The Biggest Sunk Cost in Human History

I've been reading AI funding announcements for a while now, and at some point I started noticing that the numbers stopped making sense to me as a student who's spent the last few years actually learning how markets are supposed to price things. I don't think that's cynicism. I think it's just doing the arithmetic.

The numbers I keep staring at

Here's what's real. Anthropic raised $65 billion in its Series H round this year at a $965 billion post-money valuation, with a $47 billion annualized revenue run rate behind it. Nvidia's market cap has crossed $3 trillion at various points. Microsoft, Google, Amazon, and Meta have committed hundreds of billions in capital expenditure to AI infrastructure, data centers, custom chips, power buildout, on timelines stretching years out.

OpenAI's own confidential S-1 filing tells a rougher story underneath the same excitement. The company is burning roughly $27 billion this year against revenue running at about $25 billion annualized, meaning it's losing something like $1.22 for every dollar it brings in. That's a real business with real revenue, I want to be clear about that, but it's not a profitable one, and it's burning cash at a scale that makes the "just wait for the growth curve" argument harder to make every quarter.

Anthropic's situation is more defensible in my opinion. $47 billion in annualized revenue is genuine enterprise adoption, not vapor. But $965 billion on $47 billion of revenue works out to roughly a 20x revenue multiple. For comparison, Apple and Google, two of the most dominant companies in the world, trade at something closer to 6 to 7 times revenue. The gap between those numbers is, I think, exactly where the risk in this whole thing is sitting. Investors aren't paying for what these companies are today. They're paying for a story about what they'll become.

Why I think this looks like an exit, not a bet

Here's my actual read on what's happening, and I'll admit this is the part where I'm speculating rather than citing a filing. Both Anthropic and OpenAI have confidentially filed paperwork for IPOs this year, Anthropic targeting a Nasdaq listing as early as October, OpenAI reportedly leaning toward 2027 after initially eyeing 2026. I don't think the timing of two frontier labs both moving toward public markets in the same year is a coincidence.

Scaling transformer models on ever-larger GPU clusters is showing the diminishing returns a lot of serious researchers predicted years ago. Capability gains between frontier model generations are real but visibly narrowing, and the data needed to keep pushing further is getting scarcer and more expensive to acquire. I think the honest thing for these companies to tell investors would be that the current architecture has real limits and that genuinely new research directions are needed. I also think that admitting that out loud would tank the valuations overnight, so nobody's going to say it until they have to.

An IPO, in that context, looks to me like an exit for early capital that got in at much lower valuations, cashing out at what might be peak narrative, and handing the risk to public markets, pension funds, and index investors who'll end up holding AI exposure through a broad market fund whether they chose to or not.

The comparison I keep coming back to

I think the closer parallel here isn't the dotcom crash, though people reach for that one first. It's 2008. The mortgage crisis wasn't really about bad loans, everyone already knew housing was overvalued by 2006. The actual damage came from nobody having modeled how deeply that risk was woven into instruments held by systemically important institutions.

Nvidia sits in almost every broad market index and retirement account with passive equity exposure today. I don't think that concentration has been priced properly yet. The dotcom crash didn't kill the internet, it killed the speculation, and Amazon looked nothing like Pets.com when the dust settled. I'd guess the infrastructure layer and the application layer of this AI cycle unwind very differently too, if a correction comes. Figuring out which is which seems like the actually useful analytical work right now, more useful than arguing about whether the whole thing is a bubble.

The part that doesn't show up in any pitch deck

There's a layer under all of this that I think gets almost no serious coverage relative to how central it is. A huge amount of the value these models generate was built on inputs nobody paid for: decades of writing, code, and creative work scraped without consent or compensation, plus data labeling work done by people in Kenya, the Philippines, and Venezuela for very low wages, including work involving genuinely disturbing content that's caused documented psychological harm to the people doing it.

I don't think that's incidental to the business model. I think it might be closer to the business model itself. There's ongoing copyright litigation and regulatory attention in the EU and UK that represents real financial exposure, and I notice it's almost entirely absent from the pitch being made to retail investors ahead of these IPOs.

Where I land on this

I think the ceiling on current architectures is real and getting closer, and I think the genuinely novel research that might get past it, world models, neurosymbolic approaches, whatever comes next, is underfunded relative to the money being poured into scaling the current approach, because the scaling spend is what justifies today's valuations, not the alternative research. Redirecting that capital toward long-shot research would probably be the correct long-term call. It would also collapse the story holding current valuations up, so I don't expect it to happen voluntarily.

My guess is the correction comes from something boring rather than dramatic: a benchmark plateau nobody can spin anymore, a missed capability milestone, or just enough time passing that the promised future keeps sliding further out. And when it comes, I don't think it's the people who built these companies who get hurt by it. I think they'll have already sold.

I'm a CS and financial engineering student, not a fund manager, so take this for what it's worth. But the math is the math, and I haven't found a version of it yet that makes the current multiples make sense to me.

Sources: Anthropic Series H announcement · TechCrunch · CNBC · Anthropic confidential S-1 announcement · FutureSearch OpenAI financials · Forbes