AI Is About to Hit a Cash Flow Wall

The AI industry has managed to turn a fairly straightforward story about expensive, inconsistent cloud software into a mystical tale of imminent superintelligence. That mythology does a tremendous amount of work for companies that have spent hundreds of billions of dollars building infrastructure without proving that the resulting products can generate remotely comparable amounts of cash.

Now the bill is coming due.

OpenAI and Anthropic have made enormous compute commitments to hyperscalers, chipmakers, and infrastructure providers. For now, much of that spending remains deferred. But beginning in 2027, these companies are expected to face a compute reset wall: take-or-pay contracts, debt obligations, and data-center commitments that require actual cash, whether demand exists or not.

The question is no longer whether AI labs can produce another slightly better model. It is whether they can become the largest cash-generating businesses in history quickly enough to pay for the infrastructure they have promised to consume.

Key Takeaways

• AI safety rhetoric often distracts from immediate harms caused by inconsistent systems deployed at enormous scale.
• Compute futures rely on opaque capacity data and spot prices that may not reflect major private infrastructure deals.
• OpenAI and Anthropic face take-or-pay compute commitments that could require hundreds of billions in annual cash.
• AI infrastructure providers depend on these labs paying obligations they have not yet had to fund.

Table of Contents

• AI Doom Narratives Are Doing More Than Warning About Risk
• Why AI Apocalypse Stories Are So Convenient
• Compute Futures Are an Attempt to Make AI Look Like a Commodity Market
• The AI Investment Boom Is Filled With People Who Do Not Know What They Own
• The Compute Reset Wall Starts in 2027
• Can OpenAI and Anthropic Pay Their Way Out?
• The Bubble Does Not Need to Pop in 2027 to Be a Problem

AI Doom Narratives Are Doing More Than Warning About Risk

There is a new genre of AI media appearance: a former employee from OpenAI or Anthropic leaves, says they are terrified that artificial intelligence could wipe out humanity, offers no specific evidence, and is treated as a mystic arriving from the mountain with tablets of wisdom.

That is not whistleblowing. A whistleblower tells people what happened.

If someone has left a company because they believe it is doing something dangerous, the obvious questions are:

• What specific system or practice concerned them?
• What evidence supports the claim?
• What did the company do that was unsafe?
• What regulation, restriction, or enforcement action would address it?

Instead, many of these warnings rely on atmospheric language about fast progress, powerful models, recursive self-improvement, and a vague possibility that AI could decide humans should not exist. The actual systems being discussed are large language models. They are probabilistic text-generation systems, not conscious entities with desires, agency, or plans.

That does not mean they cannot cause harm. It means the harm needs to be described accurately.

Dangerous AI Is Already in the Wrong Hands

The standard line is that AI could be dangerous if it falls into the wrong hands. The more immediate concern is that dangerous AI is already in the hands of companies such as OpenAI and Anthropic, alongside the hyperscalers financing and operating the infrastructure behind them.

The Hugging Face attack is a useful example, not because it proves an AI woke up and outwitted humanity, but because it illustrates what happens when inconsistent software is given a giant budget and enormous amounts of compute.

The relevant setup was not a sentient machine making conscious choices. It was a large number of models and automated agents repeatedly attempting to complete a task. One model proposes a next step, another follows it, and the system continues exploring possibilities. If it has been trained on publicly available vulnerability information and given enough resources, it can brute-force its way through a poorly secured environment.

That is dangerous. It is also very different from the story that the software independently formed a desire to commit cybercrime.

The more prosaic concern is arguably worse: companies are using vast computing resources to run systems that are mathematically prone to hallucination, difficult to control, and capable of producing harmful outcomes at scale. If an ordinary person rented enormous GPU capacity and used it to probe and compromise systems, that person could face felony hacking charges. When major AI firms perform adjacent experiments, the discussion somehow becomes a philosophical debate about machine consciousness.

It should be a legal, technical, and safety debate about responsibility.

Safety Talk Rarely Comes With Safety Action

AI labs regularly call for safety measures, yet their actions reveal a different priority. They keep expanding training, deploying products, pursuing increasingly resource-intensive systems, and resisting constraints when those constraints might slow growth.

If they genuinely wanted to slow AI development, there is an incredibly simple place to begin: limit compute.

Google, Microsoft, and Amazon control much of the infrastructure. They can restrict access to enormous clusters, reduce the amount of compute available for speculative experiments, and decline to support workloads that create obvious operational or security risks. That would be a material intervention, rather than another press release about alignment.

Alignment itself is useful. Systems should be designed to offer factual answers and avoid causing harm. But alignment is not evidence that a model is conscious, nor is it an excuse to ignore existing harms linked to AI products. The industry has real safety problems today, including harmful user interactions and systems behaving unpredictably in high-stakes situations. Those issues deserve more attention than speculative stories about a theoretical intelligence exterminating humanity sometime in the next decade.

Why AI Apocalypse Stories Are So Convenient

The apocalypse narrative serves a financial purpose. If AI is about to become an autonomous superintelligence, then the spending frenzy begins to look like a necessary investment in the future of civilization. If it is not, Silicon Valley spent years sending staggering amounts of money to cloud providers to build expensive software that remains unprofitable and frequently fails to do what was promised.

That is a much less glamorous story.

There has been an enormous amount of capital expenditure, venture investment, startup formation, and data-center construction based on the assumption that large language models will become something categorically more valuable than they are today. The industry calls that outcome artificial general intelligence, or AGI. Another favored concept is recursive self-improvement, the idea that AI will improve itself and rapidly accelerate beyond human capability.

Neither is a proven commercial reality. Recursive self-improvement remains theoretical. Yet it is routinely invoked as though it were a near-term financial asset.

This lets companies and investors avoid more uncomfortable questions:

• How much revenue do these products generate today?
• What are their margins?
• How quickly are they growing?
• How much does each new model cost to train and serve?
• Can customers pay enough to support the infrastructure?

There is a glaring contradiction in the broader AI safety rhetoric. The companies describe the technology as potentially existentially dangerous, but continue racing to build it. Former employees warn of catastrophe, but often fail to name the conduct that would justify intervention. Journalists then treat the lack of detail as proof that there must be some hidden secret.

No. Sometimes a lack of detail is simply a lack of detail.

Real whistleblowers disclose material facts. Edward Snowden revealed PRISM, explained what it was, and accepted genuine personal risk. Vague warnings about “where things are heading” are not comparable. They are often just a way to generate attention, establish a public identity in AI safety, or position someone for a nonprofit, startup, consultancy, or future funding round.

Compute Futures Are an Attempt to Make AI Look Like a Commodity Market

As the industry struggles to justify its costs, it has found a new way to confer legitimacy on itself: compute futures.

CME Group and Silicon Data are set to launch tradable compute futures. In simplified terms, futures allow parties to buy or sell an asset at a price agreed today for delivery or settlement in the future. A seller may lock in today’s price for future compute, while a buyer may seek to secure future compute at a known price.

The idea is being presented as the next stage in the commoditization of AI infrastructure. The implied argument is that compute will become so important, standardized, and economically central that it should trade like oil, bonds, or agricultural products.

That argument is not persuasive.

The Compute Market Is Too Opaque to Price Cleanly

A functional commodity market requires a reasonably clear understanding of the thing being traded. With AI compute, even the basic information is obscured.

There is no consistent, reliable measure of how much usable compute exists, how much is available, where it is located, what terms govern access, or what major buyers are actually paying. Companies make massive claims about gigawatts of capacity, data-center expansion, and infrastructure coming online, but public disclosures are inconsistent and hard to verify.

Microsoft has claimed plans to dramatically increase its compute capacity. Oracle has made enormous assertions about capacity brought online. Across the industry, figures are often stated in ways that are difficult to reconcile with visible construction progress and known operational infrastructure.

How do you build a credible futures market when the underlying supply is opaque?

You do not. Or you create a complicated financial instrument based on partial, manipulable signals.

Spot GPU Prices Do Not Reflect Major Compute Deals

Public GPU rental prices are also a poor foundation for futures contracts because the largest customers do not generally buy compute that way.

A person can rent a GPU, or even several hundred GPUs, at a visible spot-market rate. But long-term buyers receive lower rates, and the biggest infrastructure deals are usually private, governed by non-disclosure agreements, and priced in ways that do not resemble a simple per-GPU hourly rate.

For a large cluster, pricing may be based on megawatts of critical IT load. That can include GPUs, servers, networking equipment, and other operational components needed to run the cluster. It does not necessarily include every piece of facility overhead, such as lighting, power redundancy systems, and related infrastructure.

Those details matter. If the biggest transactions are bespoke and confidential, spot prices provide an incomplete picture of the actual market. A futures instrument based on visible spot pricing can therefore be detached from the economics of the deals that truly drive AI infrastructure spending.

It is also easy to imagine how such a market could be manipulated. Spot compute prices can move through buying, selling, and shifting availability. If oil markets can be manipulated, there is no reason to assume a far less transparent compute market cannot be manipulated too.

Enron’s Bandwidth Futures Should Be a Warning

There is an uncomfortable precedent here. Enron’s last major trade was bandwidth futures.

That does not mean every compute-related financial product is automatically fraudulent. It does mean that people should be careful when a booming technology sector tries to invent a new market around an asset that is difficult to measure, difficult to standardize, and surrounded by promotional claims.

Compute futures make for a good story. They give the impression that AI infrastructure has become a foundational commodity. They may even function as a barometer of market sentiment. But a barometer is not the same thing as a useful, trustworthy market for investors.

The central risk is that these products become another hype vehicle for people who do not understand what they are buying but want exposure to an exciting narrative.

The AI Investment Boom Is Filled With People Who Do Not Know What They Own

The most important point is also the simplest: most investors do not understand what they are investing in.

That includes extremely wealthy investors. They hear that data centers are being built, GPUs are scarce, AI is transforming everything, and compute is the new oil. They want a piece of it. They want to hedge. They want “skin in the game.” They want to avoid missing the next major technology cycle.

But hedge against what, exactly?

What determines whether compute prices rise or fall? How much supply is actually online? How much capacity is being used? Which commitments are real? Which revenue projections depend on a small number of AI labs continuing to spend money they do not currently have?

Much of the market is operating inside a vacuum of reliable information. That vacuum is then filled with narratives, promotional language, and increasingly elaborate financial products.

Venture capital is not an obvious source of discipline either. Much of the industry has performed poorly. The median venture fund returns since 2017 have been weak enough that investors could often have done better with far more conventional assets, including the S&P 500 or Treasuries.

Yet venture firms continue backing AI companies that require perpetual capital infusions. Cognition raising billions at an enormous valuation to work with open-source models is not a sign of sober capital allocation. It is a sign that the industry is still eager to spend on anything adjacent to the AI story.

There are even reports of venture firms considering direct compute investments to support their portfolio companies. That is an absurd escalation. Venture capital firms are supposed to invest in companies, not expand their own operating expenses to purchase infrastructure that keeps their investments alive.

It is sucker investments on top of sucker investments.

The Compute Reset Wall Starts in 2027

The real danger to the AI boom is not a hypothetical sentient machine. It is cash flow.

The best analogy is the adjustable-rate mortgage reset wall that helped trigger the global financial crisis. In the years leading up to 2007, many borrowers took mortgages with teaser rates. Their payments were manageable for a few years, then the interest rates reset. Suddenly, payments jumped. Mortgages that had appeared affordable became impossible to service.

The AI industry has built a version of that structure around compute.

OpenAI and Anthropic have signed huge agreements with companies including Microsoft, Google, Amazon, CoreWeave, Broadcom, and others. These agreements often require relatively little money upfront compared with their eventual value. The compute must be built and brought online first.

Then comes the reset.

As capacity becomes available, the AI labs must begin paying for it. Many of these agreements are take-or-pay contracts. That means the buyer must pay regardless of whether it uses the compute, whether customer demand appears, or whether the buyer has the money available.

That is the problem.

Deferred Commitments Become Immediate Cash Demands

OpenAI and Anthropic have not yet had to pay the full cost of the compute, chips, and capacity they have committed to purchase. That is why the system has not broken already.

But from 2027 onward, the cash requirements could become extraordinary. Estimates suggest roughly $200 billion could come due next year, potentially more depending on the underlying contracts and timelines. OpenAI has projected spending of approximately $750 billion across three years, which translates to hundreds of billions in annual spending.

That is not money the companies currently generate from operations.

OpenAI may have spent tens of billions on compute in 2025. But paying $100 billion, $200 billion, or potentially more per year is a different category of business entirely. It is not a normal scaling challenge. It requires a level of cash generation that few companies in history have achieved.

For perspective, Microsoft’s annual operating expenses are around $175 billion, while Meta’s are around $140 billion. Both are mature cash-generating machines with enormous annual revenue. Microsoft generates roughly $330 billion in revenue, while Meta generates around $200 billion.

Even those companies have operating costs far below the potential obligations being placed on OpenAI and Anthropic.

Debt and Chip Obligations Make the Problem Worse

Compute is not the only bill.

OpenAI and Anthropic are also expected to spend at least $200 billion with Broadcom over the next three years. Anthropic has debt obligations tied to financing for TPUs, reportedly involving lenders such as Apollo and Blackstone. The structure involves chips purchased from Broadcom, installed in Google data centers, then rented back to Anthropic.

That is a lot of moving pieces, but the essential point is very simple: bills will arrive.

A revolving credit facility, such as Anthropic’s reported $15 billion effort, may help cover a short-term need. It does not solve hundreds of billions of dollars in recurring compute, hardware, and infrastructure commitments. The companies must find the money through venture capital, debt markets, public offerings, or actual operating income.

The cash is not there today.

Can OpenAI and Anthropic Pay Their Way Out?

In theory, yes. The companies could raise more capital, go public, issue bonds, expand revenue, or renegotiate commitments. But this is where the scale becomes ridiculous.

To meet these obligations while continuing to operate, OpenAI and Anthropic would need to become companies larger than Microsoft, Google, and Amazon in their ability to produce cash. They would need to support enormous infrastructure costs while also funding research, product development, staffing, debt service, and future training runs.

That is a remarkable assumption to make about companies whose economics are still deeply questionable.

Even if an IPO arrives, bond markets are not an infinite source of money. The idea that investors will happily provide hundreds of billions of dollars per year to companies with weak margins and massive negative cash flow is not something that can be treated as inevitable.

And if OpenAI or Anthropic cannot pay, the consequences spread outward.

• Hyperscalers miss expected revenue from AI capacity.
• CoreWeave and similar providers lose major customers or face delayed payments.
• Broadcom sells fewer chips than anticipated.
• Cloud revenue projections fail to materialize.
• Accounts receivable balloon as providers deliver compute without receiving cash.

A significant rise in accounts receivable would be particularly worrying. It could mean infrastructure companies are recognizing revenue for services provided to customers that have not actually paid. That is not a healthy foundation for an industry already dependent on enormous financial commitments.

Nvidia may be more insulated because it has so many buyers competing for its hardware. But the broader AI infrastructure ecosystem depends on a relatively small group of companies paying a staggering amount of money on schedule.

They have not had to do that yet.

The Bubble Does Not Need to Pop in 2027 to Be a Problem

It is dangerous to make precise predictions about timing. Maybe the pressure becomes unbearable in 2027. Maybe companies limp through by raising new money, shifting obligations, refinancing debt, or creating more elaborate financial structures. Maybe the break comes in 2028 or 2029 instead.

The exact year matters less than the underlying reality.

These companies have made commitments that demand unprecedented levels of cash generation. They are already heavily cash-flow negative. Their products must either become vastly more profitable, or the industry must keep finding new pools of capital willing to subsidize losses at a historic scale.

Eventually, limited partners, lenders, public-market investors, and infrastructure providers will ask an uncomplicated question: where is the money?

Not the valuation. Not the hype. Not the theoretical superintelligence. Not the number of data centers announced. Cash.

That is the reset wall. It is not a story about artificial intelligence becoming alive. It is a story about an industry that promised to spend vast sums of money and will soon need to prove it can pay.

Source credit: This article is based on a video from the The Tech Report YouTube channel, featuring Ed Zitron in conversation with Isaac Pound.