Part III of the series “AI Infrastructure, Capital and Power”.
Part I: The Boldest Measure May Be the Safest·Part II: Apparently, $500 Billion Is Not Enough
The world needs ever more capital — for governments, defence, energy and artificial intelligence. At the same time, debt levels and financing costs are rising.
The crucial question is therefore no longer simply whether enough capital can be mobilised.
It is whether the cash flows created by that capital will be sufficient to pay for it.
In Part II of this series, we described how data centres and AI factories could evolve into a new financeable asset class — supported by institutions such as BlackRock, Blackstone, Brookfield, Apollo, Goldman Sachs and KKR.
In that scenario, the announced $500 billion would not be the end.
It would be the foundation.
But that leads to a larger question:
Does this machine still work in a world in which governments themselves are absorbing ever more capital?
When Everyone Suddenly Needs the Same Capital
The current situation is unusual.
Western governments are heavily indebted. Defence spending is rising. Electricity grids need to be rebuilt. Reindustrialisation comes at a cost.
At the same time, one of the largest private investment cycles in modern economic history is under way.
Artificial intelligence requires semiconductors, memory, optical connections, data centres, power generation, grids, turbines, cooling — and enormous amounts of capital.
Ultimately, all of them are drawing on the same global capital market.
This is now visible in credit markets as well. US technology companies have already issued roughly $220 billion of AI-related bonds in 2026 alone. At the same time, credit spreads for major technology issuers have widened, and investors are demanding greater concessions on new deals.
The companies remain exceptionally creditworthy.
But capital is no longer available without conditions.
As a result, boundaries that investors long treated separately are becoming more porous: government bonds, corporate bonds, equities, infrastructure and private credit.
The marginal investor is increasingly asking the same question across all of them:
What return am I receiving for what risk — and which future cash flows stand behind it?
Nelson’s Britain — Two Hundred Years Later
Our series began with Admiral Horatio Nelson.
Behind his ships stood not only a navy, but shipyards, trade, supply chains and, above all, a financial system capable of supporting them.
Britain went out into the world.
Switzerland developed a different strength over time: stability, institutional security and the protection of capital.
Both approaches have value.
For investors today, however, the conclusion is simple: a safe harbour protects capital. But the largest new cash flows do not necessarily arise inside that harbour.
Anyone who wants to participate in them has to look outward.
Britain itself now illustrates why not only the amount of debt matters, but also who owns it.
When the Butterfly Flaps Its Wings in London
A growing share of the gilt market is traded by internationally active, price-sensitive investors. These include hedge funds that often finance positions through repo while operating simultaneously across several sovereign-bond markets.
Such an investor does not own British government bonds out of patriotism.
They own them as long as the yield, funding cost, volatility and expected return remain attractive.
When that calculation changes, the position changes.
A crisis therefore does not have to begin with a sovereign default. It is enough for the financing mechanism to deteriorate:
rising yields → losses on leveraged positions → margin calls → higher haircuts → deleveraging
Once positions are sold, the effects can spread into other markets.
Because the same participants often own more than gilts.
The butterfly flaps its wings in London. Somewhere else, a storm begins.
Who owns the debt, and how that owner is financed, can therefore matter just as much as the amount of debt itself.
Why America Is Different
The United States is heavily indebted too. Interest costs are rising, and enormous volumes of debt must be refinanced continuously.
But the Treasury market has one distinctive feature:
It is part of the global financial infrastructure.
Treasuries serve as reserves for central banks, as collateral for banks and funds, and as the foundation of the global repo system. For institutions that need to place billions — or even hundreds of billions — of dollars in liquid, continuously tradable instruments, there is scarcely a genuine alternative of comparable scale.
That creates structural demand.
But this advantage is not unlimited either.
Movements at the long end show that even the United States pays a price for rising debt and capital demand: the 30-year Treasury yield reached its highest level since 2007 at one point in August.
When debt, interest costs and inflation rise at the same time, the long-term question remains the same:
Which future cash flows will support these obligations?
For a sovereign, those cash flows are ultimately tax revenues.
And tax revenues depend on nominal economic growth and productivity.
America Is Investing in a Possible Answer
The United States may be one of the economies with at least a realistic chance of easing part of its debt burden through growth.
Not because its debt is low.
But because a historic investment cycle is taking place at the same time.
Artificial intelligence, energy infrastructure, data centres and semiconductors are being built out on a scale not seen for decades.
In the short term, that intensifies competition for capital.
Over the longer term, however, this capital deployment could raise productivity.
Investment → Compute → Productivity → Profits & Incomes → Nominal Growth → Tax Revenues
This chain is not guaranteed.
The decisive questions will be how quickly AI actually raises productivity, how broadly those gains spread through the economy, and how much ultimately shows up in incomes, corporate profits and tax receipts.
If the chain works, America is not merely building data centres.
It is building part of the future cash-flow base from which its public obligations can also be serviced.
That is the difference between productive investment and merely higher consumption.
Two Ways to Read the Same Market
That is exactly why the current market picture is so interesting.
On one side stand gold, Bitcoin and high government-bond yields.
One interpretation is that investors are beginning to question more seriously the long-term purchasing power of paper currencies and the sustainability of rising sovereign obligations.
On that reading, gold is a warning signal.
But there is another interpretation.
Despite the recent turbulence, Ed Yardeni continues to hold to his base case that the 10-year US Treasury yield will remain between 4 and 5 per cent through the end of this year and next.
In his view, recent Treasury buybacks are primarily a tool for improving market liquidity — not an attempt to force long-term yields artificially lower.
At the same time, expected corporate earnings continue to rise. Analysts even nudged up their S&P 500 earnings estimates for the current quarter in August — unusual at this stage of the quarter. Meanwhile, 86 per cent of S&P 500 companies that had reported were beating earnings expectations.
That opens a second interpretation:
Perhaps 4 to 5 per cent on 10-year Treasuries is not the problem. Perhaps productive investment simply has to prove once again that it can earn more than its cost of capital.
That would not be a breakdown of the system.
It would be the return of capital discipline.
Does the Machine Work?
This brings us to the decisive question.
Compute, semiconductors, energy and data centres have become strategic infrastructure in the global competition with China.
America is building.
Almost like a rearmament programme.
But this modern rearmament has one unusual feature:
It is meant to be both strategically necessary and economically profitable.
The build-out does not merely have to happen.
It has to generate enough cash flow for private capital markets to support it over time.
That leaves two questions side by side:
Does America need to build this infrastructure? Very probably, yes.
Will the invested capital earn enough? That is the real investment question.
Follow the Forward Capital
Start with the capital.
Four of the largest US hyperscalers alone are planning 2026 investment of roughly $720 billion to $745 billion.
Amazon now expects around $220 billion of cash CapEx. Alphabet plans $195 billion to $205 billion. Meta expects $130 billion to $145 billion including finance-lease payments. Microsoft has most recently indicated around $175 billion for the calendar year.
The definitions differ — cash CapEx and CapEx including leases are not perfectly comparable — but the order of magnitude is clear.
Only a few years ago, such a sum for four companies would have been difficult to imagine.
Today, apparently, it is still not enough.
Despite investing $220 billion, Amazon says AWS will not have enough capacity in 2026 to meet demand fully — and expects a similar situation in 2027.
Microsoft likewise expects to remain supply-constrained at least through 2026 despite its massive capacity build-out.
This is forward capital.
It tells us where the next wave of productive capacity is going — and where bottlenecks are emerging.
But capital flows alone are not an investment thesis.
The decisive question is:
Where does the resulting cash flow accrue?
The Cash Flow Is Beginning to Become Visible
This is where the picture becomes more interesting.
The hyperscalers are investing historic sums — while substantial new revenues are already emerging.
Google Cloud generated $24.8 billion of revenue in the second quarter, 82 per cent more than a year earlier. Operating income in the cloud segment reached roughly $8.8 billion, with an operating margin above 35 per cent.
Amazon Web Services grew 37 per cent to $42.2 billion of quarterly revenue, with operating income of $16.6 billion. Amazon’s AI business alone has now surpassed an annualised revenue run-rate of $25 billion.
Microsoft reported Azure growth of 43 per cent and a cloud backlog of roughly $678 billion. Despite the enormous investment, the company still generated $19.6 billion of free cash flow in the latest quarter.
Meta also shows how large both sides of the equation have become: quarterly revenue of $60.8 billion and operating cash flow of $31.9 billion stood against roughly $31.1 billion of capital expenditure. Free cash flow consequently shrank to just $784 million for the quarter.
So the cash flows do exist.
But investment is in some cases running ahead of them.
Alphabet reported negative free cash flow of roughly $5.9 billion in the second quarter for the first time. Amazon’s trailing 12-month free cash flow fell to minus $7.6 billion despite $161 billion of operating cash flow.
That shifts the central question.
Not: Is AI generating revenue? It already is.
But:
Will the cash flow it generates grow faster over time than the capital required to create it?
NVIDIA — Where the Cash Flow Arrives Today
Few companies illustrate this second half of the equation better than NVIDIA.
While the hyperscalers invest hundreds of billions of dollars in physical infrastructure, NVIDIA remains relatively capital-light.
In the first quarter of its current financial year, NVIDIA generated $81.6 billion of revenue, of which $75.2 billion came from Data Center. Free cash flow was $48.6 billion.
That is almost 60 cents of free cash flow for every dollar of revenue.
At the same time, financial debt stood at only around $8.5 billion.
NVIDIA is therefore, so far, one of the companies where the forward cash flow from this enormous capital cycle is particularly concentrated.
For the second quarter, NVIDIA has guided to $91 billion of revenue; market consensus is around $92 billion, while more than $100 billion is already expected for the following quarter.
That is why the upcoming results matter:
Not because a single quarter will decide whether AI works.
But because NVIDIA is one of the most direct sensors of whether capital is still accelerating through the machine.
The Financing Machine
This is where the group of NVIDIA, BlackRock, Goldman Sachs, Blackstone, Brookfield, Apollo and KKR becomes particularly interesting.
Their answer to the enormous capital requirement is not for NVIDIA or any one hyperscaler to put another half-trillion dollars of debt onto its own balance sheet.
The answer is:
We create a new financeable asset class.
AI factories produce compute. Compute produces revenue.
Long-term usage agreements can support loans that are structured, distributed and securitised.
Goldman can structure such financings and place them in capital markets.
BlackRock can allocate suitable risks across client portfolios.
Private-credit and infrastructure managers can provide additional parts of the capital.
Financing therefore does not remain concentrated on a single balance sheet.
Capital is recycled. Risk is distributed.
That is the financing machine.
Securitisation Is Not Magic
Securitisation can mobilise capital and reduce financing costs.
But there is one thing it cannot do: turn insufficient cash flow into sufficient cash flow.
An AI factory with high utilisation, long-term contracts and creditworthy customers can be an excellent financeable asset.
But three specific risks deserve particular attention:
- Technological obsolescence: If new chip generations erode the value of older compute capacity faster than expected, the economic value of existing infrastructure falls more quickly than its financing can be amortised.
- Utilisation and pricing risk: If compute capacity grows faster than underlying demand, or usage prices fall more sharply than expected, cash flows come under pressure.
- Counterparty concentration: Long-term usage agreements are only as robust as the creditworthiness and economic incentives of the small number of major customers standing behind them.
If refinancing costs rise at the same time, even a technically excellent asset can come under financial pressure.
Structuring does not make the risk disappear. It merely changes who owns it.
A government borrows against future tax revenues. An AI factory borrows against future compute cash flows.
The fundamental question is the same:
Will future revenues be sufficient to pay a return on the capital invested and repay it?
From the Application to the Power Grid
NVIDIA sits at the centre of the story, but it is not the whole story.
The expected cash flow runs through the entire chain:
AI Application → Model Provider & Hyperscaler → Compute → AI Factory → Semiconductors, Memory & Networks → Energy & Infrastructure
That is precisely why Marvell is an interesting example.
In its latest quarter, Marvell generated $2.42 billion of revenue, 28 per cent more than a year earlier. Data Center accounted for 76 per cent of the total.
The new agreement with Google could generate as much as $120 billion of revenue through 2033. At the same time, Marvell expects its custom-silicon business to grow to more than $10 billion by 2029.
The capital first flows to Google.
Google builds TPUs and AI infrastructure.
But part of the resulting forward cash flow accrues to Marvell — through custom silicon, networking controllers and memory interconnect.
The same mechanism continues further down the chain: memory, optical connections, electricity, grids, turbines and construction equipment are all required.
Forward capital moves through the chain.
The investor’s task is to determine where the forward cash flow ultimately sticks.
Not Every Winner from the Build-Out Is a Good Investment
This is exactly why a powerful theme is not enough.
- Alphabet is growing cloud revenue dramatically — while temporarily reporting negative free cash flow.
- Amazon is growing AWS rapidly — while investing more than the group currently generates in free cash flow.
- Meta owns a highly profitable platform — while absorbing almost all quarterly operating cash flow in capital expenditure.
- NVIDIA is producing extraordinary free-cash-flow margins with relatively little capital of its own.
- Marvell sits at another point in the chain: lower margins than NVIDIA, but substantial leverage to custom silicon.
Five companies in the same cycle — with five completely different risk/reward profiles.
In 2000, an investor could have been entirely right about the future of the internet and still owned the wrong stock.
A good theme is not yet a good stock.
And a good stock is not a good investment at every price.
The Numbers Have to Confirm the Story
The price of capital changes. Political majorities change. Pricing power changes. Bottlenecks move.
And sometimes the market changes faster than the story investors tell about it.
Gold may be an important warning signal today — but its price reflects not only fiscal concerns, but also structural central-bank buying, geopolitical diversification and the search for a reserve asset without counterparty risk.
A strong move in gold alone therefore does not decide which interpretation of the capital markets is correct.
A 10-year Treasury at 4.7 per cent could signal the beginning of a financing crisis. Or, as Yardeni argues, it could sit within a fundamentally normal 4-to-5-per-cent regime that simply forces capital discipline back into the system.
We do not need to settle that question ideologically today.
We can observe which story is confirmed by the cash flows.
Follow the Forward Capital
Our framework therefore has two stages:
Follow the forward capital — and determine where the forward cashflow accrues.
Not the headline. Not the political narrative. Not the past share-price performance.
Capital flows show us where new capacity is being built, where bottlenecks are emerging and where new infrastructure will be required.
But the second step is what determines the investment:
Who earns from it? Who has pricing power? Who needs little additional capital to do so? Who bears the financing and obsolescence risk? And whose future cash flows are growing faster than valuation and the cost of capital?
Capital flow explains the opportunity. Forward cashflow determines the investment.
Follow the forward cashflow.
How We Know Whether the Machine Is Working
The thesis can be tested.
We do not need to know what the world will look like in ten years.
We need to watch whether five key variables are moving in the right direction:
- AI and cloud revenue growth: Is monetisation still growing fast enough to justify the rising investment base?
- Revenue and cash flow per additional dollar of CapEx: Is each new investment wave producing enough additional economic return — or is capital productivity falling?
- Hyperscaler free cash flow: How long can the major platforms finance the build-out internally before external financing becomes structurally more important?
- Credit spreads and financing costs: Does capital remain available to hyperscalers, data centres and infrastructure at sustainable terms? Or do rising capital costs begin to consume the returns on new projects?
- Compute utilisation and pricing: Does compute remain scarce and profitable? Or does capacity eventually grow faster than demand, putting utilisation prices and returns under pressure?
Together, these five variables form an early-warning system.
If monetisation, cash flow and utilisation keep pace with capital deployment, and returns remain sustainably above financing costs, the machine is working.
If the opposite happens, the investment cycle increasingly becomes a financing problem.
The Nelson Moment
Nelson had an objective. But he did not follow a rigid course.
He had to reassess the wind, the current, the enemy’s position and the strength of his own fleet continuously.
For investors today, that is critical.
We see artificial intelligence as an extraordinary economic and strategic opportunity.
At the same time, we see a world in which sovereign debt, higher capital costs, political change and new credit structures create additional risks.
Both can be true at the same time.
Our task is therefore not to predict a single future.
Our task is to adjust the course when the evidence changes.
Beyond the Safe Harbour
For Swiss investors, this may be the most important conclusion:
Switzerland remains one of the safest harbours in the world.
But a safe harbour is not automatically where the highest future cash flows are created.
A substantial share of them arises elsewhere — above all in the United States.
Anyone who wants to participate has to look outward.
But the rougher the seas become, the more important navigation becomes.
Our objective is a better relationship between expected return and the risk taken.
To achieve that, we need to follow where the capital is going. We need to understand who is financing it. We need to know who carries the risk.
And above all, we need to identify where the resulting cash flow actually accrues.
What Will These Trillions Earn?
This lets us answer the question from Part II more precisely.
The issue is not whether $500 billion is enough.
The four largest hyperscalers alone are already moving towards three quarters of a trillion dollars of investment in 2026.
The decisive question is:
What will these trillions earn?
The first answers are becoming visible: cloud revenues are growing, AI revenues are emerging, backlogs are rising, and NVIDIA is generating extraordinary free cash flow.
At the same time, warning signs are appearing: the free cash flows of the capital providers are coming under pressure, bond issuance is rising, and credit spreads are widening.
This is neither proof of a bubble nor proof that every investment will pay off.
It is the beginning of the decisive phase of this capital cycle.
The build-out phase has to become the earnings phase.
If that happens, credit can be standardised and distributed. Capital can be recycled. Productivity can rise. And technological leadership can become an economic and strategic advantage.
If the numbers do not add up, even the most sophisticated financing structure can only pass the problem on.
That is why we are not watching only how large the AI investment cycle becomes.
We are watching five things:
Monetisation. Capital productivity. Free cash flow. Financing costs. Compute utilisation.
They will tell us whether the machine is working.
Because in the end, neither the size of the narrative nor the amount of capital invested will decide the outcome.
In the end, cash flow decides.