The AI talent wars have produced a steady stream of stories that seem tailor-made to confirm everyone’s worst suspicions about Big Tech: nine-figure pay packages for star researchers, entire startup teams absorbed without a formal acquisition, and—most strikingly—reports of elite AI scientists paid handsomely to do nothing for a year under “garden leave” arrangements rather than join a rival—“hoard[ed] like Pokémon cards.”
To many observers, this looks wasteful at best and sinister at worst. Why would a profit-maximizing firm pay enormous sums for talent it seemingly has no intention of using?
Ronald Coase had a wry answer for moments like this. “[I]f an economist finds something—a business practice of one sort or other—that he does not understand,” he observed in 1972, “he looks for a monopoly explanation. And as in this field we are very ignorant, the number of ununderstandable practices tends to be rather large, and the reliance on a monopoly explanation, frequent.”
A new working paper by Shaolong Wu of Harvard Business School and Zefan Qian of Georgetown, “Talent Hoarding and Upstream Innovation: Labor Market Distortions by Large Incumbents,” supplies precisely that monopoly explanation, complete with a formal model and an empirical test. Large incumbents, the authors argue, sometimes hire and retain frontier researchers not to put them to work, but to keep rivals from doing so. Because top AI-research talent is scarce, every researcher a dominant firm keeps “on the bench” is one a challenger can’t hire. The incumbent protects its existing profits, and society loses the discoveries that researchers would have produced elsewhere.
The paper closes with policy recommendations to match: limits on garden leave, narrower noncompete agreements for publicly funded researchers, and institutional pressure to keep frontier talent “actively deployed.”
The policy audience is already primed for this argument. The Federal Trade Commission (FTC) has announced its intention to scrutinize acquihires—transactions structured around hiring a startup’s employees rather than acquiring the company outright—to ensure they aren’t used to evade merger review. Talent-centered theories of competitive harm are rapidly becoming the next front in the broader campaign against large technology firms. All the more reason to get the economics right.
It’s a clever paper, and perhaps a more careful one than many in this genre. Its headline claim, however, substantially outpaces its evidence. What the data actually show is that one group of software firms retained more skilled employees after the Supreme Court weakened their patent protections. Everything beyond that—the “idle benches,” the foreclosed rivals, the lost innovation, and the social harm—comes not from the data but from assumptions built into the model.
More importantly, nearly every one of those assumptions rules out, by construction, a far more ordinary explanation: that the same behavior reflects good management rather than anticompetitive conduct.
The Model Does the Hoarding for You
In the paper’s model, an incumbent firm earns profits from a legacy technology, while an entrant may develop a replacement. The key input into discovering that new technology is a fixed pool of researchers (an assumption we will return to) who possess knowledge of the incumbent’s technology. Hiring one of those researchers does double duty: It increases your own odds of a breakthrough while reducing your rival’s. There is, by assumption, no one else to hire.
The model also imposes another crucial condition. A researcher generates innovation only when paired one-for-one with “innovation capital”—new equipment, computing power, organizational capacity, and similar inputs. A retained researcher who lacks that matching investment is “benched”: employed and paid, but contributing nothing to the only outcome the model treats as socially valuable—the discovery itself. The model does allow benched workers to perform ordinary legacy work—it simply assigns that work no social value.
An additional assumption drives the paper’s headline results. Those results emerge from what the authors call the “no self-replacement” case, in which a breakthrough is assumed to be worth no more to the incumbent than the legacy product it already sells. The incumbent, therefore, gains nothing from innovation itself. In such a world, where the breakthrough has no upside for the incumbent, the only reason to retain a researcher is to keep her away from a rival.
Notably, retention doesn’t require leaving the researcher idle. Matching her with capital and putting her to work also reduces the entrant’s chances of success. Under the model’s assumptions, even productive employment can function as a form of exclusion. Whether she sits on a bench or works at a desk, the point is the same: she isn’t helping a competitor build the next generation of technology.
The paper thus assumes, quite literally, that an incumbent’s only reason for incurring the cost of retaining a knowledgeable researcher is to deny that researcher to a rival. Under those conditions, if a successful entrant would destroy the incumbent’s profits, the incumbent will rationally pay to keep researchers in-house even when it places no value on anything they might produce.
On the empirical side, the paper exploits the U.S. Supreme Court’s 2014 decision in Alice Corp. v. CLS Bank. The Court held that implementing an abstract idea on a generic computer is not, without more, patentable. The decision thus made broad software patents substantially harder to obtain and enforce. For firms whose competitive position depended on such patents, Alice increased the risk that rivals could imitate or displace them—what the paper terms “external displacement risk.” Talent hoarding, in turn, became a potentially more attractive way to protect an incumbent position.
The authors find that software firms holding pre-2014 software patents increased their net hiring after Alice by roughly 0.34 percentage points per month relative to software firms without patents—a not-insubstantial effect, about 63% of the sample average.
Several additional findings are noteworthy:
- The increase came almost entirely from reduced employee departures rather than increased recruiting.
- The effect was concentrated among workers with advanced degrees.
- It was not accompanied by measurable growth in assets or other complementary capital.
- A companion analysis of worker résumés found that science-and-research employees at exposed firms advanced more slowly up the occupational ladder in the years following the decision.
In short, the study finds: more high-skilled workers retained, no corresponding increase in complementary capital, and slower career progression among the researchers who stayed. The authors interpret this pattern as evidence of their central hypothesis: the defensive retention of underutilized talent.
The Data Are Not the Dispute
To give the paper its due, the distinction between retaining a worker and deploying one is genuinely useful. Too much commentary on tech hiring blurs the difference. The research design is also serious: exposure is measured before the shock, the shock itself is a judicial decision no firm controlled, and the authors are unusually candid about their limitations. Most importantly, the central empirical finding appears real. Patent-exposed software firms reduced employee separations after Alice, and those reductions were concentrated among workers with knowledge likely to be valuable for innovation.
The dispute isn’t over the finding, but over its interpretation. The talent-hoarding story depends on a series of modeling choices, each of which rules out a competing explanation under which the same data reflect efficient, rather than anticompetitive, behavior.
Foreclosure by Construction
The paper’s most significant assumption is one most readers will never notice. It has two parts. First, the incumbent and the entrant are assigned identical discovery technology, meaning a researcher generates exactly the same probability of a breakthrough wherever she works. The incumbent firm itself contributes nothing distinctive, except already being there. Second, as noted above, in the headline case, the incumbent places no value on producing the breakthrough at all.
Taken together, the model effectively assumes the incumbent is the worst possible home for research talent. It’s no better at research, and it’s uninterested in the results. The only remaining explanation for retention is blocking.
The authors would fairly answer that this isn’t a claim about the world. They set the new technology’s value to zero on purpose, to strip out the incumbent’s offensive motive—wanting the breakthrough for itself—and study the defensive motive in isolation. Fair enough. Isolating one mechanism is what models are for.
But here, the model removes the efficient reason to retain talent before the welfare verdict is rendered. Then that verdict gets carried over to a world where incumbents usually do want the breakthrough, and where retention is often offensive as well as defensive. The conclusion isn’t “found” in the research; it’s built into the model.
And that assumption elides a great deal. Firms aren’t equally good at converting researchers into discoveries. They differ enormously in management quality, in the data and infrastructure researchers can use, in their ability to select promising projects and kill bad ones, and in the colleagues a new hire works alongside. A large incumbent may get more out of a marginal researcher than a startup can, because she joins teams with experienced managers who have already solved a thousand deployment problems, and because she can work with proprietary data and installed systems no entrant possesses.
In that model—the real world—scarce talent flowing toward incumbents isn’t foreclosure. It’s the market allocating a scarce input to its highest-valued use. Eye-popping compensation is the competitive price of that input, captured by the workers themselves.
The paper’s model rules out this more realistic dynamic, and the empirical work can’t restore it. A finding that exposed firms retained more researchers is equally consistent with “those firms denied rivals an input” and “those firms are where the input is worth the most.”
An economist might respond that, in the model, the incumbent’s willingness to pay exceeds the worker’s deployment value. That’s what makes it hoarding. While true in this context, that result follows from the equal-capability assumption. Relax that unrealistic assumption, and the wage premium may reflect a productivity difference rather than a blocking premium.
That means the paper’s welfare conclusion turns on a parameter it never measures.
The Salop Problem Returns
The claim that a dominant firm will rationally outspend any challenger to preserve its position is one of the oldest moves in the antitrust playbook. A recent version of it comes from Steven Salop, who argued—in a paper pointedly titled “Potential Competition and Antitrust Analysis: Monopoly Profits Exceed Duopoly Profits”—that acquisitions of potential or nascent competitors by dominant firms raise inherent anticompetitive concerns. Because keeping a monopoly is worth more than sharing a duopoly, the incumbent will always pay more to eliminate a threat than the threat is worth to anyone else.
Wu and Qian have, in effect, transported this logic from the market for startups to the market for researchers. Here, the thing the incumbent supposedly overpays to control isn’t a nascent rival firm, but the researcher who might help create one. Same wine, different bottle.
The argument fails in the labor-market setting for many of the same reasons it fails in the acquisition context—reasons Dirk Auer, Brian Albrecht, Eric Fruits, Daniel Gilman, Lazar Radic, and I discussed in the International Center for Law & Economics’ (ICLE) comments on the FTC and U.S. Justice Department’s (DOJ) draft merger guidelines, and that Albrecht has explored at length here at Truth on the Market.
The first problem is arithmetic. Buying off a single potential entrant may be profitable when monopoly profits exceed duopoly profits. But once one challenger is paid to stand down, the next stands to enter as a duopolist, and the next after that, and so on. Each must be compensated at roughly the duopoly level, not some small fraction of it. With enough potential challengers, the cost of paying them all off exceeds the value of preserving the monopoly, and the strategy collapses. If we nonetheless observe the conduct, that suggests something other than monopoly maintenance may be at work.
The same arithmetic applies to talent. The paper’s model contains one incumbent, one entrant, and one closed pool of researchers, so cornering the input requires outbidding exactly one rival. The real frontier-AI labor market looks nothing like that. There are multiple deep-pocketed AI labs, a venture-capital ecosystem aggressively funding startups to compete for the same people, and a university pipeline continuously producing new talent.
To foreclose discovery through retention, an incumbent would have to outbid all of those rivals, for every pivotal researcher, indefinitely. The predictable result isn’t foreclosure, but an auction. And in that auction, the owners of the scarce input—the researchers themselves—capture much of the surplus. That looks less like market failure than competition doing exactly what it is supposed to do.
Two features of labor markets make the strategy even less durable than its acquisition counterpart. First, workers, unlike startups, can’t be bought outright. Because “human capital is inalienable,” a retained researcher stays only as long as she chooses to stay. Maintaining exclusivity is therefore perpetually expensive and contractually fragile.
Second, when employees leave large firms, they disperse. They don’t march single-file to the one rival most capable of threatening their former employer. As Kevin Murphy has observed, departing workers spread across many employers, with only a small fraction joining any particular competitor and many leaving the industry altogether. The model’s premise that every researcher released by the incumbent flows directly to the one firm poised to destroy it bears little resemblance to how labor markets actually function.
The second problem is the one discussed above. Salop’s result requires the incumbent to be at least as capable as any challenger—an assumption he relegates to a footnote, while acknowledging that “monopoly profits are not always higher” when an entrant has lower costs or a better product.
As Dirk Auer, Sam Bowman, and I have written:
Although it is convenient in theoretical modeling to assume that similarly situated firms have equivalent capacities to realize profits, in reality firms vary greatly in their capabilities, and their investment and other business decisions are dependent on the firm’s managers’ expectations about their idiosyncratic abilities to recognize profit opportunities and take advantage of them—in short, they rest on the firm managers’ ability to be entrepreneurial.
Once that realistic possibility is admitted, differences in firms’ willingness to pay no longer demonstrate preemption. They may simply reflect differences in productivity.
Again, the empirical prediction that retention rises after Alice doesn’t depend on the equal-capability assumption. But the conclusion that retention is driven by exclusionary motives does. The empirical test therefore confirms the portion of the model that doesn’t require the assumption, while leaving untested the assumption that does most of the work in generating the paper’s policy recommendations.
No New Assets, No New Ideas?
The paper’s cleverest empirical move is to infer foreclosure from the combination of rising headcount and flat capital investment. At first glance, the logic seems straightforward: if firms were truly putting these researchers to productive use, we should see corresponding investment in the inputs they need.
But the argument rests on a strong assumption: that a researcher paired with anything less than a full new unit of “innovation capital” produces nothing. In software, of all industries, that is a peculiar picture of the production process.
The complementary inputs that matter—data, codebases, proprietary tools, accumulated organizational knowledge, and much of the underlying computing infrastructure—are often already in place. An additional researcher can use those resources without the firm having to book a single new asset. Much of this capital is also non-rival within the firm. Another engineer working with the company’s data doesn’t consume that data. Even rivalrous inputs, such as computing power, are often available within existing capacity, with enough slack to absorb another researcher without generating a measurable increase in asset purchases.
A firm whose competitive advantage consists precisely of these intangible assets can productively absorb additional researchers with no observable capital response at all. Put differently, the paper’s headline empirical finding—more labor, flat balance-sheet assets—is exactly what we would expect if incumbents are where marginal researchers are most productive.
The authors also argue that complementary capital “can often be scaled within just days and weeks,” making its absence particularly informative. If firms wanted to deploy these researchers, they argue, they could have done so cheaply and quickly.
But that cuts both ways. If complementary capital is as inexpensive and scalable as the authors suggest, then the absence of a detectable increase in assets is weak evidence of non-deployment. Cheap capital may not show up as a measurable balance-sheet change even when researchers are fully deployed. And if computing capacity can be expanded on short notice at low cost, then computing power was never the binding constraint. Talent was.
In that case, scarce talent flowing toward the firms willing to pay the most for it is not evidence of benching. It is what efficient allocation looks like.
When the same evidence supports opposite welfare conclusions, it ceases to be especially informative.
Nor does the authors’ fallback diagnostic—that patent output per employee fell at exposed firms—break the tie. That claim is perilously close to circular. The Supreme Court had just made it more difficult for those firms to obtain patents. Of course patenting declined. That was the treatment. It is not evidence that researchers were working less effectively or producing less valuable output.
The People Are the Patents Now
The most natural reading of the paper’s central finding requires no “benches” or anticompetitive animus. Alice didn’t conjure displacement risk from nowhere; it weakened the specific legal instrument—patents—that exposed firms had used to protect their innovations. Economists have long understood patents and trade secrecy as substitute ways to appropriate the returns to invention. And trade secrets don’t live in filing cabinets—they live in employees’ heads.
When the Court devalued these firms’ patents, the rational response was to lean harder on the other instrument: retaining the people who embody proprietary knowledge.
That reading fits the paper’s evidence rather neatly. It predicts retention rather than recruitment, because the point is to protect knowledge the firm already has, not acquire more of it. It predicts concentration among advanced-degree workers, because they are most likely to carry the relevant know-how. And it predicts no corresponding capital expansion, because nothing about the firm’s investment program or resource allocation needs to change.
The paper’s own marquee example points in exactly this direction. Describing Adobe—the lead illustration—the authors write that, once Alice weakened its patents, “the strategic asset that mattered more was not just the code or the patent portfolio, but the people who embodied the know-how.” That’s a description of a firm switching from one appropriability mechanism (patents) to another (retaining employees who hold trade secrets). The paper states the benign reading in its own words, in its own flagship example, and then subsumes it under “foreclosure.”
But those are not the same thing. Preventing your own know-how from walking out the door to a competitor is a legitimate interest that trade-secret law has protected for well over a century. On the most plausible reading, it’s also welfare-enhancing here: retention substitutes for the appropriability that Alice weakened, preserving at least some of the innovation incentives the decision would otherwise have eroded.
Indeed, the paper’s own framing concedes the point without quite noticing it. The workers in the model are valuable to the entrant because of their “incumbent-specific knowledge.” What the rival wants from them, in other words, is largely the incumbent’s own proprietary information.
The Missing Foreclosed Rival
Even granting the retention findings, the paper’s welfare conclusion—that society loses when incumbents retain these workers—is built into the model rather than derived from the evidence. The social planner against whom the incumbent is judged is defined to value “the availability of the innovation” while placing no weight on the incumbent’s existing profits. By assumption, entrant innovation is socially valuable.
But a challenger’s incentive to enter often includes what economists call “business stealing.” Much of the entrant’s prospective profit comes from taking customers and profits from the incumbent. From a social perspective, that is largely a transfer rather than a gain. Indeed, the economic literature on entry has long recognized that private incentives to enter can sometimes exceed the socially optimal level for precisely this reason.
One could just as easily write down a model with the opposite welfare weights and conclude that the entrant’s poaching is the problem. Neither stipulation would constitute evidence.
Meanwhile, the foreclosure side of the story—the part that carries the antitrust implications—is never tested at all. Foreclosure requires, well, foreclosure: rivals must be meaningfully constrained in their ability to compete because they can’t access a necessary input. Yet the paper offers no evidence that startups in Alice-affected fields were starved of talent, hired fewer workers, grew more slowly, or innovated less than they otherwise would have. Perhaps they did, but the paper neither tests nor demonstrates it.
What evidence we do have points in the opposite direction. The authors themselves note that the cost of software experimentation was falling during this period, as cloud computing and related technologies made entry cheaper and more accessible.
Readers familiar with the killer-acquisitions debate will recognize the problem. Even in pharmaceuticals—the sector with the strongest evidence for the theory—the most generous estimates suggest that genuine “killer” acquisitions account for only about 5% to 7% of deals. The evidence in digital markets is considerably weaker—which is to say, nonexistent.
A foreclosure mechanism that leaves no observable trace in the supposedly foreclosed market may be an interesting theoretical possibility. It is not an empirical finding.
A Raise Is Not a Market Failure
The paper’s career-trajectory evidence is also clever, but again curiously interpreted. The slowdown in advancement for workers in the “science and research” category at exposed firms is real but tiny—about 0.018 points on a 2-to-5 occupational scale. Indeed, the scientists who are the paper’s central concern show the smallest of the highlighted effects.
The largest decline in the paper’s table of “strong negative effects” belongs instead to workers in “education and training,” at a rate of ?0.068—nearly four times the science-and-research estimate. That is difficult to square with a story about the “benching” of frontier talent.
The next-largest slowdown, among workers in “law, compliance, and public safety,” is easier to explain. The authors themselves attribute it to a collapse in demand for patent-related legal work—a direct consequence of weaker patents, not talent hoarding. And in a separate “difficult to interpret” category, “healthcare and maintenance” workers at exposed firms also show larger slowdowns than science-and-research employees.
The heterogeneous responses across job categories are supposed to support the hoarding story. But doing so requires emphasizing the category with the smallest effect while explaining away the larger effects the theory can’t account for.
There is also a simpler, and perhaps more important, problem: the data contain job titles but not wages. A worker who stays in the same role because her employer pays her substantially more hasn’t suffered a policy-relevant harm. She has been compensated. Indeed, the very study the authors cite on inventors moving to large firms finds that earnings rise by double digits after those moves.
A flatter title progression accompanied by a fatter paycheck isn’t a labor-market distortion. It’s known as a raise.
Before We Ban Good Management
If we step back and ask what conduct this paper actually condemns, the normative implications become difficult to sustain. What the paper identifies is lower employee turnover, the payment of retention premiums, and efforts to keep proprietary knowledge from reaching competitors. That describes competent management far more readily than monopolization. Every well-run firm in every industry does these things. The conduct labeled anticompetitive is observationally indistinguishable from conduct we generally want firms to undertake. And the paper’s empirical design can’t tell the difference, because the assumptions that do the distinguishing work—equal firm capability, one-for-one capital matching, and a social planner indifferent to incumbent investment incentives—are imposed rather than tested.
The paper’s reception has outpaced its findings. One economist described it as “a dark theory (with evidence)” of Big Tech hiring, in which incumbents “leave some on the bench” to “slow rival innovation.” But the bench is precisely what the paper never observes. The authors are candid that they can’t see deployment worker by worker. Benching is inferred from the absence of a capital increase, not directly measured. “With evidence” is doing a great deal of work for a mechanism the study assumes rather than demonstrates.
Despite that, the paper advances policy recommendations: limits on garden leave, restrictions on noncompete agreements for publicly funded researchers, and institutional expectations that frontier talent remain “actively deployed.” These proposals arrive amid a broader labor-antitrust push whose evidentiary foundations are already remarkably thin, resting on a small number of studies with heavily qualified findings.
Yet the mechanisms the paper would restrict perform real economic functions. Even the FTC, in the rulemaking that banned noncompetes, acknowledged evidence that such agreements can increase investment in workers’ human capital, physical capital, and research and development. Firms invest in training and entrust employees with valuable knowledge when they have some confidence that those investments won’t immediately walk across the street.
Weakening those arrangements based on a possibility theorem and an empirical pattern that admits several efficient explanations is a recipe for getting error costs backward. It risks condemning ordinary retention practices in the hope of catching the occasional genuine “idle bench.”
It is also worth remembering which firms the talent-hoarding story is meant to indict. The paper takes its empirical cue from a 2014 patent decision, but its stakes—and its policy recommendations—are aimed squarely at today’s AI incumbents. Those firms are not behaving like hoarders stashing talent on idle benches. From 2019 to 2025, the largest tech firms—Amazon, Alphabet, Meta, Microsoft, and Apple—added nearly 1 million employees while increasing capital expenditures from $77 billion to $370 billion, according to their 10-K filings. That is not a picture of firms accumulating talent they have no intention of equipping. It is a picture of firms racing to deploy labor and capital as quickly as they can acquire them. In other words, it looks like competition, not foreclosure.
The paper’s actual contribution is narrower, but still interesting—if perhaps not as headline-grabbing. It suggests that when courts weaken patent protection, firms respond by retaining the people who embody their proprietary knowledge. That’s a finding about the substitutability of appropriability mechanisms, and it counsels caution both about Alice-style doctrinal shocks and about labor-market interventions that would weaken the substitutes firms turn to in response.
Before talent retention becomes the next frontier of antitrust, we should demand evidence of the thing that matters: not that firms kept their researchers, but that those researchers sat idle, and that someone else would have put them to better use.
