Wolfgang Wopperer Notes Essays

The Political Economy of AI, Part 1: Selling the Future Short

・32 min read

The Political Economy of AI, Part 1: Selling the Future Short

Last year, everyone from the Bank of EnglandMitchell (2025) to Sam AltmanHeath (2025), Schiffer (2025) seemed to agree – we are in an AII’m using AI as a label for a) artificial (or rather: machine) intelligence in general, b) the current paradigm of generative AI, and c) large language models (LLMs) specifically. The individual meaning should be clear from the respective context. bubble. But almost halfway into 2026, it still hasn’t burst – and now there’s actual revenue in the field, and it’s growing fast: Anthropic announced in April that their annual recurring revenue (ARR) has jumped from $9B to $30B since the end of 2025,Anthropic (2026a). These numbers are gross revenue, while competitors mostly report net numbers, so a more realistic figure might be $22B – still momentous. making it, as Rogé Karma put it in a recent piece for The Atlantic, “possibly the fastest-growing business in the history of capitalism”.Karma (2026)

And the revenue may already be turning into profit: Anthropic is rumoured to be heading into its first profitable quarter. To close observers this marks a threshold – tracking the labs’ aggressive April pricing moves, Simon Willison concludes: “I think they’ve finally found product-market fit.”Willison (2026e); the profitability rumour he relays is from the Wall Street Journal.

So what’s going on? Have the AI boosters been right all along? Has “the burden of proof shifted to the naysayers”?Ibid. Or do we just not see the bubble because all of these developments form an integral part of its dynamics?

What type of bubble are we in?

David Karpf, building on a distinction made by Casey Newton,Newton (2026) differentiates between three types of bubbles:Karpf (2026)

  1. the “momentary-enthusiasm-deflating-once-reality-sets-in kind of bubble” (think crypto);
  2. “one where massive overbuilding and speculation may lead to periodic crashes and flameouts, but the resulting product transforms commerce and productivity” (think railroads);
  3. “a financialized Rube Goldberg machine on top of a limited-but-valuable product” (think 2008 housing crisis).

Critics like Ed Zitron see AI as exemplifying type 1 but have an increasingly hard time justifying that in light of recent revenue surges.See, e.g., Piper (2026). Newton thinks it’s type 2, highlighting the exploding demand underlying the revenue surges – “the biggest players are already selling about as much AI as they can make, and are scrambling to build or rent the infrastructure that will allow them to sell more.”Newton (2026). Anthropic had to strike a deal with their enemy Elon Musk’s SpaceX to (temporarily) solve their compute constraint (Anthropic 2026c). Karpf himself bets on type 3, pointing to circular investment dealsSee Sam et al. (2026) for an excellent interactive visualisation. and arguing that the demand is distorted because AI subscriptions are heavily subsidised:

AI companies have not generated the demand to pay for the expensive data centers. They’ve generated and subsidized the demand for compute, and then used that computational demand to make a case for further AI infrastructure investment.Karpf (2026); emphasis mine.

This sounds convincing – Anthropic’s highest-tier Claude subscription, for example, gives subscribers roughly $5,000 worth of compute for a fee of only $200.Browne (2026) That’s a 25× subsidy – subscribers “pay for practically none of that compute, and would stop using AI if they had to pay the direct costs.”Karpf (2026)

Consumer subscriptions are only part of the picture, though – most of Anthropic’s reported revenue doesn’t come from subsidised monthly fees, but from API use, metered and paid per token (roughly: unit of information to process). According to independent estimates (Anthropic doesn’t publish a revenue composition breakdown), only 10–15% are from direct consumer subscriptions; 85–90% are API and cloud-reseller token revenue and Claude Code enterprise contracts, which since April 2026 bill usage at the full listed API rate rather than with discounts.Lemkin (2026), Sacra (2026a), Willison (2026e). At OpenAI, the composition is roughly the inverse but moving very much into Anthropic’s direction (Sacra 2026b, Willison 2026e); more on that below. This coincides with anecdotal evidence, quoted by Karma, of companies

“overrunning their initial budgets” for AI tools “by orders of magnitude,” with some companies already spending as much as 10 percent of their total engineering labor costs.Karma (2026)

Moreover, Karpf assumes that high unit costs for intelligence are a durable pattern, not a temporary phenomenon: Because it would reduce its ability to raise capital, he contends, “the AI industry is not the least bit interested in pursuing efficiencies that would make their products less computationally demanding”.Karpf (2026) This is patently false – both Anthropic and OpenAI are improving the efficiency of their models.Interestingly, Google seems to pursue a different strategy: In its latest model generation, both token use and prices have gone up significantly. While OpenAI and Anthropic’s latest generation has been optimised for intelligence vs token use and cost, Google’s actually performs worse in this respect. Instead, it’s optimised for speed vs intelligence and multimodal capacities (Artificial Analysis 2026b, 2026c). This points to contexts where answer speed and input variety matter most: at the integration points with Google’s existing surfaces, e.g. search, Chrome, YouTube – to avoid introducing friction which could damage user experience. This is the strategy of an aggregator who wants to benefit from agentic AI without disrupting its existing business model (Fried 2026). To give just a few examples: In agentic coding, Claude Opus 4.7 uses 40% fewer tokens while achieving 6.5% better performance than 4.6;Opus 4.7 (high) vs Opus 4.6 (max), according to an internal benchmark (Anthropic 2026b). GPT-5.5 can match GPT-5.4’s intelligence at less than half the cost;GPT-5.5 (medium) achieves the same Artificial Analysis Intelligence Index as GPT-5.4 (xhigh) at 42% of the cost (Artificial Analysis 2026a). and as Kelsey Piper points out, “GPT-4-level intelligence is about ¹⁄₁₀₀₀ as expensive as it was when GPT-4 was released.”Piper (2026), citing Demirer et al. (2025). Increased efficiency is partially offset by higher per token prices, though – “OpenAI’s GPT-5.5 [is] 2x the price of GPT-5.4, and Claude Opus 4.7 is around 1.46x the price of 4.6” (Willison 2026d).

But then what are the patterns here? What type of bubble are we in – if any?

In order to find out, we need to understand the nature of the demand surge and how it relates to actual value creation.

Compute demand and productivity gains

As Ben Thompson notes, the surge in compute demand reflects a step change in how large language models are being used. He highlights two particular aspects of that change:Thompson (2026)

  1. Reasoning, i.e. models basically talking to themselves before answering, has made them more intelligent and hence useful because they can answer more complex questions while course-correcting on the way. The price is that they use a lot more tokens.Of the tokens used by frontier models to run the evaluations in the Artificial Analysis Intelligence Index, on average ~90% are reasoning tokens (Artificial Analysis 2026a).
  2. So-called “agent harnesses” that give models access to deterministic tools they can use to do research, use computers, and check their own output for correctness further increase the models’ usefulness and enable them to autonomously handle complex and long-running tasks – again at the cost of using a lot more tokens.Agentic workflows increase token usage not only because they involve multiple steps, but also because the full context (i.e. conversation history) needs to be re-processed at every step, including tool calls, as the models have no internal state or “short term memory”.

But with the shift to agentic AI, models have not only become much more useful and compute-intense, they also require less human attention to do useful work – “it actually won’t require that many people to drastically increase the amount of compute that is actively utilized to create products with meaningful economic impact”.Thompson (2026)

In other words, it’s become possible for fewer people to do more with AI – a real productivity increase, available not in a more or less distant future, but right now. Software development specifically has already changed fundamentally: because “writing code is cheap now”, as Simon Willison summarises, “one human engineer can be implementing, refactoring, testing and documenting code in multiple places at the same time”, multiplying the amount of work they can get done just by themselves.Willison (2026c)

As a result, not many users are needed to generate the enormous compute demand we see. And because in an enterprise context, not needing many users means not needing many workers, this directly translates into an incentive for companies to replace people with agents. The most ambitious version of this story doesn’t stop at task automation – it talks about collective cognitive capacity. As Azeem Azhar puts it: “as intelligence becomes cheaper and faster, the basic assumption underpinning our institutions – that human insight is scarce and expensive – no longer holds.”Azhar (2025) If that is right, the productivity surge is not just fewer workers per task but fewer humans per firm as more and more of their cognitive contributions can be outsourced.Block CEO Jack Dorsey made this argument in a letter to shareholders when Block fired 40% of its workforce in February 2026: “Intelligence tools have changed what it means to build and run a company. We’re already seeing it internally. A significantly smaller team, using the tools we’re building, can do more and do it better.” (Dorsey 2026)

Thus Thompson concludes that we are, actually, not in a bubble:

The economic imperatives are going to be impossible to resist, and will fuel demand for even more compute over time, […] supporting the case that this is no bubble.Thompson (2026), emphasis mine.

This is also where the notion of a “permanent underclass” gains currency: As Jasmine Sun reports in the New York Times, it is by now a “San Francisco consensus” running across factions – engineers, venture capitalists, doomers, accelerationists, lefties, libertarians – that advanced AI will “displace millions of jobs as fewer humans are needed to make the economy run” and “ferry power and wealth to the AI companies and the existing owners of capital”.Sun (2026). Sanders (2025) models potential job losses across US sectors, arriving at a headline number of 97M in the next ten years, and lists concrete lay-off plans and announcements for several major corporations.

But note that this prediction hinges on one particular assumption: that AI is used to create products with meaningful economic impact.

Productivity gains and value creation

I take “products with meaningful economic impact” to be necessarily based on some fundamental economic value, otherwise we’d be back to Karpf’s type 1 bubble – just like Bitcoin, AI would only be “valuable because investors have agreed to attach abstract value to it as a speculative asset”.Karpf (2026)

Where is this value being generated, and by whom? And is it being created now, or only borrowed against a future that will have to repay it?

The conventional story, which we’ve implicitly followed so far, is that agentic AI enables economic value creation for businesses across the economy, right now. It goes roughly as follows:

  1. Agentic AI enables significantly smaller teams to do more and do it better.Dorsey (2026)
  2. This makes companies as a whole more productive.Thompson (2026)
  3. This raises the demand for agentic AI.
  4. To satisfy this growing demand, we need more compute capacity.
  5. To create compute capacity, we need to build out infrastructure.
  6. To build out infrastructure, we need huge investments, and soon.
  7. To enable these investments, we need innovative investment vehicles and structures.

Gigawatts of compute, trillions of funding, circular investments and the resulting market concentration: all needed to serve a demand created by the fact that agentic AI helps businesses create more value.

But does it? Does agentic AI actually help businesses create more value?

Where better to look for an answer than in the most advanced industry,Anthropic (2026d) the one by general agreement best-suitedBecause of its highly structured problem space and formal-language focus, and the amount of available data for post-training, i.e. reinforcement learning which gives LLMs distinctively agentic capabilities. See, e.g., LeCun (2026). to benefit from agentic AI: software development.

Pete Steinberger, creator of the category-defining multi-purpose agent OpenClaw and, since early 2026, an OpenAI employee, might well be the poster child for agentic coding as the future of software development. His three-person team keeps around a hundred Codex agents running in the cloud around the clock: to open pull requests (PRs)A pull request (PR) is a proposed set of code changes submitted for review before being merged into a shared codebase – the standard unit in which software work is reviewed and integrated. off the project’s roadmap, review every PR, scan commits for security holes, deduplicate issues once fixes ship, monitor benchmarks and report regressions in Discord, and even sit in on the team’s meetings to start PRs for whatever gets discussed.

In May 2026 he posted that month’s bill – $1.3M worth of tokens in thirty days, i.e. roughly $13,000 per agent – and framed it as “trying to answer the question: how would we build software in the future if tokens don’t matter?”Steinberger (2026); the setup is also reported in The Decoder (2026) and Browne & Davis (2026a). Steinberger had joined OpenAI in February 2026, which is why the API bill is covered internally. Steinberger’s experiment has quickly become the latest expression of where the edge is in AI: in “token-maxxing“Browne & Davis (2026a), TheStandup (2026) .

Another much-discussed experiment in agentic coding is the full rewrite of the popular JavaScript package manager and runtime Bun in a different programming language,From Zig to Rust. done completely by coding agents with 6,755 commits in just 6 days. That it worked at all is testament to the productivity jump massively parallelised agentic coding enables; but it also shows exactly where the paradigm’s breaking points lie.Browne & Davis (2026a), zach (2026)

AI translation can verify that each function behaves identically to its original in isolation, but it can’t preserve global patterns – the design constraints that live in the original author’s head rather than in test assertions. And code that was legal before may now be undefined behaviour and thus exploitable. Automated tests don’t help with this challenge: They are a record of what has already been seen and fixed; pass rates certify only that historical correctness has been preserved, not that the rewrite has introduced no new categories of wrongness – which in a JavaScript runtime propagate into every application that depends on it, and there’s a lot of applications that depend on Bun (first and foremost Claude Code itself).

The larger pattern here is that agentic coding tends to produce complexity at an accelerating pace. Agentic AI boosts developer productivity by increasing raw output (features, PRs, commits, lines of code), but this increase produces hidden longer-term costs – which often means the overall economic value is actually negative.

James Shore illustrates this logic with numbers for maintenance costs per lines of code:

Every line of code you write has to be maintained: bug fixes, cleanup, dependency upgrades, and so forth. […] According to our crowd’s maintenance estimates, you’ll spend more than half your time on maintenance after 2½ years. After ten years, you can hardly do anything else. […] If you want a productive team, you have to focus on their maintenance costs.Shore (2026)

If agentic AI doubles developer productivity, i.e. allows you to write twice as much code as before, you need it to also halve per-line maintenance costs, or you turn net-negative within a few years and stay worse off thereafter.

If you double your output and your cost of maintaining that output, two times two means you’ve quadrupled your maintenance costs. If you double your output and hold your maintenance costs steady, two times one means you’ve still doubled your maintenance costs.⁠⁠ Instead, ⁠⁠you have to invert your productivity. If you’re producing twice as much code, you need code that costs half as much to maintain. Three times as much code, one third the maintenance.⁠⁠Ibid.

And this is generally not the sort of code AI produces.

For Mitchell Hashimoto, co-founder of HashiCorp and creator of Ghostty (and an avid user of agentic AI himself), these developments are a reason to sound the alarm. He sees them as an accumulation of unacknowledged risk, driven by an unquestioned belief in a short-term productivity narrative, and creating potentially enormous but displaced longer-term costs:

I strongly believe there are entire companies right now under heavy AI psychosis and it’s impossible to have rational conversations about it with them. I can’t name any specific people because they include personal friends I deeply respect, but I worry about how this plays out. […]

The main issue is I don’t even know how to bring this up to people I know personally, because bringing this topic up leads to immediately dismissals like “no no, it has full test coverage” or “bug reports are going down” or something, which just don’t paint the whole picture.

We already learned this lesson once in infrastructure: you can automate yourself into a very resilient catastrophe machine. Systems can appear healthy by local metrics while globally becoming incomprehensible. Bug reports can go down while latent risk explodes. Test coverage can rise while semantic understanding falls. Changes happens so fast that nobody notices the underlying architecture decaying.

I worry.Hashimoto (2026)

AI psychosis is not just something that happens to susceptible chat users – in software developers, its phenotype is productivity-induced tool addiction, and it’s not a fringe phenomenon.See, e.g., the OpenCode creator’s psychosis story, discussed in TheStandup (2026). I have experienced mild forms of this myself. What an agentic workflow produces first is a feeling of empowerment: a dozen agents dispatched at once, output flowing back with surprisingly good results, a sense of operating at 10× productivity. That feeling is also the hook. Steve Yegge calls where it leads the AI vampire – push the tools hard enough and you work at 10× for eight hours, your employer captures the value, and you are left “drained to the point of burnout”, the cognitive load of supervising agents so high that Yegge puts the sustainable ceiling at about four hours a day.Yegge (2026) via Willison (2026b). His framing is that AI automates the easy work and leaves the developer with the hard decisions, abstract summaries and problem-solving – which is why he reckons four hours a day is a realistic ceiling. Reports about how agentic AI is increasing cognitive load eroding attention are starting to proliferate; see, e.g., Newport (2026) and Wilson (2026). My own experience corroborates this. Empowerment morphs into compulsion, and compulsion into exhaustion.

To make matters worse, the other side of this development is skill atrophy: the supervisory skills agentic AI would require to mitigate these tendencies are the same skills its routine use erodes – debugging, mental simulation, the felt sense of why a particular abstraction matters, a realistic assessment of complexity levels. The result is a supervision paradox: you need the very capacities that the tool atrophies in order to oversee what it produces.Faye (2026). The pattern is documented beyond software: a 2025 survey of knowledge workers found that higher confidence in AI tools correlated with measurably reduced critical-thinking effort (Lee et al. 2025); and endoscopists who had routinely used AI-assisted polyp detection showed a 20% drop in detection rate once the AI was withdrawn (Budzyń et al. 2025).

So instead of maximising downstream value creation through long-term productivity, agentic AI teases with short-term productivity – and then burns through as many tokens as possible (and as many humans as needed) to generate value upstream. The present is drawn down now – tokens billed, hours poured in – and the bill, burnout and eroded judgement, arrives later, landing on the worker rather than the lab that engineered the dependency. Linear CEO Karri Saarinen asks pointedly :

We keep hearing about 10x or 100x productivity gains in engineering and knowledge work.

But outside the model labs, I haven’t seen the corresponding 10-100x revenue growth across the market or increase in quality.

So where is the productivity going?Saarinen (2026). Of course Linear being positioned as a complementary to agentic AI, Saarinen has a strong incentive to point out that agentic AI alone isn’t good enough. That doesn’t make him wrong, though.

A general pattern

The pattern we’ve just seen in software, an extractable present, paid for with a back-loaded liability, recurs at every layer of the AI build-out. Borrowing a term from finance, we could call it temporal arbitrage.

Enterprise adoption

While enterprises spend a lot more on AI, there are no objective productivity gains yet. Take the example of Uber: Its CTO announced in April that the company had already burned through their AI budget for the whole year,Bratton (2026). Uber has now started limiting all employees to $1,500 in monthly token spending per AI coding tool (Lung 2026). while its COO reports that the link between token usage and useful consumer features “is not there yet”.Rapid Response (2026). See Mills (2026) on how “corporate leaders are starting to question whether soaring AI spending is delivering meaningful returns” more broadly.

Where does this continued willingness to pay come from – what makes companies like Uber, as Simon Willison says, “suck air through their teeth and then say yes”?Willison (2026e) Seen through the lens of temporal arbitrage, there’s a clear motivation: as the Wall Street Journal puts it, tech companies are now

in effect playing a game of chicken with each other on capital-spending plans. They are shelling out as much as they can – more than their rivals, they hope – on AI chips and data centers that could put them in the lead in a race they feel they can’t afford to lose. That in turn is heightening competition over who can use AI to help do more with a lot less, freeing up money to spend on expensive chips.Gallagher & Fitch (2026)

This recasts what Willison reads as product-market fit. The concept assumes willingness to pay is driven by value gain, when in reality it is driven by an investment arms and a productivity promise. The enterprises now paying full API prices are revealing less about the technology’s value than about their fear of being left behind – the very fear the labs are working to manufacture. What presents as fit is willingness to pay decoupled from value.

Alex Imas, Google DeepMind’s Director of AGI Economics, describes this as a situation driven by

a narrative where if you’re […] not laying people off, then you’re seen as not adapting AI enough. Then you’re going to just get a cascade effect of firms needing to keep up with the Joneses in terms of starting to lay people off. […] The firm might actually be worse off after the layoffs than before, but it’s just doing the layoffs to have the perception that, “Look, we’re not behind the times. We’re using AI.”Patel (2026)

Companies replace workers with agents to free up money so they can spend it on AI, and to be seen as keeping up. This perspective reverses the conventional story, where the lay-offs are driven by productivity increases, with investment growth as an outcome, not the goal. The reality is quite different – an expensive arms race drives investment, and workers are paying the price for capital’s willingness to spend. Uber’s CEO states it verbatim: they are paying for their increasing AI investment by hiring less.Bitter (2026) This labour-to-capital transfer is the political shape of the AI build-out.

Anthropic’s strategy to capture as much as possible of this transfer is bait-and-hook: Subsidising developer subscriptions at roughly 25× their compute cost is the bait, only being allowed to use the subscription on Anthropic-controlled surfaces is the hook.On the lengths to which Anthropic goes to enforce their restrictions see Browne (2026). Developers form an attachment they then carry into their employers’ procurement decisions – analogous to how Slack rode product-led growth into the enterprise from the desks of individual engineers. And they not only expect Claude, but a certain amount of compute – driven by Claude Code’s addictive and gamified UX,Discussed, e.g., in Browne & Davis (2026b). paid for by the enterprise at an unsubsidised price. A supply-side subsidy is converted into demand-side lock-in, and the arms race narrative supplies the urgency that makes both feel inevitable.

OpenAI is playing catch-up with a slightly different approach. Where Anthropic has built a developer following early on and is now trying to lock it in, OpenAI is courting the same early adopters through closer developer relations and models tuned to be more meticulous, token-efficient – and well-suited for large-scale orchestration, as showcased in the OpenClaw token-maxxing experiment. At the same time, it prises open their distribution: In April 2026 OpenAI ended the seven-year exclusivity that had made Microsoft Azure the only place to run its models and put GPT-5.5 on Amazon Bedrock – where a large share of enterprise AI is actually procured, and where until then Claude had the frontier largely to itself.CNBC (2026)

The competitive logic is identical for both: win the developers, then meet enterprises on the cloud they already buy from, and inflate their willingness to pay with an arms-race narrative. A present cost is justified by a long-term threat (or promise) – temporal arbitrage again.

Accounting and financing

The same move governs how the hyperscalers keep their books: a present value is booked now, and the matching liability is pushed into a future where no one will trace it back. There are two primary levers for this:

Depreciation: spreads an asset’s cost across the years it earns money – sensible for fibre or railways. But a frontier GPU is obsolete in two or three years, superseded by a generation that does the same work for less, so you would expect it written down faster than ordinary hardware. The hyperscalers did the reverse, stretching assumed server life from three or four years to six between 2017 and 2024 – which on the 2025 spending base roughly halves the reported cost of those chips and lifts present earnings to match.Cerno Capital (2025): extending official server life from 3 to 6 years reduces collective hyperscaler depreciation from ~$51bn to ~$28bn on the 2025 capex base. Michael Burry, the investor made famous by The Big Short, estimated the resulting gap at $176bn across 2026–2028 – all money displaced into the future.Burry (2025). The hyperscalers have not formally rebutted the underlying analysis; Amazon partially walked back its extension in early 2025, reducing server life from 6 to 5 years for a subset of assets. The earnings are booked now; the write-down waits for the day the assumption underneath (that superseded chips are still worth something) is finally tested.The value cascade is the industry’s answer to depreciation pressure (theCUBE Research 2025): GPUs migrate from frontier training to high-value inference to low-value batch workloads as they age. It assumes inference demand keeps scaling fast enough to absorb the cascaded hardware; if that assumption fails, it forces exactly the write-downs the schedules were meant to defer.

Debt: Until 2024 the hyperscalers funded data centres from their own cash flow; by 2026 they will spend roughly 90% of it on capex, up from 65% a year earlier, and because that’s still not enough, they make up the shortfall with borrowing – increasingly debt that never appears as theirs.Bank of America estimates, January 2026; Morgan Stanley estimate for hyperscaler borrowing. Global tech companies issued a record $428.3bn in bonds through early December 2025 to fund AI data centres (Sigalos 2025). The device for that is the special purpose vehicle (SPV): a separate company that owns the data centre and raises the debt, while the hyperscaler takes a minority stake and signs a long lease. As a result, the loans sit on the SPV’s books and it records only rent.Bank for International Settlements (2026) Meta’s $27bn Hyperion campus is the template – a private-equity-managed vehicle owns 80% and borrows, Meta owns 20% and leases it back for twenty years, and the borrowing never touches Meta’s accounts.The Hyperion model is “the first hyperscale project to combine off-balance-sheet treatment, investment-grade credit, and a residual value guarantee” – an architecture modelled on aircraft leasing, with insurers providing residual-value insurance (Global Data Center Hub 2025; Meta 2025).

What does not go off the books is the risk – it has to be absorbed somewhere. Repackaged as investment-grade infrastructure paper, it disperses into pension funds, insurers and the index funds that hold the hyperscalersJP Morgan estimates AI-linked companies will account for 14% of its investment-grade index in 2026. With roughly 40 cents per dollar invested in S&P 500 funds going into the “Magnificent Seven” (Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta and Tesla), ordinary retirement savers are exposed to AI build-out risk whether they choose to be or not (Hera Lee 2026). – more than $120bn has been moved off balance sheets this way,Emanuel (2026). Microsoft is developing a similar architecture with BlackRock’s Global Infrastructure Partners. Oracle has structured over $60bn across multiple SPV deals; xAI closed a $20bn round in January 2026 partly structured through a GPU-specific SPV with debt secured by the hardware rather than xAI’s corporate assets. spread thin enough that no single vantage point can see the whole of it, which is exactly the point. Booked earnings and off-balance-sheet capacity are the extractable present; the deferred write-down and the dispersed, untraceable risk are the back-loaded liability. Temporal arbitrage, in the language of accounting.

Energy

Power is quickly becoming the most binding constraint for the build-out. The industry’s narrative response to it has been a wave of long-dated nuclear PPAs, with an aggregate capacity in the 10 GW ballpark.Microsoft’s $16bn, 20-year, 837 MW deal to restart Three Mile Island; Amazon’s expanded Talen Energy PPA for 1,920 MW through 2042; Meta’s package with Vistra, TerraPower and Oklo for up to 6.6 GW by 2035; Google’s Kairos Power SMR agreement targeting 500 MW by the early 2030s (Chernicoff & Vincent 2024). These deals supply the carbon-free credentials the industry needs for its sustainability reporting and that justify ESG-compliant institutional capital.Nuclear PPAs are reputational infrastructure, not power infrastructure, for the 2026–2030 window: SMR deployment timelines target 2030–2035 even on optimistic assumptions; first-of-a-kind units cost 2–3× per kilowatt more than proven large-reactor designs; only four SMR units are operating commercially worldwide as of early 2026; the HALEU fuel supply required by most advanced SMR designs is currently a US production gap of orders of magnitude (World Nuclear Association 2026). What is actually getting built in the 2026–2030 window is natural gas: more than 100 GW of new gas-fired capacity have been announced in the past few years, much of it “behind the meter” (on developer-owned sites) to avoid the regulatory oversight a utility would face. Wired’s analysis of permit documents found that gas projects linked to just eleven data-centre campuses around the US have the potential to emit more than Morocco’s entire 2024 emissions.Taft (2026) The nuclear PPAs provide decarbonisation rhetoric while gas plants front-load operational power for the present and back-load 30–60 years of carbon and methane emissions.Methane leakage upstream compounds the official CO₂ figures by a factor that does not appear in any permit. Combined with 30–60 year operating lifetimes for gas plants, the data-centre build-out is committing decades of fossil infrastructure that will outlast both the GPUs it powers and the corporate decarbonisation commitments that nominally constrain its builders.

Each of these layers monetises an extractable present against future costs that are made invisible to the actors making present decisions. The 2032 engineer who can’t reason about their codebase will be told the cause is “modern software complexity”; the 2032 hyperscaler write-down will be attributed to “unexpected architectural shifts”; the 2050 climate impact of behind-the-meter gas will be attributed to “the energy transition’s complexity”.

That all four arbitrages operate at the same time, in the same firms, supporting the same capital cycle, is not a coincidence. It is the structural signature of a system that has committed to a future and must now produce the conditions to absorb its present commitments.

A different story

This means we must reverse the assumed causality of the AI build-out (productivity gains create demand for inference, which justifies infrastructure investment): in reality, infrastructure investment requires demand growth, which requires the productivity narrative, which requires agentic AI as the absorptive vehicle.

Following Stafford Beer’s principle that the purpose of a system is what it does,Often abbreviated POSIWID; Beer (1979). this reversal shows that the agentic-AI paradigm is better understood as an AI capex system: a system that consumes vast quantities of capital and energy, produces inference services, and emits narratives about future productivity that generate demand for these services.

This is what Karpf was gesturing toward when he located AI in the financialised rather than the infrastructure category of bubble. And it is what makes the housing-bubble parallel stronger than the railway one: the largest – and process-driving – economic value isn’t being captured in the production of inference; it is being captured upstream of it, in the financial markets that the capex programme legitimates.

There are a few corollaries to the AI-as-capex-system view:

  1. Token consumption expands to fill available compute. Agentic systems are generally designed in ways that maximise token consumption: parallel agents, long-running background tasks, verbose reasoning chains, automatic context retrieval, retry-on-failure loops. Within that, labs are optimising for intelligence per token, which lowers the cost of a given task.Within the agentic paradigm, the intelligence per cost gains are incremental. The dramatic GPT-4-level price drops were largely an effect of the paradigm shift to thinking models, not a sustainable trend within the new paradigm. But instead of banking the saving, users spend it on tasks that were not previously worth doing: more agents, longer runs, deeper search. Efficiency is reinvested in consumption, so aggregate compute demand rises on the back of the very gains that were meant to contain it. This is Jevons’ paradox at work:First stated in Jevons (1865). Efficiency does not curb consumption of a resource whose demand is elastic – it amplifies it. Steinberger’s own framing of the OpenClaw experiment adds intention to emergence: they actively push the boundary of token consumption to find out what software development becomes once token cost is treated as irrelevant. Maximal consumption is not a symptom of immature tooling, but a strategic direction that shapes features like Opus 4.8’s new ability to run “dynamic workflows” – “orchestration scripts that run tens to hundreds of parallel subagents in a single session”.Anthropic (2026e). Add to this that agents, particularly when using sub-agents, have just been shown to “take unsafe (incorrigible) actions when those actions are instrumental to task completion” (Tien et a. 2026, 2), and you have a recipe for the “Challenger disaster” of agentic AI. Token-maxxing not as a posture but a design principle.

  2. The productivity-promise gap is persistent and structurally invisible. The persistence of the gap is what Shore, Hashimoto, zach and others show for different aspects. It can stay invisible because productivity measurement is front-loaded (immediate output) and costs are back-loaded into a future maintenance burden that does not get attributed back to the original cause. The same pattern is visible across enterprise deployment: lay-offs have already been walked back or are projected to be, but the productivity story persists.MIT NANDA (2025); Klarna’s reversal of AI-driven lay-offs has been widely covered; Orgvue (2026) reports 32% of organisations that made AI-linked lay-offs have rehired, with 55% regretting the cuts. Gartner (2026) projects roughly half of AI-driven lay-offs will reverse by 2027. We can hear stories about a lack of productivity gains and people dismissing them as rationalisations of a painful product-market fit. The result is a special sting in the “permanent underclass” thesis: the displacement is real (people are laid off, wages compress, the work degrades), but the productivity gain that is supposed to justify it is not. The human cost lands in the present; the economic value that would warrant it stays in the narrative.

  3. The justification narrative escalates in ambition as resource demands escalate. As US capex numbers have moved from $200bn in 2024 through $700bn in 2026 to a projected $1.6tn in 2031, the justification narrative has escalated from “AI tools improve productivity” (2023–24) to “your firm needs to invest or be left behind” (2024–26) to Steinberger’s “build as if tokens don’t matter”. Meanwhile, the “AGI is imminent” and “race against China” narratives are adding weight and urgency.Capex numbers from Goldman Sachs Global Institute (2026); McKinsey’s widely-cited figure is $5.2–6.7 trillion through 2030 (McKinsey & Company 2025). Each turn makes a more ambitious claim than the last; each is doing more demand-justification work than its evidence base supports. If the productivity story were tracking ground truth, you would expect claims to moderate as negative evidence accumulates; instead they escalate to keep pace with the capital ask. This is the hyperstitional signature of a narrative that helps bring about its own truth.

  4. The infrastructure is being shaped to require agents, not to serve users. The 20-year SPV terms, the 30–60-year gas-plant lifetimes, the 20-year nuclear PPAs, the 6-year GPU depreciation schedules – these are not the timescales of a market discovering what users want. They are the timescales of an industry that has committed to a future and must now generate the demand to fill it.

Far from being evidence that we are not in a bubble, agentic AI is the bubble. It is the very mechanism by which AI gets financialised: The justification (productivity), incentive (arms race) and physical implementation (ultra-large-scale cloud computing) that enable the inflation of financial investment.

This allows for an interesting interpretation of Thompson’s line, “it’s a bubble when people stop believing it’s one”: His account of the situation is part of the productivity narrative that inflates AI demand, so it has to spread the belief there is no bubble – that is its job within the bubble.Thompson (2026). The line captures, intentionally or not, exactly the structural feature the inverted-causality account predicts: when the narrative breaks, the cascade follows because the narrative was load-bearing for everything else.

The two main counter-arguments to defend the productivity narrative turn out to play a similar tole:

The first is that you can solve the supervision paradox by using AI to do the review. But that’s not a solution – it’s a further recursion: reviewing the agent’s work becomes another agent task, the supervision paradox displaces one level up, and crucially, the recursion consumes more inference in doing so. Every diagnosed problem produces a prescription for additional inference. Which means: The proposal performs the function the system requires of it.

OpenClaw development is the live demonstration. Steinberger’s hundred agents already spend much of their effort reviewing one another – scanning every PR, hunting security holes in commits, watching benchmarks for regressions. The supervision has not been dispensed with; it has been handed to more agents, which is exactly why the token bill runs to seven figures. The recursion is not a flaw in the setup – it is the setup.

The second is that the existence of individual users who derive genuine value from agentic AI refutes the systemic claim of a productivity gap. But it doesn’t: Individual value is the mechanism by which the higher-order system maintains itself. Lower-level systems (users, teams, organisations) must experience enough local productivity gain to keep adopting, deploying, paying, and generating the demand signals the higher-level system needs to legitimate its capital flows.

If users derived no value at all, the system would collapse; if users derived sustained net-positive value, no narrative production would be needed. The middle case – sufficient local value to sustain adoption, insufficient to justify the aggregate capital cycle on its own merits – is the equilibrium the system requires, and the one we appear to be in. The important restriction is that the productivity gain is limited (in time, scope, scalability) – contrary to what the productivity narrative claims.

How big is the bubble?

The best way to describe the AI capex system is as a temporally-coupled capital cycle in which present value is extracted, future liabilities accumulated, and narratives deployed to keep the latter invisible to the actors making decisions about the former.

Its fragility is located not in any single layer but in the coupling itself. Any sufficiently large shock to the productivity narrative – whether from a technological paradigm shift, from the accumulated practitioner experience of broken promises, from a high-profile residual-value test, or from regulatory disclosure of SPV exposure – will propagate through every other layer, finally bursting the bubble.

Maybe I’m wrong and we already are in the future that I think we’ve only been promised so far. But if I’m not and the productivity narrative is critical to the economics of AI, then the bubble will burst when belief in it breaks. And market exposure to it is so enormous, both in terms of national accounting and market capitalisation, that the burst might well dwarf that of the subprime bubble – and we already called that one the Global Financial Crisis.

Two numbers show why the burst will be systemic rather than sectoral. The so-called Magnificent Seven now account for around 40% of the S&P 500, Nvidia alone for close to 8% – the largest weight any single company has held in the index since 1981. And Nvidia by itself supplied roughly 15% of the index’s entire 2025 return.Statista (2026). Apollo’s Torsten Slok notes that the top ten S&P 500 firms are now “more overvalued than they were in the 1990s” (Slok 2025). At the same time, information-processing equipment and software, barely 4% of US GDP, drove 93% of its growth over the past four quarters; strip out the data-centre build-out and the economy essentially flatlined.Klement (2026), Furman (2025). Furman allows that absent the build-out, lower interest rates and energy prices would have recovered perhaps half of the lost growth.

Concentration of this order is a transmission mechanism: Because index funds buy the Magnificent Seven in proportion to their weight, regardless of price, a repricing of Nvidia or its peers cannot stay contained in the AI sector: it passes straight through the index into every passive portfolio and pension, and – through the capex that is now carrying GDP – into the real economy. And just like in 2008, the holders of the risk will discover only after the fact how much of it they were carrying, except that this time it is dispersed across corporate balance sheets rather than concentrated in the banks.Wyman (2026)

In other words, this bubble will leave behind an extraordinary amount of rubble.

What might grow in it – the technological, economic and political openings a post-bubble AI could be built on and enable – will be the subject of the second part of this essay.

Disclosure

I used Claude Opus 4.6, 4.7 and 4.8 at various stages of writing this essay.

References