A $1.6 trillion figure appeared in my feed last week, attached to a Morgan Stanley framework about AI capital expenditure. It had no time horizon. No statistical perimeter. No methodology note. Just a very large number, forwarded through a Web3 vertical that had clearly never modeled a depreciation schedule in its life. I spent three hours trying to reverse-engineer where the number came from. I could not confirm it. What I could confirm is far more interesting: the narrative structure of that article tells you more about where this cycle sits than the number ever could.
Let me be precise about what I am claiming. The original research behind this chain โ assuming it exists in the form described โ reached me as third-hand information: sell-side note, to crypto media rewrite, to a thin summary with five bullet points and one unverifiable scalar. When a Web3 outlet crosses over into traditional capital markets coverage, the motive is almost always traffic, and the technical depth is almost always absent. So everything below treats the $1.6 trillion as a directional signal, not a data point. Check the supply schedule. Always. And check the perimeter of the number before you check anything else.
Here is what the perimeter should have told you. The four hyperscalers โ Microsoft, Alphabet, Amazon, Meta โ ran combined capital expenditure around $230 billion in 2024. The 2025 guidance aggregated to somewhere between $380 and $400 billion. Consensus for 2026 has already pushed past $500 billion. If $1.6 trillion is a three-to-four year cumulative figure, it reconciles almost exactly with published guidance. Which means the number is not a forecast. It is an extrapolation of guidance dressed as independent analysis. That distinction matters enormously, because extrapolations inherit the assumptions of the people doing the spending, and those people are not disinterested parties in the outcome.
Now the structural fact that does hold up under scrutiny, and holds up hard: index concentration. The top ten S&P 500 constituents carried roughly 35 to 40 percent of total index weight through 2025, the highest band in half a century. The Magnificent Seven contributed more than half of index earnings-per-share growth across 2024 and 2025 by most credible estimates. So the equation โ index growth equals AI capex growth โ is statistically defensible. This is not sell-side fabrication. It is arithmetic.
But arithmetic is where the analysis stopped, and that is the failure. Three links in the chain were simply omitted, and each one is more dangerous than index concentration.
The first is depreciation policy. Capital expenditure does not hit the income statement directly. It arrives, years later, through depreciation. Several hyperscalers have quietly extended the useful-life assumptions on servers and networking equipment from three-to-four years toward five-to-six. Extending useful life during a period of steep capex acceleration is mathematically equivalent to inflating current-period EPS. Some portion of the earnings growth that investors are treating as operational improvement is an accounting choice. I have audited enough infrastructure budgets to know that nobody extends an asset life because the hardware got better. They extend it because the number needs to work.
The second omission is circularity. Chip vendors take equity stakes in customers. Customers reciprocate with purchase commitments. Cloud providers sign multi-decade compute contracts worth tens of billions with model labs. Data centers get pushed off balance sheet through special purpose vehicles and joint ventures. The effect is that AI revenue and AI spending become partially self-causing. Circular revenue structures inflate the apparent independence and durability of every participant in the chain simultaneously. This was the central controversy of 2025 and the article did not mention it once.
The third omission is the migration of risk from equity to credit. Hyperscalers and the new-class cloud providers have moved into large-scale debt issuance โ corporate bonds in the tens of billions per transaction, private credit, asset-backed structures financing the physical plants. The AI capex cycle is no longer a technology equity story. It is an investment-grade credit story with a technology wrapper. For anyone who believes their portfolio is diversified because they hold investment-grade fixed income alongside their tech exposure, this is a materially uncomfortable realization that almost nobody has priced.
Which brings me to the part of the narrative that is genuinely well-founded, and I want to give credit where the structure is sound. The $1.6 trillion represents real, cross-sector physical capital formation. It has produced measurable revenue pull in electrical equipment, construction, cooling, real estate, and utilities. Multiple institutions estimate that data center construction and information processing equipment investment contributed on the order of one full percentage point to US GDP growth in the first half of 2025. US economic growth has, in a statistical sense, become partially AI-ified. That is the strongest evidence for the dependency thesis, and it is stronger than the index concentration argument the article led with.
But the distribution of that benefit is brutally uneven. Bottleneck segments capture excess margin. Generic segments capture volume without pricing power. Transformers, switchgear, gas turbines, and CoWoS-class advanced packaging carry order books extending past 2028. The pricing power in those nodes is exceptional. It is also fully reflected in valuations that have already re-rated these traditional industrials into high-beta AI proxies. If you bought a utility for defensive characteristics and it now trades like a semiconductor, you did not diversify. You levered up.
The employment profile deserves its own warning. Data center construction is trade-intensive โ electricians, pipefitters, HVAC specialists, structural crews. Operations require a hundred people per site, maybe. The employment impulse is a pulse, not a plateau, and it is geographically concentrated in Virginia, Texas, Arizona, and Ohio. Meanwhile the white-collar displacement effect from AI adoption is continuous and distributed. These two effects do not overlap in time or in population. They cannot hedge each other. Any regional economic model that treats them as offsetting is wrong.
Now the contrarian angle, because this is where I part company with both the bulls and the bears. The bear case says the capex is bubble spending that will collapse. The bull case says the revenue is coming. Both are constructing the same model with different coefficients, and both are wrong about the mechanism. The binding constraint is no longer capital. It is electrons. Grid interconnection queues, transformer lead times, gas turbine production slots โ these are three-to-five year rigid physical limits. The physical ceiling on this capex cycle will be reached before the financial ceiling, and the timing of the peak year will be set by substation schedules, not by boardroom confidence. Almost no one is modeling this. It means the cycle extends longer than bears expect and terminates harder than bulls expect, because the cliff arrives on a fixed calendar that management teams cannot negotiate with.
And underneath all of it sits the question the entire narrative framework exists to avoid: what revenue is required to service this capital? Run the arithmetic yourself. If $1.6 trillion is deployed over five years and depreciated over five, annual depreciation lands near $320 billion. At a 60 percent gross margin, covering depreciation alone demands roughly $530 billion in annual AI-attributable revenue. Add a cost of capital and the requirement climbs further. Current directly attributable AI revenue โ model APIs, AI application subscriptions, compute rental โ sits in the hundreds of billions. The gap is somewhere between five and ten times. That gap is not proof of a bubble, because a large share of this capex funds internal efficiency rather than sold product, and internal efficiency returns are real but unmeasured. That measurement problem is precisely why the gap is unfalsifiable, and precisely why sell-side research avoids the reverse calculation entirely. Missing that calculation means missing the load-bearing wall of the entire argument.
Watch the budget migration instead. Enterprise AI spending today comes predominantly from innovation budgets, not from reallocation of existing IT spend. The order-of-magnitude step change requires the second. That is the single most important observable over the next twelve to eighteen months, and it will not appear in a headline earnings number. It will appear in enterprise software budget surveys and in the quiet deterioration of legacy vendor renewal rates.
Code does not lie. People do. The code in this case is a depreciation schedule, a debt maturity wall, an interconnection queue, and a revenue line that has not yet arrived. Read those four artifacts and you will know where this cycle turns before the narrative does. The narrative, after all, has already reached the traffic-driven verticals. Historically, that is not where the bottom forms.