Last week I ran the first part of my What If…We’re In An AI Bubble? Series, where I asked questions and posed scenarios as to the consequences of the many, many questions I’ve asked over the last few years. It quickly became one of my most-read articles I’ve ever written, and for those of you who joined me for the first time last week, here’s a quick list of what we’ve covered already:

• What If The AI Industry Moves To Entirely Token-Based Billing? 
• What If Organizations Can’t Afford To Keep Spending On AI?
• What If The AI Capacity Crunch Never Ends (And Data Centers Aren’t Getting Built)?
• What If CoreWeave Can’t Keep Up With Its Capacity Demands?
• What If Hyperscalers Can’t Build Data Centers Very Fast?
• What If Hyperscalers Have Warehouses of Uninstalled GPUs?
• What If Hyperscalers Write Off A Large Chunk of GPUs?
• What If Data Center Construction Demand Collapses? 

As I mentioned last week, I believe one of the many problems with the analysis of the AI bubble is that people are willing to consider individual facts — like that AI is too expensive for everybody involved and data centers are not being built at the speed that we believed — but never the gestalt of their consequences. 

For example, if data center construction slows to a crawl (as I’ve discussed is already the case) there’s a cascade of events that will occur:

• OpenAI and Anthropic can’t expand much further than their current capacity.
• As they both make up 50% of Amazon, Google and Microsoft’s revenue backlogs, hyperscalers will be unable to make the majority of the revenue they’ve promised their shareholders.
• The $178.5 billion in US data center debt from 2025 will go mostly unpaid, as a great deal of it is project financing that’s dependent on revenue from data centers that won’t be built and thus won’t be making any revenue.
• NVIDIA, which claims to have shipped over 3 million Blackwell GPUs in 2025, will have trouble selling its next-generation Vera Rubin GPUs, as nobody will have anywhere to put them.
  • Alternatively, we’ll see write offs of billions of Blackwell GPUs that will now be considered obsolete.
• Banks that are already afraid of “choking” on data center debt will stop issuing it, because these investments will not be paying off.
• It will become very difficult for anybody to afford to buy more NVIDIA GPUs, because AI data centers — which cost around $44 million per megawatt — require massive amounts of upfront capital expenditures, making it unlikely-to-impossible that somebody has the money lying around.

It’s really easy to say “wow, this stuff needs a lot of debt!” and “wow, this stuff takes a while!” but actually sitting and thinking about what that means logically leads you to some gruesome outcomes. 

And to be clear, there’s not really an alternative to that scenario if data center construction slows. Even in an optimistic scenario, if data centers that started being built in 2024 don’t get finished until 2027 or 2028, that means that NVIDIA’s “latest” GPUs are perennially two or three years in the future. 

While some capacity exists, I believe there are at least one million Blackwell GPUs sitting in warehouses waiting to be installed years into the future, which means that projects are going to launch in a year or two with potentially three-year-old GPUs, or said projects are going to have to either replace their orders with Vera Rubin or dump aged capacity onto a market saturated with Blackwell GPUs.

The argument against what I’m saying is that there’s “insatiable” demand for AI compute — that “any viable compute on the market will be used,” which is true in measurements of days or months, but breaks down in the space of a year. As I mentioned a few weeks ago, AI’s demand story is a lie, because capacity is mostly taken up by Anthropic and OpenAI, creating the illusion of demand by absorbing most available inventory, while simultaneously obfuscating the fact that other sources of demand are simply non-existent in any meaningful numbers.. 

Many are conflating “there’s not much available” with “there’s so many people that want GPUs” without quantifying what “so many” means or how much they want, when the remaining performance obligations from Google, Amazon, and Microsoft have, outside of OpenAI and Anthropic, effectively plateaued, as is also the case when you remove these companies from CoreWeave order book. 

If there were incredible, insatiable, indisputable demand, RPOs would be exploding across the board. Instead, nobody seems interested in buying capacity at scale outside of Anthropic, OpenAI, and the hyperscalers supporting them — or, in some cases, the likes of NVIDIA providing backstops to compute providers, agreeing to buy surplus compute in the case that they’re unable to sell it themselves. This is, to be clear, something that shouldn’t happen if there was genuine, distributed demand.  

The sheer scale of the supposed AI data center buildout is in the tens of gigawatts of capacity, which translates to  $10 billion to $15 billion per gigawatt in annual revenue. I can find no examples of anybody but Anthropic and OpenAI spending billions on compute. 

Both companies need to make or raise a combined $1.25 trillion in the next four years to afford their compute commitments across Oracle, Microsoft, Google, Amazon and CoreWeave. 

The counter-argument to everything I’m saying is effectively two points:

• Nuh uh!
• That the amount of revenue flowing to both NVIDIA and associated hardware companies making CPUs, RAM, and solid-state storage is proof that there’s demand for…services run on them.

The latter is far from compelling, but I can see how somebody would believe it. 

So much money appears to be flooding into companies like AMD, Samsung, and Sandisk — tens of billions of dollars to the point that it’s creating shortages across basically every component imaginable — which naturally might make you think that demand would exist at the other end.

For the consumer, that perception becomes even more believable when you notice how consumer electronics are getting more expensive. Certain games consoles, nearly six years after their initial release, are more expensive than they were at launch. Typically, the inverse is true. 

Meanwhile, smartphones and PCs are expected to ship with weaker specs or high prices, in part because of shortages of key components, caused by demand for AI data center hardware. 

The thing is, demand for AI compute doesn’t have to exist for AI data centers to get built. While some have clients signed up in advance, said deals were signed so many years before construction will complete that it’s hard to guarantee that they’ll be willing — or solvent enough — to pay. 

I also imagine most clients have signed contracts that have milestone dates for delivery of compute capacity. If data centers are delayed, clients likely have a contractual out, much like Microsoft does with its $17 billion compute deal with Nebius.

In any case, in a frothy debt market full of desperate speculation, these projects are being funded by the very same private credit firms that piled into SaaS companies between 2018 and 2022 under the assumption that every software company will grow in perpetuity. When due diligence is so weak in private equity and private credit that Apollo’s John Zito says that their valuations are “all wrong,” it’s hard to believe that the same financiers are diligently making sure that enough revenue exists to justify these massive data center debt deals.

The same questionable attention to detail applies to venture capital, which has seen (much like private equity) its investment model slow to a crawl since 2018, with an average TVPI (total value paid in) slow to a horrifying 0.8 to 1.2x since 2018, meaning that for every dollar invested, you’re at best likely to get even money in return. 

These are the very same investors telling you that every AI company is worth perpetually-growing amounts of money, that everything will work out perfectly, that somebody will work out how to make AI profitable, and that AI is both here to stay and doing incredible things, even if they can’t really explain what those things might be.

In reality, none of these people have any idea how to turn around these rotten economics. Data centers are massive money-losing operations that in the best case scenario take five years to make a single dollar of margin, and their customers are eternally-unprofitable AI startups that rely on a constant flow of venture capital dollars. 

The AI bubble is entirely built by people who hope somebody else will solve their problems. AI labs depend on venture capitalists to fund them, hardware providers to invent silicon that makes their businesses profitable, and their AI startup clients to find ways to make profitable businesses using their APIs. In turn, AI startups rely on AI labs to work out a way to make their models cheaper so that AI startups can make their business models profitable. 

Put another way, everybody’s response to “how does this become profitable” is “don’t worry, somebody will work it out, but don’t worry, they’re going to at some point.”

Today, I want to explore what happens if they don’t. 

Time. Space. Reality.

It's more than a linear path — it’s a prism of endless possibility. I am the Watcher, and I am well aware of how AI generated that sentence sounds. 

I am your guide through these vast new realities.

Follow me and dare to face the unknown.

And ponder the question…

What if…We’re In An AI Bubble? 

In Today’s Where’s Your Ed At Premium…

• What if venture capital funding stops flowing to AI startups?
  • What would make venture capital stop funding AI startups?
• What if most AI startups go to zero?
• What if OpenAI and Anthropic became AI’s lender of last resort? 
• What if AI broke venture capital’s back? 
• What if inference isn’t profitable?