Neuvottelija.com

2026-07-15 · AI Lab · Insight

Is AI a Bubble? Inference Pricing, Temu-Intelligence, and What Would Actually Have to Break

One question, three arguments

People ask me “is AI a bubble” as if it were a single question with a single answer. After recording three separate conversations on exactly this, I don’t think it is. Talking with Petri Kajander, with Ville Tolvanen and Timo Savolainen, and with Mikko Alasaarela, I kept noticing that “bubble or not” was really three separate questions stacked on top of each other: is inference priced below what it will eventually cost, does the underlying capability mean anything without genuine understanding, and does the rest of the economy actually adopt the technology fast enough for any of this to matter. My guests don’t even agree on which of the three is the real question, let alone the answer.

The wage bill argument

Kajander’s episode opens as a running joke about his own hardware habit — a Mac Studio running local models, a Mac mini humming through Gemma and Qwen while he waited for Peter Steinberger’s OpenClaw to ship — but it lands on a genuinely load-bearing economic argument. His claim is that the “bubble” framing uses the wrong denominator. The comparison that matters isn’t what investors have paid for compute; it’s what the same capability would cost as human labor. His back-of-envelope math: the global wage bill runs to roughly €54,000 billion a year, and Finland’s alone is about €135 billion. Measured against that, a subsidized AI subscription costing tens or a few hundred euros a month isn’t dot-com-style overvaluation — it’s underpricing.

He’s clear-eyed, though, that today’s token prices won’t last. “I’m currently in the position of thinking these tokens will probably get more expensive, since they’re being subsidized so heavily right now. It could be five times as expensive, could be 500 times, could be a thousand times” (41:07). He ties that directly to his own experience — burning through $50, then another $50, in a single morning once Anthropic moved off flat-rate pricing, and watching GitHub Copilot and Google’s Antigravity do the same repricing dance, sometimes by banning heavy users outright rather than charging them more. His conclusion isn’t that the technology is a bubble; it’s that mispriced access is a bubble, and it will get repriced. He does not think the data centers will simply sit idle — his argument is that it’s the untalented, who don’t know how to direct the intelligence properly, who end up with Temu-AI instead of real value.

Temu-intelligence: capability without understanding

Kajander’s other idea is a diagnosis, not a forecast, and it’s the one he’s most attached to — he literally calls AI “Temu-AI,” after the bargain e-commerce site. His argument is that these models guess extremely well on statistical grounds, but there’s no comprehension behind the guess, and the gap shows up wherever the guardrails are thin. His example is a wave of hijackings of well-known Instagram accounts, after Meta reportedly cut its AI trust-and-safety staff: attackers logged into AI-run customer support using a VPN that merely matched the victim’s rough location, and the AI reset two-factor-protected accounts on that basis alone. “These AI machines have no notion whatsoever of what they’re doing. It’s statistical guessing, completely random” (16:17).

What makes this more than a security anecdote is that Kajander doesn’t think more scale closes the gap. Stacking more data or bigger models onto the same predictive architecture makes the guesses better-calibrated, in his view, not more genuinely understood — cram in more of the same, he argues, and you get better predictions of what already exists, not comprehension of anything new.

Alasaarela, recorded five months earlier in a very different register, is arguing something close to the opposite trajectory — that the reasoning layer is closing exactly the gap Kajander says can’t be closed. His evidence is Humanity’s Last Exam, a benchmark researchers built specifically because they expected it to hold as a ceiling. His claim on air: OpenAI’s o3 model scored 13% publicly, but its internal testing had already reached 35% within roughly a month of the test’s release. He puts real weight behind the public statements from Anthropic’s and OpenAI’s founders that “according to the current mathematical modeling the AGI moment is in 2026 which is next year” (01:20) — a date that, as I publish this, has already arrived.

I don’t think the two men are strictly contradicting each other. Kajander is talking about brittle, deployed products — customer-service bots, hallucinated comparison pages. Alasaarela is talking about frontier reasoning benchmarks. But they can’t both be describing the same trend line. Either the reasoning layer is manufacturing something that functions like understanding, or it’s manufacturing better-calibrated guesses that still fail exactly where Kajander says they fail — quietly, at the edges, in whatever wasn’t in the benchmark.

Two investors, one stage, opposite conclusions

The Tolvanen and Savolainen episode is the most explicit “is this 1999” conversation of the three, and it’s a real disagreement, not two people nodding along. Tolvanen, who ran Radiolinja’s WAP services and later built the “Digitalist” movement through the dot-com years, hears late 2025 as an echo of 1999: massive valuations without matching revenue, and strange cross-investments — Nvidia putting a billion dollars into Nokia, Nokia investing four billion back into Nvidia. He’s specifically skeptical of how hyperscalers account for their chip investments, arguing the hardware is being depreciated over five years even though, in his view, “these circuits are relevant for about two years,” which he says makes total depreciation for some of these companies larger than their annual profit. His verdict, stated plainly: “In all my skepticism, I hope I’m wrong, but I’m afraid this is like November ‘99” (28:31).

Savolainen pushes back directly, calling Tolvanen out by name mid-argument. His position is that whether the market is overvalued and whether the technology matters are two separate questions, and only one of them is interesting: “the real question is what to do with this technology. Even if all companies went under just like that… we still get those services from across the big pond here to us” (29:45). His own complaint isn’t about valuation at all — it’s that institutions are barely using any of this. Doctors won’t let an AI sit in on a consultation and change a treatment plan. Schools default to phone bans rather than AI literacy. He names the pattern directly: technology monetization, he says, has been one of the slowest parts of the last twenty years, across every previous wave, not just this one.

That’s the disagreement in miniature. Tolvanen is worried the financial layer is overbuilt relative to revenue. Savolainen is worried the adoption layer is underbuilt relative to the technology’s actual capability. Both can be true simultaneously, and neither one rebuts the other.

The internet is already eaten

The third thread, and the one I find hardest to shake, is about the raw material running out rather than the money. Alasaarela cites an estimate that roughly two-thirds of internet traffic was already bot-generated at the time of recording, heading toward 90% by year’s end, and flags the resulting quality collapse directly: newer models trained on a web already contaminated by older models’ hallucinations come out worse than the ones they replace, in his account, not better. Kajander makes the identical point from the supply side, quoting Oracle’s Larry Ellison: the public internet has already been picked clean, down to piracy archives like Anna’s Archive, and what’s left to train on is private data nobody has handed over yet.

Neither of them treats this as evidence that the bubble is about to pop. It’s a different kind of constraint from the financial one Tolvanen worries about — not “is this overpriced” but “does the input run out regardless of price.”

Where I land

I don’t think “is AI a bubble” survives as one question, and I said as much on air. When Kajander predicted that half of the announced data-center projects won’t get built and that the AI bubble is probably going to start rotting, I told him directly, on the record, that I didn’t really believe that. The valuation layer Tolvanen worries about can absolutely correct — some announced data centers won’t get built, some of the trillion-euro chatter is exactly the kind of number Buffett would tell you to distrust — without the underlying capability being worthless, because Kajander’s own wage-bill math doesn’t depend on any single company’s share price holding up.

I’m also not fully with Kajander’s flatter “always Temu-AI” framing, even though I take his security examples seriously. I told him as much in the moment — that I sensed a bit of Luddism and AI-nihilism in his position, and that my own instinct is to let all the flowers bloom and let those who know how to direct the intelligence do it. Savolainen’s complaint that hospitals, schools and pension funds barely touch any of this is, if anything, an argument against a bubble having already formed: you can’t pop a bubble in something most institutions haven’t adopted yet.

What would actually change my mind is the one thing none of the three conversations resolved: whether reasoning models are closing the understanding gap Kajander describes, or just producing better-calibrated Temu-AI. Alasaarela’s Humanity’s Last Exam numbers are the strongest evidence I’ve heard that it’s closing. Kajander’s account-hijacking stories are the strongest evidence I’ve heard that it isn’t. Until one of those two trend lines clearly wins, I’d bet on repricing over collapse — tokens get more expensive, some vendors don’t survive the repricing, and the capability underneath keeps compounding regardless of what happens to any single valuation.

Source episodes

Every claim in this essay is grounded in the following episodes; quotes carry timestamps linking to the original video. English subtitles and full transcripts are on each episode page.

← AI Lab · Markdown mirror: index.md