Substrate Lag

Key Takeaways

• Neither AI winter was caused by the ideas failing. Both were funding withdrawals that followed the substrate becoming uneconomic relative to what had been promised.
• Substrate lag, the gap between when an algorithm is published and when the electronics economy can afford to run it, was 54 years for the perceptron, 26 for backpropagation, and about 3 for the transformer.
• That collapse to near-zero was produced by one thing: the price of a FLOP falling for fifty straight years.
• The price of a FLOP largely stopped falling around 2017. The cheapest published single-precision GFLOPS fell 625× from 1984 to 1997 and 16,667× from 1997 to 2011, then about 2.2× from 2017 to 2025.
• Which means capability since 2017 has been bought with capital, not with cheaper substrate. That is what a $600B capex year actually is.
• If 2026 turns, it would be the first AI winter caused by the substrate being too abundant rather than too scarce, a failure mode the previous two teach us almost nothing about.

The standard history of artificial intelligence is a history of ideas: symbolic reasoning gives way to connectionism, connectionism to statistical learning, statistical learning to transformers. Winters arrive when the ideas disappoint, and springs arrive when someone has a better one.

I think this history is mostly wrong, or at least badly mis-weighted. Run the same seventy years against what a unit of computation cost, and a different and much duller story appears. The ideas were nearly always there early. What moved was the price of the substrate they had to run on.

This essay proposes one construct to make that claim measurable, plots it, and then argues that the thing which produced every AI spring since 1958 quietly stopped working around 2017, which is a more interesting way to read 2026 than the bubble argument currently on offer.

What Is Substrate Lag?

Substrate lag is the interval between the year an algorithm is published and the year the electronics economy can afford to run it at useful scale.

Formally:

SL(algorithm) = year_viable − year_published

Where year_viable is the year of the first deployment that demonstrated the algorithm working at a scale people cared about, and year_published is the year the method entered the literature in essentially its modern form.

The construct is worth having because it separates two questions that are constantly collapsed into one: was the idea any good, and could anyone afford to find out. The history of AI reads very differently depending on which one you think the winters were about.

Three measurements, using the least controversial dates available:

Rosenblatt’s perceptron and the backpropagation algorithm did not become correct in 2012. What happened in 2012 is that Krizhevsky, Sutskever and Hinton trained AlexNet on two consumer GTX 580 graphics cards over about six days. The decisive input was not an insight. It was a pair of gaming GPUs costing a few hundred dollars each, doing work that would have required a national laboratory twenty years earlier.

Why Did the Price of Computation Determine All of This?

Because substrate lag is not an independent variable. It is a shadow cast by one curve.

Here is that curve, against the events usually credited with causing AI’s booms and busts.

Two features of that curve do all the work.

The first is its sheer magnitude. A GFLOPS cost roughly $1.27 trillion on ENIAC in 1945 and about 1.2 cents on a consumer graphics card in March 2025, fourteen orders of magnitude in eighty years. No algorithmic advance in the field’s history is remotely that large. When a single input to a system falls by a factor of a hundred trillion, most of what happens downstream is that input.

The second feature is the one nobody talks about, and it is at the right-hand end. I’ll come back to it.

What Actually Caused the AI Winters?

In both cases the money left after the substrate proved uneconomic. The ideas were still standing when the funding stopped.

The first winter is usually pinned on the 1973 Lighthill report, and the report is usually described as a verdict on AI’s intellectual promise. Read what it actually attacked. Lighthill’s central charge was combinatorial explosion and intractability: the claim that AI algorithms which worked on toy problems would fail on real ones as the search space grew. That is not an argument that the algorithms were wrong. It is an argument that running them on anything that mattered would cost more than anyone could pay. It was a compute argument wearing the clothes of an intellectual one. In 1974, the year DARPA money became hard to find, a GFLOPS still cost somewhere in the region of tens of millions of dollars. Lighthill was, on the economics available to him, right.

The second winter is cleaner still, and it is the case I’d point to if I had to defend this whole thesis with one example.

By 1985, corporations were spending over a billion dollars a year on AI. Expert systems worked. XCON is estimated to have saved Digital Equipment Corporation some $40 million over six years. The technology was delivering. And then, per Wikipedia’s account, “an entire industry worth half a billion dollars was replaced in a single year.” Symbolics, Lisp Machines Inc., Lucid, Texas Instruments and Xerox all left the field.

Replaced by what? Not by a better idea about intelligence. By Sun, Apple and IBM workstations that ran the same software on general-purpose silicon for less money, and by portable Lisp environments that removed the last reason to buy a dedicated machine.

The second AI winter was a hardware business model losing to a cheaper hardware business model. The intelligence had nothing to do with it. That is as literal a hardware recession as the field has ever had, and we have spent forty years filing it under “AI disappointed.”

The Inversion: When the Substrate Ran Ahead

Between roughly 2012 and 2022, for the first time, the hardware arrived before the algorithms needed it.

Every era before 2012 had ideas queued up waiting for affordable silicon. Then the queue emptied. Substrate lag went from 54 years to 26 to 17 to about 3, and at three years it is barely a lag at all. GPT-3 followed the transformer paper about as fast as a large engineering effort can physically move.

This is what the compute overhang argument describes from the other direction: not scarcity of compute relative to ideas, but a surplus of compute relative to ideas about what to do with it. The scaling era is exactly what you would predict from a surplus. When the substrate is abundant and cheap, “use much more of it” is the highest-return research direction available, and for a decade it was.

But notice the dependency. The collapse of substrate lag was not a fact about AI research getting faster. It was a fact about hardware getting cheaper faster than ideas could consume it. Substrate lag went to zero because the denominator went to zero.

Has the Price of a FLOP Stopped Falling?

On the cheapest-commodity-hardware measure, very nearly, and that is the most consequential fact on the chart.

Take the same source series and read it as rates rather than as a curve:

Between 1997 and 2011 the price of computation fell by a factor of sixteen thousand. Over a comparable stretch ending in 2025, it fell by rather less than a factor of three. On this measure the engine that drove seventy years of AI history has, in the last decade, close to stalled.

I want to be careful about what that series does and does not show, because the caveat is real and it cuts both ways. These are cheapest-published-configuration datapoints for commodity consumer hardware at single precision, in nominal dollars. They are not a controlled index, the later entries are not perfectly like-for-like, and the 2020 RTX 3080 row is actually worse than the 2017 entry it follows. Most importantly, an enormous amount of the last decade’s genuine efficiency gain came from places this series cannot see: specialised accelerators, and above all the move down through FP16 to FP8 and below. The industry did keep getting more arithmetic per dollar. It substantially did so by redefining the arithmetic.

So the honest claim is narrower than “compute stopped improving,” and more interesting. What stopped is the thing that made every previous AI spring possible: general-purpose computation getting radically cheaper for everybody, without anyone having to do anything. Fifty years of free substrate deflation lifted every researcher with a workstation. The gains since 2017 have gone disproportionately to organisations that can co-design silicon, retrain in exotic numeric formats, and buy at hyperscale. That is a different economy, with different entry conditions.

What Does This Say About 2026?

That the current capex numbers are not evidence of irrational exuberance so much as the arithmetic consequence of the price curve flattening.

If a FLOP is no longer getting meaningfully cheaper, and capability still scales with total compute applied, there is exactly one remaining lever: buy more of it. So the industry does. US big tech capital expenditure is projected to climb another 50% past $600 billion, roughly 23% of revenue, more than double pre-ChatGPT capex intensity.

That is what a $600 billion capex year is. It is the bill that arrives when capability has to be purchased rather than inherited from the substrate. From 1958 to 2017, progress was substantially subsidised by hardware getting cheaper on its own. That subsidy has largely lapsed, and someone has to pay the difference in cash.

Which reframes the bubble question. The usual argument is that spending has run ahead of returns, and the warning signs are real enough: two-thirds of hyperscaler spend on assets depreciated over five to six years that critics argue have an economic life of two to three, and a credit window where investor orders per dollar offered fell from about 5× oversubscribed in February 2026 to about 1.6× in July.

But hold 2026 against the shape of the two winters that actually happened. Both began the same way: capability plateaued, and then the money left. Lighthill described a wall before DARPA cut the contracts. Expert systems had visibly stopped scaling before the Lisp market died.

2026 has the opposite shape. Capability is still climbing. Capex is at a record. The stress is in the financing, not in the results.

So if this does turn, it will not be a repeat of anything. It would be the first AI winter caused by the substrate being too abundant rather than too scarce, a capital-structure failure, where the compute exists and works and simply cost more than the revenue arriving to service the debt behind it. The two previous winters teach us almost nothing about that, which is precisely why they keep getting cited badly.

What Does Substrate Lag Predict for World Models?

Genuinely: it predicts the construct’s own trend may be about to reverse, which would be the first reversal since 1958.

World models are the loudest candidate for the next paradigm: systems that learn to predict how an environment evolves rather than which token comes next. The name and the modern formulation date to Ha and Schmidhuber’s 2018 paper, which is the publication year I’ve used on the chart. The money now agrees: Yann LeCun left Meta after twelve years and founded AMI Labs, which raised $1.03 billion in March 2026 at a $3.5 billion valuation, and DeepMind shipped Genie 3.

The naive extrapolation is easy. Substrate lag went 54 → 26 → 17 → 3, the trend points at zero, so world models should be viable more or less on arrival.

I don’t believe that, and the reason is the whole argument of this essay. That sequence was not a law of AI research. It was a shadow of the price curve. Remove the falling price and you remove the mechanism that produced the sequence. Extrapolating substrate lag while ignoring what generated it is exactly the error that made “AI is dead” the consensus in 1988 and “scaling solves everything” the consensus in 2022.

The specific worry for world models is that they are compute-hungry in a way that lands squarely on the flattened part of the curve. Video and simulation-based training consumes far more computation per unit of learned structure than text does, and it cannot be rescued by the same trick that rescued the last decade, since much of the precision headroom has already been spent.

So the honest projection, and I want to label it as a projection, since a four-point trend extrapolated to 2030 is a hypothesis rather than a finding, is this. World models may be the first major AI paradigm in fifteen years to have to wait. Not for a research breakthrough. For someone to make the substrate cheap again.

If that is right, the interesting question for the rest of this decade is not which architecture wins. It is whether anything replaces Dennard scaling and Moore’s law as a source of free deflation: advanced packaging, optical interconnect, analogue or in-memory compute, something not yet on the roadmap. Every AI spring in history was downstream of that question being answered yes.

For most of seventy years, we mistook the substrate for a constant and argued about the ideas. It may be time to invert the habit.

Further Reading

• FLOPS: cost of computing, Wikipedia. The historical price series used for the chart above.
• AI winter, Wikipedia. Source for the Lighthill report, the DARPA cuts, and the 1987 Lisp machine collapse.
• ImageNet Classification with Deep Convolutional Neural Networks, Krizhevsky, Sutskever & Hinton, 2012. The two-GPU training run that ended a 54-year substrate lag.
• The AI capex cycle, Allianz Research, March 2026. Capex projections and intensity figures.
• Why world models are AI’s next frontier, InfoWorld. Background on the paradigm this essay declines to date.