Cognition got dramatically cheaper this year. The physical layer holding it up got sharply more expensive. Twenty years of operating plans encode the first assumption implicitly, and almost none encode the second.
We argued in an earlier piece that the collapsing cost of cognition reshapes every business model it touches. That remains true, and the numbers have only gotten more emphatic since — inference pricing fell roughly 43 percent over ten weeks this summer.
This piece is the other half, and the two do not contradict each other. They describe different layers of the same stack moving in opposite directions at the same time. One of those layers has been getting cheaper for a generation and is still doing so. The other quietly reversed, and it is the one nearly every operating plan assumes will keep falling.
The cleanest number in the AI economy
Microsoft guided to roughly $190 billion of capital spending in calendar 2026, up about 61 percent year over year. Buried in that guidance is a figure that does more analytical work than any other single datapoint available this year: chief financial officer Amy Hood attributed $25 billion of it purely to higher memory and storage component prices.
Not to more capacity. Not to more sites. To the same components costing more.
That is a clean isolation of input-price inflation from volume growth, disclosed by one company, for one component category, in one year. Almost nothing else in the AI spending debate is that well-specified — which is why it is worth more than the aggregate industry figures that circulate without any filing behind them.
For twenty years, "it will be cheaper next year" was the safest assumption in business planning. It is now true of exactly one layer of the stack and false of the layer holding it up.
Where the money is actually going
Gartner's July revision put 2026 worldwide IT spending at $6.37 trillion, up 14.2 percent. The composition matters far more than the total.
Data center systems grow 62.5 percent, to $822 billion. Services grow 5.3 percent, to $1.57 trillion. Software grows 15.5 percent, devices 9.8 percent, infrastructure-as-a-service 29.3 percent.
Sixty-two and a half against five point three. That is not a market where all boats rise. Gartner's own analyst said so explicitly, and the same forecast has now been revised upward four separate times in ten months — from 9.8 percent last October, to 10.8, to 13.5, to 14.2.
A forecast revised up four times in under a year is not really a forecast. It is a series of corrections chasing something the model did not anticipate.
Visual 1 — Two curves, opposite directions
Layer | Direction in 2026 | What your plan probably assumes |
|---|---|---|
Inference and model access | Sharply cheaper — roughly 43% in ten weeks this summer | Nothing. Most plans have no line for it at all. |
Memory and storage components | Sharply more expensive — $25bn of one company's capex | Flat or gently falling, inherited from two decades of experience |
Data center systems overall | +62.5% | Steady replacement-cycle spending |
Hardware refresh cycles | Costs rising; lead times lengthening | Same cost as last cycle, or slightly less |
IT and consulting services | +5.3% — barely ahead of inflation | Continued growth, if you sell them |
Physical capacity availability | Constrained; permitting now a factor in some US regions | Effectively unlimited, available on demand |
How to read it: Row one is the good news and is genuinely large. Rows two through six are the bill for it. Very few three-year plans contain an explicit assumption about either — which means they contain an implicit one, and the implicit one is that technology gets cheaper.
Three consequences worth planning around
The hardware refresh line is wrong. Most multi-year plans carry equipment replacement at flat or declining cost, because that is what the last two decades taught. If component prices are rising and lead times lengthening, that line understates. It is a small line in most businesses, which is exactly why nobody re-examines it, and it compounds quietly across a three-year plan.
If you sell services to enterprises, look hard at 5.3 percent. Consultancies, agencies, systems integrators, managed service providers, staffing firms — the demand signal for your category is growing at roughly the rate of inflation while the category next to you grows at 62.5 percent. That is budget cannibalization visible in market data. Money is moving from people to iron, inside your customers' budgets, right now. Whether it comes back is a question, not an assumption.
Capacity is becoming a contracted input rather than a utility. The physical constraint has started to bind in ways that are new. Some US regions have introduced permitting friction on data center construction this summer, and the largest buyer of capacity in the world has told investors it expects to be capacity-constrained through the year. For most businesses this shows up indirectly — in what your cloud provider will commit to, and on what terms — but the direction is toward scarcity pricing in a market that spent fifteen years behaving like an unlimited utility.
One figure we are not usingAggregate 2026 hyperscaler capital spending is widely quoted at $690 billion, sometimes $725 billion. Neither appears in any of the relevant companies' filings — both trace only to secondary aggregation. Microsoft's roughly $190 billion is company-guided and citable. We mention this because the aggregate is currently doing a lot of unearned work in strategy decks.
What actually changed, in one sentence
The cost of thinking fell and the cost of the machinery that does the thinking rose, and because the first is visible on a pricing page while the second is buried in someone else's capital expenditure line, only the first has made it into how businesses plan.
The deflationary assumption was never written down anywhere. That is what makes it difficult to correct. Nobody's three-year plan has a line saying "we assume technology costs fall" — it is embedded in the refresh schedule, in the growth-without-proportional-cost assumption, in the confidence that scaling up is a procurement exercise rather than a negotiation.
The practical work is short. Find the places in your plan where a falling or flat technology cost is assumed rather than stated, and state it. Then decide whether you still believe it, layer by layer. For inference and software you probably should. For anything with metal in it, this year says otherwise — and this year is the first in a very long time where that has been the case.
Sources and method. A LookatBusiness original. Microsoft's approximately $190 billion calendar 2026 capital spending guidance, up about 61 percent year over year, and CFO Amy Hood's attribution of $25 billion of that figure to higher memory and storage component prices, from company guidance and earnings materials, July 2026; see Microsoft investor relations. Worldwide IT spending forecast of $6.37 trillion for 2026, up 14.2 percent, with data center systems at $822 billion (+62.5 percent) and services at $1.57 trillion (+5.3 percent), and the forecast's revision history across October 2025, February, April and July 2026, per Gartner, July 27, 2026 — a forecast, not a measurement. Inference pricing decline of approximately 43 percent between May 31 and August 8, 2026 per Jefferies research using Silicon Data, reported by the South China Morning Post, August 10, 2026. Aggregate hyperscaler capital spending figures of $690 billion and $725 billion could not be traced to any company filing and are not relied on here. This piece is a companion to our earlier analysis of falling cognition costs and is intended to be read alongside it.
