AI, Quantum and the Market

As I write this, the AI complex is having a bad week. The Nasdaq 100 is down 2.0% on the day at 661.73 with an RSI of 32 and a stochastic reading under 10 — mechanically about as oversold as it gets. Semiconductors are down 4.8%. Nvidia is 19.7% below its 52-week high. The quantum names are down 8–9% today and Rigetti has given back 77% from its high. The VIX is up 13.5% to 20.65.

Two things can be true simultaneously: this is a sharp, disorderly repricing of the AI trade, and none of it tells you whether AI itself will prove transformative. Markets price expectations against positioning, not truth against time. This piece separates the two questions — what is being repriced, and what the underlying evidence supports — because they have different answers and different time horizons.

1. The scale of the capex commitment

The numbers are genuinely without modern precedent. Hyperscale cloud providers have collectively guided to somewhere between $600bn and $630bn of capital expenditure in 2026, with roughly three-quarters earmarked for AI infrastructure. Amazon alone plans around $200bn; Alphabet has guided to $175–185bn. Some estimates put the total above $700bn once the wider ecosystem is included.

Goldman Sachs projects total hyperscaler capex from 2025 through 2027 at $1.15 trillion — more than double the $477bn spent across the preceding three years. To put that in perspective, this is a small number of companies committing capital at a rate comparable to national infrastructure programmes, on a three-year view, against a demand curve nobody can yet measure precisely.

2. The part that actually matters: how it's being financed

For most of the last decade, big tech capex was funded from operating cash flow. That is changing. FactSet's analysis shows hyperscalers tapping external financing as AI capex outruns cash flow, and declining free cash flow across the group.

This is the single most important structural shift in the story, and it gets far less attention than the headline spending figures. Capex funded from surplus cash is a choice that can be reversed cheaply — you simply stop. Capex funded by debt is a commitment with a coupon attached. It changes the behaviour of the borrower: the incentive to keep spending in order to justify prior spending increases, and the tolerance for a slow revenue ramp decreases.

It also introduces a genuine transmission channel to the rest of the market. If AI infrastructure were financed entirely from Alphabet's and Amazon's cash piles, a disappointing return would be a shareholder problem confined to a handful of stocks. Financed with credit, a disappointing return becomes a credit-spread problem, and credit-spread problems propagate.

3. The capex-to-revenue gap

The central question is simple to state and hard to answer: is the spending producing revenue at a rate that justifies it? The honest current answer is not yet, and the gap is widening. Forbes reported in June that the AI capex-to-revenue gap is widening and that markets have begun to notice — which is a reasonable description of what the tape has been doing since.

There is a further concentration problem underneath the headline concentration. The hyperscalers' AI revenue is itself concentrated in a small number of customers — OpenAI and Anthropic prominent among them. So you have concentrated capital expenditure, serving concentrated customer demand, funded increasingly by concentrated credit. Each layer is individually defensible. Stacked, they describe a system with limited redundancy.

The bull case is not naive, and it deserves stating properly: infrastructure historically gets built ahead of the applications that use it, and the returns accrue over a longer horizon than quarterly reporting accommodates. Railways, electrification and fibre all involved a capex phase that looked reckless in real time and reasonable in retrospect — with the important caveat that in several of those cases the first generation of investors was wiped out and the second generation collected the returns from the assets they left behind.

The distinction that matters for an investor is not "is AI real" but "does the entity I own capture the value, at the price I paid." Those are different questions and only the second one is answerable with a spreadsheet.

4. Quantum: real milestones, unreal valuations

Quantum computing has moved from speculative to demonstrable this year, and the distinction is worth being precise about.

IBM's CEO Arvind Krishna has said he expects the first examples of quantum advantage — a quantum machine doing something useful that classical hardware cannot practically do — to appear during 2026. That's a meaningful statement from someone with a reputation to protect. IonQ reports 99.99% two-qubit gate fidelity, Q1 revenue of $64.7m (up 755% year on year) and roughly $3.1bn of cash. Rigetti tripled quarterly revenue to $4.4m, launched a 108-qubit system, and secured $100m as part of a $2bn US government quantum package.

Now the other side of the ledger. Rigetti's quarterly revenue of $4.4m supports a market capitalisation in the billions. That is not a valuation supported by cash flows; it is a valuation supported by a narrative about the 2030s. And the market has begun to notice: the stock is down roughly 77% from its 52-week high and fell another 9% today. D-Wave fell 8% alongside it.

My own read: quantum is a genuine long-duration technology story with a real chance of mattering enormously, attached to equities whose prices currently embed an outcome that is neither confirmed nor timed. Government funding provides a floor under the research; it provides no floor under the share prices. If you own these, own them in a size that survives a 70% drawdown, because you have already had one.

5. Employment: the data so far

Here the evidence is more concrete than the discourse suggests, and more limited than either camp claims.

What the data shows: in April 2026, roughly 21,400 job cuts — about 26% of all announced cuts — were directly attributed to AI. On a global basis the net twelve-month employment effect is mildly negative, with the share of firms reporting AI-related job losses running about 5 percentage points above those reporting gains. Surveys of firms adopting AI cite process efficiency (64%) and productivity (59%) as the main goals; headcount reduction is cited by only 24%.

That last statistic is the interesting one, and it's routinely misread. A company doesn't need to intend to cut jobs for employment to fall. If one person with AI tooling can draft, summarise, research and route work faster, the team simply hires fewer juniors next year. No redundancy is announced. Nothing shows up in the layoff data. The effect appears as an absence.

And that's precisely where it is appearing. Anthropic's own labour-market research finds fewer people aged 22–25 starting jobs in highly AI-exposed occupations relative to low-exposure ones. The mechanism isn't mass displacement of existing workers; it's a quiet closing of the bottom rung of the ladder.

If that persists, the consequences are slow-moving and serious. Entry-level roles are how professions train their next cohort. A profession that stops hiring juniors for five years discovers a decade later that it has no mid-career practitioners. That's a labour supply problem that no productivity gain fixes, and it won't be visible in the monthly payroll print.

For balance: the World Economic Forum's projection is that by 2030 structural churn affects 22% of jobs, with roughly 170 million roles created and 92 million displaced — a net gain of 78 million. Historically, that pattern of net job creation through technological transition has held. It has also historically involved a painful interval for the displaced cohort, and the aggregate net figure conceals that entirely.

6. The economy: a disagreement of an order of magnitude

On the macro question, the range of credible professional opinion is extraordinarily wide — which is itself the most useful piece of information available.

The spread between Acemoglu's roughly 1% and Korinek's 18% is not a rounding difference — it's a disagreement about whether this is a useful new tool or a change in the character of economic growth. When the professional range is that wide, any confident market narrative built on "AI will add X to GDP" is selecting a number to suit a position.

On present evidence, the sceptics are ahead on points. US Census Bureau data puts AI implementation below 10% in most sectors. Adoption is slower than headlines imply, hard tasks remain hard, and institutional friction in large sectors like healthcare is substantial. Penn Wharton's middle path — +1.5% by 2035, around 3% by 2055 — is the estimate I'd weight most heavily, because it takes adoption friction seriously and doesn't require anything unprecedented to be true.

The critical nuance: the optimists are largely not making a claim about 2026. They're making claims about trajectories — that compute scaling unlocks task complexity, that capital overwhelms institutional friction, that recursive innovation compounds. Those claims can't be falsified by this year's data, which means this year's data can't validate them either.

7. Short term versus long term

Short term (0–24 months)

The dominant force is financial, not technological. What matters is whether capex growth can be sustained while free cash flow declines, and how credit markets price that. The economic contribution in this window is mostly the capex itself — construction, power, semiconductors, cooling, grid connections — which is real GDP regardless of whether the models ever earn a return. That's why industrials and energy have been carried along, and why energy is one of the few sectors green today.

The risk in this window is a straightforward positioning risk: an enormous amount of capital is in a small number of correlated names, and the marginal buyer has to be found at ever higher prices. That's what's unwinding this week.

Long term (5–15 years)

The dominant force is diffusion, not invention. General-purpose technologies deliver their gains when they're embedded in ordinary processes across ordinary firms, which historically takes one to two decades and requires reorganisation, not just adoption. Electrification took roughly forty years to show up in productivity statistics because factories had to be physically rebuilt around it.

My working assumption: real but unspectacular aggregate productivity gains arriving later than consensus expects, with a highly uneven distribution — large gains in a few sectors, near-zero in most, and material labour-market disruption concentrated in entry-level knowledge work. That's a world in which AI matters a great deal without justifying every current valuation.

8. What would change my mind

Stating this in advance is the only way to keep the analysis honest:

• Bullish revision if: hyperscaler AI revenue growth accelerates to close the capex gap while free cash flow stabilises; AI adoption in Census data breaks convincingly above 20% across sectors; measured productivity growth exceeds 2% annually for consecutive quarters.
• Bearish revision if: depreciation schedules on AI hardware get extended; a major hyperscaler cuts capex guidance; investment-grade tech spreads widen materially; entry-level hiring contraction spreads beyond the most AI-exposed occupations.
• Quantum reassessment if: a verified commercial quantum advantage result appears with a paying customer attached — that would move it from narrative to cash flow, and the valuation question would become answerable.

Conclusion

The technology is real; the timeline is uncertain; the current prices embed a specific and aggressive version of that timeline. This week's selloff isn't evidence that AI has failed — it's evidence that positioning had run ahead of demonstrable cash flows, which is a statement about markets rather than about machines.

The most defensible position, given a professional forecast range spanning an order of magnitude, is to size positions such that being wrong about the timeline is survivable. Both the enthusiasts and the sceptics are making claims that current evidence cannot settle. Accepting that is more useful than pretending otherwise.

For educational purposes only. This is analysis of publicly reported figures and my own interpretation of them — not investment advice, and not a recommendation on any security mentioned. Market data as at 30 July 2026 via TradingView; figures move. Verify all data against primary sources before acting on it.

Sources

• FactSet — Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow
• Forbes — The AI Capex-to-Revenue Gap Is Widening
• Futurum — AI Capex 2026: The $690B Infrastructure Sprint
• CoBank — Why the AI Capex Cycle May Just Be Beginning
• Motley Fool — IonQ vs Rigetti
• Quantum Stocks Rally Ahead of 2026 Advantage Catalyst
• S&P Global — AI Impact on Employment 2026
• Goldman Sachs — How Will AI Affect the US Labor Market
• Penn Wharton Budget Model — Projected Impact of Generative AI on Productivity
• AI Frontiers — The Quadrillion-Dollar Disagreement
• AEI — Goldman Sachs Responds to Daron Acemoglu

← All reports