Vault
research

If the Bubble Collapses: Which Way Do Token Prices Break, and Do the Labs Survive?

Created

If the Bubble Collapses: Which Way Do Token Prices Break, and Do the Labs Survive?

Builds-on: token-cost-velocity-2023-2026, anthropic-subsidy-stress-test, ai-circular-financing-and-banking-exposure-audit Related: the-efficiency-counterthesis, ai-infrastructure-endgame-indicators, ai-survival-theater-and-the-bubble, anthropic-unit-economics-and-the-power-user-loss, the-data-center-convergence, why-the-market-refuses-to-crash, the-shadow-bank-household-channel

The Question, and Why It's Actually Two Questions

"If the AI bubble collapses, do token costs go up or down?" sounds like a single question with a single answer. It isn't. It bundles two prices that move in opposite directions during a collapse, and most of the confusion in the discourse comes from not separating them:

  1. The price of compute — a GPU-hour, a rack, a data center's worth of capacity. This is the input cost.
  2. The price of a token — what OpenAI, Anthropic, or Google charges you per million tokens for a specific model. This is the retail cost.

In a real collapse, compute craters and frontier tokens spike — at the same time, for the same reason. The reason both happen together is the resolution to the whole puzzle.

This doc is the discontinuous companion to two existing vault docs. token-cost-velocity-2023-2026 mapped the smooth deflation curve and found commodity and frontier tiers already decoupling. anthropic-subsidy-stress-test modeled a gradual subsidy normalization (its base case: 500–1,000bps GM compression, no IPO break). This doc asks the harder version: what if it isn't gradual — what if the equity/credit event in ai-circular-financing-and-banking-exposure-audit actually fires?

Part 1 — The Compute Side Craters: The Dark-Fiber Replay

The cleanest historical analog is the telecom crash, not the dot-com equity crash. In the five years after the 1996 Telecommunications Act, carriers poured $500B+ (mostly debt-financed) into fiber on the belief that internet traffic would triple every few months. When demand grew merely fast instead of impossibly fast, the overbuild was exposed: by 2001, an estimated 95% of laid fiber was "dark" — lit by nothing, earning nothing. Bandwidth prices collapsed up to 90%. Corning, the largest fiber maker, went from ~$100 to ~$1. WorldCom, Global Crossing, 360networks all went bankrupt.

The AI version has a name already: "dark GPUs." The mechanism is structurally identical and the early tells are live in 2026:

So the input price of intelligence — raw FLOPs — falls hard in a collapse. If you own a GPU cluster and can pay the power bill, your marginal cost of serving tokens goes down, not up. Hold that thought; it's half the paradox.

Part 2 — The Frontier Token Side Spikes (Or Just Disappears)

Now the retail side. Today's frontier token prices embed two stacked subsidies (token-cost-velocity-2023-2026 decomposed roughly half the headline deflation as subsidy):

  1. Below-cost retail pricing. OpenAI is on a path to negative cash flow through 2030; Anthropic doesn't expect even a small profit before 2029. One frontier provider posted a –94% gross margin in 2024. Prices sit below the cost to produce, funded by VC and hyperscaler capital.
  2. Below-market compute. anthropic-subsidy-stress-test quantified the implicit Trainium/TPU discount at $2–4B/yr — Anthropic pays ~$0.50/chip-hour against a $2–5 reserved-H100 counterfactual.

A collapse is precisely the event that removes the capital funding both subsidies. VC risk appetite evaporates; hyperscalers under equity pressure stop selling compute below cost. What's left is survival economics. The independent estimates already converge on the gradual version of this: 30–50% API price increases over the next ~18 months as vendors move to sustainable unit economics (Arize, AnalyticsWeek). A discontinuous collapse compresses that into a step-change and pushes the top tier further.

Two forces make the frontier tier spike more than the mid-tier:

Part 3 — The Paradox Resolved: A Scissors, Not a Level

Cheaper compute and pricier frontier tokens, simultaneously. How? Because the GPU crater doesn't pass through to the token buyer in a collapse. Three reasons:

  1. The fire-sale GPUs get bought by survivors who no longer need to subsidize — they capture the cost saving as margin, not as lower prices.
  2. The competitive discipline that forced below-cost pricing is gone.
  3. The number of frontier sellers shrinks toward a pricing-power oligopoly.

But there's a powerful floor pulling the other way, and it's the same floor the telecom analogy actually predicts. Dark fiber didn't stay dark — survivors bought it for pennies, and that cheap infrastructure became the substrate for broadband and then the cloud. The AI equivalent of "fiber that got bought for pennies" is open-weight models running on fire-sale GPUs.

Open weights are the crucial asymmetry the fiber crash didn't have: the capability can't be un-released. DeepSeek V4 Flash is at $0.14/$0.28 per MTok; Kimi K3, GLM-5.2, Qwen are all downloadable and self-hostable, breaking even against API pricing at ~1–5M tokens/day. Fable 5 is reportedly >50x more expensive per token than DeepSeek V4 Pro. Even if every US frontier lab vanished tomorrow, GPT-4-to-GPT-5-class capability is permanently loose in the world, self-hostable by anyone with a cluster — and in a collapse, clusters are cheap.

So the collapse doesn't produce a single price move. It produces a scissors:

flowchart TB
    C[AI bubble collapse:<br/>capital + competition evaporate]
    C --> G[Compute prices CRATER<br/>dark-GPU fire sale 30-50c]
    C --> S[Subsidies removed<br/>price war ends]

    G --> OW[Open weights + cheap GPUs<br/>set a hard low floor]
    G --> SURV[Survivors capture<br/>cheap compute as margin]

    OW --> LOW[COMMODITY/MID TIER<br/>stays cheap or gets cheaper<br/>fiber-becomes-broadband]
    S --> SURV
    SURV --> HIGH[FRONTIER TIER<br/>spikes / rations / freezes<br/>oligopoly pricing]

    LOW --> SCISSOR[The spread WIDENS:<br/>good-enough AI abundant,<br/>frontier AI a luxury good]
    HIGH --> SCISSOR

The blade that opens: the commodity/mid tier stays cheap or gets cheaper (floored by open weights on cheap compute), while the frontier tier spikes, rations, or freezes. This is the token-cost-velocity-2023-2026 commodity/frontier decoupling, but a collapse doesn't close the gap — it rips it wide open.

The one-line answer to "up or down": down for good-enough intelligence, up (or gone) for the best intelligence. A crash makes cheap AI cheaper and abundant, and makes the march to the next capability level expensive or impossible. What a collapse kills is not access to useful AI — it's the frontier's forward motion.

Part 4 — Do Anthropic and OpenAI Survive, and How?

Separate corporate survival from capability survival — they're not the same thing, and conflating them is the second big error in the discourse.

Capability always survives. Weights, talent, and brand are durable assets. Even in the worst corporate outcome, the models keep running — the way airplanes keep flying through an airline's Chapter 11. The take-or-pay compute obligations are what a restructuring sheds; the IP is what a buyer keeps. So "does the technology survive?" is almost always yes. The real question is which corporate entity owns it afterward, and at what price to you.

Anthropic — survives smaller, pricier, enterprise-first

OpenAI — more fragile on cash, more likely to be rescued

The dark-horse winner: Google

The telecom-crash survivors weren't the pure-plays — they were the balance sheets that bought distressed assets for pennies (the "asset-light wins" pattern). Google is the AWS-of-this-cycle: it owns its TPUs (no Nvidia margin), owns the cloud, owns distribution (Search, Android, Workspace), and funds AI from a river of unrelated cash flow. In a real collapse, Google is best-positioned to buy the wreckage and set the post-crash price. anthropic-subsidy-stress-test already noted Google escalating its Anthropic commitment ($40B) precisely when the naive read said pull back — that's a company positioning to own the substrate.

Part 5 — Mapping to the Endgame Archetypes

ai-infrastructure-endgame-indicators's four archetypes tell you which kind of collapse you're in:

Which archetype dominates decides the token outcome. Sovereign absorption → frontier stays available but state-shaped and pricey. Efficiency cliff → commodity abundance, frozen frontier. Slow deflation → the smooth curve continues and this whole doc stays hypothetical.

Part 6 — What This Can't Resolve

Personal Note — The Bet Underneath the Workflow

This one has a real referential hook, so it's worth one honest paragraph rather than a reflexive tie-back. The async, Opus-by-default, high-throughput working style — running Claude Code AFK, ~$1.5–2k/mo in work API — is implicitly long the frontier tier staying cheap-ish. The scissors says that's the exact bet a collapse breaks: commodity tokens get cheaper (fine for most work), but the frontier tier — the Opus-class reasoning the workflow leans on — is the blade that spikes or freezes. The hedge is already the right one and it's not financial: it's the the-orchestrator-premium / the-2030-landing-posture thesis that the durable edge is shipped systems + eval harnesses, not raw token spend. If frontier tokens ration, the person who has internalized how to get frontier-quality output from mid-tier models (good scaffolding, tight evals, open-weight fallbacks) is advantaged by the scissors, not hurt by it. Worth actually building the open-weight fallback muscle now — DeepSeek/Kimi/GLM in the loop for the tasks that don't need Opus — as a live rehearsal, not a someday plan. That's the one concrete action this analysis argues for.

The One-Liner

A bubble collapse doesn't move token prices in one direction — it's a scissors: compute craters (dark-GPU fire sale), which floors good-enough intelligence cheap and abundant via open weights, while the best intelligence spikes, rations, or freezes as subsidies and competition die and survivors consolidate pricing power. The technology always survives (weights and talent are durable, like planes through Chapter 11); the question is which entity owns the frontier afterward — likely Anthropic smaller and pricier, OpenAI absorbed or backstopped, and Google buying the wreckage to set the post-crash price.

Sources

Collapse scenarios and lab economics

The dark-fiber / telecom analogy

Stranded GPUs / neocloud distress

Subsidy unwind and token repricing

Open-weight floor

Bailout / backstop / sovereign absorption

Backlinks