The AI Act Moved Sixteen Months. The Capex Didn't.

The most consequential AI regulatory fact of 2026 is a date that moved. According to the European Commission’s own implementation timeline, adjusted for the Digital Omnibus on AI, the obligations for high-risk systems listed in Annex III of the AI Act now apply from 2 December 2027 — not August 2026. High-risk systems embedded in products covered by Annex I follow on 2 August 2028.

Sixteen months. Every enterprise buying cycle, governance product roadmap and compliance budget built on an August 2026 deadline was built on a date that no longer exists, and the adjustment was made in the name of competitiveness.

What did apply on 2 August 2026 was real but narrower than the market narrative: the majority of the Act’s rules came into force, Article 50 transparency obligations started to apply, innovation-support measures began, and enforcement started at national and EU level for general-purpose AI models, the prohibitions, the transparency rules and AI literacy. New prohibitions covering non-consensual sexual deepfakes and child sexual abuse material follow on 2 December 2026, with a transition deadline the same day for providers of synthetic-content systems already on the market.

So the regime that arrived in the summer of 2026 was a labelling-and-transparency regime plus enforcement machinery. The regime that requires conformity assessment, risk management systems and technical documentation arrives 2 December 2027. Those are different products sold to different buyers at different prices, and the sixteen-month gap between them is the space in which a lot of AI governance revenue was supposed to be earned.

Meanwhile the other clock ran on schedule

Goldman Sachs Global Institute’s May 2026 analysis puts baseline AI capital expenditure at roughly USD 7.6 trillion cumulatively between 2026 and 2031 — USD 765 billion in 2026 alone, rising to about USD 1.6 trillion in 2031. Its assumptions are unusually explicit for a number this size: NVIDIA at roughly 75 per cent of total compute spend, the VR200 generation as baseline at about USD 80,500 per GPU including node costs and 3,000 watts per package, a power usage effectiveness of 1.2, USD 15 million per megawatt for data centres and USD 2,500 per kilowatt for new power.

JPMorgan’s 2026 technology outlook, using a different method, puts global AI infrastructure spending above USD 400 billion in 2025 with 2026 projections exceeding USD 600 billion. Two banks, same direction, materially different magnitudes. That is a range, and anyone quoting one of them as a point estimate is quoting an assumption set as though it were a measurement.

The two numbers are not comparable, and get compared constantly

Pheonix Research’s forecast for the artificial intelligence market runs from USD 315.80 billion in 2025 to USD 1,285.60 billion in 2033 at a stated 19.20 per cent; the endpoints compound at 19.18 per cent, so the headline is sound. Software is the largest segment, generative AI the fastest growing, and BFSI the leading end-use segment.

Set the two in the same sentence and something uncomfortable appears. Annual AI capital expenditure in 2026, at USD 765 billion on the Goldman baseline, is 2.42 times the entire 2025 revenue of the AI market it is built to serve. That is not a demand forecast. It is a financing statement: infrastructure is being installed against revenue that has to arrive later, and the arithmetic requires the revenue line to compound faster than 19.2 per cent for a period if the 2026 spend is to look proportionate in hindsight.

There are innocent explanations. Some of that capex serves non-AI workloads. Some is replacement of depreciating kit inside existing facilities. Some is being built for customers who are not in the “AI market” as defined by a revenue segmentation, because they are banks and manufacturers buying AI inside something else. And the useful-life question — the one that moved Microsoft’s capex guidance earlier in this cycle when it extended data-centre lives — changes how much of this is even a cost in any given year.

But the ratio is the thing to hold on to. A capital-goods cycle that is twice the size of the market it serves is either very early or over-earning on someone else’s balance sheet, and the 19.2 per cent forecast is not the number that tells you which.

What a 19.2 per cent average conceals

An AI market growth rate is a weighted average of businesses that do not behave alike. Software sold on subscription at 19 per cent is a healthy, mature enterprise category. Accelerators, networking and data-centre capacity growing at thirty-odd per cent is a capital goods boom with a different buyer, a different margin structure and a different cycle.

Blending them produces a number that is defensible as arithmetic and dangerous as a planning assumption, because the two halves fail differently. Software demand can soften without anything physical being stranded. Capacity demand softening leaves assets in the ground.

The regulatory shift interacts with this in a way worth thinking through carefully. If a material share of enterprise AI spending in 2026 was going to be compliance-driven — conformity assessment, documentation, monitoring — then that spending has moved to late 2027 and 2028, and it will be bought by a narrower set of buyers because the transparency duties that did apply are largely documentation duties that can be discharged inside an existing product.

That does not make AI governance a bad business. It makes it a business whose revenue arrives in 2028 rather than 2026, into a customer base that will by then have been through one procurement cycle of the transparency regime and will have opinions about what it actually needs. Vendors who sold the deadline are now selling into a vacuum of their own making.

The honest read

We do not put an independent market size alongside our own here, and we do not use market-research competitors as sources. What the primary record supports is narrower and more useful than a second opinion on the total.

Three statements survive the evidence. European high-risk AI obligations now begin in December 2027, per the Commission’s own timeline as amended by the Digital Omnibus. Global AI infrastructure capex in 2026 is somewhere around USD 600 billion to USD 765 billion depending on whose assumption set you accept, which is between roughly two and two and a half times the 2025 revenue of the AI market itself. And the transparency regime that did start in August 2026 is a documentation obligation, not a conformity assessment one.

For a supplier, that means the next eighteen months are a software market — solving problems customers will pay for now — and the compliance market is a 2027 and 2028 story with a smaller addressable base than the one that was being priced in 2025. For a buyer, it means the deadline you were told about in a sales meeting has moved, and anyone who has not updated their plan since the Omnibus is working from a document that no longer governs.