Artificial Intelligence

Overhyped today. Possibly underestimated tomorrow.

Both of those can be true at once, and most arguments about AI happen because people insist on choosing one.

The mental model I keep returning to is the internet around 1997. By then nobody serious doubted it mattered. But if you looked at what the internet actually was for most people, dial-up connections, basic websites and email, you had almost no way to see cloud computing, smartphones, streaming, or the advertising businesses that would eventually fund half the web. The technology was visible. The applications weren’t.

I’m not claiming AI sits at the same point on the curve. I have no way of knowing that. What I take from the comparison is narrower: it is genuinely difficult to price a foundational technology while the things built on top of it don’t exist yet. So when someone argues AI is expensive based on what it does today, I don’t dismiss them. They may be right about today and still be measuring the wrong thing.

The obvious problem with that analogy

If I’m going to use 1997, I have to accept what came next.

A large number of the companies that most obviously led the early internet did not end up capturing the value it created. Some faded, some were acquired, and some are still substantial businesses that never came close to justifying what investors paid at the peak. Several of the businesses that eventually captured the most value were either tiny at the time or didn’t exist yet. Being early and being obvious turned out to be very different things from being the eventual winner.

That history is the strongest argument against the way I invest, because it says identifying the trend correctly tells you remarkably little about which companies to own. I don’t think the answer is to pretend I can pick the winner now. It’s to accept that I can’t, and build around it.

So I select companies that, on the information available today, appear to have the highest probability of remaining competitive as this develops. That is a probabilistic judgement, not a prediction. It also means today’s leaders are not permanent holdings simply because they lead today. I reassess as evidence arrives, and if stronger businesses emerge, I’m willing to move capital toward them. The internet era is a strong argument for adaptability. It isn’t an argument for certainty in either direction.

Why I take the direction seriously

As of the second-quarter 2026 reporting round, Amazon, Alphabet, Microsoft and Meta had guided to somewhere between roughly $710 and $745 billion of combined capital expenditure for 2026. Every one of them revised that figure upward during the year. Alphabet alone moved its range three times, ending at $195 to $205 billion.

The revisions matter more to me than the level. A single large number could be one budgeting cycle. Four companies repeatedly raising the number, in public, against visible investor discomfort, is four of the most commercially disciplined organisations on earth restructuring their balance sheets around one technological assumption.

This doesn’t prove they’re right. Companies overinvest, and whole industries have destroyed capital while being technically correct about the future. What the number tells me is how seriously the people closest to the technology are treating the transition, with far better information than I have.

I’d also generally rather own a business willing to spend uncomfortably to stay relevant than one defending this year’s margins while the ground shifts underneath it. I don’t mind investment hurting short-term profitability, provided I can see the underlying business moving in the right direction.

What the evidence actually shows

The economic case rests on productivity, and this is where I try hardest to stay honest.

I’m not going to tell you workers will become five or ten times more productive. Nobody knows that. What I can point to is measured evidence from workplaces already using these tools. A study of more than 5,000 customer-support workers found that access to a generative AI assistant raised productivity by roughly 15% on average, with the largest gains going to the least experienced staff. A separate workplace experiment across more than 7,000 knowledge workers found frequent AI users spent around 3.6 fewer hours per week on email.

Neither number is spectacular, and that’s rather the point. Both studies were run on an earlier generation of these tools than the ones available now. I don’t know where the ceiling sits. But a measurable effect showing up this early is a different starting position from a technology that only works in demonstrations.

Is AI actually becoming a business yet?

Capital expenditure tells me what four companies believe. It doesn’t tell me whether anyone is paying. So the question I actually want answered is duller than the one the industry likes to debate: are customers spending money on this, and are they spending it repeatedly?

On adoption, the most useful source I’ve found is the US Census Bureau, because it surveys around 1.2 million businesses rather than asking executives at a conference. Its Business Trends and Outlook Survey put national AI use at 19.8% of American businesses in the collection period ending 3 May 2026. In the AI supplement covering November 2025 to January 2026, 18% of firms reported using AI in a business function, rising to 32% when weighted by employment.

That gap between 18% and 32% is the finding, not the headline rate. Adoption is concentrated in large firms. Around 37% of businesses with at least 250 employees reported using AI, against under 20% of firms with four or fewer employees. Between December 2025 and May 2026 use rose among firms with 20 or more employees and didn’t change significantly among the smallest. Information was at 39.7% and finance and insurance at 33.9%, both well above the national rate.

So this isn’t universal. It’s roughly a fifth of businesses, weighted heavily toward the large and the digital. What matters to me is that it’s measured, it’s growing, and it’s growing in the places with the budget to pay.

On monetisation I’d rather use audited disclosure than private valuations. In the quarter ending June 2026 Microsoft reported that Azure had passed $100 billion of annual revenue for the first time, growing 43%, and that Microsoft 365 Copilot had surpassed 30 million paid seats. Its commercial remaining performance obligations reached about $678 billion, up 84% year on year. Google Cloud grew 82% to $24.8 billion in the same quarter, with roughly $514 billion of backlog. Amazon said its AI and chips businesses had each passed run rates of more than $25 billion. Those are paid seats and contracted commitments, not usage of a free tier.

The private model companies are growing faster than that and I trust the numbers less. Anthropic’s annualised revenue run rate was reported above $65 billion in late July 2026, up from roughly $9 billion at the end of 2025. OpenAI’s was reported around $40 billion. Both figures come from investor updates and press reporting rather than audited filings, and the two companies appear to calculate revenue differently, so I wouldn’t compare them directly or treat either as settled.

Now the part that keeps me honest. Money arriving at the vendors is not the same as value arriving at the buyers. MIT’s NANDA work found that the overwhelming majority of enterprise generative AI pilots delivered no measurable profit-and-loss impact. Consultancy surveys through 2026 put the share of companies generating meaningful financial value anywhere from 5% to around a quarter, and they disagree with each other badly enough that I treat all of them as directional at best.

That combination is the accurate picture. Real revenue, real contracted commitments, real adoption concentrated among large employers, and a customer base that mostly cannot yet demonstrate what it got. Those two things can stay true together for a while. They can’t stay true together forever, and which way they resolve is what the rest of this page is about.

Where the economics actually land

This is the question I care most about.

There are three separate things here and they get collapsed into one constantly. A technology can succeed. That technology can create enormous economic value. And the shareholders of the companies involved can still do badly. Each step has to happen for an investment to work, and none of them guarantees the next.

Value gets competed away all the time. Models may commoditise. Competition may push prices toward cost. Customers may keep most of the productivity benefit rather than paying it to a vendor. Infrastructure providers may spend extraordinary sums and earn poor returns on them. AI could be one of the most significant technologies of this century and still leave large parts of the investment chain producing mediocre returns.

So my job isn’t to work out that AI will be important. That’s close to consensus and there’s no edge in it. My job is to form a view on where the economics are likely to settle, and keep updating it as evidence arrives. Value could accumulate in semiconductors and compute, infrastructure and cloud, the models themselves, distribution, applications, or eventually physical AI. Each layer has a different competitive structure, the profit won’t spread evenly across them, and I don’t claim to know the final answer yet.

How I decide what to own

Because I can’t identify the winner, I manage that uncertainty through portfolio construction rather than through conviction in a single name.

Established leaders get greater consideration because they’ve already demonstrated the things that are hard to demonstrate: execution, capital, talent, customers, distribution, and the financial strength to keep investing while the picture is still forming. That’s not an argument that large means safe. It’s an argument about demonstrated capability and staying power.

Earlier-stage companies can offer genuinely asymmetric upside when the technology, the management and the vision line up with where I think things are heading. They also carry far greater uncertainty, so I treat them differently within the portfolio. I accept that several of them may fail, provided one exceptional outcome can contribute meaningfully to overall returns.

I don’t need every position to win. I need the portfolio to win.

Across both, what I weigh is the size of the future opportunity, competitive position, management’s grasp of where technology is going, execution, adaptability, the ability to fund continued investment, and whether that investment is translating into business results rather than announcements.

Price matters. False precision doesn’t.

I’ve never argued price is irrelevant.

What I don’t do is make traditional valuation metrics my primary decision tool when the eventual products and markets may not exist yet. A discounted cash flow model can be precise to two decimal places while resting entirely on assumptions about a world nobody can describe. The further a case depends on markets that don’t yet exist, the less weight I place on long-range models built on guesses about them. Precision isn’t the same as accuracy.

So I’m willing to own expensive businesses when I believe the future opportunity, competitive position and execution justify it. If an expensive multiple is a hard disqualifier, there’s a real risk of spending an entire technological cycle waiting for the strongest companies in the world to become cheap, and sometimes they don’t.

That doesn’t mean price stops mattering once I own something. It means my discipline sits somewhere other than the buy decision. I can’t point to a company I wanted to own and refused purely because it looked expensive, because that isn’t how I operate. What I do instead is manage it through position size, through trimming when a holding runs hard, and through rotation when another opportunity offers better prospects from here. Every position competes with every other use of the same capital, and that competition, rather than a valuation threshold, is what keeps me honest.

What if the boom is real but the economics aren’t?

This is the version of the bear case I take most seriously, and it’s much harder to answer than “AI might fail.”

Assume AI works. Usage explodes, productivity gains are real, demand for compute keeps climbing. Shareholders can still do badly, because providing AI is extremely capital-intensive and competition doesn’t care how important your technology is. Prices fall as capability spreads. Assets depreciate on short cycles. Energy costs are real and rising. Capacity gets built ahead of demand and then competes on price. The boom is genuine and the returns are still disappointing.

Technology has done this before. The infrastructure got built, it changed the world, and the capital that funded it wasn’t the capital that got rewarded.

So what I’m watching, quarter by quarter, is fairly ordinary. Are revenue and demand still growing. Is management delivering what they said they would deliver. Does the money going out appear to be strengthening the company’s position rather than just defending it. Over a longer horizon, the test is whether all this investment converts into sustainable revenue, real cash generation, attractive economics and durable competitive advantage. If capital expenditure keeps climbing while those don’t follow, that isn’t a detail. That’s the thesis failing on its most important axis, and I’d rather notice early than defend a story.

When intelligence leaves the screen

The longer-term opportunity becomes considerably more interesting if intelligence stops living inside a computer and starts doing useful physical work.

If digital AI becomes extraordinarily capable but physical AI stays expensive, unreliable and economically inferior to human labour, then a meaningful part of my long-term outlook simply hasn’t materialised. That wouldn’t invalidate everything above it, but it would cap it.

That argument deserves its own page. See T·02 Robotics.

Long the trend, flexible on the company

If a position has an exceptional run and grows into an outsized share of the portfolio, I may trim part of it. That isn’t a verdict on the company. It’s usually concentration management, or a judgement that another opportunity offers better prospective risk and reward from current prices.

The distinction matters, because the whole point of owning exceptional businesses is that a small number of them produce a disproportionate share of long-term returns. Selling a compounder because it went up is one of the more reliable ways to miss the outcome the thesis depends on. Trimming concentration is a different decision from abandoning the position, and I try not to confuse the two.

My view on the technology can hold for decades while the portfolio behind it keeps changing.

Every dollar competes

None of this is an argument that AI is the future and therefore you should buy AI. It’s narrower. I believe this is one of the most significant technological transitions currently underway, which makes it a serious competitor for my capital rather than an automatic winner of it. Every dollar can only be invested once, so every position I hold is a decision not to own something else.

I don’t need to predict the destination to invest in the direction. I need to keep checking the evidence, and stay willing to change my mind about the companies while holding on to the trend.