By Darren Katz 

Every great capital cycle eventually produces two sets of books. One is kept by the optimists, who measure the size of the prize. The other is kept by the lenders, who measure how long it takes to get paid. For most of the cycle the two ledgers agree. The interesting moment arrives when they stop agreeing, and investors have to decide which one to believe. 

AI infrastructure spending has reached that moment. This week two pieces of research landed on my desk within days of each other, both looking at the same companies and the same spending, and both reaching very different conclusions. 

The first came from ARK Investment Management, whose newsletter of 3 November asked why market fears about AI infrastructure spending are misplaced. Its answer is an elegant piece of arithmetic that ends, with characteristic confidence, in the line “math is math”. The second came from Charlie Bilello’s weekly chart pack, which asked whether the AI spending bill is coming due and answered with free cash flow statements, bond issuance tallies and credit default swap spreads. 

Both are interesting. Neither is the whole story. The useful work for a long term investor sits in the gap between them. 

The optimist’s ledger: a very large prize 

ARK’s case, set out by its Chief Futurist Brett Winton, starts with the size of the market rather than the size of the bill. Global knowledge work wages total roughly US$40 trillion. Businesses have historically paid software vendors something like 10% of the productivity gains they deliver. If AI systems can make knowledge workers ten times more productive, the arithmetic produces a revenue opportunity of around US$40 trillion over the next five to ten years. 

Against that, ARK estimates that OpenAI and Anthropic combined have penetrated less than 0.4% of their addressable market. On its numbers, the two labs’ combined run rate revenue could rise from about US$150 billion at the end of 2026 to roughly US$700 billion by the end of 2027 and more than US$2 trillion by the end of 2028. The 2031 figure, which ARK itself concedes seems outlandish, is north of US$10 trillion. 

The more interesting part of the argument is not the size of the market but the efficiency trend. ARK points to Anthropic’s newly public S-1 filing, which shows revenue per dollar of compute rising from about US$0.40 in early 2025 to roughly US$2.08 at its peak, before settling at about US$1.72 as new short term capacity came online. Anthropic’s first half 2026 revenue was US$16.3 billion, its gross margin before research costs was about 85% on ARK’s calculation, and it reported its first quarter of operating profitability in the June quarter. It also holds about US$132 billion of cash and short term investments, having raised around US$141 billion since its founding. 

ARK’s conclusion is that the binding constraint is compute, not demand. Rental rates for GPUs that are more than five years old are still rising. Give the frontier labs more capacity and they will fill it. In hindsight, it argues, today’s fears will look misplaced, if not laughable. 

It is a coherent argument, and it should be read with the respect it deserves and the scepticism it invites. ARK runs funds built around disruptive innovation. It is not a disinterested observer of this debate, any more than a property developer is a disinterested observer of interest rate policy. 

The lender’s ledger: a very large bill 

Now turn the page. Bilello’s evidence is less about what AI might earn and more about what it is costing right now, and who is paying. 

Meta’s third quarter result, released on 21 October, is the cleanest example. Revenue rose 26% to a record US$64.6 billion. Advertising pricing was strong. Yet capital expenditure of US$34.9 billion, up about 80% on a year earlier, exceeded operating cash flow of US$32.8 billion. Free cash flow was negative US$2.1 billion, the first negative quarter in the recent run of reported results. Meta repurchased no Class A stock in the quarter, narrowed its 2026 capex guidance to US$135 to 145 billion, and announced a new non-voting share class that analysts widely read as a funding tool for the build. 

Alphabet told a similar story a week later. Google Cloud revenue rose 103% to US$30.7 billion, a remarkable number by any standard. But capital expenditure of US$51.4 billion overwhelmed operating cash flow of US$42.3 billion, leaving free cash flow negative US$9.1 billion for the quarter, the second negative quarter in a row. Alphabet lifted its 2026 capex guidance to US$195 to 205 billion and flagged a significant increase again in 2027. Amazon’s trailing twelve month free cash flow is now an outflow of US$42.4 billion, against an inflow of US$14.8 billion a year earlier. Oracle does not expect free cash flow to turn positive until the second half of fiscal 2028. 

When cash flow no longer covers the build, the gap has to be financed. Alphabet, Amazon, Meta and Oracle had issued more than US$228 billion of bonds this year as of mid October, according to LSEG data. Bilello’s broader tally of AI related bond issuance puts the year to date figure at about US$419 billion, around 90% higher than all of 2025. Trading in credit default swaps on AI linked companies rose from under US$2 billion a week in July to more than US$9 billion a week in September. People do not buy insurance in those quantities for decoration. 

CoreWeave has become the poster child. Its 9.625% notes due 2032 traded at yields of around 15% in mid October, and five year credit default swaps were quoted between 1,000 and 1,200 basis points, implying roughly a 60% probability of default within five years on market convention. This is a company with a contracted backlog of more than US$100 billion and a capacity backstop from Nvidia. The bond market is not disputing that the contracts exist. It is asking whether the cash arrives before the debt falls due. 

The quiet admission in ARK’s own newsletter 

Here is the part that most readers of ARK’s headline will miss. Buried at the end of the same newsletter is a separate note from ARK’s own cloud research team, titled “As Cloud Growth Accelerates, Hyperscalers Face A Cash Flow Squeeze”. 

Using Epoch AI’s analysis of company filings, it shows that in the first quarter of 2026 cash capital expenditure at Microsoft, Amazon, Alphabet, Meta and Oracle combined reached about US$148 billion, absorbing 94% of their operating cash flow, up from 65% a year earlier. ARK’s analysts describe this as a timing challenge, in which companies must fund infrastructure before collecting the revenue that justifies it, and they nominate cloud margins and cash generation as the measures to watch. 

That is, almost word for word, the bond market’s argument. The two ledgers are closer than their headlines suggest. ARK’s futurist is right that the prize is very large. ARK’s cloud analysts, and the credit market, are right that the bill comes first. The disagreement is not really about whether AI will generate revenue. It is about timing, and about who carries the risk if the timing slips. 

Not all balance sheets are equal 

Once the debate is framed as a timing problem, the investment question changes. It stops being “is AI real?” and becomes “who can afford to be early?” On that question, the AI complex divides into three quite different groups. 

The first group is the frontier labs themselves. Anthropic, on its own filing, is sitting on a very large cash cushion and is now generating operating profits from a revenue base growing at extraordinary rates. Its second quarter revenue of US$11.6 billion was up more than 1,300% on a year earlier. The catch is the size of its forward commitments. Estimates range from roughly US$265 billion of fixed and minimum commitments disclosed at 30 June to nearly US$900 billion on ARK’s reading of the amended filing, much of it non-cancellable and running out over many years. The filing itself notes that compute commitments generally must be paid regardless of revenue or usage. A company growing this quickly can carry such obligations comfortably, right up until it grows a little less quickly. 

The second group is the hyperscalers. These are the businesses investors have spent fifteen years valuing as capital light compounders that throw off enormous free cash flow. That description is now under strain. Microsoft’s cash from operations still covers its capital spending; Alphabet’s, Meta’s and Amazon’s currently do not. None of these companies is in any financial danger. Their balance sheets are formidable. But a business that consumes all its operating cash to stay competitive is a different business, with a different appropriate multiple, from one that returns it to shareholders. The market has not entirely made that adjustment. 

There is also a revenue side wrinkle worth noting. ARK’s own newsletter observes that Alphabet’s Search revenue growth slowed from 17% to 12% in the September quarter, and that queries in Google’s AI Mode are monetising at a lower rate than traditional search. AI is not only a new revenue line for the incumbents. In places it is quietly cannibalising the old one. 

The third group is the neoclouds and the debt funded builders. This is where the timing risk is most concentrated and least cushioned. Companies like CoreWeave borrowed heavily at high coupons to build capacity for customers who are, in many cases, themselves cash flow negative. If anything in that chain slips, whether a customer’s funding round, a utilisation assumption or the resale value of a GPU, the equity absorbs it first and quickly. That is what a 15% bond yield is telling you. 

Where the market may be mispricing AI infrastructure spending 

Equity markets have largely priced AI as a single trade. When the theme is in favour, everything rises together. When it falls from favour, as it did in the selloff ARK describes at the start of its note, everything falls together. That is rarely how capital cycles actually resolve. 

History suggests the more likely outcome is dispersion. In the fibre optic buildout of the late 1990s, the internet turned out to be every bit as transformative as its promoters claimed. The investors who lost money were not those who doubted the internet. They were those who financed the capacity with debt and assumed demand would arrive on schedule. The companies with the strongest balance sheets bought the distressed assets afterwards at a fraction of their cost. 

If that pattern repeats, the important variable for investors is not the size of the eventual AI market, which may well be as large as ARK suggests. It is the price paid for exposure, the cash flow profile of the business while it waits, and whether the balance sheet can survive the gap between the bill and the payment. 

Risks and counterpoints 

The bullish case deserves a fair hearing in the risks section too. Demand indicators remain strong. Cloud growth rates of 39% to 103% at the three largest providers are not the numbers of a fading theme. Older GPUs are still renting at higher prices than a year ago. Anthropic’s improving unit economics are real and documented in an audited filing. If ARK is even half right on the trajectory of frontier lab revenue, today’s credit spreads will look like an opportunity rather than a warning. 

On the other side, several risks are under appreciated. ARK’s revenue model rests on assumptions, including a ten times productivity gain and a 10% value capture rate, that are illustrative rather than measured, and its forecast relationships are fitted to very few data points. Anthropic’s first half net loss of about US$80.7 billion, although largely driven by non-cash items, is a reminder that the accounting is complex. Concentration risk runs through the whole system, with the same handful of labs, cloud providers and chip suppliers sitting on both sides of many contracts. And as I noted in a recent piece on choke points, overcapacity rarely announces itself in advance. It usually shows up first in falling utilisation and then in falling prices. 

The macro backdrop is also less forgiving than it was. With the US 10 year Treasury yield having traded above 5% for much of September and October, the cost of financing long duration projects has risen for everyone, not only for CoreWeave. 

What long term investors should actually do 

This is not a call to abandon AI exposure, nor to make heroic bets on which side of the debate is right. It is a reminder to look at the ledger the market is currently ignoring. 

For long term investors, the practical questions are straightforward. Does the business fund its AI ambitions from operating cash flow, or from the bond market? Is its revenue from AI incremental, or is some of it replacing existing, higher margin revenue? What multiple is being paid for a company whose free cash flow profile has changed? And if the timing slips by two years, does the balance sheet still work? 

The opportunity may sit with businesses that benefit from AI spending without having to finance it, and in high quality companies whose share prices have been dragged down by the theme’s periodic selloffs without any deterioration in their own cash generation. The risk sits with leveraged capacity builders priced as though the timing is certain. 

TAMIM Takeaway 

ARK and the bond market are looking at the same AI boom through two different ledgers. ARK measures the size of the prize, and on its arithmetic the prize is enormous. The credit market measures the size and timing of the bill, and right now the bill is arriving faster than the cash. Meta and Alphabet have both posted negative free cash flow quarters, Amazon’s trailing free cash flow is negative, hyperscaler bond issuance has surged and the most leveraged builders are being priced like distressed credits. 

Tellingly, ARK’s own cloud analysts concede the timing problem. The real debate is not whether AI will create value, but who can afford to wait for it. For long term investors, that shifts the focus from the theme to the balance sheet: favour businesses that can fund the journey from their own cash flows, be wary of those relying on the kindness of lenders, and remember that in every great buildout, the technology can be right and the financing can still be wrong. 

____________________________________________________________________________________

Disclosure: Alphabet (NASDAQ: GOOG) and Meta (NASDAQ: META) are held in TAMIM Portfolios as at date of article publication. Holdings can change substantially at any given time.