Skip to content
Markets & Trends

The AI Trade Is Now 41% Of The S&P 500: Reading The Capex Debate Without Picking A Side

The 'AI Big 10' now make up around 41% of the S&P 500 - comparable to tech and telecom at the peak of the dot-com bubble, and on some measures more concentrated. US Big Tech capital expenditure rose roughly 60% in 2025 and is expected to climb another 50% past $600bn. Those two facts have produced the loudest argument in markets, and most of it generates more heat than light because the two sides are answering different questions. Here is the steel-manned version of each case, the specific things that would settle it, and what an honest observer should actually watch - including why the strongest bear argument is not about valuations at all.

AlchmAI Editorial12 min read

41%

Share of the S&P 500 made up by the 'AI Big 10' - comparable to tech and telecom at the dot-com peak, and higher on some measures

+60% → +50%

US Big Tech capital expenditure growth in 2025, with a further ~50% rise expected, surpassing $600bn

Cash-funded

Microsoft, Alphabet, Meta and Amazon are financing data centre investment from operating cash flow rather than debt or equity raises

The gap

Between infrastructure spending and observed monetisation - the number that actually decides this argument

There is an argument running through markets in 2026 that has become almost impossible to have productively, which is a shame, because it is the most consequential one available. On one side: AI stock concentration has reached levels comparable to previous bubbles, with the AI Big 10 accounting for roughly 41% of the S&P 500 - similar to the combined tech and telecom share at the peak of the dot-com era, and by some measures more concentrated than that peak. On the other: after rising around 60% in 2025, US Big Tech investment is expected to climb another 50% and surpass $600bn, funded largely from operating cash flow by companies with genuine and growing revenue.

Both of those are true. The reason the debate goes nowhere is that the two camps are answering different questions and treating the answers as contradictory. We are not a research house and have no position to sell; we build AI systems for financial firms and therefore have an unusual vantage point on the one variable that actually settles this, which is whether enterprises are getting returns. So rather than adding another verdict, here is an attempt to lay out both cases at full strength and identify what would genuinely resolve them.

The Bull Case, At Full Strength

The strongest version of the constructive case is not 'AI is amazing'. It is a set of structural differences from the dot-com comparison that are genuinely material:

  • The spending is funded from cash flow. Microsoft, Alphabet, Meta and Amazon are building data centres out of operating cash, not debt raises or equity issuance against a story. This matters enormously for the shape of any downturn: a cash-funded overbuild produces disappointing returns on invested capital, while a debt-funded one produces defaults, forced selling and contagion. Those are very different events.
  • There is revenue underneath, at real scale. The dot-com comparison struggles here. The companies at the centre of this are among the most profitable in history, with the AI spending representing a fraction of cash generation rather than a bet against future revenue that does not yet exist.
  • Adoption indicators keep improving. Survey-based measures of corporate adoption have continued rising through 2026, and what has been learned over recent months has generally been positive for the sector. Enterprise deployment is broadening rather than stalling.
  • Concentration partly reflects earnings, not just multiple expansion. A meaningful share of the Big 10's index weight is justified by actual profit share. That is a different situation from 1999, when concentration was driven substantially by companies with no earnings at all.
  • The infrastructure has residual value. Unlike much dot-com investment, a data centre with power and interconnect is a durable asset with demand from many sources. Even a serious AI disappointment leaves something behind.

The Bear Case, At Full Strength

The strongest bear case is not 'this is a bubble' either. It is considerably more specific, and the most interesting version has very little to do with valuation:

  • The monetisation gap is large and not obviously closing. The scale of AI infrastructure spending is enormous relative to observed monetisation. If useful lives turn out shorter or utilisation lower than assumed, returns on invested capital disappoint - and depreciation schedules on this scale of asset are an assumption, not a fact.
  • Concentration is a systemic risk independent of whether AI works. When ten names are 41% of the index, every pension fund, tracker and 'diversified' portfolio holds the same concentrated bet, usually without intending to. A drawdown in those names is not a sector rotation; it is everyone's portfolio at once.
  • Correlated capex is a fragile equilibrium. Each hyperscaler's spending is partly justified by the others' spending - nobody can afford to be the one that under-invested. That is rational individually and unstable collectively, because the first credible signal of a pullback changes the calculus for all of them simultaneously.
  • Circularity in the revenue is real and under-examined. A non-trivial portion of AI revenue involves companies in the ecosystem buying from each other, sometimes financed by investment from the seller. That flatters headline growth in ways that unwind quickly if end demand disappoints.
  • The financing edge has begun to fray. The Bank of England's July 2026 Financial Stability Report noted AI companies turning increasingly to debt financing to fund infrastructure through the first half of 2026, warning that an adverse shock affecting their ability to service that debt could materially affect global financing conditions. The cash-funded argument is strongest for the largest four and weakens rapidly beyond them.

What Would Actually Settle It

Rather than another prediction, here are the specific observables that would move a reasonable person's view - the things worth watching instead of the narrative:

  1. 01Enterprise AI spend converting to measurable operating margin, in ordinary companies rather than technology firms. This is the load-bearing assumption of the entire bull case, and it shows up in non-tech earnings calls before it shows up anywhere else.
  2. 02Depreciation schedule changes on AI infrastructure. If useful-life assumptions are revised downward, reported earnings across the complex change materially, and that revision would be the clearest possible signal about utilisation.
  3. 03The mix of capex funding. Continued cash funding supports the benign case; a shift toward debt, vendor financing or special purpose vehicles moves this from a returns problem toward a credit problem, which is the scenario with contagion in it.
  4. 04Whether any hyperscaler blinks. The first credible guidance-down on capex from a major player would test whether the correlated equilibrium holds, and the reaction would be more informative than the announcement.
  5. 05Token economics at the model layer. Falling inference costs per unit of useful work supports the productivity thesis; rising costs per unit as tasks get more complex undermines it. This is measurable and rarely discussed in market commentary.

The View From Inside The Deployments

Here is the one thing we can add from where we sit, and we offer it as a data point rather than a thesis. Across the financial services AI work we do, the pattern is consistent and does not fit neatly into either camp. The technology plainly works, and where firms deploy it against well-scoped, high-volume, checkable processes, the returns are real and measurable - not speculative. But the binding constraint is almost never model capability. It is integration: connecting systems that were never designed to talk, cleaning data, and building the controls that let a regulated firm put automation into production. Research finding that 95% of firms have not implemented any cross-system integration for AI matches what we see precisely.

That cuts both ways, which is why it is a useful observation rather than a side. For the bulls: the value is real and there is an enormous backlog of un-captured return, which implies a long runway. For the bears: the bottleneck is unglamorous, slow, firm-by-firm integration work that no amount of additional compute accelerates. Which means whatever the eventual adoption curve looks like, it is likely flatter and longer than capex schedules priced on rapid uptake assume - and the gap between those two curves is where the risk lives.

“The bottleneck on enterprise AI returns is not intelligence. It is integration - and integration does not scale with compute.”

For Financial Firms, Not Just Investors

If you run a financial institution rather than a portfolio, this debate has a practical edge that is easy to miss. Your AI strategy probably assumes the current price and availability of frontier models, and both have been shaped by an extraordinary investment cycle that is not guaranteed to continue on the same terms. A sensible institution should be able to answer what happens to its automation if inference pricing doubles, or if its primary provider becomes unavailable. That is not pessimism about AI; it is the same operational resilience question you would ask about any critical vendor, and the Bank of England has now effectively asked it on the record.

The Bottom Line

The AI capex argument is unresolvable in its current form because the participants are answering different questions, and the honest position is that both halves are substantially true: this is a real technology shift producing real value, and concentration is at genuinely unusual levels with a monetisation gap that has not closed. The most useful reframing is to stop asking whether it is a bubble and start asking which risk you are actually exposed to. For investors, the concentration point stands on its own merits regardless of the fundamental view. For financial institutions deploying this technology, the integration bottleneck we see every day suggests the returns are real but slower to arrive than the spending curve implies - which is an argument for starting the unglamorous integration work now rather than for waiting to see how the argument resolves. The firms that captured value from the internet were not the ones who called the top correctly. They were the ones who had already done the plumbing.

References & Further Reading

AI capexmarket concentrationS&P 500AI bubbleFintech AI Agency LondonAI Agency UKmarkets 2026
Share Email
AI

AlchmAI Editorial

Research and analysis, London

The AlchmAI team writes about the markets, technology and regulation we work with every day. We build trading platforms, real-time charts and AI analysis tools for brokers, prop firms and fintech teams from our office in Mayfair, London. Every article lists its sources. Nothing we publish is investment advice.

This article is general information and commentary. It is not investment advice or a recommendation to buy or sell any investment. Important information

Keep reading

All articles