Published on

The AI Cost Reckoning, Revisited: Capex Doubled Anyway

Authors

In June I wrote that the industry's founding assumption - that the AI cost curve would bend down and the capex would justify itself - was being tested on four fronts at once. I said it was not a post about the bubble bursting, just one noting that the cost curve was doing something different from what the slides predicted, and that people building real products were starting to plan accordingly.

Five months on there is enough data to mark that post. Parts of it hold up. One part was straightforwardly wrong, and the way it was wrong is the interesting bit.

Table of contents


What I got wrong

The June post's implicit assumption was that capex discipline would arrive roughly when the economics stopped justifying it. The argument ran: frontier training runs are getting expensive enough that the marginal model has to clear a higher commercial bar, open-source has closed the gap on non-frontier tasks, local inference is viable, and buyers are starting to price senior human engineers against AI for critical work. Put those together and you would expect spending to moderate.

It did the opposite. Spending roughly doubled.

That is worth being direct about, because it means the June post correctly identified the pressures and incorrectly predicted the response to them. The pressures were real and are still real. They simply did not translate into restraint, and the reason is not that the people spending the money disagree with the analysis. It is that the incentive structure of a land grab does not reward being the first to stop.

The numbers, as of November

The 2026 capital expenditure guidance from the four large hyperscalers:

Company2026 capex guidance
Amazon~$220B
Google~$200B
Microsoft~$170B
Meta$135-145B (raised 21 October)

That is north of $720 billion from four companies in a single year. Morgan Stanley has aggregate global hyperscaler capex up 99% in 2026 to $917 billion, reaching $1.467 trillion in 2027 before growth slows sharply to 12% in 2028. IDC measured $121.4 billion of AI infrastructure spend in Q2 2026 alone, up 53% year over year, and raised its full-year forecast to $521 billion.

Hyperscaler AI capex trajectory - 2026 and beyond are projections, earlier years approximate

Two caveats on that chart. The 2023-2025 bars are approximate reconstructions, not reported figures, and the forward years are a projection from one bank. The shape is the point, not the precision: a near-doubling into 2026, another large step in 2027, and then a forecast cliff in growth in 2028. A forecast that has spending decelerate from 99% growth to 12% growth in two years is not really a forecast of a smooth landing. It is a forecast that something changes.

The gap that is actually widening

The June post argued that application revenue was concentrating in a few categories rather than spreading across every business. That part held, and the consequence is now being stated plainly by people whose job is to be boring about it.

Gartner has flagged a widening gap between data centre capital spending and realised enterprise and consumer AI revenue, calling the current pace of spending unsustainable and putting more than $3.5 trillion of potential impairment charges on the table by 2030 if the buildout continues at this rate.

Impairment is the word to notice. It is not a prediction that AI is useless. It is an accounting claim: that a material share of these assets will not generate the returns their carrying value assumes, and the difference will eventually be written down. You can believe AI is transformative and still believe a trillion dollars of it was built in the wrong place, at the wrong density, on the wrong depreciation schedule.

Physical constraints showed up on schedule

The June post said land, power, water, transformers, network fabric and chips were binding constraints rather than line items. That has become visible in delivery rather than in planning documents.

IREN told investors on 5 November that some infrastructure deliveries for Microsoft had run late. A single vendor's slipped schedule is not a trend, but it is the texture of what a supply-constrained buildout looks like from the inside: not a dramatic halt, just everything arriving later than the plan assumed, at every link, simultaneously. The phrase now being used for the buildout is war of attrition, which is a notable shift in register from the language of eighteen months ago.

Last week the market started voting

On 6 November, AI stocks sold off hard, with the midterm result cited as part of the trigger alongside growing scepticism about the buildout economics.

The politics connection deserves one paragraph, because it is new. Data centres have become locally unpopular - they consume power and water, they raise residential electricity prices in the regions that host them, and they employ comparatively few people once built. That was a manageable local story while federal and state policy was uniformly accommodating. The midterms changed the composition of who writes that policy, and the market appears to have marked AI infrastructure down partly on the expectation that permitting, power allocation and tax treatment get harder from here.

That is a genuinely different risk from the one the June post was tracking. I was watching unit economics. The thing that moved the price was the political economy of where the hardware physically sits.

Three ways this resolves

I do not know which of these happens, and I would be suspicious of anyone who claims to.

The revenue arrives late but arrives. Enterprise adoption is slower than the capex assumed but the direction holds, and by 2028-29 the utilisation catches up with the footprint. In this version the $3.5 trillion impairment risk partially materialises as a few painful quarters rather than a structural break, and the people who built early own the capacity.

The write-down happens and is survivable. Utilisation disappoints, a meaningful share of the buildout is impaired, and the losses concentrate in the most leveraged participants - the neoclouds and the debt-funded operators rather than the companies funding capex from cash flow. The hyperscalers absorb it; the capital structure around them does not.

Capex discipline arrives abruptly. The 2028 deceleration in that forecast happens in 2027 instead, triggered by a single bad earnings cycle. This is the version that looks most like a traditional capital cycle, and the one the June post implicitly expected to begin already.

The thing all three share: the correction, if it comes, lands on the financing layer before it lands on the technology. The models keep working either way.

What I would plan around

For anyone building on top of this rather than funding it, three things seem robust across all three scenarios.

Inference pricing is not a stable input. It has been subsidised in both directions - down by competition for market share, up by the need to recover training costs - and neither force is settled. Architect so that a 3x move in either direction is a configuration change rather than a redesign.

Keep a credible non-frontier path. The open-source and local-inference arguments from June got stronger, not weaker, and the strategic value of being able to run acceptable-quality work without a frontier API call rises in exactly the scenario where frontier pricing has to start recovering real costs.

And treat capacity availability as a planning variable, not a given. The delivery slippage above is the leading edge of that. If your 2027 plan assumes you can buy as much inference as you want at the price you want on the date you want, it has three assumptions in it, and the buildout is currently struggling with all three.

The June post said the cost curve was doing something different from what the slides predicted. It still is. I just had the direction wrong: the slides were not too optimistic about spending, they were too conservative, and the reckoning got deferred by roughly a trillion dollars rather than resolved.


Sources and further reading: