A friend recently sent me Groundbreaker’s The Teaser Period: Why the AI Boom Is Built to Break. They found it deeply worrying, and I can understand why. The article makes a serious argument around a set of facts that deserve attention.
Frontier AI companies have signed enormous commitments for compute capacity that does not yet exist. Data centers take years to build. Payments begin as that capacity comes online. Some of those commitments behave economically like debt because the customer has to pay whether or not it ultimately needs all the capacity.
The article argues that this creates a 2027–28 “reset wall” resembling the adjustable-rate mortgage resets that exposed the fragility of the housing market in 2007 and 2008. OpenAI receives particular attention because its future compute commitments are huge relative to its current revenue.
I think that financing risk is real. OpenAI could have overcommitted. Other AI companies could have overcommitted. Data-center operators, neoclouds and lenders may discover that they financed projects at assumptions that no longer hold when the capacity finally comes online.
All of that seems reasonable to me.
The 2008 analogy still bothers me because it makes the outcome feel much more likely than I think it is. The article begins with a useful question about financing risk and gradually makes the situation feel like a forecast about AI demand itself.
Those are different questions.
A company can fail because it financed useful infrastructure badly. Investors can lose enormous amounts of money while society continues consuming more of the underlying product.
I think that distinction matters enormously here.
A house stays a house
The 2008 analogy starts to break down when I think about the speed of change.
A house built outside Las Vegas in 2006 remained essentially the same asset in 2008. It did not become ten times more productive while the mortgage aged. Its operating cost did not fall by orders of magnitude. New technology did not suddenly create hundreds of new things people could do with it. Someone in India could not start using it because its price dropped.
Housing demand also changes slowly. Population, household formation, migration and income move over years. A house cannot instantly move from an unsuccessful owner to a more productive user in another market.
AI compute occupies an entirely distinct category.
A data center financed today may open into a market with different chips, better models, lower inference costs, new applications and a much larger population of users. The contractual obligation can remain fixed while almost everything that determines the economic value of the underlying compute changes.
That difference seems important enough that I am reluctant to treat a three-year compute commitment like a three-year mortgage reset.
I see at least five forces that could increase demand while today’s capacity gets built. I do not know how strong each will be, and some may be disappointing. My concern with the Groundbreaker analysis is that these forces receive much less attention than the fixed commitments, even though they will help determine whether those commitments are excessive.
1. Better and cheaper AI should create more demand
The first force is a version of Jevons paradox. Efficiency does not necessarily reduce total consumption of a useful resource. Lower prices often make more uses economical, and total consumption can rise even though each individual use becomes cheaper.
AI may amplify that effect because it is becoming both cheaper and more capable.
A tenfold reduction in the cost of performing the same task should increase consumption if demand has a reasonable elasticity. A tenfold reduction in the cost accompanied by a significant increase in capability does something more interesting: it creates demand for tasks that were not previously possible.
A cheaper version of a model that can summarize a document expands an existing market. A model that becomes reliable enough to perform an accounting workflow, write a substantial software feature, operate an application or reason through a scientific problem can create an entirely new one.
Global adoption adds another dimension that I think deserves more attention. AI remains much more heavily used in wealthy countries and wealthy companies than throughout much of the rest of the world.
A useful AI service that costs $10 per task has one market. A better version that costs ten cents has a very different market. Smaller companies can use it. Lower-value applications become worthwhile. Consumers who could never justify the old price can participate.
The service can also reach much of the world without anyone building a new physical distribution network.
The relevant future demand curve therefore does not consist only of today’s American enterprises buying more AI. Falling costs can move entirely new populations and categories of work onto that curve.
I do not know how elastic that demand will be. I would hesitate to assume that efficiency gains translate directly into lower aggregate compute spending.
2. Capability thresholds can create sudden booms
I also doubt AI demand will grow smoothly.
Many applications have thresholds below which they simply do not work well enough to matter. A system that performs a task correctly 80% of the time may require so much human supervision that it creates no economic value. If you raise reliability enough then the same application can suddenly become useful.
Coding provides an obvious example, but the idea applies broadly. Customer service, accounting, research, document processing, medicine, robotics and administrative work all contain tasks where a modest improvement near the right threshold can produce a large change in economic value.
An agent that can reliably work for five minutes creates one set of applications. An agent that can work for several hours without losing the plot creates another. A system that still needs a human watching every step is a tool. A system that can reliably complete meaningful work on its own begins to compete with labor.
Those transitions probably will not happen gradually from an economic perspective. Capability can improve steadily and then demand can increase very fast once you cross that threshold.
I find it easy to overlook that we as humans can also improve in how we use the AI because we spend so much time measuring models. I use AI much more effectively than I did a year ago. Some of that comes from better models, but some comes from learning where AI helps, where it fails, how to structure problems for it and how to incorporate it into my ordinary work.
Companies go through the same process. Organizations may already have access to capabilities they barely know how to use. This also moves the demand curve.
3. AI reduces the cost of creating software
Software has always had a large fixed cost. An application has to create enough value to justify the engineers, designers, infrastructure and ongoing maintenance required to build it.
Companies consequently have enormous backlogs of software they would like to have but cannot economically justify.
An internal application that would cost $1 million to build and save a company $400,000 never gets built. Reduce its development cost to $100,000 and it makes sense.
That new application then consumes cloud infrastructure, databases, storage, networking and probably more AI compute. This same calculation occurs across millions of potential projects in parallel..
One possible future has companies using AI simply to produce today’s quantity of software with fewer programmers. I expect some of that. Another future has us producing vastly more software because the marginal project becomes much cheaper. I find that outcome at least as plausible.
Software has never suffered from a shortage of things people would like computers to do. We have suffered from the cost of telling computers how to do them.
That creates in interesting positive feedback loop:
Better AI → cheaper software creation → more software → more cloud consumption.
4. Much of that new software will use AI
This creates a second-order effect that I initially forgot, despite encountering it in my own work nearly every day.
AI does not just help us create more conventional software. It helps us create software that uses AI.
Better AI makes software development cheaper and faster. That allows more applications to exist. Many of those applications then call models, run agents, process images, generate speech, use embeddings, reason over documents or invoke other forms of machine intelligence.
I think this is an even more interesting positive feedback loop:
Better AI → cheaper software development → more applications → more AI-enabled applications → more inference demand.
AI may also reduce the minimum market size required to justify creating software. Traditional development costs encourage us to build products for thousands or millions of users. Much cheaper development makes applications for one company or one department more reasonable.
Highly personalized software may become common. People may routinely create applications for one project, one team or even one person because the engineering cost has fallen enough to make that sensible. I know in my day to day work I regularly write small applications to do tedious things for me, and most of them have some AI component to them.
That could expand the number of software applications by a very large amount. We do not need every new application to consume enormous amounts of AI for the aggregate effect to matter.
Efficiency cuts the other way, of course. Many applications will use small models, deterministic code, caching or local inference rather than continuously calling frontier models in the cloud. Engineers will probably become much better at using expensive intelligence only when they need it.
The important question is whether the number and sophistication of applications grow faster.
My guess is that they will.
5. Better AI can help build better AI
The last loop carries more uncertainty, but it may eventually become the most important one.
Semiconductors already provide a modest version of this phenomenon. Better chips give engineers better tools for designing better chips. Faster computers allow more sophisticated simulation, verification, optimization and physical design. Each generation of computing helps create the next.
AI could develop a similar recursive character. AI already helps engineers write code, analyze experiments, generate tests, explore alternatives and optimize systems. Better AI should make at least some of that work faster.
No intelligence explosion is required for this to matter.
An AI researcher who can run three times as many useful experiments will consume more experimental compute. Some of those experiments will produce better models. Better models then become better research tools.
The loop starts again:
More compute → better AI → more productive AI research → better AI → greater economic value → more demand for compute.
Efficiency improvements work against this loop too. Better algorithms can reduce the resources required to reach a particular level of capability.
That does not settle the demand question. It simply returns us to the same problem: will falling cost per unit of intelligence outrun growth in the amount of intelligence we find worthwhile to consume?
I just do not know on this one.
Fast technological cycles create a strange kind of risk
The speed of AI development does not eliminate the possibility of a glut. It may actually make temporary gluts more likely.
Data centers take years to build. Forecasts made today will inevitably be wrong in important ways when the capacity arrives. Hardware may depreciate faster than expected. A cluster that looked extremely valuable when someone committed to it may compete against substantially better equipment by the time it reaches full utilization. Owners can lose a great deal of money in that environment. However, speed also makes a prolonged glut harder to sustain.
A data center may take years to build, while software can change in days or weeks. Providers can alter prices almost immediately. Workloads can shift far faster than physical assets can. A model release can make a previously marginal application economically interesting in months. Cheaper AI can become accessible to new users around the world without anyone building anything new.
Compute is not perfectly fungible. Hardware generations matter. Networking matters. Geography, latency, power and data locality matter. Specialized training clusters cannot always become generic inference capacity with the flip of a switch.
The comparison with housing remains weak despite those constraints. Much of the demand for compute can move electronically toward available capacity, and falling prices can itself stimulate new demand. We may also use AI to help us make better use of any older hardware or hardware that was targeted to an older use case.
That makes me expect repeated mismatches rather than one great 2008 style reckoning.
The distinction might be stated this way:
Long construction times and large forecast errors increase the probability of a glut. High demand elasticity, fungibility and rapid application creation reduce the probability that the glut persists.
AI is exposed to the first two factors, but it is also likely that gluts will be short lived.
Electricity offers a useful analogy
Electricity provides one way to think about this.
Electric motors became more efficient. Lighting became dramatically more efficient. Appliances became more efficient. Computers became more efficient.
Total electricity consumption still rose because technological progress kept creating more uses for electricity. Air conditioning spread. Factories automated. Refrigeration expanded. Computers arrived. Data centers followed. Electric vehicles created another source of demand.
The industry hardly traveled along a smooth path. Power markets have experienced shortages, excess capacity, bankruptcies, price spikes and terrible investments.
A particular generating plant can make no economic sense while the world still needs more electricity over time.
Efficiency can reduce the amount of compute required for one task without reducing aggregate demand for compute. Better and cheaper intelligence can simply make more tasks worthwhile.
AI compute may behave similarly, although I suspect its cycles will run much faster.
Semiconductors may provide an even closer analogy
The semiconductor industry has spent decades moving generally up and to the right while experiencing repeated and sometimes brutal cycles.
Chip companies make enormous capital commitments long before they know future demand. Shortages encourage investment. New fabs arrive. Inventories grow. Prices fall. Companies cut production and capital spending. Demand eventually catches up.
Memory has produced some spectacular examples. Those gluts were real. They also occurred inside one of the great secular growth industries of the last half century.
The semiconductor industry did not avoid gluts because its technology improved quickly. Fast improvement probably contributed to them. Companies made forecasts against a moving target and sometimes built too much of the wrong thing.
The interesting part is what happened next.
New applications appeared. Better chips lowered costs and created new capabilities. Demand absorbed the capacity. The cycle repeated from a larger base.
Semiconductors also contain the recursive mechanism I mentioned earlier. Better chips give engineers more powerful tools for designing better chips. AI may magnify that process because AI can help design not only the next generation of hardware but also the software and AI systems that consume it.
I expect gluts. I am less convinced by dark fiber.
This distinction has become central to how I think about the question.
Long construction times and bad forecasts make temporary oversupply likely. Rapid technological improvement increases the odds that someone builds the wrong capacity, buys the wrong generation of hardware or pays a price that later looks foolish.
I would be surprised if AI did not experience periodic gluts.
Persistent oversupply requires more.
Imagine the industry builds enough capacity for expected demand of 100 and only 70 arrives. Prices fall. Some leveraged owners fail. Older hardware gets written down.
Cheaper compute then makes another set of applications worthwhile. Developers build more software. More of that software uses AI. A new model crosses an important capability threshold. Enterprises increase usage. Lower prices expand adoption in countries and applications that could not previously afford it.
Demand reaches 110.
The original glut really happened. The financial losses were real. It simply did not become a long-lived shortage of useful things to do with the infrastructure.
That outcome seems considerably more plausible to me than treating all excess capacity as the AI equivalent of unused fiber. The supply side moves slowly but the demand side can adapt unusually fast.
Financial losses do not prove that the infrastructure was unnecessary
This point changes how I think about OpenAI and some of the more aggressive infrastructure commitments. OpenAI may have committed to more compute than its eventual business can support. I do not know. The numbers certainly justify asking the question.
The company could fail. A neocloud could restructure. Lenders could lose billions. Hyperscalers could write down projects. AI stocks could suffer a severe correction.
None of those outcomes necessarily tells us whether aggregate AI compute demand remains healthy.
The original buyer of a GPU can earn a terrible return while someone buying the same capacity after a restructuring earns an excellent one. A data-center company can fail financially while its facilities remain full. A lender can lose money even as total consumption of compute rises.
Financial return and physical demand are related, but they are not the same thing.
Technology history contains plenty of examples where investors financed useful infrastructure badly. Telecommunications provides one. Railroads provide another. Semiconductors provide them repeatedly.
The capital structure can fail without the underlying demand thesis failing.
What would have to be true for a real AI-wide dark-fiber moment?
I find this a more useful way to evaluate the bearish case.
A stock-market correction does not qualify. Neither does the failure of OpenAI or several data-center companies.
A true AI-wide dark-fiber moment would mean that the world built substantially more useful AI compute than customers wanted even after prices adjusted, and that the excess persisted for years.
Several things would probably need to go wrong together.
AI capability would have to improve much more slowly. Falling prices would need to generate surprisingly little additional usage. New capability thresholds would need to create relatively few important applications. Global adoption would need to expand slowly even as AI became more affordable.
AI coding tools would need to be used to reduce the number of programmers rather than increase the amount of software we produce. The additional software we did create would need to consume relatively little AI. AI-assisted research would need to contribute little additional demand for experimentation. Companies would need to approach saturation in their ability to find valuable uses for cheap machine intelligence.
Excess capacity would also need to prove relatively difficult to redirect, and falling prices would need to do relatively little to stimulate new uses.
Any one of those outcomes seems plausible to me.
Their conjunction seems considerably less so.
That is why I find the Groundbreaker argument useful as a warning about financing but less persuasive as a forecast for the industry.
My base case is a secular boom with painful corrections
I do not expect AI to follow a smooth path upward.
I expect shortages and overbuilding. I expect localized gluts. I expect hardware to depreciate faster than some owners anticipate. I expect companies with fragile financing to fail. I expect at least one large AI company to discover that a commitment looked much better when it signed the contract than when the invoice arrived.
I would also be surprised if AI equities made it through the next decade without one or more severe corrections.
Real booms can occur between those corrections.
A sufficiently reliable coding agent could create one. General-purpose business agents could create one. Personalized software could create one. Scientific applications, robotics or a dramatic decline in inference costs could create others.
Companies becoming much better at using capabilities that already exist could create substantial demand without any technological breakthrough at all.
The likely pattern strikes me as something like:
shortage → investment → excess capacity → falling prices → new applications → rising utilization → shortage again.
Different regions, providers, workloads and generations of hardware will probably occupy different points in that cycle at the same time.
Each correction will make a compelling story about the end of the boom. Some of those stories will correctly identify terrible investments while still misunderstanding the long-term trajectory.
The semiconductor industry has demonstrated this pattern for decades. Cyclical downturns and secular growth are not opposites. The former can form part of the latter.
That is the possibility I think the 2008 analogy misses.
The Groundbreaker warning still matters
The article gets several things right from my perspective.
Contractual backlog does not guarantee economic value. Large fixed commitments deserve careful scrutiny. Transformative technologies attract capital, and some of that capital will inevitably make poor decisions. AI does not receive an exemption from ordinary economics because the technology is exciting.
My disagreement concerns the inference we should draw from those facts.
Mortgage resets occurred inside a slow-moving system whose assets and uses changed little while the clock ran.
AI compute contracts mature inside a system where hardware, models, efficiency, software, applications, prices and users may all change substantially before the capacity comes online.
That does not guarantee good investment returns. It does make the outcome much less mechanical than the 2008 comparison suggests.
The article examines the fixed financial obligations in considerable detail. It gives much less weight to the processes continually creating new demand for cheaper and more capable intelligence.
I think those processes belong near the center of the forecast.
The question is not simply whether today’s AI businesses can support tomorrow’s compute capacity. The harder question is whether efficiency and supply can grow faster than our ability to invent valuable uses for increasingly capable and inexpensive machine intelligence.
I may be wrong.
AI progress could slow. Efficiency could outrun demand. We may discover that many promising applications produce less economic value than expected. The industry may have committed far too much capital based on extrapolations that cannot survive contact with reality.
Those possibilities deserve serious attention.
I simply do not think a comparison with the 2008 mortgage-reset cycle establishes them.
I expect corrections. I expect some of them to be severe. I also expect the underlying market for useful machine intelligence to continue expanding.
Those beliefs are not contradictory.
One final disclosure feels especially appropriate for this article: I used AI extensively to help me think through the argument, challenge my assumptions, research the underlying data and help write this piece.
I did that because AI has become intrinsic to my day-to-day work. I use it to explore ideas, write software, analyze problems and improve my own thinking. I am better at using it than I was a year ago, and the AI is better too.
That does not prove the bullish case. It is one tiny observation about the demand side.
But it illustrates why I am reluctant to model that demand as a fixed thing.