It’s not quite schadenfreude, because it’s not funny yet. It’s not quite dread, because you’ve already done the grief work once. It’s more like sitting in a theater for a sequel you didn’t ask for, watching the protagonist walk toward the basement again, and being genuinely unsure whether to laugh or leave.
That’s where I am with the AI venture numbers.
Let me give you the data first, because it’s genuinely staggering. According to Carta, AI startups accounted for 41% of the $128 billion raised by VC-backed companies last year. Record high. In a single month earlier this year, OpenAI and Anthropic alone accounted for a significant chunk of $189 billion in global venture capital raised. OpenAI closed a $110 billion round. Anthropic closed a $30 billion Series G at a $380 billion valuation. xAI raised $20 billion in January.
To be clear: these are private companies. None of them have IPO’d yet.
And here’s the part that should sound familiar: 10% of startups captured 50% of the funding.
If you’ve spent any time in crypto — if you watched 2020 and 2021 turn into 2022 — you know this shape. You know what a market looks like when capital stops making bets and starts making monuments. When the numbers get so large, they stop corresponding to anything you can touch or verify. When everyone is very confident that, this time, the fundamentals are real and the valuations are justified, the technology is different.
It might be. It also might not be.
A brief history of concentration risk, for the newcomers
If you weren’t around for what the crypto community calls Operation Choke Point 2.0, here’s the short version: between roughly 2021 and 2024, crypto companies — founders, exchanges, infrastructure builders — started quietly losing their bank accounts—no formal charges. No explanation. Just a letter, or sometimes not even that, and then a closed account and a very unhelpful customer service line.
More than 30 crypto founders lost banking access. The mechanism wasn’t dramatic. It was bureaucratic. Regulatory pressure flowed through the banking system, which responded by simply dropping clients who were too complicated to defend. The message was clear even when the paperwork wasn’t: you are not welcome in this infrastructure.
The thing about infrastructure is that you don’t notice how dependent on it you are until it’s gone. Banking felt like a given until it wasn’t. You couldn’t rent an office, pay a contractor, or make payroll without a bank account. Getting cut off wasn’t just inconvenient. It was existential.
The crypto community’s response was predictable in retrospect: build alternatives. Decentralize. Remove the single points of failure that could be leaned on.
Now look at AI.
The new concentration
The venture market for AI is what analysts are calling “K-shaped.” At the top of the K: Anthropic, OpenAI, xAI — companies raising at valuations that would have seemed like science fiction five years ago, now apparently table stakes. They are, functionally, the infrastructure layer for an entire industry. The models everyone builds on. The APIs everyone calls.
At the bottom of the K: everyone else. Founders building on top of those models. Startups trying to differentiate at the application layer. Companies that can’t raise a $30 billion round but still need compute, still need access, still need the infrastructure to keep running.
This is the part where it gets interesting — and where the pattern starts to feel less like coincidence.
Earlier this month, Forbes ran a piece about Claude users getting their accounts suspended. A fintech CEO — someone who builds tools for the unbanked, which is its own irony — uploaded a travel itinerary to plan a business trip—used the Google Calendar integration. The tool worked great. The next morning, his account was disabled. Reason given: “suspicious signals.”
On Reddit, paying subscribers described similar experiences. Threads with titles like “Claude Max Subscription Silently Revoked” and “Account Permanently Banned, $300 Charged, No Explanations.” A developer who built one of the most popular third-party Claude tools had his personal account suspended the same way, reinstated only after the story went viral.
Anthropic’s position isn’t unreasonable on its face. Competitors were allegedly routing massive query volumes through the API to train their own models — effectively using Anthropic’s infrastructure against it. Compute costs at $200/month subscriptions were being used to run agent tasks worth multiples of that. The math didn’t work.
But here’s what the crypto community learned the hard way: the reasonableness of the underlying business logic doesn’t protect you when the enforcement mechanism is opaque, automated, and has no appeals process.
False positives are the cost of doing business at scale. Every bank, every payments platform, every fraud detection system produces them. The question isn’t whether they happen. The question is what happens after the flag goes up — and right now, for a lot of users, the answer is silence.
Someone on X coined a term for it: “DeCognizing.” The idea being that if AI access becomes essential infrastructure — and it’s rapidly heading there — then losing that access looks a lot like losing your bank account did. Not immediately catastrophic for everyone. But quietly, steadily, irreversibly isolating.
What the IRR doesn’t tell you
Back to the venture numbers. The early returns look good. Carta data shows that funds raised in 2023 and 2024 — post-ChatGPT launch — are posting higher internal rates of return than funds from the 2017-2020 vintage.
Here’s what that means in practice: a fund invested at seed, the company raised a Series A at a higher valuation, and on paper the fund made money. No exits. No liquidity. No actual cash returned to LPs. Just a higher number in a spreadsheet, waiting for an IPO that’s been teased but not scheduled.
Crypto had great paper returns, too, for a while.
I’m not predicting a crash. I genuinely don’t know what happens when OpenAI or Anthropic eventually tries to go public, or whether the valuations survive contact with public market scrutiny. Nobody does. The technology is real. The applications are real. The productivity gains are real. This is not the 2021 JPEG season.
But the structure — the concentration, the opacity, the infrastructure dependency, the paper returns, the we’ll-figure-out-exits-later energy — that structure is one we’ve seen before.
The Web3 community spent years building alternatives to centralized infrastructure because it learned what happens when you’re entirely dependent on systems you don’t control and can’t appeal.
The question worth asking now is whether AI needs to learn that lesson the same way.
Or whether we can just remember.
CryptoJazzHands is a field dispatch from the edges of the AI and Web3 overlap. If someone forwarded this to you and you’d like to receive it directly, you know what to do.