The Bubble Question, Taken Seriously

As usually asked, the question has no answer, because “bubble” is doing no work: it means prices are high, and high compared to what is exactly the thing in dispute. Made precise, it becomes a calculation with inputs you can go and find.

Making the question answerable

A useful definition: current prices imply expectations about future cash flows that no plausible path of fundamentals can deliver. That version is testable in principle, because it asks you to state the implied expectation and then to test it, and it separates two things that usually get merged: whether the technology is transformative and whether the assets are correctly priced. Those are different questions with different answers.

It also rules out the two lazy positions. “Prices have risen a lot” is not evidence — prices rise when expectations improve. “The technology is real” is not evidence either; a real technology can be accompanied by capital deployment far in excess of what it will earn, and historically often has been.

The implied-expectation calculation

The method is standard and the arithmetic is simple. Capital deployed must eventually be recovered with a return; you can back out what revenue that requires and compare it to a demand path built from the bottom up.

Required annual revenue from an investment I:

  R_req = I * (1/L + k) / m

  L   useful life of the asset in years
  k   required return on the capital
  m   gross margin on the revenue it supports

Worked with I = $1 (so the answer is a ratio):

  L = 5, k = 0.10, m = 0.50
    R_req = 1 * (0.20 + 0.10) / 0.50 = $0.60 per year
    per dollar invested

  L = 3, k = 0.10, m = 0.50
    R_req = 1 * (0.333 + 0.10) / 0.50 = $0.87

  L = 5, k = 0.10, m = 0.30
    R_req = 1 * 0.30 / 0.30 = $1.00

Then compare against a bottom-up path:

  achievable revenue = tasks per year
                       * price per task
                       * share you capture

Both sides must come from sources you can cite, with
dates. This page supplies neither.

The value of writing it this way is that it localises the disagreement. Two people who disagree about whether current investment is proportionate are, almost always, disagreeing about L, about m, or about the number of tasks. Once that is visible the argument becomes tractable, because each of those is a separate question with its own evidence — and L in particular is a disclosed accounting assumption, discussed in the depreciation page.

What the optimistic case rests on

Stated at its strongest, the optimistic case is a chain, and it is only as strong as its weakest link:

• Capability keeps improving at a rate that justifies continued frontier spending — the condition under which the frontier curve in the four cost curves keeps rising rather than stalling.
• Demand is elastic, so falling prices raise total spending rather than lowering it. This is the Jevons condition and it is an empirical claim about a specific elasticity, derived in the elasticity page.
• Adoption follows capability without a long institutional delay — the assumption most exposed by the adoption term in an enterprise business case.
• Assets last long enough for the depreciation assumption to hold, and capacity constraints resolve.

What the pessimistic case rests on

• Depreciation is optimistic. If economic life is shorter than assumed, required revenue per dollar invested rises immediately and mechanically. This is the highest-leverage bear argument because it does not require capability to disappoint.
• Value accrues to buyers, not producers. Competition can compress margins so that the technology creates enormous value while almost none is captured by those who financed it. This is the subtlest and most historically common outcome, and it is entirely consistent with the technology succeeding — the mechanism is in where value gets captured.
• The bottleneck is not capability. If integration, trust, workflow change and verification are what limit deployment, more capable models do not unlock proportionate revenue, and spending aimed at capability is aimed at the wrong constraint.
• Financing is circular or reflexive. Where suppliers finance their own customers, or where revenue at one layer is capital spending at another, revenue growth can be partly an accounting circuit rather than end demand. The test is whether the money originates outside the industry.

Why both can be right at once

The historical pattern for large infrastructure build-outs is not “transformative” or “bubble”. It is both, in sequence: capacity is built faster than demand arrives, the assets are repriced downward, the capacity is then used for decades by companies that bought it at a fraction of its construction cost, and the technology’s eventual importance is not in question. Investors and users experience the same episode as, respectively, a disaster and a gift.

The mechanism behind that pattern is worth stating because it is not irrationality. Capacity has long lead times, so it is ordered against expected demand rather than observed demand; every participant orders against the same expectation; and the orders all arrive at once, some years later, into whatever demand actually materialised. Nobody has to be foolish for the result to be an overshoot. That is also why the correction, when it comes, is usually a repricing of assets rather than an abandonment of the technology — the capacity exists, it works, and somebody will operate it at the new price.

So “is it a bubble” and “does it matter” are orthogonal, and a debate that treats them as one question cannot resolve. The precise version is: is the capital deployed at these prices likely to earn its required return — which is a question about L, m, timing and competition, and not about whether the technology works.

The observables

No values are given here, because any would be wrong within a quarter and inventing them would be exactly the failure this page is arguing against. These are the quantities that move the answer, each with a place to find it:

Collect those with dates and sources and you will be able to answer the question for yourself at whatever moment you are asking it, which is the only form of the answer with any shelf life. Anyone offering a verdict without stating their L and their m is offering a mood.