Cloud Computing

I came at this from the wrong end

I should be honest about how this thesis started, because it didn’t start with cloud computing.

It started with companies. I was interested in Microsoft, Amazon and Alphabet for unremarkable reasons: they’re established technology businesses with long track records of executing through several waves of change. If you’d asked me to explain what their cloud divisions actually did in any depth, I’d have given you a vague answer.

What changed my thinking wasn’t a valuation screen. It was understanding one thing about what these businesses sell, and realising it explained something I’d been missing about how the rest of my worldview works.

So this page has a narrower job than the three before it. T·01 is about intelligence becoming scalable, T·02 about intelligence acting physically, T·03 about compute becoming infrastructure. This one is about who gets to use that infrastructure, and who ends up owning it.

What actually clicked

A company can have an excellent idea and no ability to build the machinery underneath it.

Twenty years ago, a startup with a good idea had to buy servers, storage, networking, databases and security before writing anything a customer would ever see. Most of its early capital and talent went into building the floor it intended to stand on. Plenty of good ideas died there, not because they were wrong but because the entry cost was infrastructure rather than product.

Cloud removes that. You rent the floor, and point limited capital and talent at the thing you actually want to build. I think of it the way I think about premises: plenty of good businesses need an office without it making any sense to own the building. Needing something and owning it are different questions, and conflating them has probably killed more companies than bad ideas have.

This matters more now, because of what T·01 and T·03 describe. AI is expensive to run, and the compute it needs is exactly the kind of asset almost nobody can build for themselves. If the next twenty years belong partly to companies building on top of artificial intelligence, most of them will be renting that layer rather than constructing it. Cloud is how a small company gets access to infrastructure that would otherwise belong exclusively to the largest firms on earth.

The idea I had to stress-test

My instinct was that this creates an unusual investment position. If thousands of companies compete to build AI applications, most will fail, but successful and unsuccessful attempts both consume infrastructure on the way. Owning part of what they all need looks like a way of participating without having to identify which of them wins.

There’s something in that, but the research knocked a hole in it and I’d rather show it than hide it.

The problem isn’t the failures. Providers have credit controls, prepayment arrangements and the ability to switch off service, so the customers who die aren’t the real risk. The problem is the ones who succeed. A customer that grows large enough gains leverage and uses it. They negotiate. They spread workloads across a second provider to keep the first one honest. Eventually some build their own infrastructure, because at sufficient scale that starts making sense. The customers you most want are the ones most able to erode your economics.

So the honest version is narrower than my instinct. Infrastructure gets used by nearly everyone during the experimental phase, but owning it doesn’t guarantee capturing the economics of what gets built on top. That depends on whether the provider still has leverage once its customers are no longer small.

Renting doesn’t always win, and I shouldn’t pretend otherwise

The intuition that businesses will keep preferring to rent is mine, and it’s the part of this thesis where I was most likely to be lazy. So I went looking for the counter-evidence, and there’s plenty.

A late-2024 Barclays CIO survey found around 86% of chief information officers planned to move at least some workloads from public cloud back to private or on-premises infrastructure, and IDC has reported something similar. I’d rather not lean on the exact percentage, partly because the survey predates the AI surge and is now the oldest evidence on this page. The conclusion that holds is that the evidence points to selective workload repatriation rather than wholesale abandonment.

The sorting is what matters, and the sorting doesn’t depend on any survey. Renting tends to win when demand is uncertain, workloads are spiky, scaling quickly matters, or the infrastructure needed is specialised and hard to obtain. Owning tends to win when demand is predictable, utilisation is high, workloads are stable, and the scale justifies the capital. A company running the same load every day for five years is renting flexibility it no longer uses.

That’s not a problem for the thesis, just a correction to a lazy version of it. Cloud isn’t universally cheaper, and providers who implied otherwise were overselling. What cloud does is let each workload sit where it makes sense, and the growing, uncertain end of the economy is exactly the end that suits renting.

The detail that convinced me this runs deeper than I’d assumed came from Alphabet itself. On its second-quarter 2026 earnings call, the company said it would expand its use of third-party cloud capacity as a bridge to meet demand, and warned this would put modest pressure on its own cloud margin. Meta has separately signed large capacity agreements with specialised providers. These are companies that own more computing infrastructure than almost anyone alive, and they still rent when speed, flexibility or access to scarce capacity matters more than owning the asset. If renting makes sense for them, it isn’t a close call for a small company.

Access expands. Ownership concentrates.

That’s the whole page in four words, and it’s the part I find most interesting.

Access is expanding fast. Synergy Research put global cloud infrastructure spending at $143 billion in the second quarter of 2026, up more than $43 billion year on year, a growth rate of 43% and the eleventh consecutive quarter in which that rate increased, with the market doubling over that stretch. A founder in any city on earth can now rent compute that would have been unavailable to a large corporation twenty years ago.

Ownership hasn’t followed it outward. The same data puts AWS at 28% of that spending, Microsoft at 20% and Google at 15%. Roughly 63% between them, and the striking thing is how stable that share has been. It sat in the low sixties several years ago and has drifted slightly upward since, so if anything the three of them have been very gradually taking share while the market doubled underneath them. Position moves between them. Their combined position doesn’t.

That’s the shape of this layer. The market roughly doubles and the number of companies who own it doesn’t change. As an investor I’m on the ownership side of that, and I’d rather say so plainly than dress it up as democratisation.

But do the economics actually work?

This is where I stop reasoning and start looking at numbers, because a good story about infrastructure is worth nothing if the businesses underneath it don’t earn anything.

They currently do. In the second quarter of 2026, AWS generated $42.2 billion of revenue growing 37%, its fastest in eighteen quarters, with operating income of $16.6 billion and an operating margin around 39%, up from about 33% a year earlier. That single division produced roughly 60% of Amazon’s total operating income from about a fifth of its revenue. Google Cloud grew 82% to $24.8 billion, with operating income of $8.8 billion and a margin of 35.6%, up from 20.7%. Azure grew 43% and passed $100 billion of annual revenue for the first time.

I want to be careful about what that proves. Margin expansion can come from utilisation, product mix, scale, custom silicon, lower unit costs or depreciation assumptions, and it isn’t the same as pricing power. What it suggests is that these businesses are currently capturing attractive economics rather than behaving like undifferentiated infrastructure sold on price. A meaningful observation about a moment in time, not a guarantee about a decade from now.

The forward commitments point the same way. Google Cloud reported roughly $514 billion of backlog, up more than $50 billion in the quarter, with management expecting just over half to convert to revenue within two years. Microsoft’s commercial remaining performance obligations reached about $678 billion, up 84% year on year, though that covers its broader commercial business rather than cloud alone, so the two aren’t directly comparable. Both indicate customers making very large forward commitments to compute capacity rather than buying month to month.

These three aren’t the same business wearing different logos. AWS has the longest operating history, the broadest service catalogue and the deepest developer ecosystem. Microsoft’s advantage is distribution: it already sits inside most large enterprises through software relationships that predate cloud entirely. Google’s is technical, built on its own data infrastructure heritage, custom TPU silicon and model ecosystem. Cloud makes each more attractive to me, for genuinely different reasons, and those differences will matter more as growth normalises.

The test that matters to me

I don’t react badly to heavy capital spending. If the future requires infrastructure, a company willing to spend uncomfortably to build it may end up better positioned than one protecting this year’s margins. Companies that underinvest through a genuine technological shift tend not to get a second chance.

But spending has to be answerable to demand, and my test is ordinary supply and demand. Demand rises, capacity gets constrained, the company invests, capacity expands, usage and revenue follow. That sequence is rational however uncomfortable the numbers look. The sequence I don’t want is capital expenditure rising, capacity expanding, and utilisation and revenue failing to follow.

Through the second quarter of 2026 the first sequence is what the evidence shows, and it’s expensive. Alphabet spent $44.9 billion in a single quarter, roughly double a year earlier, and raised full-year 2026 guidance to $195 to $205 billion, up from $180 to $190 billion a quarter earlier. It reported negative free cash flow of about $5.9 billion for the quarter, its first negative quarter since going public in 2004, and management said plainly that free cash flow would remain under pressure because of infrastructure investment. Amazon’s trailing free cash flow swung to a $7.6 billion outflow, driven primarily by a $66.1 billion year-over-year increase in property and equipment purchases, which Amazon said primarily reflected AI investment.

So the cloud divisions generate mid-to-high-thirties operating margins while the parent companies consume cash building them. Both facts are true and need holding together. That’s not a contradiction, it’s what building ahead of demand looks like. Whether it’s wisdom or overreach depends entirely on whether the revenue arrives.

That’s where the real bear case lives, and it isn’t that cloud stops mattering. It’s the opposite. Cloud becomes enormously important, and the providers spend so much capital earning the right to provide it that the returns never justify the outlay. An industry can be indispensable and still be a poor place to have put money, which is the same trap I described in T·03.

Where I think the real risk sits

Not in demand. In depreciation.

These companies are spending enormous sums on hardware, and how long that hardware stays economically useful determines whether the reported profits mean what they appear to mean.

The mechanism states in a sentence. Assume an asset lasts longer and you spread its cost across more years, lowering annual depreciation and raising reported operating profit. Assume it lasts less time and the reverse happens. When a company is buying hundreds of billions of dollars of equipment exposed to fast technological change, that assumption stops being a technicality.

The assumptions have been moving, and not all in the same direction. Alphabet extended the useful life of certain servers from four years to six in 2023, and Microsoft made a comparable change in 2022. Oracle moved from five to six years in 2025 and Meta extended to around five and a half. Amazon went the other way in 2025, shortening the life of some server and networking equipment from six years to five, and said explicitly that it was because technology was developing faster, particularly in AI and machine learning.

I don’t read that divergence as one of them being wrong. These companies run different hardware, workloads, utilisation and refresh policies, and increasingly different custom silicon, so they can reasonably reach different answers. What it tells me is that the economic life of AI infrastructure is uncertain enough that the largest operators in the world are making materially different assumptions about it, and those assumptions feed straight into reported profitability.

Amazon’s disclosure is the one I find most informative, because shortening an asset’s life is an admission against your own reported earnings. That’s management saying obsolescence is arriving faster than they previously assumed.

I can’t resolve where the truth sits and won’t pretend to. But it’s the single most important variable I’d want to keep watching.

The specialised AI cloud providers

A group of specialised providers has emerged renting AI compute almost exclusively, and they’re growing quickly enough to take seriously. Synergy counted nine of them among the forty largest cloud providers in the world in the second quarter of 2026. Whether they end up as genuine competitors to the hyperscalers or as complements filling capacity the big three can’t supply fast enough is open, and both outcomes can be profitable. I’m not ideologically attached to the incumbents; if a smaller company develops genuinely better technology or a real competitive advantage, I’ll invest.

What makes me cautious is the financial structure rather than the opportunity. These businesses typically fund rapidly depreciating hardware with borrowed money, much of it maturing in a narrow window over the next few years. Customer concentration tends to be severe, and the largest customers are frequently the same hyperscalers who could eventually bring the work in-house. A company can be growing spectacularly and still be fragile if refinancing conditions turn.

That’s not a reason to avoid them. It’s a reason to start small. My approach with an unproven challenger is a modest initial position, then watch execution. If evidence accumulates and conviction rises, the position can grow. Buying the story at full size before execution is proven is how people lose money in exactly this kind of business.

What would make me wrong

The clearest failure mode is overbuilding. If capacity gets built for demand that doesn’t arrive on schedule, the industry ends up with expensive assets depreciating against weak utilisation, and margins that look structural turn out to have been a shortage. Related: AI compute prices could fall faster than usage grows. T·03 covers why I think falling prices have so far expanded this market rather than shrinking it, but that relationship isn’t guaranteed forever.

Then the competitive risks. Multi-cloud strategies and containerisation make workloads more portable, which erodes the lock-in supporting pricing. Large customers gain bargaining power as they grow. Capable local hardware and better small models could pull inference toward the edge. Data sovereignty rules are fragmenting what was designed as a global system. And power availability may cap how much useful capacity can be deployed where customers want it.

There’s also a plainer risk with nothing to do with technology. These are now enormous companies. Growth that impressive gets harder to sustain from a larger base, and the shares already price in a great deal going right. The technology can be exactly as important as I think while the returns disappoint, because I paid for a future already in the price.

How this actually sits in my portfolio

I don’t have a cloud allocation, and I won’t invent one for the sake of a tidy page.

Cloud is a lens rather than a slice. When I look at Microsoft, Amazon or Alphabet, I’m not buying a cloud position. I’m buying established technology companies, and the fact that each owns a dominant cloud business makes them substantially more attractive than they would otherwise be. It’s one of the reasons those companies earn a place, not a category I’m trying to fill.

That distinction matters because it changes what I’d do if the thesis weakened. I wouldn’t be selling a sector. I’d be revising my view of specific companies, which is what I’d be doing anyway.

What I want across all four of these pages is exposure to companies building the parts of the future that are hard to replace. Cloud is one of those parts, and it’s the one that determines who else gets to participate.