Roughly half of one company's $664 billion backlog and at least a third of another's $678 billion order book rest on the same customer. That is about $558 billion of contracted revenue, at two companies, depending on one counterparty — which is itself funded by the same capital cycle.
Enormous sums are moving. The useful question is not whether artificial intelligence will matter, which is unanswerable and mostly irrelevant to a portfolio. It is simpler and it is checkable:
Follow the money to where it settles, rather than where it is announced.
Who collects, and who commits
| Selling the tools | Building the claim | |
|---|---|---|
| Who | ASML · TSMC | Oracle · Microsoft |
| Cash position | TSMC: net cash, 67.7% gross margin, revenue +40% | Oracle: free cash flow −$23.7bn, debt $156bn, rated BBB− |
| What is rising | ASML guidance €36–40bn → €43–45bn; capacity +30% | Microsoft capex at 35 cents of every revenue dollar — and free cash flow fell |
| The constraint | How fast they can build machines | How much they can raise, and from whom |
| Who owes whom | Paid on delivery | ~$558bn of backlog on one customer |
The third row is the one to read twice. Microsoft's capital spending reached 35 cents of every revenue dollar — and free cash flow fell. That is a company in excellent health choosing to convert current cash into future capacity, which is a legitimate and possibly brilliant decision.
It is also, mechanically, a transfer. The money leaves Microsoft and arrives at the companies in the left-hand column.
In a gold rush the outfitter is paid on the day. The claim pays later, if it pays at all — and both are in the same industry, reported in the same headlines.
The number that decides this, and it is not a technology question
Our Oracle report scored the company 5.2 — our lowest technology score — while its revenue grew 30% and its cloud infrastructure business grew 121%. Those are not contradictory findings. They are the finding.
The growth is real. The backlog is real. What our report objected to was the combination: borrowing heavily to build capacity for orders that are roughly half dependent on a single customer, with free cash flow at minus $23.7 billion and a rating one notch above speculative.
Remove any one of those three and the position is ordinary. Together they mean the outcome is decided in a refinancing conversation rather than an earnings report.
If Oracle's largest counterparty and Microsoft's largest counterparty were different companies, these would be two independent risks and ordinary diversification would work on them.
They are the same name. So the two positions are correlated — and an investor holding both, believing themselves spread across two providers, is holding one bet twice.
This is the same shape as the risk we described for Taiwan: something that looks like diversification, measured by ticker, and is not, measured by what it depends on.
What each side costs against a 5.12% Treasury
| Multiple | Earnings yield | Against 5.12% | |
|---|---|---|---|
| ASML | 40× 2026 | 2.50% | −2.62 pts |
| Microsoft | 28.8× fiscal 2026 | 3.47% | −1.65 pts |
| TSMC | 27× 2026 | 3.70% | −1.42 pts |
| TSMC (forward) | 21× 2027 | 4.76% | −0.36 pts |
Every one of them yields less than a government bond on current earnings. That is what paying for growth means, and it is not an argument against any of them — but it is the thing being claimed, stated arithmetically.
Note the last row. TSMC on next year's consensus comes within 36 basis points of the risk-free rate — the only company in this article whose cash generation nearly clears the hurdle without requiring the story to work. And it is the one carrying the geographic risk, which is not a coincidence: that is precisely what the market is discounting.
Three questions, for any build-out
Does cash arrive when the product ships, or does it leave now against revenue contracted for later? Both can be excellent investments. They fail in completely different ways, and a downturn arrives at them years apart.
Concentration is disclosed. Then check whether the same name appears behind more than one of your holdings — because if it does, you are not diversified between them, and no sector classification will tell you that.
These are the two questions and almost every argument about AI confuses them. The technology being genuinely transformative and the shares being expensive are entirely compatible — and our verdict on the most prominent name has been "too hard" twice, which remains the honest answer when both can be true.
The only company that makes the machine, the company that runs most of them, and one borrowing to build against a single customer.
Backlogs, counterparties, capital spending and ratings, worked through.
Frequently asked
Who is making money from the AI build-out?
So far, the companies selling the equipment. TSMC earns 67.7% gross margins with revenue growing over 40% and net cash on the balance sheet; ASML raised its 2026 guidance by a fifth in three months and is adding 30% more capacity. The companies building the data centres are in the opposite position — Oracle's free cash flow was minus $23.7 billion and Microsoft's capital spending reached 35 cents of every revenue dollar while free cash flow fell.
Is the AI capital spending boom sustainable?
It depends on one thing more than any other: customer concentration. Roughly half of Oracle's $664 billion backlog and at least a third of Microsoft's $678 billion order book rest on the same customer — around $558 billion of contracted revenue at two companies, dependent on a counterparty funded by the same capital cycle. That is not a judgement about the technology; it is a statement about who owes whom.
What does negative free cash flow mean for a cloud company?
That it is spending more on capacity than the business produces, and funding the difference. For a regulated utility that is normal because a regulator allows a return on the spending. For a cloud provider there is no such guarantee — the return depends on contracts holding, which is why Oracle's minus $23.7 billion, $156 billion of debt and BBB− rating sit together in our report rather than separately.
Are AI stocks in a bubble?
We have declined to judge the most prominent one twice, in June and September, with the verdict "too hard" — which is an honest answer rather than an evasion. What can be said without forecasting is where the cash is going: toward the companies with capacity constraints and away from the companies with construction schedules. That distinction holds whether or not the enthusiasm is justified.