Lecture by Marek Poliks & Roberto Alonso Trillo
Transcription
Marek Poliks: Is this working? Yeah, it's working. Awesome. I guess we should continue the litany of thanks. Thanks to Hugh — our boss, the boss of Disintegrator at this moment. Big fan of Hugh. Big fan of Ellie and Hugh, the other Hugh here at Index, and the general Index project. Big thanks to Garden 3D. Big thanks to Montes Radio. Who else am I forgetting? I'm sure I'm forgetting people. Um — Yale, Dena. Well, Dena is not just a thanks, you know. It's more than a thanks. Dena's in the footnotes of all of our work. We're just in our cups at this moment to be here with Dena.
Do you want to kick it off?
Roberto Alonso Trillo: Yeah, sweet. And thank you to all of you for being here today.
So the title of this talk is The Data Center Does Not Exist — which is not a dismissal of the very real social and environmental implications of AI or of data centers. But we do mean the title literally, because when we attempt to understand what AI actually is on economic terms, we begin to understand that the economic engine of AI has more to do with leveraging the idea of the data center for speculative financial gains than it has to do with actually building data centers — and ultimately even AI.
And from there, perversely, we'll argue that the financial objects of AI would actually prefer that this physical infrastructure is never built and never actually used, because financial speculation alongside delays, cancellations, and restructuring operations are ultimately more profitable than building and running data centers.
AI, for us, is the name for capital achieving escape velocity from the real world. The name of the epoch in which capital has started to successfully separate itself from any dependencies on building things, selling products, and managing operations. And this separation, we believe, has huge implications for where we all fit in this AI thing — as humans, as workers, as nominal participants in the economy, and also of course as political subjects.
So let's start with the central thesis of our book, Exo-capitalism, produced with our beloved friends at Becoming Press here in Berlin. One of the main ideas — the TL;DR of Exo — is that capital seeks independence from its material base. We call this tendency lift. Capital tends to lift away from physical dependencies, from infrastructure, from manufacturing, from production — ultimately from the soil, the earth, and the sphere of human affairs. It does so through abstraction, by moving from doing things to managing things.
Lift describes the tendency of businesses to sell licenses instead of commodities, to become software businesses instead of hardware businesses, to become insurance companies, to become banks, and then ultimately to make their money primarily through financial engineering within the business sectors they used to occupy — which is, in aggregate, how most money is made. Every step higher on the abstraction ladder protects them from risk, expands their total addressable market, and plugs them into larger and more freely moving flows of capital.
Lift means: take your business and make it about something more abstract. That could mean a shipping company getting more and more interested in shipping management software, or Tesla making more money from selling carbon emissions credits than selling cars. Or it could mean how Starbucks is in the top third of biggest banks in the United States simply by reinvesting the $25 you have to hold in the app. Or how Monsanto doesn't really sell seeds — they ultimately sell licenses to genomes and crop insurance.
An example we always like to give is the airline industry. Basic consumer air travel hardly contributes to the airline industry's razor-thin margins. In a sense, the real business of an airline is not transportation. The real business is the loyalty points marketplace. The Delta SkyMiles program, for example, generated almost $60 billion from the sale of miles to American Express in 2025. The airline manufactures a currency from nothing. They sell it at a 600% markup. And the entire productive apparatus of flight — the planes, the fuel, the pilots, the different boarding groups, the human beings herded through the jet bridge — they just function as loss-leading infrastructure for this token economy.
The plane still flies, but the plane is incidental to the business model. The plane is the underlying asset for a superstructure of financial engineering that generates returns independently of whether any passenger sits in any seat. The objective, of course, is to get farther and farther and farther away from the plane itself. But Delta is only halfway to achieving real escape velocity. The SkyMiles economy dominates the airline industry, but it still needs the plane to fly — even if the flying is always a loss.
What we're going to show you today, with the case of AI, is an economy that has discovered it does not need the plane anymore.
Marek: Hell yeah.
Roberto: But first we're going to get a little depressing.
Marek: It's kind of fun, you know.
Roberto: So we're going to talk about a couple things first to couch it, and then we'll go directly into AI.
The first big point here is that we don't live in an information economy. Information is not the commodity form of the moment, despite what people like to say. It's very easy — and people tend to very often make the argument — that Meta, Spotify, and Google are engaged in the sale of personal information to advertisers or other actors who are interested in exploiting that information to target specific consumers and produce specific behaviors. But that's kind of where the critical curiosity ends.
Critical accounts of technology almost entirely focus on Meta, Spotify, Amazon, and Google — on consumer tech. But consumer tech represents about a seventh of the industry in terms of revenue, a fourteenth in terms of new investment, and by far the slowest growth rate in terms of CAGR. It's the most visible layer, but you really should not be allowed to create a general theory of technical political economy by looking at the smallest and least important component of it. It's really like writing a book about the oil industry by looking at gas stations. You just shouldn't be allowed to do it. I really believe that.
The most extreme version of this kind of folk analysis in critical spaces is the allegation that big tech software — and specifically the AI economy — is either funded by or deeply integrated into state military and surveillance programs. And it is absolutely reasonable to condemn any company for doing anything with the US military. Don't cooperate with state violence. But the theory of military-industrial collusion or even control over the tech economy is a dangerous and crazy misreading of the situation. Economically speaking, military participation in the software economy is at best a rounding error. It represents a fraction of a percentage of aggregate investment in AI.
In the news recently: OpenAI and Anthropic engaging in $200 million deals with the US military. That's a marketing campaign. OpenAI loses that $200 million in a few days. That figure represents about 0.7% of their revenue this year, and 0.02% of the capital they raised this quarter. That kind of matters. It's really important to mention that so everyone understands the scale of what we're about to talk about.
Digital capital is something that is so much bigger than even the scale of Empire — larger, easily, by two orders of magnitude if you're measuring in terms of complexity, energetics, and human resources.
Once you take a serious look at the business of technology — the really, really big business of technology — you see a very different relationship to data and to information. Tech companies don't actually sell data as a primary generator of value. Honestly, that data is not really that interesting and it's really difficult to do anything useful with it. Instead, what they sell is themselves.
The business of technology is a very strange world. It's something that theories of political economy haven't quite built the muscles to understand. But in brief, the software business figured out — sort of before anyone, or better than anyone — that products are less valuable than the connections between products. If you take a tech company and cut it open, you'll find all kinds of other little tech companies inside. Every tech company has a stack: the series of third-party providers, platforms, tools, and data structures they use to do business. And every tech company is assembled out of that stack. The particular assemblage — the combinatorial web of things they put together — is what makes them unique as a business.
If you cut open any of those little third-party companies living inside a tech company, you might find it's also full of other third-party companies. And if you keep cutting, you might go all the way back to where you started. At some point, you come to the realization that all tech companies are decomposable into each other. They are just isomorphic expressions of one monadic general tech company.
You don't really buy or sell software anymore. You instead facilitate formal arrangements to integrate, co-market, and sell aggregated software solutions to third parties, increasing your footprint — or what we call surface area — in the business. Surface area is the opportunity space you can sell into, basically your total addressable market. And the goal is to always have a play, meaning you always want to create new opportunities to generate new business by doing slightly different things, which you do by integrating further with more and more partners.
Once you have an agreement in place, you can subsidize that customer's purchase of your stuff through the exchange of mutual usage agreements or credits, immediately transforming that new customer into a partner. You don't buy or sell software. You enter a partnership with a software provider. You trade credits to each other. You fall in love. You become one another.
You clot, drift apart, cozy against the hyperscaler for a while, disperse back along the infinite red curtain of BlackRock or Morgan or Centerview or Catalyst. Disappear and take millions with you. Pop. Money isn't a liquid, and it's absolutely not subject to the laws of thermodynamics.
So yes, the goal of a tech company is to never quite be a customer and never quite be a vendor, but rather to be an elastic membrane through which other tech companies touch each other, fold into each other, and collapse into each other.
The final point — contrary to contemporary expectation — this actually doesn't really resolve into monopolies, but rather into extremely diffuse, differentiated webs of interconnection that span across even the most direct competitors: nodding into mergers, budding into acquisitions, fizzing into acqui-hires, imploding into PE rollups or fire sales. This incredibly sticky, thick web of collapsing is where the real software economy is lurking — concentrated around the institutions that financialize the acquisition of one company by another or by the public: consultancy firms, venture capital, large banks, private equity, M&A advisorships and their instruments.
Tech is just a particularly activating substrate for finance. Or to put it another way: finance is the only real consumer of the tech sector. Software has no commodity form. Its product is not software and certainly not information. It is instead a particularly supple vehicle for the continuous injection and re-injection of investment capital into itself.
Roberto: AI time.
So let us take a first tour of the data center from this perspective. Every time you use the internet, you're almost always connecting with a piece of software running on a server in a data center. AI has dramatically increased the need for these data centers, because AI ultimately means using the internet to access a very powerful application that is usually working across several servers together to execute.
You can think of a data center in the conventional way — essentially a large, hot warehouse filled with Nvidia GPUs gorging on local resources for both processing and cooling, screaming infrasound. But we like to think of data centers differently. We like to think of data centers as collections of contracts.
As we were describing before, the theory of lift identifies the business's need to get away from the data center itself — from vertical integration into the data center — into what is called a value chain: an array of partnerships that progressively push out their commitments away from the actual data center itself. In this sense, we want to help you understand the physical data center in terms of its financial surface area: the aggregate series of connections between businesses and business units, the handoffs between those businesses, and the radiant stream of debts and contracts incurred at each connective joint.
Marek: Can I just interrupt and say — this is my favorite kind of diagram. There are like a thousand diagrams like this. Just get a little drunk, go on Google Images, and take a bath. There are so many different value chain diagrams, and they all have different partners in them. This is just a tiny, tiny, tiny piece of all the buzzing little hives of businesses intermingling with each other. Sorry.
Roberto: So going back to the idea of a collection of contracts — you can start very low level. The first contract is between the intellectual property owner of the chips themselves (that's almost entirely Nvidia these days) and a web of contract manufacturers in Taiwan, Malaysia, South Korea, and China, who actually fabricate, package, test, certify, and pre-program those chips.
When a physical server arrives, it is racked into a physical data center facility. That facility is owned and operated by so-called colocation providers. Colocation providers are companies structured — and this is very important — as real estate investment trusts, which means they are pass-through entities designed to avoid corporate taxes by distributing income directly to investors. They are basically very weirdly financialized hotels for servers that they do not own but are kind of liable for. And what's interesting for us is that the insurance that spins out of this business is highly lucrative — its own kind of separate lifted economy.
Now, at the facility administration level, you get a new type of contract: software contracts with infrastructure management, remote management, network and compliance security, environmental monitoring, and lifecycle management platforms. Beneath these facilities, another layer of contracts governs the physical resources required to keep GPUs alive — water rights, backup generation contracts, grid interconnection agreements, and much more.
If the data centers are like hotels and hotel administrators, then the people renting the rooms are the cloud providers. But the relationship with the data center providers they nominally rent from is less like a landlord-tenant relationship and more like a partnership. They exchange credits, they co-market, they co-sell, they avoid as much as possible the indignity of a direct sale. Direct sale is something they're trying to avoid all the time.
These cloud providers then resell that compute in a variety of different ways — through on-demand spot purchases, enterprise contracts, synthetic vehicles, and secondary markets. Their customers might be model providers like xAI or Anthropic, who purchase that compute from the layers below, use it to train and serve models for inference, and then further resell access on a per-token basis. Above this, you might have one or many contracts with an application layer, who then resells that compute again in the form of subscription fees. Only after a number of resale loops at this layer does the generic end user pop up at the surface of a tremendous, mind-boggling, hugely scaled pile of contracts.
All of that activity is worked into a frenzy before breaking ground. It comes to a head in negotiations between all of the above entities at a level of scale that has nothing to do with this or that physical installation — and it is the initiation of these negotiations that signals capital to move.
Marek: Let's go. Okay. So we have this spool of contracts that makes up a data center, and what we're really interested in is how that spool of contracts can actually generate capital on its own — without ever doing anything.
There are three main ways we've isolated that this occurs. The first — which we're about to talk about now — is the announcement of building a data center as a means to generate capital. The second is the speculative financial markets around energy and compute, which gets really wild. And the third — the craziest — is the recursive relationships between the businesses that sign these contracts and how they're all ultimately financing each other in a very beautiful, loving way.
So let's start with the announcement as a value-generating activity — announcement as production, or announcement as labor.
Just over a year ago, on January 21st, at the White House, President Trump announced the $500 billion AI data center initiative called Stargate. The announcement felt like a nationalization project — big Trump national AI thing. But it's actually just a generic multinational joint venture investment company, featuring a few core players: SoftBank, OpenAI, Oracle, and — not in the room — the Emirati sovereign fund MGX. (I wonder why.)
On day one, the group had committed about $52 billion — that's how much they actually had at the moment, even though the announced figure was $500 billion. That's about 10% of the announced figure. The remaining 90% — the $450 billion — was left really ambiguous. And even that $52 billion seemed shaky. Elon Musk was tweeting that none of this money was actually real. But what's cool is that it didn't matter, because the market was really excited about Stargate — and particularly about these specific partners.
Oracle, for example, rose 13% the following day, generating $60 billion in market cap between market close and market open. More value was generated by the announcement itself for Oracle in 24 hours than the $52 billion in committed equity across the entire group that the announcement described. It's so beautiful.
SoftBank, whose CEO was standing beside Trump at the podium, would nearly triple in value over the following year — an almost $60 billion aggregate gain attributed to the market treating the company as a proxy for a project that still only really exists as a press release.
And this phenomenon has become so reliable that markets now discriminate based on the narrative quality of otherwise identical announcements. When Meta announced $120 billion in 2026 AI capex a couple months ago, it surged 10% overnight — seen as: okay, they're doing AI, good, buy Meta. But when Amazon announced its own $200 billion AI capex investment two days later, it lost 7%, because it was perceived as late to the trend.
And while these announcement-driven surges are totally going to dissipate and self-correct through the market, they nevertheless create opportunities for tangible extraction: creditworthiness for debt raises, wildly overvalued acquisitions, and cash — lots of cash — from sold insider shares.
Another example: September 2024, Constellation Energy announced a 20-year power purchase agreement with Microsoft to restart Three Mile Island's Unit One reactor, rebranded carefully as the Crane Clean Energy Center. The restart will cost $1.6 billion and the plant will not produce a single watt until 2028 at the earliest — if that. But the ultimate cost and construction timeline is kind of irrelevant, because the financial function of the announcement was fully served on day one. Constellation stock surged 22% the day the deal was announced — about $50 billion in new market capitalization within 12 hours, nearly 10 times the cost of the restart itself. Within four months, Constellation leveraged that inflated valuation to acquire Calpine for $16.4 billion, making itself America's largest power generator.
Another example: CoreWeave is a neo-cloud that pioneered the use of Nvidia GPUs as debt collateral. They were able to raise billions of dollars against hardware they were still in the process of acquiring — a recursive signal in which the announcement of purchasing GPUs itself underwrote their creditworthiness to finance that very purchase. They literally said, "We're going to buy a ton of GPUs," and all the credit agencies said, "Okay, you're using this for GPUs, so you must be creditworthy enough to buy GPUs." And it created itself. It's amazing. The company went public in March 2025 and allowed its investors to extract 800x returns before its debt obligations even came due — even as the business was highly unprofitable and losing more money the more revenue it generated.
The interesting thing here is that the announcement is all you need. There is something beautiful about this. It's not about hype per se — it's more about an autotelic process wherein the signal alone is eminently productive of value, often more productive than any debt incurred through the signal generation process.
If the narrative craft of the announcement is a value-producing function, then it follows that the contractual instruments through which these announcements are made and ultimately formalized are going to exhibit a kind of weird autonomy from the physical assets they're nominally describing.
A great example is how Meta buys power — through VPPAs, virtual power purchase agreements, which is a type of energy contract in which no electricity actually changes hands between buyer and seller. Instead, the energy utility sells power directly into the wholesale market at a floating price. The corporate buyer — in this case, Meta — commits to a fixed price in advance. If the actual price is higher when they later collect energy from the grid, the utility pays them back. It's a very weird way to buy power.
The Meta-Constellation Clinton deal of June 2025 is a particularly crazy example. Meta bought the full output of the Clinton Clean Energy Center in Illinois — nuclear power. But what Meta is ultimately purchasing is the fact that clean energy is produced, not the energy itself. They're not buying power for the data center. They're actually committing to raise energy prices, and they pay out only if they fail to do so. They've purchased an incentive to raise energy prices. And counterintuitively, that's how they purchase energy.
But it's also a win-win, because VPPAs generate renewable energy certificates (RECs) — a kind of tokenized environmental attribute. It's like a carbon credit, but tradable and retirable on secondary markets regardless of any actual usage of energy. Meta gets RECs that the Clinton plant produces. Those RECs transfer to Meta regardless of whether Meta actually uses any of that energy. And if the data center is never built, Meta holds RECs in excess of any specific facility's consumption — so they can retire them against consumption elsewhere, bank them for future use, or just sell them as surplus on secondary markets.
This gets even weirder at the compute layer. GPU hours are becoming a commodity class traded with the same market infrastructure as crude oil. This could be as simple as buying GPU hour futures — essentially committing to a price for a GPU compute hour at some future point in time, then selling that contract after you've driven up the price by announcing the very purchase of those futures. It's an amazing, good free-money printer right there.
Or it could be as elegant as playing with the useful life estimate of GPUs — the accounting assumption about how long a GPU remains productive. Between 2020 and 2024, basically every major hyperscaler came to the totally asynchronous decision that GPUs lasted 6 years instead of 3. This meant they could move out the depreciation schedule and cash out $18 billion in annual boosts to reported profits simply by saying, "Actually, these things last 6 years" — which is actually even crazier because GPUs depreciate like crazy in a real sense, since every new generation is orders of magnitude better in terms of performance and energy consumption. Meta did this last year in Q2. Same kind of thing for a quick billion.
It matters less and less how many GPUs exist, how fast they run, or what they're actually computing. What matters is simply that GPUs are surface area for all kinds of complicated gambling. The thing being computed is irrelevant, as is the GPU itself as an object. It only matters insofar as it's interesting enough to gamble on.
Roberto: October 2025. Meta and funds managed by Blue Owl Capital entered a joint venture to develop and own the Hyperion Data Center campus in Richland Parish, Louisiana. The campus is Meta's largest data center worldwide, scheduled for completion in 2029.
To finance the build, Morgan Stanley arranged over $27 billion in debt and about $2.5 billion in equity into a special purpose vehicle. Because it is structured as an SPV, Meta is not actually borrowing the funds itself — and therefore does not have to show any large debts on its balance sheet. Instead, Meta gets a $30 billion data center while keeping $27 billion in debt invisible to its credit rating agencies and investors. It remains the developer, the operator, and the tenant. It controls the thing itself, but it doesn't own the debt.
As Fortune put it, the deal stands out for its scale — the largest private debt offering ever — and for its A+ rating from Standard & Poor's, which reflects Meta's backing. Blue Owl, the other actor in the SPV, contributed $7 billion in cash to Hyperion, three billion of which went into Meta as a one-time payout. Which means Meta actually got paid to have someone else take the debt for building their own data center. Kind of crazy.
And this is not a bad deal for Blue Owl, because they didn't put up $24 billion of their own money — they put up $7 billion, and the SPV raised $27 billion in debt from Pimco (which has already generated $2 billion in paper profits by selling bonds on this deal), BlackRock, and other companies. The longer the construction timeline extends, the longer Blue Owl collects management fees on the capital under its administration. Its incentive is not to finish the data center itself, but to manage a large, complex, long-duration financial structure. If the project needs to be restructured, refinanced, or renegotiated — all of which become more likely with delays — each event is a new fee event.
And now of course the big one: the Nvidia-OpenAI-Microsoft-Oracle loop — the most extreme example of lift in the contemporary economy.
In 2025, Microsoft invested $13 billion in OpenAI, but much of it was delivered as Azure cloud credits rather than cash — meaning Microsoft paid OpenAI with its own product. Then OpenAI committed to purchasing $250 billion in Azure services, which Microsoft got to book as backlog and revenue. Win-win.
At the same time, OpenAI committed $300 billion to Oracle for Stargate data center infrastructure. Oracle used that commitment to purchase $40 billion in Nvidia GPUs and raise $50 billion in new debt. Nvidia in turn announced an investment of up to $100 billion back into OpenAI, whose CFO acknowledged that most of the money would flow back to Nvidia to buy — again — GPUs.
SoftBank bridged the gaps by selling its Nvidia stake and borrowing $16.5 billion to invest $41 billion in OpenAI equity, then booking $16 billion in paper profits in a single quarter based on OpenAI's appreciating valuation — a valuation sustained by the very commitments that SoftBank capital helps fund.
The total committed infrastructure spend across these interlocking contracts is $1.15 trillion. And its nucleus is a startup with $12.7 billion in annual revenue that expects to lose $47 billion by 2028.
The combined cloud backlog across the three major providers now exceeds $1.1 trillion — a figure that justifies its own creditworthiness simply because of its scale. What's interesting is that each dollar circulates through a series of contractual handoffs: investment becomes cloud credit, which becomes infrastructure commitment, which becomes GPU purchase, which becomes hardware revenue, which becomes further investment. And at every joint, new surface area is created: new debt instruments, new equity stakes, new consulting needs, new compliance requirements, new financial products.
Marek: Most data centers are never built.
Meta's former director of energy strategy expects only 10% of currently underway projects to complete. Sources that look at grid hookups cite about a 23% completion rate. More generous sources go up to 50%. We like the more charitable estimate of around 20%.
In 2025, 25 data centers were just straight up nixed due to community opposition — a fourfold increase in a single year. But they're largely canceled because of two really perverse incentives.
The first — and this is something people really don't understand — American utilities infrastructure can't handle data centers, but not because they consume too much power or water. We absolutely have power for data centers. We don't have a modern utilities infrastructure to deliver power in this country. American utility companies have also lifted away from actually building, maintaining, modernizing, and operating infrastructure. They operate more like financial institutions. They've lifted. They're not utility providers anymore. And they're able to leverage the demand for data center production into major surges of capital without the desire, let alone the ability, to actually build and service that power infrastructure.
Then there's perverse incentive number two: the process of data center pre-development is really generative. The environmental permitting, the wetlands mitigation analyses, the grid studies, community engagement consultancies, the lawyers, the compliance officials, design and architectural services, partnership and underwriting work — if only a conservative 25% of planned data centers actually reach operation, and each project requires three to five years of pre-operational services, not only is the majority of data center-related services revenue generated by projects that never serve a single AI model, but revenue generation actually compresses at the very moment you turn on the data center and start to serve AI.
The financial infrastructure literally rewards not building data centers — or building them really slowly and kind of never really getting there. Each construction delay keeps assets from depreciating and generates new intermediary layers, new financial instruments, new speculative investments in utilities and resource extraction — which is honestly the real game here — that are never actually realized.
This loop does not actually need to close. And in fact, it would prefer not to close. It does not need OpenAI to become profitable, or Oracle to finish building data centers, or Nvidia's GPUs to be used for anything in particular. The AI financial cycle has already completed.
You might justifiably gesture toward the risk to rate-payers in Virginia whose bills are projected to double by 2039. The pension funds holding 24-year bonds backed by buildings that do not exist yet and may never exist. The insurance companies whose policyholders are the ultimate holders of Pimco's remaining $80 billion in Hyperion debt. The taxpayers backstopping billion-dollar DOE loans for nuclear restarts that will not produce power until 2028.
But it's worth remembering: the Virginia rate-payer's doubled bill is itself doubly financed through consumer credit payment plans and federal mechanisms, and that debt is securitized, bundled, and sold. The pension fund's exposure is one node in a reinsurance chain that nests itself through BlackRock and hedging instruments back into itself. Every real-world consequence you might point to has already been converted into another financial instrument.
Even the water table, the raw lithium supply, atmospheric carbon capacity — these constraints have already lifted into representational vehicles that can be quantized, hedged, sorted, and resold as representation. And the higher we get into that representational space — into the types of financial instruments that represent a representation of a representation of relations — the more elasticity and tolerance is built against the revenge of the real.
In the 2008 crisis, you could argue that the collateral layer that popped the bubble was the revenge of the real. It was real houses in real places that already existed, with real people paying real mortgages. But it was not their reality as such that created the crisis. It was, to us, a moment of mass forgetting — a moment where finance forgot that it was never real, where finance forgot that its downstream dependencies only become dependencies when finance writes them in as such within the book of fiction.
The present system has learned from that forgetting — and it has done so not by returning to material prudence, not by lending more responsibly, but by doing the opposite: by engineering materiality out of the collateral stack altogether.
What's been collateralized at every level of the AI economy is not physical infrastructure or human debtors. It is the future — and that futurity is bound to the narrative power of AI as an apex civilization-scale event. In a sense, it's the promise of AI itself that holds this all together. And that promise needs to be sustained as something so huge, so transformative, so powerful, so dangerous, so ethically compromised, that any form of ultimate delivery would result in an underperforming performance. So the objective must be to never actually perform — but instead to continue to sustain the intensity of the promise.
And the promise is sustained, in some small part, by a critical apparatus that has done the financial sector an enormous favor by treating AI's physical footprint as apocalyptic. In reality, AI-specific compute currently accounts for roughly one-tenth of 1% of global energy consumption. The entire data center sector — AI and non-AI alike — consumes about 1.5%, which is less than the world spends powering air conditioners and a fraction of what electric vehicles are projected to add to the grid by 2030.
But AI is narrated as the first user, the only user, the one that's ultimately going to swallow the grid. And I want to be clear: this narrative is manifestly useful to the industry that seeks to profit from the promise of AI.
The Ohio family paying 60% more for electricity. The Oregon town that lost a quarter of its water supply to Google's evaporative cooling. These are real issues — absolutely, genuinely bad. But they don't have anything to do with raw supply constraints on electricity or water. They have to do with the limitations of American utility infrastructure to distribute resources at scale. These are meaningfully, powerfully local problems — and their locality is precisely what makes them useful.
A crisis that is locally devastating but globally negligible is the most powerful narrative asset. It's visceral enough to sustain the feeling of apocalyptic transformation. It's small enough to not actually constrain the system's expansion. And it's distributed in exactly the communities least equipped to understand, let alone contest, the financial engineering that ultimately produced their suffering.
While the critical apparatus continues to valiantly tilt against the dark monopolistic militaristic Palantir core of big tech — a vision of big tech that big tech is very happy to serve them as advertising collateral, and which they are very happy to gobble up — the financial sector has built a fractal latticework of interconnections. It has transformed the backstop collateral of the present moment into a web of knots. It has inverted and everted the physical asset through itself hundreds of times to the point of senselessness. There is no ultimate backer of this bubble, because that backer contains itself multiply as a financial object.
We're not saying the system is stable. Not at all. Not even a little bit. Rather, that its instability is literally the way it continues to make money. And sure, it's absolutely going to collapse. And that collapse will be financialized as well.
So let us conclude with the promise. What is the promise of AI?
Obviously AI is powerful and is changing the software business and business in general. We are not AI skeptics. The software platform is dying. The software application is dying. Code is dying as a craft. Data analytics is dying as a craft. Writing is dying as a craft.
On one hand, these deaths are due to the incredible capability of contemporary models to automate not only toil but even ideation, integration, communication, and even the creative process itself. But on the other hand, when we think about layoffs — like the Block layoffs two weeks ago, Jack Dorsey cutting 4,000 employees, which is 40% of his staff — that was not based on the present capabilities of these tools for labor automation, but on the future threat they represent for the question of labor in general.
Among the many dependencies capital wants to lift away from, capital of course wants to lift away from labor. And it's important to understand that in our present moment — where neither products, commodities, production, consumption, nor transaction are the media through which value is ultimately produced — labor actually already has no power in this equation, because it is not the thing that produces value.
The economic promise of AI has little to do with opex labor reduction, time-saving, or efficiency gains. After all, capital is very good at equalizing any optimizations across its entire surface. So while job numbers, the threat to the software industry, the threat to general white-collar labor might feel like an economic collapse, what's actually happening is an explosion of economic activity — not because of the specific affordances and utility of AI, but simply because of the change it represents.
Every change means a new financial opportunity. This could mean an expansion or upsell within an existing relationship, an opportunity to renegotiate a contract and extract additional value, an opportunity to create a new kind of business, an excuse for an announcement, or an opportunity to capitalize on a future implied by that change. You also kick off the proliferation of businesses that specifically monetize change management — the business of business transformation. A change is anticipated; they help you anticipate that change; and at scale, they make the change manifest. They work with businesses to change the technologies those businesses use, and in so doing change the technologies through which business is done.
The longer this transaction goes on, the more value it produces. And we're reaching a level of velocity between technological innovation and business re-uptake that could sustain a permanent transitional state — forever.
So the promise of AI is the promise of a permanent transitional phase that requires permanent transition management. To us, this is the real singularity.
So, wrapping up — thanks for bearing with us.
Let's conclude by reaffirming the promise of software, of which the apogee is AI. Software is the medium through which capital has best been able to detach itself from the real — from labor, from production, from products, and from the customer. AI, the data centers it flows through, the labor that touches it, the intellectual property it contains, is distinct only in the potential energy it creates for the motors of speculative finance.
And the moment that AI actually arrives — full force — is the moment it dies.
The data center does not exist. Or at least for now, it doesn't need to. And at the same time, it's never been more productive.
Thank you.
[Q&A with Dena Yago]
Dena Yago: Well, thank you for the clarity. The experience of reading the book and hearing you speak has made these very complex, obtuse, sort of invisible forces a lot more tangible. So I think that's very important work. Thank you for your service.
Marek: Hell yeah. Thank you. We're doing good work.
Dena: Yeah, we're helping the people.
I think there's something — and you guys write about this a lot — in terms of holding and time dilation and what that does in terms of expanding the surface area and opportunity for other voices to enter the chat. But I do kind of want to focus a little bit on this approaching moment of truth, which we've talked about a bit in terms of prediction markets. I am not a B2B SaaS person. In my personal work and professional work, I end up working in consumer tech and dealing with people and their emotions.
I'm thinking about — you know, strong grift, but Anna Delvey ended up getting caught because she tried to buy a Beaux-Arts building in the Flatiron. I feel like that's a lot of what hitting the soil looks like and becomes a problem for people.
I'm curious if you can talk a little more about the moment of truth — and if, in your argument, that moment actually never arrives.
Marek: Yeah. I mean, I think the goal is to defer the moment of truth as much as possible. That's fundamentally what lift is all about — moving away from physical infrastructure, moving away from labor, but also very much moving away from the moment of great retribution, the revenge of the real. It could be a temporal kind of lift as well.
And I think there's a temporality to it too. It is very much in the interest of business to not only push itself away from any floors or ceilings that stand in its way, but also to push other things into that space — and those businesses are going through the same exercise, also trying to push themselves out. So you have this beautiful kind of shepherd-tone of expansion.
Dena: Can we talk a little bit about the pressure that's coming in on that moment of truth — with prediction markets like Polymarket and Kalshi, which made some cameo appearances in the slides here. And the sort of one-two punch of something we saw recently, like the Catrini research report, where all of these markets are pegging specific dates where some prediction is going to come to fruition.
Marek: Yeah, exactly. Like, I think — if you think about it — there were a lot of predictions around when Trump would actually strike Iran. That was the thing you would ultimately bet on, and that meant there was actual commitment to a date, and money directly tied to a moment. There's a lot of lift in that gesture. You've created a commodity out of something very complex and violent. And all of a sudden that's been dissociated from this thing that happens in the world into a kind of commodity — temporalized in a specific way. There's a window where you can either win or lose based on whether the thing comes to pass within that time.
Roberto: I mean, Polymarket — I might be wrong here — but I think of it as a financialization of divination, which is a very basic human impulse. There's a kind of basic definition of life as projective thinking, and this is taking it to the next level. But it also has to do with hyperstition — the idea that by betting on something, it will become real. And it has to do as well with an inversion of the statistical-algorithmic elements of prediction that are now built into computers, which we are absorbing back as a normalization of human relationships.
Dena: I mean, in my background in trend forecasting I was never really doing predictions as much as prognosticating the present. And I find the idea of "by 2028 we're going to see X, Y, and Z" a little bit corny. The capacity to be wrong in such a binary way is interesting, I guess.
Roberto: Yeah. Like — is it Nicholas Luhmann? Elena Esposito references this — the idea of the future-present and the present-future. Like, 2028 in the present. So Catrini released a Substack post — God, time is crazy, like two weeks ago but it feels like a year — essentially a kind of state-of-AI-2028: nobody has a job anymore, the economy is messed up, everything is messed up. And what's interesting is that it created a couple of accelerative moments toward that future. It was hyperstitional in a very real but also very local sense. Within a week, a bunch of people dumped a bunch of SaaS stocks. People were like, all right, we're not going to buy Workday or Salesforce anymore. But also, we're going to buy TSMC, which I think doubled in a week or something insane. It created this very local mini-version of 2028 financially in 2026 that we experienced for a couple of days and then forgot about.
Dena: Yeah. I mean, I think that kind of segues into the slide that the announcement is the event in a way. And I feel like that time gap between a hot take and it coming to fruition — the moment of truth — creates more surface area for discourse and yapping to create or extract value from actual markets.
And thinking about the emergent function of go-to-market within all of these companies — and how your comparison between the Amazon announcement and the Meta announcement kind of falling flat — maybe we can talk about glazing as a service.
Marek: Oh, I love it.
Dena: I think I might have seen that in a tweet, so I don't hold this one close to heart. But I feel like the current announcement playbook is very ahistorical. I want to tease apart what the composite parts are. One tends to be this sort of cinematic brand film that looks like a wannabe Terrence Malick blue-graded 35mm — we've all probably seen the Anthropic one, and everyone's trying to do that exact same move. Then there's the thought piece — what was, in crypto days, the ICO white paper. And then there's the cacophony of yappers who are either gassing people up or deflating them. I think that's pretty much the current playbook. There's also the UGC makers, the content farm surrounding all of it.
Marek: The only thing I'd add is industry white papers — they're always two months behind, but when a DORA report comes out, everyone in the industry is like, pull over, read the DORA report right now. So yeah, industry whitepapers are a big part of this economy too.
Dena: And then I kind of want to talk about something that adds an inverse pressure — the discovery, as opposed to the announcement. I want to talk about the correlation between something like GameStop — that Indiana Jones "I found it" moment — and something that happens all the time with curators and dead artists, where it creates a speculative market around a discovery for a minute. I'm wondering if you've seen more of that in this emergent space — if people are kind of micro-betting on smaller players.
Marek: Yeah. I mean, the idea of glazing as a service is great — the idea that there's a cottage industry not capitalizing on the announcement itself, but nesting itself inside the announcement. They're not necessarily putting money where their mouth is, but they're contributing to the discourse around it, trying to move it in a specific direction. And obviously monetizing that through Substack subscriptions, Patreon, or being connected to some industrial organization. To me that's really beautiful, and I think we'll see a dramatic expansion of that sector because it seems unbelievably productive in terms of its net impact on things.
The part that's also interesting is that a lot of these announcements are either simultaneously shorted or simultaneously capitalized by insiders within the same organizations making the announcement — which creates a little pre-wave. Which kind of brings us back to the divination or hyperstition point: they make it real by doing insider trading in real time, like the Iran example.
I think the distinction between discovery and prognostication is interesting. The announcement as prognostication would be something like the Catrini report — or Leopold Aschenbrenner's "Situational Awareness" paper from a couple years ago, another very dramatic, Terminator-is-happening-in-2027 document. Immediately after releasing that report, he was dismissed from OpenAI, immediately started a hedge fund, and started investing in nuclear — which was obviously directly built up by the very narrative he had constructed. And he's killing it right now. Super young, doing very well.
But I'm getting off topic. That's the announcement as prognostication as active generative value — and that's different from the discovery. The discovery involves some prognostication, in that it has a kind of "I found this thing, and I affirm it's valuable" quality. But the GameStop thing is: I found a hack, I found a loophole in the system. I found a money printer — either in the form of a deceased artist's estate I can exploit, or in the form of GameStop being actively shorted by financial institutions. I'm trying to think if they're the same thing. I think they're kind of different things.
Roberto: I think what you've said makes sense. But the way we're trying to present it — there's a sort of emptiness to the announcement, to the extent that it's only part of a chain trying to reproduce itself. It's just a gesture trying to propel something that moves without any friction. Vibes.
Marek: Yeah, that's right.
Roberto: But vibes are real. We're all living in this economy fueled by them. That's the substrate we as consumers are dealing with and interacting with.
Marek: Yeah. I mean, it's a great question, because there's a part of me where I'm like, there is no substrate — that's the whole point, it wants to move away from the substrate. But at the same time, all these announcements are based, in a weird way, on aggregate human behavior. A signal happens and people do things. In the case of 2008, a signal happened and a certain type of behavior responded to it, which ultimately created the collapse. It was a very psychological thing.
So there's the question of whether psychology at some scale — vibes at some scale — represents a kind of substrate for capital. I think it does, but everyone wants to lift away from vibes. If you are subordinated to the vibes economy, you are subordinated to it. If you leverage the vibes economy, you're still subordinated to it. You want to get to that space of pure autotelic generation.
When we had a conversation with Hito Steyerl and Simon Denny, they raised this really nice point: maybe vibes is labor. Maybe vibes is the agency we have. Maybe that's actually the last grasp of agency we have — very distributed, a weird position of scale, but the agential substrate of the moment.
Roberto: To me the way we frame this divide in the book is between lift and drag — the human attempt at embedding ourselves into the lifting process. Psychology relates very much to what happens with drag — the underpinning drives of human relations. So it is built into the system as a productive substrate, but to the extent that humans are engaged in this in a performative way —
Like the example we give in the book of the coaching service for OnlyFans managers: it is very lifted, but it is also the drag. And that ends up being the kind of communication and affect labor that more lay people — the ones also making bets on Kalshi and Polymarket — are performing.
Dena: There was a nice one-to-one relationship there between active Kalshi users and active OnlyFans managers.
Roberto: It's just a circle.
Dena: Yeah. I think that's what ends up being more roles for people in knowledge work, cultural work, or affect-based work to directly participate in this economy — they become the people getting paid under $100 per video clipping, UGC creating, ring-lighting their life to oblivion.
Marek: And again, the goal of the OnlyFans manager is to not do that. The goal is to find some way to outsource all of that work — to AI generation, which we're now seeing at serious scale, or overseas — but ultimately they don't want to do that stuff either. They want to essentially coordinate OnlyFans activities, deal with social media strategy in the most broad and ambiguous sense.
I'm obsessed with marketing operations — the idea of marketing operations within any company is amazing to me. It is a type of job whose entire function is essentially to prove that marketing activity did something. That's exactly what we mean by drag. It's this way of finding some way to make this make sense to a board of directors or an executive panel. The problem with marketing — and the reason I'm so interested in it — is that ROI on marketing is the most difficult problem to solve. So what you have to do is build a funnel with like 19 stages, where stage 19 is conversion and stage one is maybe they know who you are. And you build these stages and these narratives. I just think that's so beautiful — the human animal being like, hey, I got this, I'm doing something here. Yeah.
Dena: I think everyone here living in New York, dealing with the janky brand space of this city, comes into contact with the physical reality of hype and vibes in this economy. You're jostling through street vendors trying to sell things, bumping into hype lines. During COVID there was this moment where you couldn't figure out if something was a line for a popup or a food line. And now you have people starting Kalshi, and all these roles — chief vibes officer, chief taste officer — all these bluebird-wearing people thinking that taste is a moat.
Marek: Yeah. And I think that janky capitalism concept feels so immediately real, especially living in a walkable city. A shout out to Daniel Felstead and Jenny Young who have a great video series called Janky Capitalism — they're articulating exactly that: the human experience of moving through the city. The archetypal example is the Deliveroo driver on a bike with five cell phones duct-taped to the front, these weird hand-warmers not provided by any service provider, on an e-bike with a knockoff battery because the provider doesn't want to actually provide anything. They're not employed by anybody. They're working through five different jobs at the same time.
That experience — being an on-the-ground person in a lifted world — is janky. That's their terminology. And to me that's the material implication of all of this: we're all moving around a world that has no interest in helping us reproduce ourselves socially in any way. The result is these improvised, local ways of dealing with it.
Roberto: The way I would frame their idea of jankiness is what happens when technological possibility hits the reality of the human layer. There's an expansion — a kind of noise expansion — of a necessary technology that starts surrounding the immediacy of daily actions. You can see it at different levels and in different places. I come from Hong Kong. We went through a phase — now illegal — where taxi drivers had 20 phones. You get out of that in economies where technology hits the ground level in a way that hasn't been properly articulated. It's the collapse, the collision, between the lifted impulse and the liminality of being a human at the soil level.
Marek: Yeah. There was a great Instagram post — it was a joke, but it was like: here's a free, super ethical money hack. You go to a parking meter that's been converted to an app — that's already a lift, right? Our governments said, okay, no more meter attendants, we're going to make it easy by partnering with this app provider, putting the sticker there, and making you pay for it through the app. So already there's this beautiful chain: you to this application to this payment processor to your credit card company. And the Instagram joke was: I'm going to make my own app with a really nice front end for that parking thing. I'll put my sticker directly on top of that. You scan it — it actually interacts with the API for that payment processor, so I'm not stealing anything — but I produce $1 in value-add by making the UI really nice. And you can just keep popping stickers on forever.
And it sounds like a joke. But then we were parking with Hugh, I forget where, and I'm like, all right, I have to download another app because I'm in another small town with its own individual relationship to this payment processor. That's the jank. The government opens up the surface area of an engagement with some corporate partnership, and all of a sudden you have this prolapsation of weird intermediaries that leave individuals having the craziest time.
But I feel like we didn't answer your question around hype. Like, where do you think there's a point of disagreement? I'd love to understand that.
Dena: Maybe it's just a different zone of focus. I live in that space. My job is to coordinate that work. It's a very affect-based job and you're just sort of reliant on the yappers and the clippers and the content creators and the gassers. But I think it creates an interesting cacophony where you realize that critique ends up becoming part of the flywheel as well. And I think that's the world everyone in this room is kind of living in — what function does critique have in value-add, speculation, and prediction? The Substack-X-Polymarket recent partnership is like the event horizon, in a big way.
So yeah. I think it's just about where our focuses lie. But the janky capitalism concept feels so sort of immediately real, especially living in a walkable city. Very different from living in LA.
Marek: Can we say — we're not very good at Q&A and don't particularly enjoy it, but we would love to meet you guys and chat.
Roberto: He is very good at Q&A and enjoys it.
Marek: We would love to meet you guys because otherwise we just get really nervous.
Dena: Well, I don't know if we can convince you to stick around and maybe sign some books. The Exo-capitalism books will be for sale for $20 — or if you hate it, we can rip it up with you. Dena's book will also be for sale for $15. Let's have a round of applause. Thank you so much.
©2026, Berlin/Nicosia
hello@becoming.press
@becoming.press
Prolegomena
Impressum
Submissions
Distribution
Imprints
(un)Becoming
Agenda
Contact
Newsletter
Subscription
Events
Radio • On-Demand
ÊT/RE • Soundcloud
Materials
Immaterials
Seminars
Cinematics
Nodes
Search