With Anthropic having hit a $44 billion run rate, up from “just” $9 billion three months ago, and on a trend line to a $100 billion run rate by the end of the year, they are putting their business on par with some of the most cash-generating business models of all time. OpenAI’s growth with Codex is just as impressive.

And while I have my counterarguments, one way or another, AI has found some sort of product-market fit, and people have finally put to rest the idea that AI is a bubble.

Well, wrong.

The economic picture in AI is much more complicated than it meets the eye; it’s bubbly in ways people in San Francisco, too smart for their own good, fail to identify.

We’re tackling all angles today. CapEx, power, memory, advanced packaging, financials, and all else there is to know about AI in an effort to truly understand the state of the industry, as always in words you will understand.

Let’s dive in.

The first mistake in analyzing the AI buildout is pretending the technology is still fake.

It was a reasonable argument in 2023, when the core product was a chatbot, gross margins were horrendous, and the most visible use cases were better search, homework help, marketing copy, and Studio Ghibli selfies.

But it is no longer a reasonable argument at all because in late 2025 and early 2026, something broke loose: AI models hit a capability threshold, especially coding agents and research/analysis agents, starting to cross from demo into production, at least in those particular areas.

The clearest evidence is revenue. SemiAnalysis estimates that Anthropic’s ARR has exploded from $9 billion to over $44 billion in just a few months, while its inference infrastructure gross margins rose from 38% to over 70%.

Whether that exact number is perfect or not, it’s impossible to ignore that AI Labs are no longer just burning money to prove capability, and they are beginning to offer actual business models in return.

SemiAnalysis also argues that the value per token has risen sharply while the cost of producing tokens has fallen, because new hardware and software optimizations are increasing throughput faster than system cost.

In layman’s terms, the performance per dollar is growing because performance (the amount of operations per second our hardware can output) is increasing much faster than the prices of new servers.

As you can see below, the prices users and companies are paying for every “ounce” of compute is dropping fast, and the rate of acceleration of this decrease is only becoming more apparent.

As mentioned, OpenAI’s numbers tell the same story. Currently, enterprise now account for more than 40% of revenue, APIs process more than 15 billion tokens per minute, Codex serves more than 2 million weekly users, Codex usage is growing more than 70% month over month (this is already a multi-million-user product), search usage nearly tripled over the prior year, and its ads pilot reached more than $100 million in ARR in under six weeks.

Cursor, soon to be part of the SpaceX ecosystem, is another good example because it is not a frontier lab. TechCrunch, citing Bloomberg, reported that Cursor surpassed $2 billion in annualized revenue, doubling its run rate in three months, with large corporate buyers accounting for roughly 60% of revenue.

The reason I’m mentioning all of this is that the analysis of whether AI is a bubble cannot be based on whether the technology is useful; this is not a tulip mania bubble.

If anything, it’s a dot-com bubble type, one that places a bet on the correct technology but might be/become overextended. Thus, the question is more about whether we have gotten ahead of ourselves like we did that time, or not.

Therefore, if what we’re discussing is whether we bit more than we could chew, the next question is: how large is the real AI revenue market today?

But I want to set a very clear disclaimer here before we proceed: not all AI revenue counts as real revenue in my book.

I’m just going to cut to the chase and call bullshit: I don’t care about (most) Hyperscaler revenue.

The reason is two-fold:

I’m also not counting NVIDIA and semis revenues, because they are not proof that AI is real; they are just benefactors and enablers of this great buildout.

Knowing all this, using the available numbers, the answer is probably already in the ballpark of $80 billion to $120 billion annualized right now, an amount that has more than tripled since the beginning of the year.

Anthropic alone is at $44 billion. OpenAI is plausibly above $40 billion in annualized revenue, seeing how Codex is growing at 70% month-over-month over a million-user base.

As mentioned, Cursor is around $2 billion. Add GitHub Copilot, Claude Code, API usage across enterprises, Perplexity, Midjourney, Runway, Lovable, Replit, Cognition, Harvey, Glean, enterprise copilots, vertical agents, inference providers, and the growing long tail, and a rough $100 billion/year real AI revenue estimate is no longer absurd.

All in all, this is not a debate between “AI is fake” and “AI is not a bubble.” The technology can be real, and the financing structure can still be bubbly.

But is the AI buildout at the scale it is happening justified?

Once you accept that the technology is real, the second mistake is to pretend the buildout is just vibes. It is not.

For what it’s worth, the scale is indeed massive and dizzying.

According to recent estimates, the boom could reach as high as one trillion per year in 2027, driven by just five companies: Meta, Microsoft, Amazon, Google, and Oracle. For reference, that’s more than Taiwan’s entire GDP, and more than Portugal, Greece, and Finland combined.

Feels quite unprecedented. But is it? Well, actually, not at all.

As we discussed recently in this newsletter, you would be mistaken to think this is unprecedented. In fact, it doesn’t even come close to the rate of investment other supercycles saw relative to world GDP at the time (what percentage of world GDP was a certain investment cycle).

The historical comparison matters. Even if, in nominal terms, this is the largest investment boom ever, if we discount inflation, there have been several investment cycles in history, like the dot-com and rail-and-auto expansions, which required much more investment relative to the size of the economy at the time.

The point is not that AI infrastructure is small. It is very large indeed. The point is that, compared to other global infrastructure booms, it is not “metaphysically impossible.”

Societies have financed giant physical transitions before. We financed the cloud. We financed the internet. We financed railroads, telecom, oil, gas, electricity, and clean energy.

But the AI buildout certainly has a different shape. Clean energy, oil, and gas were attached to obvious final-demand categories: electricity, fuel, heating, mobility, industrial energy, and national energy security.

AI infrastructure is attached to a revenue curve that is much younger, smaller, and harder to observe. Therefore, the relevant question is not “is the buildout large?” The relevant question is “Is the buildout large relative to the speed at which real AI revenues can compound?”

And this is precisely why the gap matters. A $100 billion real AI revenue market can support significant investment if it is growing 5x or 10x.

But it cannot support a multi-trillion-dollar buildout unless it keeps compounding for years, margins hold, utilization stays high, and enterprises move from “AI as productivity software” into “AI as operating layer.”

Now, finally, we’re talking. This is the question that needs answering. People analyzing this technology predicate their entire thought process on whether the buildout is too large or whether the technology is real or false.

All this brings us to the first big insight of this piece: to identify the bubble, we need to stop being casuals and start looking at this technology from the right framing: the difference between what gets announced and what actually gets done.

AI is real. But I never said this industry doesn’t have a lot of fiction worth a Hollywood debut. And most of that fiction comes from the announcements themselves.

The AI infrastructure market has become addicted to announcements. Ten gigawatts here. Five gigawatts there. A $100 billion partnership. A $300 billion cloud deal. A $500 billion Stargate plan. A 1 GW campus. A 2 GW campus. $1.4 trillion in commitments from a single company. A 7 GW site in New Mexico. A 10.8 GW power-development pipeline.

The numbers are so large and the deals so many that we’ve become numb to them, and our minds stop caring whether they are scheduled, financed, permitted, interconnected, under construction, energized, or useful.

But those distinctions matter. A lot, because here’s the thing:

Announced capacity is not funded capacity.

Funded capacity is not under-construction capacity.

Under-construction capacity is not energized capacity.

Energized capacity is not useful capacity.

And useful capacity is not profitable capacity.

Therefore, the state of the AI buildout must be measured as a conversion funnel, not as a press-release database.

And here’s where things start to get sketchy: the gap between what gets announced and what's actually getting done is daunting, and it very well explains who’s winning and who’s losing.

According to Sightline Climate’s Q1 2026 Data Center Outlook, which tracks 190 GW across 777 hyperscale projects above 50 MW announced since 2024, there’s a huge gap between the announcements and reality: 148 GW announced, 20.6 GW in construction, 12.3 GW operational, 3.0 GW delayed, and 1.4 GW canceled.

The 2026 funnel is particularly revealing. Sightline says at least 16 GW of data-center capacity is slated to come online in 2026 across 140 projects, but only 5 GW is currently under construction.

Furthermore, it expects 30% to 50% of 2026 projects to be delayed. It also says 11 GW of 2026 capacity remains in the announced stage with no signs of construction, despite typical build times of 12 to 18 months.

But to actually energize the site, someone has to secure land, power, interconnection, transformers, switchgear, turbines or PPAs, cooling, permits, water, construction labor, tenant commitments, debt, equity, GPUs, memory, networking, storage, and a cluster-operations team. Each step can fail. Each step can slip.

Now, with all the defenses I’ve made regarding AI up to this sentence, this is a real problem; I believe it’s hardly priced in by markets.

The reasons for the delay aren’t mysteries either. Sightline reports that 26% of 2025 capacity was delayed, and another 10% of projects quietly shifted their commercial operation dates back as power, permitting, and construction constraints bit.

For 2026, it explicitly identifies power constraints, increasingly effective community opposition, and grid-equipment shortages as delay drivers.

And to be clear, we’ve known this would happen for almost two years now, as I wrote about this particular issue back in June 2024.