either machines or spare parts.

Moreover, in May 2026, the Chinese semiconductor trade press outlet ICSmart (芯智讯) debunked many of the claims surrounding the yield performance of SMEE’s SSA800 and the status of its acceptance by local Chinese chipmakers. However, the ICSmart report did claim to have confirmed some reported performance details as accurate, stating that the SSA800

“is a 193nm ArF immersion lithography machine, comparable to the ASML TWINSCAN NXT:1950i. The [1950i] has a dedicated chuck overlay accuracy of 2.5nm and a hybrid matching overlay accuracy of 3.5nm, and can process 200 wafers per hour. […] the SSA800 series has an overlay accuracy controlled between 2.3 and 2.5nm and can process 150 wafers per hour.”

7. Can Chinese chipmakers produce advanced AI chips using ASML’s immersion DUV machines?

If “advanced” means comparable in performance to the latest and greatest Nvidia chips, then the answer is clearly no. If “advanced” means “useful to some degree for modern AI applications,” then the answer is yes. Using ASML immersion DUV machines, SMIC has demonstrated the ability to manufacture Huawei AI chips at the SMIC equivalent to the TSMC 7nm (aka N7) node. TSMC first began running the 7nm node in 2018, also using immersion DUV. While Apple began using that node as soon as it became available and moved on to TSMC’s 5nm (aka N5) node in 2020, Nvidia was still using the TSMC 7nm node for its A100 data center chips when they were introduced in 2020. The successor generation of Hopper chips, which used TSMC’s custom 4N node (confusingly, part of TSMC’s 5nm process family), went on sale in late 2022.

In simple terms, each new generation of AI chips brings two sources of performance improvement: a superior manufacturing process node—which allows more transistors to be crammed into a given chip wafer area—and improved chip design, in which designers make better use of that manufacturing process for the applications their chip is intended to serve. When effectively enforced, U.S. export controls cut Chinese chip designers off from most of the post-2020 process node performance improvements. However, they can still improve their designs within that constraint. This means that Chinese 7nm chips can be superior to what Nvidia’s 7nm era chips offered because Chinese chips can incorporate some of the AI-chip optimization design lessons that the industry has learned since 2020.

Increasingly, however, AI performance depends not just on the logic chip, but also on the surrounding high-bandwidth memory and advanced packaging for the full module, both of which are key bottlenecks for China. China also faces major limitations in the number of AI chips that it can produce with its existing inventory of advanced machines and the amount of AI-relevant computation that these chips will deliver.

8. Can Chinese chipmakers produce advanced AI chips using Aishengna’s immersion DUV machines?

The above relates to what China can accomplish with ASML’s immersion DUV machines. The remaining question is how much China can accomplish with the machines provided by Aishengna. As mentioned above, Chinese sources describe the local machine as comparable to the ASML TWINSCAN NXT:1950i, which is an older tool than SMIC currently uses for its 7nm process. Immersion DUV tools can only reach below the 28nm node by using multi-patterning, where the binding constraint is the overlay budget. An overlay of 2.5nm, which the 1950i has, is generally not good enough to enable quadruple patterning for 7nm-class critical layers. 7nm-class critical layers run on ~1.6nm dedicated-chuck-overlay tools (ASML’s NXT:1980Di class, the machines SMIC uses or better). Even the SSA800’s claimed 2.3nm is nearly a nanometer short. This suggests that, without further improvements, the first generation of Aishengna immersion DUV machines will (in the best case for China) match 1950i performance and thus not be useful for making AI chips. More likely, they will be used to make legacy node chips, to improve Aishengna’s rate of learning in partnership with Chinese chipmakers, and to refine future lithography machine designs that will be more relevant to the AI sector.

To make a dent in the U.S.–China AI race, China needs to produce both more lithography machines and better lithography machines. In the absence of direct insider information about the trajectory of the Chinese program, we can look to the precedent of ASML to get a sense of what is possible. Presumably China will reach the milestones ASML hit 18 years ago faster than ASML originally did. After all, China has many significant advantages:

• Access to many copies of working ASML machines for reverse engineering

• Former ASML technicians and engineers now working for Chinese companies

• Access to working foreign components and subsystems to fill local gaps in supply chain availability and performance

• Massive government financial support

• Help from government cyberespionage and industrial espionage

9. What is the U.S. government’s policy on China buying vs. locally producing immersion DUV lithography machines?

Since the October 7, 2022 export controls, the U.S. government has had a clear policy against U.S. companies and U.S. persons supporting the advancement of the Chinese semiconductor manufacturing equipment sector in any way. This included adding an additional 140 firms to the entity list in December 2024, of which 20 were Chinese semiconductor equipment, materials, and component makers.

At the same time, the U.S. government has allowed the export of DUV immersion lithography machines to China provided that their performance (including chuck overlay) is not above a certain threshold. The idea was that allowing continued sales of such “legacy” machines would reduce China’s interest in producing local equipment alternatives while contributing little to China’s ability to produce AI-relevant chips. The original reasoning behind this is not obviously unsound, but it does appear that the United States has allowed China to massively stockpile these machines (and thus increase China’s global share of chipmaking) without diminishing China’s commitment to locally producing lithography machines. Embarrassingly, the Financial Times reported in December 2025 that Chinese fabs were successfully importing in components that allowed them to upgrade their ASML lithography machines to boost performance past the levels allowed by export contols.

At the same time, the U.S. government has been, by its own admission, unable to fully align the export controls posture of other key countries in the semiconductor value chain. Under Secretary of Commerce Jeffrey Kessler admitted this explicitly in his Congressional testimony last month. The fact that the United States is attempting to block sales of key categories of semiconductor manufacturing equipment while allowing key allies to continue selling key components and subsystems to the Chinese companies working to defeat these export controls is absurd.

The second Trump administration, for its part, suggested in its July 2025 AI Action Plan that it would end this contradiction. Specifically, it stated that

“The United States must also prevent our adversaries from using our innovations to their own ends in ways that undermine our national security. This requires new measures to address gaps in semiconductor manufacturing export controls, coupled with enhanced enforcement. […] America must impose strong export controls on sensitive technologies. We should encourage partners and allies to follow U.S. controls, and not backfill.”

To plug loopholes in existing semiconductor manufacturing export controls, the Plan called for policy actions such as:

• Develop new export controls on semiconductor manufacturing sub-systems. Currently, the United States and its allies impose export controls on major systems necessary for semiconductor manufacturing, but do not control many of the component sub-systems.

• Coordinate with allies to ensure that they adopt U.S. export controls and work together with the U.S. to develop new controls.

These prescribed policy actions directly address the regulatory disconnect between the U.S. and its allies that has accelerated China’s effort to develop local immersion DUV machines, and the lack of progress on them is what Representative Young Kim pointed out to Under Secretary Kessler in the Committee hearing. The “incomplete cooperation” Kessler describes is disappointing given that this Trump administration has not shied away from making requests of allies for other issues, from increasing defense spending to negotiating new trade deals. It is surprising that the Trump administration has not been more forceful and direct in aligning export controls on such critical inputs for strategically relevant technologies.

10. What Would the MATCH Act Do About This?

The Multilateral Alignment of Technology Controls on Hardware (MATCH) Act is a piece of bipartisan legislation that has been introduced in both the House and the Senate to codify export controls on covered semiconductors and semiconductor manufacturing equipment. One way of understanding the MATCH Act is that it would fill the gap between what the Trump administration said that it wanted to do in the AI Action Plan and what it has actually done as described in Under Secretary Kessler’s recent testimony. It would dramatically expand country-wide prohibitions on semiconductor manufacturing equipment. The bill’s text explicitly includes “all deep ultraviolet immersion photolithography machines” in its definition of “Covered Semiconductor Manufacturing Equipment.” The Act would also designate all subsidiaries, affiliates, and chipmaking facilities run by the most technologically sophisticated Chinese chipmakers (CXMT, Hua Hong, Huawei, SMIC, and YMTC) as being covered by the entity list. It would also give allies 150 days to adopt equivalent controls before the application of the Foreign Direct Product Rule, effectively making foreign exports of covered machines, components, and subsystems to China illegal under U.S. law. Finally, it would treat servicing of installed equipment as a first-class control, meaning that ASML and other equipment vendors could no longer repair or provide upgrades to existing advanced equipment in China.

11. Conclusion: How Big Is China’s Reported DUV Semiconductor Lithography Breakthrough?

Smaller than the one-day drop in ASML’s share price implied, but bigger than China’s official silence suggests.

The milestone itself appears genuine. Twenty-two years after immersion DUV lithography first entered commercial production, a Chinese firm has reportedly begun volume production of a domestic equivalent — and measured against China’s own stated ambitions, it is arriving roughly on schedule. Made in China 2025 called for domestic production of immersion DUV machines before 2025. Aishengna appears to have delivered about a year late.

Measured against self-sufficiency, however — the goal Beijing actually cares about — the breakthrough is modest on every dimension that matters.

On quantity, Aishengna’s expected 20 machines in 2027 compare to the 131 immersion systems ASML shipped in 2025 alone and a global installed base of more than 1,400.

On quality, the machine’s claimed performance is comparable to what ASML was selling 18 years ago, with overlay accuracy that falls short of what 7nm-class AI chip production requires. And the performance data that matters most, defect rates and reliability once the machines are running in actual fab production, does not yet exist in the public record. If the machines prove unreliable, the true story may be less that Chinese chipmakers are eagerly buying Aishengna’s products and more that the Chinese government pressed chipmakers to accept them, possibly at no cost, to help Aishengna mature.

And on independence, the machine still depends on critical imported components that China cannot yet make. China has not built an ASML rival. It has cut ASML out of final assembly while remaining reliant on the foreign suppliers beneath it. That is a huge gap in the existing U.S. and allied export controls architecture, which has always underestimated the importance of key components.

But judging this development by its current level would repeat the error that export control skeptics make about Chinese chipmaking: the right measure is trajectory, not position. A firm that can integrate a working immersion scanner can likely also manufacture a growing share of the spare parts that keep China’s existing fleet of hundreds of ASML machines running — arguably the more strategically important capability. And with working machines to study, poached engineers, imported components, and massive state financial support, China should be expected to climb ASML’s learning curve considerably faster than ASML did. The binding constraint on that climb is now components rather than finished machines — which is precisely where U.S. and allied export controls remain weakest, and precisely the gap the AI Action Plan promised and the MATCH Act would actually close.

For the past several years, China has been buying everything it can from the store before it closes. The Aishengna reports show something new: China is learning to run a store of its own. For now, its shelves hold last decade’s models, stocked partly with imported parts. The task for U.S. policy is to keep it that way for as long as possible.

So, it’s a bit disappointing that I practically never hear anyone mention Kurt Vonnegut’s 1952 debut novel Player Piano when we talk about AI taking our jobs. Player Piano is not a masterpiece of the same caliber as Vonnegut’s two best-known works, Slaughterhouse-Five and Cat’s Cradle, but the novel is a fun, funny, and occasionally insightful book about a dystopian future in which widespread mechanization has destroyed effectively all demand for working-class labor. Only the engineers and managers who design and repair the machines can enjoy steady employment.

The book really has quite a lot to offer:

• Technology-driven unemployment and the resulting loss of human dignity;

• Genius engineers and other economic elites blaming the victims of automation only to later see themselves automated; and

• Make-work government jobs programs and guaranteed material security (sort of like universal basic income) as necessary policy outcomes.

Beyond its virtues as a novel, Player Piano is also a helpful reminder that concerns about automation and its effect on unemployment have been with us almost uninterrupted since the violent crushing of the machine-breaking Luddite rebellion in the early 1800s. Even 20^th century American politicians like John F. Kennedy—whose Apollo moon program led to phenomenal growth in science and engineering—acknowledged that accelerating automation was a policy problem.

During his 1960 campaign for president, Kennedy gave a speech at a major labor union convention that, with some light tweaking, would work nearly perfectly as a speech for a politician today. Here’s how the speech opens:

“Today we stand on the threshold of a new industrial revolution – the revolution of automation. This is a revolution bright with the hope of a new prosperity for labor and a new abundance for America – but it is also a revolution which carries the dark menace of industrial dislocation, increasing unemployment, and deepening poverty.

Already entire automobile engines are being manufactured, untouched by human hands. Modern lathes and drills are turning out parts machined to the closest tolerances, guided only by electrical impulses which make the settings and automatically correct all errors. Electronic equipment is sorting material as it enters a warehouse and carrying it – without human guidance – to its proper place of storage. And in the future, as the complexity, the versatility, and the precision of modern technology continues its inevitable advance, thousands of processes and functions now performed by men will be done, more cheaply and more efficiently, by machine.”

The full speech is a fascinating, eloquent artifact. And just as Kennedy’s framing of the problems rings familiar, so too do his prescriptions: government aid to encourage companies to adopt machines that augment workers rather than replace them, combined with welfare benefits and retraining for displaced workers.

Between Vonnegut and Kennedy, one could be forgiven for thinking that there’s nothing new under the sun when it comes to the debate on automation and labor. AI is just the mechanical loom, steam engine, tractor, and electronic equipment debates all over again.

At least in terms of the public, mass-media debate on these issues, I think this is correct. But that’s only because the public debate on automation and labor remains stubbornly stuck at the Econ 101 level. What I’m going to cover next are the four case studies and six economic intuitions that most helped me to get past the 101 level. Each of the case studies—agriculture, horses, music, and chess—offers illuminating and instructive lessons for the changing shape of the labor market in the AI era.

What actually happens will, of course, depend on the rate of technological progress and adoption on a sector-by-sector basis. What these cases give you is a way to see what would have to be true for each AI scenario to be plausible. I have deliberately left out policy recommendations. That will be the subject of a future essay.

I’ll begin with agriculture, unpacked in two parts.

Agriculture Part 1: Sectoral labor demand vs. overall labor demand

Why automating jobs doesn’t have to increase unemployment

In 1900, 41% of the American labor force (11.9 million people) was employed in agriculture. Seventy years later, in 1970, only 4% of the labor force (3.1 million people) was employed in agriculture. That is a 10-fold reduction in labor share at a time when the number of mouths to feed (the overall American population) nearly tripled from about 75 million to about 205 million. Despite using far fewer human workers, the agricultural sector produced far more food and other outputs. Over that same period, the number of operational tractors on American farms increased from basically zero to 4.7 million.

The increased productivity story comes down to three main factors:

1) Automation: Growing adoption of tractors (and later mechanical pickers) which increased in both number and per-tractor power. Tractors were not only stronger and faster than the animals and men they replaced: they also freed up the significant share of land on most farms reserved for growing food to feed working horses and mules;

2) Sector reorganization: Consolidation of many small farms into fewer, bigger, and more efficient farms; and

3) Introduction of modern farming techniques: Improvements in fertilizers, pesticides, crop strains, and irrigation.

Combined, these delivered astonishing gains in U.S. agricultural productivity, providing more food, more cheaply. That’s the good news.

The bad news is that the agricultural workforce shrank by nearly 9 million.

A 1937 Dorothea Lange Photo captioned by the Department of Agriculture. The caption reads “Texas tenant farmers who have been displaced from their land by tractor farming.”

The most remarkable part of the story, however, is that overall American unemployment was basically unaffected. Those 9 million workers successfully shifted from jobs in agriculture to jobs in the booming industrial and service sectors.

These two trends—booming farm productivity and rising non-farm labor demand—aren’t a coincidence: they’re the same trend seen from opposite ends. When mechanization made food vastly cheaper, American households didn’t respond by eating vastly more. There’s only so much a stomach can hold. Instead, Americans spent the savings on everything else, first manufactured goods and later services. The shift is stark in the data: in 1901 Americans spent 42.5% of their expenditures on food; by 1970, it was 19.3%. Every percentage point that fell off the national food bill reappeared as demand — and jobs — somewhere else.

Key insight for AI: automation can drive job losses in a specific sector without meaningfully contributing to overall unemployment.

Over the past year there has been a wave of companies, mostly in software, announcing layoffs as part of a strategy of replacing workers with AI. At the same time, economists are scouring every new batch of unemployment data looking for evidence that AI is driving job loss in specific sectors. The early 20^th century story of farm automation, however, tells us that even if a productivity-enhancing technology does drive unemployment in a specific sector or job category, the savings from that productivity still cycle back into the overall economy and labor demand. AI might be short-term painful for specifically affected groups of workers or geographies, but we should not confuse that with evidence that AI is driving up long-term unemployment.

Agriculture Part 2: Demand Elasticity

How much more can you eat?

There’s a key aspect of the early 20^th century tractor revolution in the U.S. agriculture sector—and particularly the food part of the agriculture sector—that I only touched on briefly in the first section: there’s only so much a stomach can hold. In more general terms, an increase in a sector’s supply potential does not automatically mean that there will be adequate demand to absorb the increased production.

Conceptually, a big increase in per-worker productivity can be “spent” in two ways:

1. Fewer workers to produce the same (or greater) output

2. More production with the same (or greater) number of workers

20^th century food production falls into the first category, but the second story has also played out in the U.S. agricultural sector. It happened more than two centuries ago with the invention of the cotton gin by Eli Whitney in 1793.

What is a cotton gin? It’s the machine that separates cotton seeds from cotton fiber. The reason why the cotton plant grows those white tufty balls of cotton fiber is to catch the wind and disperse the plant’s seeds. Dandelion flowers evolved a different version of the same trick. But cotton seeds are much bigger and tougher than dandelion seeds, so the seeds have to be removed from raw cotton before the fiber can be processed into textiles.

Before the invention of the cotton gin, separating the seeds from the tightly attached fibers of short-staple cotton (the kind that grows in the climate of the American South) was a slow, labor-intensive affair. A single worker could typically remove the seeds from only 1 pound of raw cotton per day. The costs were high enough that growing cotton in the South was, for the most part, economically non-viable, even with slave labor. Whitney’s cotton gin increased the per-worker processing output to 50 pounds per day, a 50-fold improvement in productivity. But the cotton gin did nothing to automate growing and picking, which stayed entirely manual. Picking would continue to require the dexterous human hand (a huge human advantage!) for another 150 years.

The cotton gin smashed the deseeding bottleneck while making every pound of picked cotton far more valuable. It also made cultivation of Southern short-staple cotton suddenly profitable, opening an enormous new growing region. The massive decrease in the cost of a complementary service (deseeding) drove an explosion in demand for the field labor the gin didn’t replace. Specifically, the gin increased demand for enslaved labor.

It is no exaggeration to say that the cotton gin changed the course of history for slavery in America. Part of the reason why some of America’s founding fathers thought (perhaps naively) that slavery would die out naturally was that southern agricultural slavery wasn’t actually all that profitable and faced unclear future prospects. Chesapeake tobacco was exhausting soils and facing sustained low prices, such that many farms were switching to growing far less labor-intensive crops like wheat and sometimes freeing slaves in the process.

The cotton gin changed all of that. Labor was now desperately needed.

So why did 20^th century tractors reduce demand for farm labor while 19^th century cotton gins increased demand for farm labor? The explanation lies not on the supply side, but on the demand side. In response to increased availability of cheaper food, people can eat more, but only to a certain point. An adequately fed person can perhaps double the amount that they eat per year, but they cannot increase the amount that they eat by 50-fold. Their stomachs would explode.

By contrast, consumers (including those outside the United States) could and did increase their consumption of cotton by more than 50-fold. In fact, U.S. cotton production increased 1000-fold from about 3,000 bales in 1790 (nearly all of which was long-staple cotton that wouldn’t grow in the Deep South) to 3.8 million bales in 1860—all of it happily purchased to meet near-insatiable demand. The short-staple cotton crop that wasn’t worth growing before 1793 became the dominant fiber of the world.

How? The American cotton gin (combined with a near-simultaneous British revolution in mechanized spinning) caused one of the steepest price collapses in economic history. The price of one pound of cotton yarn fell by 95% between 1784 and 1832. Quite rapidly, cotton went from a semi-luxury to the cheapest fabric on earth. As a result, cotton displaced wool and linen as the everyday fabric for most people.

Before the cotton gin, wool and linen were expensive but the only real options for most clothing. Common people might own only a few sets of clothing or perhaps just one. After the cotton gin, the common wardrobe went from sparse, durable, and rarely washed wool and linen to abundant, varied, and laundered cotton. You can guess what happened to employment in the English wool industry.

Average Americans could suddenly afford to enjoy the thrills of following fashion, buying new clothes to match changing tastes. Moreover, household textiles that had previously been luxuries—bedsheets, curtains, towels, tablecloths—became affordable cotton goods. Finally, cotton expanded to industrial and other uses: thread, bags, and eventually bandages.

This is the essence of the economic concept of demand elasticity. How much does market demand change in response to a change in supply and price? Food demand was relatively inelastic: a big decrease in prices didn’t cause a big shift in how much people ate. By contrast, cotton demand was highly elastic: the 20-fold drop in cotton prices caused a far, far greater increase in cotton demand.

Cotton growers didn’t just expand production with the existing labor pool: they also increased the size of the labor pool (almost entirely slaves) and acquired more land for cotton cultivation. Production exploded.

As a side note, the Southern short-staple cotton industry was not profitable even with slave labor before the cotton gin, and slavery wasn’t necessary to make cotton profitable after the invention of the cotton gin. Within fifteen years of abolition, the South was growing and exporting more cotton than it had under slavery. What slavery provided was not viability but coerced, below-market-rate labor and thus extra profit for landowners.

Key insight for AI: whether automation leads to job losses or job growth in a sector depends heavily on that sector’s demand elasticity.

Put simply, how much more will consumers buy if prices go down? We appear to already be experiencing a version of this in the software sector. Today’s AI-coding systems are excellent at producing many kinds of computer code, but demand for some kinds of software engineers (especially at companies making AI) is increasing.

The reason is that demand for software appears to be highly elastic. People can’t increase how much they eat by 50-fold, but consumers, businesses, and the government might be happy consuming 50 times more software if creating and changing custom software is ridiculously cheap and convenient. And as long as humans are still needed for some part of creating software—at least for now—overall demand for software engineers might go up, not down.

Human software engineers can move to higher-level tasks (design, planning, oversight) while machines handle churning out the raw code. The data on this is not yet crystal clear, but anecdotes about how experienced software engineers are in higher demand than ever while entry-level software job openings have cratered are consistent with this interpretation. The challenges of that potentially bifurcated outcome will be addressed in the Music and Chess case studies below.

Horses Part 1: The Limits of Retraining and Labor Mobility

Horses can’t work in hospitals

In the first section, we saw that the introduction of the tractor led human farm employment to crash. Despite major sectoral pain and disruption, there was no aggregate impact on human unemployment. But the key word in that claim turns out to be “human.” There was a different farm labor population for whom the introduction of the tractor did lead to large population-level increases in permanent unemployment: horses and mules.

Today it might seem silly, but in the 19^th and early 20^th centuries, economists had no trouble thinking of horses and mules as part of the labor force. In 1915, the American economy still ran on hooves. The United States was home to roughly 26 million horses and mules, about one for every four people, and they were not pets. They pulled the plows, hauled the freight, pumped the water, and powered the streetcars. The federal government counted them the way it counted any other input to production: the Census Bureau tracked the nation’s working horses and mules, herd by herd, year after year, until 1960. They stopped because horses and mules stopped being economically important.

When the Bureau stopped counting, about 90% of that population was gone (dead, not merely unemployed) — displaced not by a war or a plague but by a machine.

Why didn’t the horses follow their human coworkers into the factories and the offices? In short, because they had nothing to sell that employers wanted. A horse offers physical power, not mental power, and by the 1920s, machines offered more power, more cheaply, and more precisely in every task a horse could do. And for the parts of the economy where machines weren’t yet superior in every way to horses, humans were. A horse can’t run a drill press (no dexterous hands), can’t converse with customers (no vocal cords), and can’t handle bookkeeping (no big brain). When the machine beat the animal on both price and performance, the demand curve for horse labor didn’t shift to some new sector. It fell off a cliff.

According to economic theory, in a functioning market there is always some wage low enough that a worker gets hired. That is the logic behind comparative advantage, and it’s one of the standard reasons economists insist automation can’t cause permanent mass unemployment: even if machines have an absolute advantage and are better at everything, humans can accept jobs at lower and lower wages until employing them somewhere is still worthwhile, and the market clears. The horse example provides the asterisk on that theorem. There was, in fact, a theoretical wage at which a horse was still worth employing. The problem was that this theoretical wage had fallen below the cost of feeding and stabling the horse. Comparative advantage guarantees that some theoretical wage exists; it does not guarantee that the price covers the cost of staying alive. When it doesn’t, you don’t get unemployment. You get the glue factory.

And the glue factory is roughly what happened. The surplus horses were slaughtered — for meat, for pet food (horse meat was a staple of canned dog food well into the 20th century), and yes, rendered into glue and fertilizer. The rest simply weren’t allowed to have offspring. A working population of 26 million was culled to about 3 million in 45 years.

Key insight for AI: Even if increasing automation in one area of the economy leads to increased demand for labor in other areas, there is no guarantee that the automation-displaced workforce will be able to meet the new demand.

Much of the policy conversation on AI job displacement focuses on retraining displaced workers for new jobs. According to some reporting on this, a cohort of AI-spooked Gen Z students is forsaking college and its weakening promise of post-graduation knowledge work to learn blue-collar trades (plumbing, carpentry, welding, etc.) that appear unlikely to be automated anytime soon. This may be a perfectly reasonable approach to handling an AI transition period in which humans still have employable strengths in areas where machines still have weaknesses. However, horses demand that we confront the harder case: what if at some point AI enables machines that are legitimately better and cheaper at everything that humans can do?

There are at least two answers to this question. First, if AI leads to a total collapse in employable labor and wages, humans might still have other potential sources of non-wage income, including from returns to owned capital (horses can’t own property) and from government redistribution (horses didn’t have a large voting population calling for new social safety nets). The second answer is covered in the next section.

Horses Part 2: The Relational Economy

Why mules had it much worse than horses

There’s a twist that should give an AI-anxious human a little comfort: the horses did not go extinct. The population bottomed out around 1960 and then did something strange — it grew back. Today there are 6.6 million horses in the United States, more than double the low point. Almost none of them “work” in the 1900s sense of that word. They are shown, raced, ridden for pleasure, and adored. By the American Horse Council’s reckoning, something like two in five American horses are competitive athletes — a show jumper, a barrel racer, a Thoroughbred — and most of the rest are pleasure and companion animals. An entire economy, worth well over a hundred billion dollars a year, now exists to keep horses around purely because humans like having them around—even though they’re slower than dirt bikes.

So, what are all these horses doing? 92% of them are devoted to three activities:

• Recreation (trail, pleasure, backyard)

• Showing (competitive show horses)

• Racing (both active and for breeding stock)

True working horses (mounted police units, carriage operations, equine-assisted therapy programs, and lesson programs) are only 8% of the population, and even there it’s debatable whether horses are truly “best” for the job, or just sentimentally preferred.

That is the optimistic horse story: their economic value was almost entirely annihilated, but then a second life was born as a sort of luxury good. Horses are “workers” in a luxury industry that exists specifically because of intrinsic consumer preferences about who the workers are.

Now here’s the pessimistic horse story: mules — the sterile, sturdy, deeply unglamorous donkey-horse hybrids that did much of the South’s plowing. The U.S. mule population peaked at nearly 6 million in 1925. Today, there are perhaps 20-30 thousand in the United States.

That’s not a count of how many are working: it’s how many are left alive. Mules experienced a 99.5% population collapse with no rebound.

Why did the horse get a luxury second act and the mule didn’t? Because the mule’s only value was work. No one enters a mule in the Kentucky Derby; there is no mule dressage at the Olympics. In America, little girls might dream of someday owning a pony, but practically none dream of owning a mule. Strip out the mechanical utility and there’s no consumption demand left to catch the fall.

Survival after absolute functional displacement depends on residual consumption demand — on whether anyone wants what you provide because you provide it. After the 1960 low point, horses still had charisma, sport, and beauty, and they got a hundred-billion-dollar leisure economy. Ugly, stubborn mules had only output, and they got turned into glue, leather, and dog food. Then they were mostly forgotten (at least in the United States).

Key insight for AI: Arbitrary preferences can sustain “relational economy” demand for human labor if consumers intrinsically care about the source and nature of the labor.

The human version of this sorting is already visible: we will likely keep paying a premium for many kinds of human-made objects and human-performed services — live music, craft, care, the handmade, the “a person did this” — long after a machine could reproduce them, sometimes even at superior quality. A group of economists, including Alex Imas, the director of AGI Economics at Google DeepMind, call this “relational economy,” and it’s perhaps the most optimistic scenario for how humans might still earn wages even in a world where machines are stronger, more dexterous, and smarter than humans in every way. The hopeful future for human labor may look less like the much-hyped “everyone becomes a blue-collar worker” scenario (do you really care that your plumber is a person?) and more like “two in five horses are competitive athletes.”

It bears repeating, however: the “horse rebound as a luxury” story happened after the 90% horse population collapse. For humans, the post-AI relational economy will almost certainly exist to some degree, but whether it will be big enough to employ all the workers who want a job at livable wages is deeply uncertain. We’ll explore this question in the Music and Chess case studies.

Music: Automation vs. Augmentation and Superstar economics

Winners take all

Tractors and cotton gins gave history-shaping boosts to productivity in the agriculture sector, but they pale in comparison to the productivity gains related to copying and transmitting data. The cotton gin’s 50-fold improvement in productivity is nothing compared to the printing press or digital music, which increased the output of text and audio performances by millions or billions of times.

A very loud, unamplified medieval singer could still only perform to people within earshot, perhaps a thousand people in a large, acoustically well-designed cathedral. By contrast, during the 1985 Live Aid concert, which was broadcast around the world, perhaps 1.5 billion people listened to Freddie Mercury of Queen sing in a single day. Freddie Mercury’s voice could reach a million times more people. Now that’s augmentation! And Freddie’s limiting factor was not the technology, but the ability to attract an audience. The actual scaling factor is almost unlimited.

Just like cotton, perhaps more so, demand for music is highly elastic. As microphones, radios, record players, and eventually MP3 players made music ever cheaper and more convenient to enjoy, people consumed a lot more of it.

So, what happened to professional musicians? Well, there was certainly a lot of job loss. Predictably, this created musical versions of the Luddite movement. For example, before synchronized sound arrived in film in 1927, all music in movie theaters was provided by a live band. In 1930, the 22,000 live musicians who traditionally performed the music that accompanied “silent” movies were thrown out of work by the adoption of what they called “canned music.” According to the anti-robot advertisements they ran in newspapers around the country, canned music would allow “300 musicians in Hollywood [to] supply all the ‘music’ offered in thousands of theatres.”

The musicians’ union’s ads were quite creative in pursuit of a kind of “relational economy” edge, but they were not enough to persuade the movie-going public to care. Elastic demand for the output didn’t translate into demand for the workers, because the output could be reproduced infinitely without most of them.

But even for those musicians who managed to keep their jobs, the situation resembled the opening of Charles Dickens’ novel A Tale of Two Cities: “It was the best of times, it was the worst of times.” The invention of high-quality recorded music led to a major split in incomes among professional musicians, with the most successful musicians’ incomes exploding and the average musician’s plummeting (among those that kept their job at all).

In economics, this is called the superstar phenomenon, a “winner-take-all” outcome that allows a small subset of the producers—hugely augmented—to meet nearly all the demand.

There are two ingredients required for the superstar effect to apply:

• Imperfect substitution in demand: where lesser talent (even a lot of it) is a poor substitute for greater talent. Think about it this way: if decently performed music is rare, as it was before the invention of recorded music, then a local musician of any talent at all is a real treat. However, if music can be recorded and re-played infinitely, why would you listen to any but the best musicians who play your favorite style of music? If you like classical cello music, would you rather hear ten recordings of ten average high-school cellists or one piece performed by the cello master Yo-Yo Ma?

• Large economies of scale in distribution (also called joint-consumption technology) such that providing the product or service costs about the same whether ten or ten million people consume it.

The superstar effect offers an important caveat to the idea that workers in sectors where demand is highly elastic will be safe. With modern music streaming services like Spotify, people today can listen to almost anything, around the clock, at a marginal cost near zero. Music consumption has exploded into the trillions of streams per year. However, about 90% of music streams are for songs produced by the top 1% of artists. Perhaps the most extreme version of the music superstar effect is The Beatles in the first three months of 1964. During that period, six out of every ten records sold in the United States were by The Beatles.

Key insight for AI: The much-hyped automation vs. augmentation distinction doesn’t do much to clarify whether AI will create or destroy jobs in a given sector. Exclusively “augmentation” technologies might still result in both high job loss and extreme income inequality among workers in that same sector.

Chess: Centaur Economics

Mere humans can help superhuman AI, but who knows for how long

In 2005, a website called PlayChess ran a “freestyle” tournament with one rule: bring whatever help you want. Grandmasters, AI chess engines, teams of grandmasters running multiple engines on serious hardware — all allowed. The favorites were exactly who you’d expect. The winners were not. First place went to two American amateurs, Steven Cramton and Zackary Stephen, with USCF chess ratings of 1685 and 1398 (not even expert level), hunched over a couple of ordinary PCs. They beat grandmasters, standalone chess supercomputers, and teams that included both.

Garry Kasparov, the former world champion who had been brooding on human-machine chess teams since famously losing to IBM’s Deep Blue in 1997, drew the lesson that would launch a thousand keynote slides. A weak human plus a machine plus a better collaboration process, he observed, was superior not only to a strong computer alone but to a team of strong humans and machines running an inferior process. The decisive skill was no longer exquisite chess. It was knowing how to collaborate with and manage your silicon partner. People started calling these hybrid teams “centaurs,” and “Kasparov’s Law” became the optimist’s favorite proof that AI would augment us rather than replace us. The future of work, the story went, was centaurs all the way down.

It’s a lovely story. Unfortunately for chess players, within three years it was conditional and narrow; within 12 it was over entirely.

Rewind to the setup. Deep Blue’s 1997 win over Kasparov was treated as a stunt — expensive custom supercomputer hardware beating one exhausted human. But AI chess engines never stopped improving. By the early 2000s an ordinary PC could play at elite grandmaster strength, and engine ratings kept rising straight through the human ceiling and out the other side. The strongest human chess player who ever lived, Magnus Carlsen, peaked at an Elo of 2882 in 2014; Kasparov topped out at 2851. Today’s top engine, Stockfish 18, is not rated on the FIDE Elo scale but would be off the chart. One has to occasionally lose a game to the (in FIDE’s case exclusively human) competition pool to correctly calculate Elo, and Stockfish never would.

Today’s Stockfish is so far beyond Carlsen that across a thousand games he would lose nearly all of them, manage only the rarest draw (with some luck), and never once win. In a recent podcast interview, Carlsen himself said that he would have “no chance” competing against the chess app on his phone.

Even at its 2008 height, the centaur advantage was conditional. In blitz (with only three to five minutes of total thinking time per side), the human added nothing. In hour-long games, the human-machine team didn’t clearly beat the machine alone. Only at the very longest time controls — where a person had room to second-guess the engine and occasionally catch something — did centaurs reliably win. The window was real, but it was narrow even while it was open.

As the engines pulled away after 2009, the human’s job shrank to a single task: deciding which engine to trust when two top programs disagreed. Then even that window closed by 2017. The blunt version, now conventional wisdom among strong players, is that human intervention yields only negative returns — override Stockfish or its top AI competitors today and you are almost certainly introducing a mistake. The centaur did not evolve into a higher being. The human half became a handicap.

That is the pattern worth watching for in other domains. For a brief, real moment, the best way to play chess was a human and a machine together. That moment was not the destination. It was a waypoint on the road from “machines can’t” to “machines do it better without us.” Optimists thought that the centaur chess era looked like the future of work. If they were right, the future of work isn’t a very hopeful picture. The centaur chess era was brief and, even at its peak, not that reassuring.

Of course, both the relational economy and the superstar phenomenon are still at work in professional chess. In general, humans want to watch humans play chess, not AIs, so there are lots of tournaments with prize money for human players, and no prize money for AI-versus-AI tournaments. But the distribution of human chess earnings is overwhelmingly concentrated at the top. In 2025, Magnus Carlsen earned nearly $1.5 million in chess tournament winnings, plus millions more from sponsorships and appearance fees. Yet only 26 professional chess players earned more than $100,000 in 2025. And for everyone outside the elite, there’s no big hidden sponsorship upside: prize money is most of it.

Key insight for AI: A period in which humans and AI work better together than either does alone is not evidence that the partnership will last.

It may be nothing more than the short overlap between the moment a machine gets good enough to help and the moment it gets good enough to work alone. Plan for the collaboration. Don’t bet your career on its permanence. We may be in the peak centaur era of software development right now. How long it will last is deeply unclear.

This is also where the comforting analogies from the earlier sections begin to run out. The displaced farmhands had factories and offices to shift to. The horses got a luxury second act because some people happen to love them (even if others prefer dirt bikes). In some circumstances, live musicians still have the “a human made this” premium based on an arbitrary (but potentially durable) consumer preference. Every one of those escape routes depends on humans keeping some edge — a new sector, a buyer who cares who did the work. Centaur chess is the cautionary tale because it shows what’s left when the last human edge is collaboration itself, and then that goes too. And to the extent that the relational economy offers an island of safety, that may be a small island with room for only a few superstars.

“How to love people who have no use?”

I started this essay with one Vonnegut novel, so let me end with another.

I’m in the camp that thinks that, given enough time (maybe a decade, maybe a century), AI will be able to do nearly everything humans can do that has economic value not rooted in an arbitrary, intrinsic preference for a human doing it. In Player Piano, even the engineers and managers who built the machines eventually became surplus.

But in God Bless You, Mr. Rosewater, Vonnegut posed what I think is the question sitting underneath all four of these case studies: “How to love people who have no use?” The horses had charisma, sport, and beauty, so we loved and kept them. The mules had only their labor, so we didn’t. Strip out the economic value, and what’s left is whatever we’ve decided is worth keeping for its own sake.

Vonnegut had an answer. Later in that same novel, as a benediction over newborns, he delivered it with characteristic bluntness: “There’s only one rule that I know of, babies — ‘God damn it, you’ve got to be kind.’”

It isn’t an economic policy. It may not be much. But in a world where fewer and fewer of us are useful in the old sense, it might be the only thing that scales.

• Q1: What is Maven Smart System?

• Q2: What problems is MSS designed to address?

• Q3: What capabilities do large language models (LLMs) provide in MSS?

• Q4: What is Anthropic’s role in Maven Smart System?

• Q5: What is Anthropic’s future with Maven Smart System?

• Q6: What future does the Pentagon envision for Maven Smart System?

Full article available here. Excerpt below.

What Is Maven Smart System, and What Does It Do?

The Flagship Software Platform of AI-Enabled Warfare, Explained

by Matt Mande and Gregory C. Allen

June 2, 2026

During the first 24 hours of the war in Iran, the United States used Maven Smart System (MSS) to help strike more than 1,000 targets, a tenfold increase over what was possible in the pre-MSS era. However, few outside the government and contractor ecosystem involved in creating and using MSS are familiar with what it is, how it works, and the capabilities enabling a revolution in military intelligence and targeting.

Fortunately, a combination of old and newly available public and unclassified information sources allows for a much more complete understanding of MSS. Government and industry partners are sharing more than ever before about the platform, including through video demonstrations of its capabilities. At the same time, Katrina Manson’s recently published book Project Maven: A Marine Colonel, His Team, and the Dawn of AI Warfare includes interviews with critical stakeholders, many of whom were, until recently, remarkably tight-lipped. Reporting by various news outlets on the war in Iran has also revealed new details.

The following analysis consolidates and organizes this information to answer six key questions about Maven Smart System and its future in the Department of Defense (DOD).

Q1: What is Maven Smart System?

A1: MSS is an output of Project Maven, a Pentagon initiative founded in 2017 to bring AI capabilities to warfighters. Project Maven had a turbulent early history: Google was an early technology partner but withdrew in 2018 following employee protests over the company’s involvement in military AI development. Palantir, a data integration and analytics company, stepped in shortly after and has remained Project Maven’s primary industry partner. The company’s DOD contract ceiling for Maven surpassed $1 billion in May 2025, with MSS work added to that specific contract in fall 2025.

Thus, “Maven” refers to the current DOD program, which includes many other lines of effort and AI activities—such as AI-enabled Automated Target Recognition (ATR)—beyond just MSS. “MSS” refers to DOD’s flagship AI-enabled software platform, which draws from many data feeds, including Maven ATR. Palantir is the prime software integrator for MSS and stated in unpublished documentation reviewed by CSIS that “MSS is powered by the Palantir Platform.”

MSS is already used by the Joint Staff, Combatant Commands, and various elements throughout the DOD and Intelligence Community, as well as NATO allies. Palantir claims that MSS’s user base has doubled every six months for more than two years. In May 2025, MSS reportedly had more than 20,000 users, implying a current user base of roughly 80,000.

MSS provides a graphical user interface (GUI) for military intelligence and targeting purposes as well as a growing list of Combined Joint All-Domain Command and Control (CJADC2) functions. As described in the unpublished Palantir documentation, “MSS interfaces share a live, synchronized view of the Battlespace to imbue Warfighters and decisionmakers with real-time understanding in support of the overall CJADC2 mission.” The MSS user interface also allows for user tasking and human validation of AI labeling done by other Maven AI systems. In practical terms, this means that the use of MSS can speed up targeting decisions without sacrificing analytical rigor or judgment quality.

The platform aggregates, organizes, and visualizes vast streams of intelligence, surveillance, and reconnaissance (ISR) data. According to Manson’s reporting, CENTCOM’s deployment of MSS drew from 179 distinct data sources in 2024, a figure that industry sources told CSIS has increased significantly since. AI models, including those from other Maven data feeds, identify patterns in this data, which MSS visually displays. For example, computer-vision models find and label potential targets in live satellite or drone video feeds, which then appear within a yellow box on a user’s screen.

MSS also embeds workflows directly into that interface, allowing operators to move from observation to action without switching systems. The DOD’s public demonstrations show that in a targeting scenario, a user can select an AI-detected target, evaluate available strike assets in the vicinity, and compare options using operational constraints such as time to strike, distance, and fuel requirements. Once an asset is selected, the user can order the strike and subsequently monitor its effects using ISR feeds.

Palantir’s documentation of MSS provides a nonexhaustive list of six specific baseline capabilities:

• Battlespace management

• Target management

• AI-enabled deliberate planning and execution

• Computer-vision detections

• Machine-assisted disclosure

• Generative AI
  

Oversight of Project Maven and MSS has shifted locations within the Pentagon several times over the past decade. Project Maven originally sat within the Office of the Undersecretary of Defense for Intelligence and Security before the Biden administration split responsibilities between the National Geospatial-Intelligence Agency (NGA) and the Chief Digital and AI Office (CDAO) in 2022. Under this configuration, the CDAO paid MSS’s licenses and managed its text-based functions, while the NGA handled geospatial-intelligence work, including the production of computer-vision models. In March 2026, a memo from Deputy Secretary of Defense Steve Feinberg ordered oversight of MSS to be fully relocated to the CDAO, a decision that this article discusses in greater depth in Q6.

Read the rest of the rest of the article at CSIS.org

“So how much do you think semiconductor export controls contributed to China’s push for self-sufficiency? It can’t be zero, right?”

It definitely isn’t zero, but the size, contribution, and timing of the effect are different than the most vocal critics of export controls claim. Get the answer to this question wrong and you get the policy wrong. This includes, urgently, the policy this administration is now deciding in the wake of Donald Trump’s trip to China, with Nvidia CEO Jensen Huang riding along.

Ming-yen Ho is a non-resident fellow at a Taiwanese think tank, the Research Institute for Democracy, Society, and Emerging Technology (DSET). Last year, Ho published a paper on China’s semiconductor industry. Among many other virtues, the paper provides an interesting dataset to apply to the export controls question. Specifically, Ho has published the best dataset that I have seen thus far on how much the Chinese government is putting into semiconductor industry subsidies and equity investments (which have many of the same effects as subsidies). He also obtained good data on how much investment Chinese semiconductor firms are receiving each year from foreign investors, which can shed some light on how they responded to the controls.

If we compare Ho’s dataset to the timeline of the biggest events in U.S. semiconductor export controls and Chinese semiconductor industrial policy, we can test the hypothesis that U.S. export controls led to a big spike in Chinese government support to the industry. While government subsidies are not the only Chinese reaction that should interest U.S. policymakers, they’re still a hugely relevant test case. A mostly complete timeline of key export control events is included at the end of this article. But for our purposes, there are three events that loom largest:

• Made in China 2025, the 2015 policy in which China made semiconductor self-sufficiency a core goal of its industrial policy;

• The 2018 ZTE export controls, which were the first big semiconductor move by the first Trump administration; and

• The October 7, 2022 export controls by the Biden administration on AI and semiconductor tech.

As you review the subsidy and investment data, keep two things in mind:

1. Causality can’t run backwards in time: export controls cannot cause trends that existed before the export controls do (or at least before the perceived threat exists).

2. The change in the trend matters more than the change in the values: An observation that China’s investment in self-sufficiency is increasing or decreasing after the export controls doesn’t by itself tell us that the export controls had an effect. We need to know whether the rate of increase or decrease has changed. Knowing the change in trend helps us understand the difference between where China is now and where China would have been in an alternative scenario in which the export controls were never applied (what historians and social scientists refer to as the counterfactual analysis).

Chinese semiconductor self-sufficiency has been a stated goal of Chinese policy for decades, long preceding the modern U.S. approach to export controls that began in 2018. Moreover, this is not merely a goal, but one that the Chinese government has been willing to invest enormous amounts of money in. The river of money began with the first “Big Fund” in 2014, when the Chinese government started pouring tens of billions of dollars’ worth of investment into the semiconductor industry. The first Big Fund ended up being about $21 billion of investment, and was followed by a ~$29 billion second Big Fund in 2019 and a ~$48 billion third Big Fund in 2024. The first Big Fund was paired with a simultaneous dramatic expansion of the subsidies available to chipmakers, though the second and third weren’t.

Now let’s take a look at the data itself.

Looking at the timeline, there are two big investment spikes, in 2016—after Made in China 2025 and the first “Big Fund” of government semiconductor equity investments got underway in earnest—and again in 2020, after the second Big Fund was announced. When 2025 and 2026 data is available, it will likely show another spike reflecting investments by the third Big Fund, which was announced in 2024.

Based on the explanations given by Chinese policymakers and documents at the time (which expressed fear of dependence on the U.S.), it is at a minimum plausible that the 2018 ZTE export controls caused the spike in 2020. However, the government subsidy and investment levels were broadly similar to those present before the export controls.

The “this is just the continuation of the previous trend” explanation is even more obvious when you look at the government equity investments received by Chinese companies in the form of paid-in capital stock. Again, government equity investments are a very important tool of Chinese semiconductor industrial policy. In terms of their impact, government equity investments are similar to direct subsidies.

The Big Funds used large single-year government outlays to establish the investment vehicles themselves (hence the big spikes), but those financial resources are disbursed to semiconductor companies at a more consistent rate over time. The second Big Fund was evidently established when the first one ran out of money with the goal of continuing the same rate of industry investment, not increasing the level as a response to export controls.

Thus, the Chinese government’s overall rate of accumulation of semiconductor equity ownership investment was almost unchanged in the period after the ZTE export controls. After the 2022 Biden export controls, the rate of government investment actually decreased a bit. China’s semiconductor industry’s access to foreign private sector capital investment dropped to almost zero in the wake of the controls, which makes sense as a reaction to the new U.S. policy.

In short, available data on Chinese government subsidies and equity investment does not provide strong support for the hypothesis that export controls were a deciding factor in the Chinese government’s decisions about the appropriate level of investment.

My favored interpretation of the data is that, in 2018, the government had already reached or exceeded what it perceived to be the maximum productive rate of subsidization in pursuit of its previously decided policy of self-sufficiency. After all, the first billion dollars of subsidization (from 0 to 1) helps a lot. The 100^th billion dollars (from 99 to 100) helps far less. At a certain point, all the additional subsidization and investment doesn’t buy you a more robust semiconductor industry: it just fuels further corruption and waste. The first Big Fund already had corruption and waste in abundance.

Perhaps the Chinese government would have liked to subsidize and invest more if it had additional attractive options available, but those options did not seem to exist. The exodus of foreign capital probably made things worse, since interest by sophisticated foreign investors can help provide persuasive evidence to not always especially sophisticated Chinese government bureaucrats that a given Chinese company has real merit and is not merely a scam.

Government investment is not the only data that we’re interested in to answer this question, and in a future post, I’ll share some additional interesting data on the change in China’s chip sales, chip production capacity, and the rate of technological progress of China’s leading chipmakers.

So let’s return to the question from the former senior U.S. policymaker:

“How much do you think semiconductor export controls contributed to China’s push for self-sufficiency? It can’t be zero, right?”

Actually, it might be zero, at least in terms of government investment in the semiconductor industry.

Of course, the Chinese government has other tools to achieve its goals besides subsidies—facilitated smuggling, state-backed industrial espionage, coercion of international firms, stockpiling of manufacturing equipment, etc.

And the Chinese government is not the only relevant actor: the Chinese memory chipmaker YMTC launched a very large and well-resourced de-Americanization campaign in 2019 that was clearly a response to U.S. export controls on ZTE and Huawei and the fear that YMTC would be next.

But for Chinese government investing and subsidies specifically, export controls do not seem to have moved the needle. Made in China 2025 had already directed the government to do everything in terms of investing and subsidizing that it could. By 2022, that had been true for nearly a decade.

I’ve spent more than a decade working on AI, semiconductor, and defense tech policy. Across all that time, one pattern keeps showing up. The public version of an export controls action, an AI regulation fight, or a U.S.–China technology dispute is usually three or four analytical steps removed from what’s actually going on. The headlines reflect the press conference. The substance lives in the footnotes — in the corporate financial filings, the Chinese state-media coverage, and the procurement and budget documents that are mostly public but mostly unread.

Decision Tree is where I write down what I find when I do that close reading and number crunching, plus drawing upon a global network of smart folks I’ve met over years of doing this.

A few specifics on how I plan to do this.

The publishing pace will be set by the analysis, not the news cycle. I’ll publish when I have something to say that’s worth your time, and not before. AI policy is awash in takes. What’s in short supply is patient reading and careful work. That will probably be every one or two weeks.

My focus is the geopolitics of strategic technologies, broadly drawn. That covers U.S. export controls on advanced semiconductors and the AI systems they enable. It covers frontier-model policy (including safety, regulation, and economics), the defense industrial base, and the U.S.–China contest over compute and chip equipment. It also covers the harder long-run questions — where the next generation of leading AI labs and fabs get built. Subjects rotate, because the debate does. The same factories, ministries, and labs usually matter for several topics at once.

Hit reply to any issue, or write to contact@decisiontreeresearch.com. I read both. If you’d rather hear me than read me, I host weekly episodes of The AI Power Podcast.

Welcome. Glad you’re here.

— Gregory C. Allen, Founder and CEO, Decision Tree Research