Streetwise Professor

R.A.B. (@LibertyLynx) and I have completed a document (attached) that fully details the October “pinning” of ICEYE X36 and the (serious and disturbing) implications thereof. The document also examines the depletion of US magazines in the ongoing war with Iran. That latter angle is an important story in its own right, but as will be seen, it also intersects with the X36 incident in important ways.

The document is extremely dense–purposefully, because the details are important and because it’s essential to support extraordinary findings with extraordinary evidence. Crucially, all of this information is derived from public sources. We also indicate types of non-public information that could potentially falsify our conclusions, though that would be awkward for SpaceX in particular as it would require it to falsify its own public data.

For those who prefer video–here you go. More fine work by R.A.B.

If you don’t want to dive into the details, there are a couple of key takeaways. First, in the midst of a Trump effort to flatter Putin and coerce Zelensky into a peace deal, SpaceX made possible Russia’s pinning and compromise of a satellite that both Ukraine and the US rely on for crucial intelligence. Second, Hegseth and Michael embedded xAI/SpaceX in Pentagon targeting, shutting out others (namely Anthropic); relied almost exclusively on AI for targeting, claiming that the resulting speed was war-winning, when in fact it did not result in Iran’s defeat but did succeed in speeding up the depletion of US weapon stocks. Both have materially weakened the US.

As we note, timing is a decisive causal determinant. And here the timing is quite disturbing. The SpaceX throttling of defensive Ukrainian drones and its essential role in the pinning of a US intelligence satellite (which we show could have occurred by chance with a probability close to zero) occurred in the very middle of Trump’s very public courting of Putin and very public demonization of Zelensky, all part of his frenzied pre-midterms effort to get a peace deal, shift the blame for high fuel prices, and gain some fuel price relief. It takes a Herculean effort to conclude that these events are not connected.

Pause and think about that for a moment, and think about it again after I’ve gone through some of the implications.

It is particularly appalling that the throttling and pin occurred in the context of a vain and futile attempt to seduce Putin into a peace deal. All that Trump and his minions (notably Witkoff and Kushner) have dangled before Putin are but mere baubles to him. The man has sacrificed hundreds of thousands of Russian lives and has inflicted severe short term and long term damage on the Russian economy, and the well-being of Russians, in pursuit of his obsession to conquer Ukraine. Only the delusional would think sanctions relief or business deals will deflect him from pursuing that goal. Doing things that harm US national security in pursuit of a chimerical deal is disgraceful.

The document also details several post-pin events that like waves in a pond demonstrate that something big hit the water. These are:

• Putin stated that the Russia was coordinating with “Mr. Musk” in space.
• Finland announced that it is taking control of three ICEYE satellites-X36, X37, and X38, which just happen to be the ones compromised by the pin.
• The day after the pin ended, a contractor (Stellar Exploration) that provided crucial systems that we show were compromised by COSMOS 2612’s 40 plus hour sub-kilometer stalking of X36, engaged in a restructuring transaction in the Czech Republic that reduced the ownership of the US party from 80 percent to 33 percent and gave the largest share to a Czech individual, which conveniently places the contractor under the shield of Czech and EU commercial-secrecy law. Said law protects it from any US investigation or audit short of a criminal investigation by the DOJ.

The Finnish story is woefully under-reported because it clearly implicates what happened just a few days before the Finnish announcement. Note what the article says:

The threat does not necessarily have to involve a direct attack on a satellite. Other possibilities include interference with navigation signals, attempts to take control of systems, cyberattacks on ground infrastructure and activities by other space objects that approach satellites and interfere with their operation. 

“Approach satellites and interfere with their operation.” Which is exactly what happened in mid-October.

Again, the timing of the Finnish announcement is telling.

We uncovered the information about Stellar. We only became aware of its existence when its CEO–the US citizen of Czech extraction who charitably gave up a 47 percent stake in the company the day after the pin ended–lost his mind at us on X. Not smart!: That outburst prompted us to do a little digging, which uncovered that Stellar provided the very electrical systems that were compromised by COSMOS 2612. Further digging uncovered the stock switcheroo (filed in the Czech Republic) that is a transparent attempt to impede US inquiries: Czech law shields companies with a foreign owner holding less than 33.3 percent of the stock, and after the switcheroo the CEO owned 33 percent. Only a DOJ investigation (not the Pentagon, for instance) can get past that.

This transaction is a red flag: why would anyone do it except to try to stymie an investigation? A lack of a DOJ investigation would be another.

The timing is telling here too.

Another important takeaway is what links the pinning and US magazine stories, namely the information that the Russians obtained via the pin. Information that they could not have obtained otherwise. And it involves Stellar.

Specifically, from a sub-kilometer distance COSMOS 2612 could monitor and electronically map X36’s power and radar systems in ways impossible from the 400+ kilometer distance 2612 maintained before the pin: close proximity allowed COSMOS to overcome the clutter that prevented it from extracting signals from afar. The information gained allows it to determine X36’s power and battery state, and to monitor its radar use, from its current 66-90 second trail of X36. COSMOS 2612 can therefore determine whether X36 has sufficient power before recharging to conduct surveillance over the Persian Gulf. It can consequently give Iran several minutes warning of whether surveillance is impending. Minutes Iran can use to determine whether it is safe to break concealment to launch missiles–which depleted US Patriot and THAAD stocks limit our ability to shoot down. Minutes that it can use to scurry into concealment if it is aware that eyes from the sky are gazing at them.

And as we show, the pin did not compromise X36 alone: it compromised its sister satellite X37. Which, lo and behold, is being trailed by two other COSMOS satellites.

Thus, the pin degraded US capability in the Iran theater: the two non-SpaceX satellites on which the US relies on in the Gulf are compromised, and SpaceX played a necessary role in the compromising.

That US weapons stocks were depleted by the February-April conflict is what has leveraged the adverse impact of the pin. Which leads to another portion of the document: the highlighting of the role of Hegseth and his sidekick Emil Michael in relying on AI to accelerate targeting, which resulted in the magazine depletion which leverages the adverse consequences of the pin.

Until, well yesterday, Hegseth and Michael constantly preached that speed is lethality, and bragged about how the US had used AI to dramatically speed up targeting in the Gulf, allowing more targets to be struck in less time. Indeed, they were so giddy about the efficacy of AI that they slashed the number of humans tasked to targeting in CENTCOM from 16 to 1.

But it is evident that speeding up a bombing campaign did not speed up Iranian capitulation. Nor did it deny Iran the use of important capabilities, namely asymmetric attacks that closed the Gulf of Hormuz, and missile attacks that hammered US bases in the Middle East (killing servicemen and women, and destroying many valuable intelligence and refueling planes).

What the targeting acceleration did accomplish was to speed up the expenditure of scarce US weapons. But crucially it did not speed up the replacement timeline: it will take years to replace the JASSMs, Tomahawks, Patriots, and THAADs shot off so quickly. US capability in other theaters, notably Asia-Pacific have therefore been seriously compromised.

In focusing on speed in a single theater, Hegseth and Michael ignored opportunity (“shadow”) cost–the value of these weapons in other theaters. This makes their firing of two JCS planning (J-5) chiefs look all the more suspicious. It is precisely such individuals who are responsible for taking a global perspective in designing military operations. We have little doubt that the fired (and traduced) generals raised hell about the extravagant expenditure weapons in a single theater, other theaters take the hindmost.

This all makes recent news even more striking. At a Wall Street Journal tech conference in Napa, former AI evangelist Michael suddenly turned all cautious and conservative and sober about it. He claimed the military “remains conservative about using artificial intelligence in combat,” insisted the U.S. “is not very sophisticated at AI-generated targeting,” and lectured that “there must be appropriate human judgment at every level” right as his shop floated new “AI compliance guidelines.”

“AI? AI? What’s that? Never heard of it. Does it have something to do with computers? I don’t even know how to use a computer!” Mr. Michael is apparently shocked, shocked! to find AI going on in here.

The stories we have documented are important in their own right. But they are just one element of extreme dysfunction within the Pentagon, dysfunction that will be turned up to 11 by the looming Hegseth impeachment. (Yes, the initial attempt failed: wait until January). Hegseth has allegedly threatened Trump that he is going to the mattresses rather than resign–with the message being delivered by a lawyer who is a protege of and former defender of John Gotti’s mouthpiece, and a mouthpiece for Bonnano and Geneovese associates in his own right, no less. This internecine conflict will make the dysfunction even more extreme.

And who is in the middle of all this but the suddenly sotto voce Mr. Musk. SpaceX played a central, indeed necessary, role in the pinning of X36. xAI is the central player and the main beneficiary in the suddenly amnesiac Mr. Michael’s drive to inject AI into crucial DoD decision making.

We wish we could say that this document is the end of the story. Unfortunately, it appears to be just the beginning.

AI hysteria has predictably led to calls to regulate! regulate! regulate!, seemingly from all corners except for the Trump administration and NVIDIA’s Jensen Huang. Those baying for “regulation” include frontier model leaders Anthropic and OpenAI.

I put “regulation” in quotes because as is often the case, those calling for regulation are imprecise on just what regulation should entail, and have evidently given little serious consideration to how regulation works in practice. The word is usually more of a magic incantation than a description of something real that has any chance of realizing the goals of those using it.

So what should AI regulation look like? To answer that question, it’s important to start at the basics.

Decades ago Steve Shavell wrote two important articles that speak to the subject: “Liability for Harm Versus Regulation of Safety” (JLE, 1984) and “The Optimal Structure of Law Enforcement” (JLS, 1993). Shavell laid out several criteria for determining the best way to regulate.

Broadly speaking, there are two alternative methods of regulation. Harm-based (“ex post“) sanctions that attempt to deter harmful conduct, and preventative (“ex ante“) regulations. The former includes the common law tort system, and civil and criminal laws. The latter includes things like government approval for the sale of a product, with FDA approval of drugs being an example. Many of the calls to regulate AI fall in the latter category, and some members of the Trump administration basically advocate the former.

So which is best here?

Shavell identifies several factors that determine the optimal legal structure. On balance, these favor ex post/harm-based tort-like approaches.

The first criterion is “is the possibility of a difference in knowledge about risky activities as between private parties and a regulatory authority.” That is, information asymmetry between the regulator and the regulated.

This is quite intuitive. If an actor (an AI firm, in this instance) has better information about the likelihood its actions will cause harm, and the amount of harm it will cause than does a regulator, if it knows it will pay damages commensurate with the harm it causes, it can take the efficient precautions. It will appropriately trade off the cost of mitigating harms (including the cost associated with reduced performance) with the cost of the harms because the prospect of paying damages based on costs and probabilities of harms will induce it to internalize the costs of its actions.

In contrast, an ignorant regulator knows less about the probabilities and magnitudes, and is therefore likely to implement measures that are either overinclusive or underinclusive. That is, regulators are far more prone to Type I and Type II errors.

It is clearly the case with AI that the developers-the labs/frontier firms in particular-will be far better informed about their creations and the risks that they pose than any government regulator would be.

Shavell’s second criterion is “that private parties might be incapable of paying for the full magnitude of harm done.” That is, in the present context, an AI firm might be “judgment proof.” In the event, the actor will not internalize the costs it imposes on third parties, and will tend to take excessive risks and impose excessive harms. Here, external interventions that constrain actions in order to reduce the magnitude and frequency of harms can be superior to deterrence.

The inability to impose monetary damages commensurate with harms is one justification for criminal sanctions, including imprisonment or even execution. These are non-monetary penalties that enhance deterrence.

This is one factor that might-might-favor ex ante regulation. Anthropic or OpenAI (or any other firm) could not pay damages commensurate with harm if that harm is the end of humanity!

But this brings us back to the first issue. In deciding whether judgment proofness is indeed a problem, we need to have some assessment of the distribution of harms, and here the AI firms have a decisive information advantage. And as I discuss below, we absolutely cannot take at face value the pronouncements of Amodei, Altman, Musk, et al, on the potential harms posed by AI, or their likelihood.

And more, even though judgment proofness may preclude AI firms from fully internalizing costs of their products (especially the harms they may create), bureaucrats operating subject to low powered incentives clearly don’t internalize them. Nor do they internalize the benefits of more capable AI. Therefore, although judgment proofness may be a necessary condition for ex ante regulation, it is not a sufficient one.

Shavell’s third factor is “the chance that parties would not face threat of suit for the harms done.”

In the US legal system? As if! Indeed, the likely bigger problem with relying on the tort system to deter would lead to suits for harms not done, i.e., frivolous or baseless litigation.

Shavell’s last factor is the relative administrative costs of the two systems, i.e., the cost of a regulatory bureaucracy vs. the costs of litigation.

Both litigation and bureaucracy are costly. The advantage of ex post, harm-based regulation is that the litigation costs are only incurred when a harm incurs: no harm, no foul, no legal expense (with the caveat related to frivolous litigation mentioned above). In contrast, the bureaucracy must be fed continuously, even in the absence of harms.

And note that regulation would necessarily have to be international in scope. Look at the post-Great Financial Crisis history of banking and financial regulation for a demonstration of the monumental costs and challenges of international regulation. AI regulation would put even that in the shade, especially since it adds China and strategic competition into the mix.

To which, I would add another factor: influence costs, including regulatory capture. It is well known that the regulated exploit their information advantages and financial resources to induce regulators to take actions that benefit some regulated entities, at the expense of other entities and the public at large. This is especially likely to work when the regulated have a big information advantage (check) and a lot of money (check).

Overall, to me the scorecard tilts heavily in favor of ex post, harm-based, liability mechanisms as opposed to ex ante regulation. This is especially true given the radical uncertainty and information asymmetries inherent in such a dramatically new technology. We are learning the harms as we go, and the developers learn more and faster than any bureaucrat can. Indeed, being liable for harms gives the developers a strong incentive to identify those potential harms-their magnitudes and likelihoods. A far stronger incentive than bureaucrats have.

And that provides a segue into a discussion of one of the most fascinating aspects of the ongoing debate/hysteria: the fact that the frontier labs are the loudest and most strident voices calling for regulation. In my opinion, this is overdetermined.

A major reason is that Anthropic et al are deathly afraid of being liable for the harms their products cause. Palantir’s Alex Karp perhaps overstates the case when he says they are effectively calling for nationalization in order to avoid liability, but I think it’s pretty clear that avoiding paying for the harms their products cause is a major driver behind the calls–pleas, really–for regulation.

Which, to be honest, another reason to favor a liability-based system. It suggests they understand that even precluding the end of the world or singularity scenarios, they are aware that their products can do great damage. The only way to get them to act on this knowledge is to force them to pay the costs, and not be able to say “don’t look at me! The regulators approved what I did!”

Another reason-the one that has received most commentary-is that the frontier firms know that they would be able to capture their regulator(s) by exploiting their information advantage and vast wealth. And as George Stigler pointed out long ago, the biggest benefit a regulator can bestow on its industry masters is a reduction in competition, including by the erection of entry barriers. As I’ve written with respect to financial regulation, regulatory costs are largely fixed and therefore benefit large incumbents at the expense of new, smaller, disruptor entrants.

There is another reason that is laid bare by some of the specific rhetoric and actions of Amodei, Altman, et al: a belief that intense competition between the frontier firms, let alone between them and open weight models, small models, etc., will be ruinous to them.

The cabal has been negotiating in public for “pacing,” i.e., a coordinated slowdown of model development. Competition on the capability dimension is hugely expensive, especially with respect to training. I view the various public statements-including the apocalyptic rhetoric justifying a slowdown-to be “cheap talk” intended to coordinate reduced investment. A way of cartelizing the industry.

Which leads to another motive for favoring regulation–protection from antitrust laws.

A coordinated slowdown would also address something else that I think looms over them. The whole AI sector is a bet on future demand, future prices (which depend on demand and competition from non-frontier implementations), the costs of production (e.g., data center investment and operational costs, including notably electricity), and the execution timeline (e.g., will bringing data centers online be delayed by supply chain bottlenecks or local political opposition?) Doubts have been increasing on all these dimensions, and I strongly suspect that the industry insiders fear that their massive investments (at all levels of the value chain) will not pay off if the current pace and growth are maintained. That is, the “pause” is not driven by safety fears, but by fears of a financial bloodbath.

It reminds me of Stalin’s pausing collectivization because he was “dizzy with success.” AI has been so successful that it has led to dizzying excesses that need to be reined in.

So there are many reasons for AI executives to want a pause. I think all are operative, and all push in the same direction–regulation by means other than liability.

But the nature of the industry and the product and the economics of harm reduction mean that a liability mechanism is the best way to address the safety concerns surrounding AI. That mechanism would give AI developers the strongest incentive to control the risks that they lecture us about daily. Anthropic doesn’t need OpenAI to mitigate the risks of its models, and vice versa. Internalize costs via a liability system and they will unilaterally have the incentive to do so.

Elon Musk enterprises are deeply embedded in the US national security establishment, most notably via SpaceX’s space ventures, but now via its xAI subsidiary and its Grok AI. The Department of War’s/Defence’s set-to with Anthropic demonstrates that an AI enterprise’s interests can potentially conflict with the Department, which could pose a security issue if the enterprise attempted to hold it up by restricting use in some way. The same risk could arise with Grok.

The embedded documents, prepared by @LibertyLynx and I, analyzes in detail the holdup potential for Grok. The basic conclusion is that the DoW has access to multiple AI providers, and can switch out at short notice–but not immediately. That matters, because in war, mere moments can count (what Patton called “the unforgiving minute”), meaning that even a fleeting withholding or even degradation of Grok (or another AI) could have dire consequences. In an optimistic case, a switch out would take 96 hours–pretty much the time that the ground campaign in Desert Storm lasted. A military eternity, in other words. That is, there are acute “temporal specificities.”

As the US military becomes more reliant on AI, especially for operational functions (e.g., targeting), the model relied on will become a potential single point of failure. As we analyze in the attached documents, failures can arise for purely technological reasons, or because of human interference with employment of the models–the latter being the holdup problem.

The reason that the Pentagon attempted to axe Anthropic was that the company demanded the power to withhold deployment of its models for purposes that it deemed incompatible with company values. That is, a deployed model posed holdup risks that the Pentagon deemed unacceptable.

It is troubling enough to extend someone as mercurial and megalomaniacal as Musk serious holdup power that the Pentagon assumed unacceptable in the case of Anthropic. If Anthropic’s values are problematic, what about Musk’s?

What is even more troubling is that Musk is deeply embedded in China. Specifically, Tesla is very dependent on China, to an extent that is barely appreciated in the United States–in large part because Musk hardly says anything about it, in contrast to his pontifications on pretty much every other subject known to man.

A majority of Tesla production is in China. Tellingly, Tesla’s output per unit of fixed assets is far higher in China than in the US. Moreover, Tesla gets appreciable funding from China. It hardly seems prudent to give an individual with such conflicting interests as Musk control over a vital defense function that requires complete reliability 24/7, and which is characterized by acute temporal specificity.

There is a fundamental economic challenge here. The government does not have a comparative advantage in developing these models, and is inherently reliant on the developers for continuous support of their operation. But given the specificity of these assets, what contracting and governance mechanisms can be put in place to protect the government against opportunistic behavior by the model developers? It’s an inherently daunting challenge, and those challenges are particularly acute when the counterparty is one of the world’s most egregious opportunists who is known for using his leverage to benefit himself. Of all the people to give leverage to . . .

Some problems could be mitigated by requiring Musk to sever his China connections. That would change his incentives, but would not change him.

Our research also raises questions about how disinterested Secretary of War/Defense Hegseth and his Chief Technology Officer, Emil Michael, were in their Anthropic decision, and their subsequent embrace of Grok.

The embedded documents analyze these issues in detail, and highlights issues that have received too little attention amidst the ongoing hype over space, digital data centers, and AI. Not that those are unimportant, and will the the subject of future similar posts.

It’s a story as old as time. An allegedly genius investor has a winning streak, which burnishes his genius image even more. He starts to believe it. Levers up. Then the wave he was riding crests and falls, the iron logic of leverage exerts itself. He becomes illiquid-if not insolvent-and either goes bust or gets bought out. And he’s no longer a genius, but an excuse maker.

One particularly famous example of this is Long Term Capital Management. LTCM had bona fide geniuses, Nobel Prize winners Myron Scholes and Robert Merton. Their genius apparently paid off for quite a while, with LTCM showing stellar risk-adjusted returns. Then, you know, stuff happens. Stuff like Russia effectively defaulting on its debt. Then your supposedly uncorrelated trades become suddenly correlated, and all start going in the wrong direction.

All that leverage that amped your returns now turns on you with a vengeance, and amps your losses. What’s more, your success has allowed you to grow to the point where you effectively are the market, which (a) moves the markets strongly against you when you try to liquidate some positions to meet margin calls and/or reduce leverage, and (b) pushes up the correlations even more.

Then, you are transformed from Golden Child to Systemic Problem Child. In LTCM’s case, it was so large that the Fed had to intervene to coordinate a liquidation of its positions into the tender loving care of the likes of Goldman.

Fast forward 28 years. The Golden Child and putative genius is Leopold Aschenbrenner, head of the oh-so-ironically named Situational Awareness hedge fund. The soi disant “Nostradamus of AI” implemented the picks-and-provisions strategy in the AI space, investing heavily in stocks of companies that provide the inputs to data centers and other AI-related activities. (SA also has private investments in AI firms like Anthropic).

He did time his entry right, and rode prices upwards. He (sound familiar?) amped his returns through extensive leverage (reputedly 4x). But-and there’s usually a but-the market peaked, and began to drop. And as has happened time and time again (it wasn’t just LTCM) the iron logic of leverage took over. Wrong way market moves beget margin calls on leveraged positions. Margin calls mean cash. Mark-to-market paper gains aren’t cash. So the imperiled fund was forced to sell positions to raise money-which only exacerbated the price declines.

Worse, the financial distress of a big position holder can’t be kept a secret. Evidently many others were apparently more situationally aware than Situational Awareness, and started to unload positions in stocks in which it was heavily invested. This front running exacerbated the price declines, and hence severely leveraged SA’s distress.

It happened so fast!, Aschenbrenner is no doubt thinking.

Sure did, (ex) Golden Child. Leverage comes at you fast.

And so do sharks. Jane Street, Millennium, and Citadel circled. In a single night, the last of these negotiated a deal to acquire all of SA’s publicly traded holdings at a discount estimated to be 10 percent.

The markets breathed a collective sigh of relief when news of Citadel’s purchase eliminated fears of a dumping of SA’s shares on the market: tech stocks rallied on the news.

Though no doubt his motives were purely self-interested, and are not as public spirited as J. P. Morgan’s intervention in the Panic of 1907, Citadel’s Ken Griffin did the markets a big favor. It is not from the benevolence of the butcher, the brewer, or the baker–or the hedge fund mogul–etc. (NB: I am not a reflexive Ken Griffin fan, as this old post demonstrates).

But as the saying goes, no good deed goes unpunished. Many-mostly Leopold fanboyz apparently-have argued that Griffin plotted SA’s demise (a reprise of CZ and SBF, maybe) by releasing a note predicted a Fed rate increase some time before the last Fed meeting. When the Fed did meet, and held rates steady-not raise them, mind you-the markets dropped, delivering the coup de grace to SA’s strategy that was already bleeding profusely. So per the fanboyz: See! Citadel caused this and then swooped in to take advantage!

Er, no. SA was already teetering. Citadel’s prediction-wrong, in the event-did not cause the final, fatal sell off. The Fed did.

For his own part, in a letter to investors Aschenbrenner rounded up the usual suspect-short sellers! It’s kind of pathetic. I have more respect for the natural gas fund operator who gave a blubbering, cringing apology when his widow maker strategy imploded.

If you are looking for other culprits here-beyond Aschenbrenner-look at SA’s prime brokers. Why did they supply so much leverage to back a poorly diversified strategy focused on a sector where sentiment had started to turn prior to the initial declines?

And if you want to do an It’s a Wonderful Life homage and wonder what would have happened if Ken Griffin/Citadel hadn’t been born, look at the disastrous unwind of another highly leveraged hedge fund-Archegos. Another example of an undiversified, highly leveraged fund caught in a doom loop when the markets turned against it.

In that instance, its prime brokers got together and agreed to assume the fund’s positions, and undertake a controlled liquidation. So far, so good. But one of them-Morgan Stanley-played “defect” in the prisoners’ dilemma facing the PBs, and sold $5 billion of Archegos’ holdings. Then the others jumped in and a full-scale fire sale occurred, leaving all the PBs badly singed. With Morgan Stanley (a la Margin Call) coming out best by getting out first (and not telling its counterparties what it knew about Archegos’ impending demise).

One firm-Citadel in this instance-didn’t face a prisoner’s dilemma. It-with the support of SA’s prime brokers-took the whole portfolio. Fire sale averted-to the markets’ relief (as noted above).

In sum, the Situational Awareness story is news, but not news. It is a story that has been repeated many times in many places. Leverage makes you look like a genius when the markets move your way. But with probability bounded away from zero, it will kill you when the market turns.

Much of the hysteria and euphoria about AI is implicitly based on the idea that AI will be free, or alternatively, that the marginal cost of AI will be close to zero. This is the basis for the belief that AI will replace everybody, rendering humans worthless, with all the benefits accruing to the creators of AI. (This is based on incomplete, not to say idiotic, economic reasoning, but I leave that for a later post). It is also the basis for the Always Look on the Bright Side of Life way of looking at the same prediction:

There won’t be dollars or scarcity of goods & services for individual consumption in the future

— Elon Musk (@elonmusk) July 17, 2026

But these predictions for ill or bliss are obviously, metaphysically wrong. AI is incredibly resource intensive. I just posted on power and water. But there is also the huge physical capital requirement–GPUs, with all the attendant capital costs they entail, buildings, wire, electronics, elaborate cooling systems, and on and on.

And as I’ve also posted on recently, businesses have been rudely awakened from their AI revery to realize that it is anything but free. It is damned expensive, actually, and will therefore require them to make discriminating choices on using it. “Fire everybody” is not a discriminating choice.

It seems there has been a ubiquitous category error in analyses of AI. It has been analyzed as if it is software. Software, once created, has effectively zero marginal cost. Hell, at one time the marginal cost was a floppy disk or CD: now it can just be downloaded for virtually nothing. Using software is also effectively costless. The cost of software is in its creation, and is largely fixed.

In contrast, although the creation of an advanced AI model is highly costly–that is, involves high fixed costs–using an AI model is also very costly. That is, the marginal cost of use is positive, because it involves consumption of scarce resources like power, water and GPUs.

Moreover, it is highly likely that AI is an increasing cost industry because the supply curves for AI inputs are upward sloping.

Software: high fixed cost, effectively zero marginal cost. AI: high fixed cost, positive and arguably high marginal cost.

All meaning that whereas people like Elon and other AI “visionaries” apparently believe that AI will invalidate the laws of economics–which are predicated on scarcity–this is a category error. Scarce resources are necessary to produce the AI that firms and individuals actually consume. The costs of these resources will drive the costs of such consumption, which will in turn lead to economizing decisions regarding where, how, and how much AI is used.

Almost all of those forecasting the future of AI (e.g., Elon) have ignored the constraining effects of this scarcity. Therefore, their prognistications–both dire and utopian–are worthless. Ignore anyone who does not place such scarcity front and center when attempting to discern how AI will reshape the world.

Like the fracking wars, but only much bigger: the current war on data centers. This confict has achieved some notable results, including an outright ban on new data centers in New York, and restrictions even in Texas.

Much of the opposition is organic. Some is astroturfed, and traceable back to China: it provides fertilizer for even of some of the organic growth. As with all these things, a lot of the rhetoric and analysis is fundamentally confused, but there are important economic issues here that deserve serious consideratoin.

First among these is the impact of data centers on power prices. The industry and its backers have pushed back with studies claiming that the impact of data centers is benign, or actually beneficial. For example, an Institute for Energy Research policy paper does a cross sectional analysis of state electricity prices and concludes that states with the highest concentration of data centers do not have higher electricity prices, and that data center penetration is only weakly positively correlated with power price increases.

This is a superficial, not to say stupid, analysis. With respect to the cross section, data center siting is endogenous: developers tend to site in lower cost areas. With respect to time series, a state level analysis is obviously too crude. Furthermore, the analysis does not take into account capacity utilization.

Furthermore, it is based on EIA data on “Average Price of Electricity to Ultimate Customers by End-User Sector.” These are effectively regulated retail rates that roll in historical capital costs, that are effectively fixed costs, meaning that retail costs are based on average costs. As the authors state: “Electricity systems have large fixed costs. If demand increases, those fixed costs are spread over more kilowatt-hours, reduction average costs.”

The problem is that these are backwards looking measures, and that when demand starts pushing towards capacity, those declines in average cost will stop abruptly and be replaced by sharp increases. As Herb Stein famously said, “If something cannot go on forever, it will stop.” And we are at–or even past–the stopping point. Certainly, given projected data center growth, if the stopping point hasn’t been reached, it will be soon. Meaning that hyperscaling will drive up power prices, unless power producing and transmission capacity is increased commensurately.

It has been recognized for some time that the US electricity generation and transmission system is becoming dangerously overstretched. New capacity–reliable capacity (more on this below)–is essential. Data center growth will only stretch the system more.

In the short term, electricity supply curves–the marginal cost curve, which is what is really relevant for pricing–become vertical, or close to vertical. Additions to capacity push out the supply curves. Those additions are vitally needed.

Want evidence? PJM–the Independent System Operator that manages the power grid in the region where much of the largest data center growth has occurred–uses a capacity market to incentivize investment sufficient to meet power demand. Roughly speaking, PJM determines how much capacity is needed to meet anticipated demand, and holds an auction for that quantity. High capacity prices are an indication that capacity is scarce.

Here are the last 5 PJM auction results:

Delivery Year	BRA Held	RTO Clearing Price ($/MW-day)	Change

2022/2023

2021

$50.00

–64%

2023/2024

2022

$34.13

–32%

2024/2025

2023

$28.92

–15%

2025/2026

July 2024

$269.92

+833%

2026/2027

July 2025

$329.17*

+22%

Oh, 2026/2027 doesn’t look so bad, right? Er, the price hit the FERC-imposed price cap. God only knows what would have cleared the market.

What should you be thinking reading this? Here’s an idea:

Looping back to the study mentioned earlier, the days of declining average cost are over–unless capacity increases to accommodate the higher demand.

The long run effect of data center-induced demand growth on prices depends on the long run supply curve, which depends on a variety of factors, including technology, capital costs, the long run energy supply curve (especially the long run supply curve for natural gas)–and policy.

All but the last of these are largely out of the hands of policymakers. So what policies can make the long run supply curve as flat as possible given these other factors?

Perhaps the most important policy lever is to incentivize efficiently investment in “behind-the-meter” capacity, i.e., vertical integration of new capacity and data centers that will consume the power it produces. This kills a couple of birds. First, it adds to generating capacity. Second, it doesn’t add to the demand for transmission. Given that transmission is the main vulnerability of the grid at present, this is an important consideration.

Some months ago, I analogized an xAI (now SpaceXAI) investment in gas generation co-located with its Memphis data center to mine mouth coal generation plants, and argued that transactions costs considerations provided an incentive for such vertical integration and coordinated co-investment. Those economic considerations apply with data centers, but the electricity market is highly regulated at the state and federal level, and those regulations can affect the incentives for BTM investment.

The crucial issue to provide the appropriate incentives to data center developers. One of the enduring economic lessons that I have learned (from my high school economics teacher, no less) is that a maxim of business success is to “externalize costs, and internalize benefits.” If the costs of connecting a data center to the grid, or the costs of transmission, or other factors that determine the opportunity cost of a BTM investment do not reflect actual costs, developers will have too weak an incentive to vertically integrate because they can externalize their costs. The opportunity cost to colocating generation with a data center is the cost of electricity drawn from the grid. If data centers do not pay the full cost of doing so, they will invest too little in building their own generation.

Data centers don’t want to pay high costs for power, obviously. But if their consumption from the grid is subsidized, they will consume too much from it, and build too little behind the meter capacity.

So, as always: Get the prices right. I say again: Get the prices right. Unfortunately, the politicization of electricity rate setting means that seldom happens.

Some of the biggest subsidies for data centers come from tax benefits granted by states. Just stop. Please. And if you can’t break the habit, limit the subsidies to generating capacity.

The other big policy change that can mitigate the impact of data centers on power prices is to terminate, with extreme prejudice, the monstrosity of “climate crisis”-driven laws and regulations that severely distort incentives to invest in generating capacity. These penalize investment in fossil fuel generation and and encourage investment in renewables. Absent subsidies, renewables are not cheaper than conventional generation. If you want cheaper electricity-where you strip out subsidies in determining cost-reverse these destructive policies.

This is particularly important when one considers that price is not the only relevant consideration here: reliability is too. Greater renewables penetration reduces reliability, all else equal. When you consider the true cost of power, which includes the cost of service interruptions (which can be extreme) and the cost of lower quality power, reliability is an important consideration that has been largely ignored in the force feeding of renewables onto the grid.

Ironically, the unreliability of a grid with a decidedly suboptimal generation mix may encourage behind-the-meter investment. Data centers need reliable power. If they can’t get it from the grid, they’ll decouple from the grid. But there can be too much BTM investment just as there can be too little. Both result from getting the prices (and incentives) wrong.

Some states have responded to impending overwhelming of the grid by data centers in a predicable way: by restricting or stopping altogether data center development. In the case of states like New York this response was foreordained by their previous perverted policy choices which were largely driven by climate hysteria. If they want a long run fix, they need to jettison their hysteria-induced policies.

Will they? As if.

This post has focused on power. Another sensitive issue is water. I don’t know much about the technical issues involved here, so I’ll just make one observation. Markets for water don’t exist, for all practical purposes. It’s almost certain that the prices are wrong. Meaning that it is almost certain that data center water consumption is inefficient. Knowing that doesn’t require knowing anything about the technology.

Unfortunately, it is necessary to remember Chesterson’s Fence in this context. That is, to ask “why aren’t there markets for water?” The answer is ultimately a political one, and as a result hopes for a market-based water pricing mechanism are in vain. Meaning that concerns about data center water consumption are likely to be addressed with a meat cleaver–no data centers!–or not at all–go ahead, build what you want! The discriminating, balancing decisions about water consumption that would occur when water is at least remotely priced correctly will therefore not occur, and the water issue will become a bone of political contention.

In sum, the correct approach to the data center war is Econ 101: GET THE PRICES RIGHT. The prices of connecting to the grid. The price of transmission. The price of energy. The price of reliability. The price of water.

And that’s why I am pessimistic about the ultimate outcome of this war. The products involved-power and water-are highly regulated and highly politicized. Regulation and politicization are inimical to getting prices right. Regulation and politicization were not imposed by aliens. They are the equilibrium outcome of political processes that are unlikely to change. Furthermore, the political salience of energy prices in an era in which “affordability” has heightened political valence means that pricing perversities are likely to get worse, not better.

Sorry to break it to you. But you probably didn’t come here to get sunshine blown up your backside.

Palantir made a long post on X regarding “sovereignty” from AI:

Our thoughts on the importance of AI sovereignty.

1. Your AI sovereignty dictates your institution’s future. Sovereignty is the precondition for choice. Relinquishing sovereignty transfers the future choices of your institution to others, who are likely to exploit it for their…

— Palantir (@PalantirTech) July 1, 2026

The post is a little esoteric, and requires an almost Straussian reading, but the basic theme is clear: protect your IP from prying AI. Fools and their IP are soon parted. This is especially true if you are lured by the supposed cheapness of Chinese models like DeepSeek.

Palantir’s Alex Karp was more straightforward in an interview (see especially starting around 9:20):

BREAKING $PLTR Dr. Karp Full interview CNBC@PalantirTech Dr. Alex Karp new $NVDA Deal

Palantir is profiting from Tokenmaxxing, because AI Labs can't deliver the value they charged customers.

Palantir product is 5 years+ ahead.

We deliver the best products to warfighters and… https://t.co/qtlbrSrcp5 pic.twitter.com/ooglNRobQL

— Mike (@MikeLongTerm) July 1, 2026

Karp is essentially warning against model lock-in, more than having your data scraped and used for training: there are contractual safeguards for that. This brings to mind the interoperability debates involving operating systems and applications (e.g., Office) in the late-90s and early-00s. But the issue with AI is much deeper, given the complexity of the models. Palantir is offering interoperability. But won’t that get you locked in to Palantir?

One area where I part ways with Karp is over pricing. Karp criticizes pricing AI on a usage basis, i.e., through charging for tokens. Instead, he envisions deals in which the frontiers get a share of he value created by their AIs.

Great thought. But such contracts require measurability and verifiability to a third party (a court). Maybe this would be possible for, say, a drug created by AI. You can measure the sales of the drug. But for something like improving the efficiency of an internal business process, measuring the savings and perhaps more importantly proving the value to a court (in the event of a dispute between the parties) is likely impossible. So for many of the AI use cases, such contingent contracting is economically infeasible.

And even the drug case is hard. How do you verify the portion of a company’s AI that is attributable to that project? Companies would have incentives to try to attribute AI usage for other things to the project.

Charging for tokens reminds me of the old case of IBM requiring users of its machines to buy punch cards from IBM. It was a way for IBM to price discriminate–extracting more value from intensive users. IBM couldn’t measure the value an individual computer user captured–which basically was determined by where the user’s demand curve was located–but figured that more intensive users had higher demands and more surplus to extract. By tying cards and computers, and pricing cards at above marginal cost, IBM was able to extract more surplus from the bigger demanders. That is, it could price discriminate.

Charging for tokens works the same way. Indeed, the price discrimination is more subtle, because of the different pricing levels and token restrictions based on access to different models, usage intensity, etc.

The industry is new, and changing rapidly. No doubt pricing models will change as well. There will be attempts to implement Karp’s vision, and some of those will work, but it is not likely to be the default pricing mechanism. Instead, some sort of token-based pricing, likely tweaked from its current implementation, is likely to remain in place. This is especially true for non-siloed, general business purpose uses of AI.

The biggest innovations will occur on the buy side. I’ve referred recently to some of the sticker shock that AI users had experienced, and how they are cutting back on AI usage. Meta (surprise, surprise) was arguably the biggest idiot.

Facebook (Meta) burned through an astronomical 60 to 73 trillion AI tokens in a single month—costing an estimated $221 million—because employees intentionally wasted them to climb a gamified internal leaderboard. Rather than a software glitch or an accidental AI loop, this massive token burn was the result of a corporate culture trend known as “tokenmaxxing.” [1, 3]

To encourage its 85,000 employees to adopt AI tools, Meta launched an internal, voluntary dashboard called “Claudeonomics” (or Clawonomics). [1, 4, 5]

• The Metric: The system ranked employees solely on the raw volume of AI tokens they consumed.
• The Incentive: Meta mistakenly treated high token usage as a proxy for employee productivity. Top users were publicly showcased on the intranet and rewarded with gamified corporate titles like “Token Legend” and “Cache Wizard.”

Because token consumption was tracked as an “input metric”—measuring how much AI you used rather than what you actually accomplished—employees figured out how to fake productivity:

• Ghost AI Agents: Engineers left autonomous AI agents running continuously in the background for hours to execute completely pointless, infinite research tasks.
• Massive Waste: The top individual “Token Legend” at Meta managed to burn 281 billion tokens in 30 days entirely on throwaway work—enough data to reproduce the entirety of Wikipedia more than 33 times over.
• System Strain: The unchecked AI overuse became so extreme that it even caused internal system outages and tech disruptions (SEVs) inside Meta.

Once details of the massive, multi-million dollar waste leaked to the public, Meta quickly shut down the leaderboard. The company has since clamped down on “tokenmaxxing” by introducing a centralized dashboard called AI Gateway, which enforces strict spending controls, real-time tracking, and automated alerts for unexpected token spikes.

Incentives matter! Who knew?

Apparently Meta didn’t, despite the fact that its whole business model revolves around manipulating its users’ incentives.

Other companies (e.g., Uber) have experienced similar, though less extreme outcomes due to incentivizing token usage. They are now adapting.

Which means that the demand curves facing AI firms will change dramatically, as companies experiment with different internal incentive structures–the design of which will be anything but trivial, for similar reasons to those cited above. How can a company determine the value a given employ can generate with AI, and design a system that incentivizes realization of that value rather than mindless waste.

As companies experiment, AI firms will adjust their pricing. Which will induce companies to modify their incentive structures. This will take a long time before things settle down. There will be learning on both sides, and that learning takes time.

No doubt companies will use AI to try to design incentive structures. And the AI firms will use AI to adjust their pricing. The use of AI might accelerate the learning process.

But at the end of the day I predict some form of usage-based pricing, with structures that incorporate a substantial degree of price discrimination. A long way from IBM and punch cards, but it rhymes.