AI Works. The Returns Might Not.

I started out trying to understand whether the AI investment boom could possibly generate an adequate return. I ended up somewhere rather different.

Amongst all the doomer noise, LinkedIn talking heads and finance podcasts, there is a wonderfully inconvenient possibility hiding. What if the AI optimists are right? Not in the "Copilot saved me twelve minutes writing an email" kind of right, but really right.

I am living proof of the extraordinary productivity AI promises. I've done the math on my own output. In the last six months I have produced somewhere between four and six years' worth of work. If others have my experience, then the cost of producing software will collapse, organisations will automate enormous amounts of work, intelligence becomes cheap and businesses genuinely become much more productive.

Fantastic news for the companies currently spending hundreds of billions of dollars building AI infrastructure, right?

I'm increasingly unsure that it is.

Why? There is a critically important distinction between creating economic value and capturing economic value.

AI is exceptionally good at the first while making the second considerably harder, and the implications are profound.

Let me explain.

I started with software

The thing that originally sent me down this particular rabbit hole was software development.

For most of my career, enterprise software has been expensive. If you wanted an ERP system, CRM, billing platform or some other serious workload, you essentially had two choices: buy one or build one. Building was expensive, slow and risky, so most organisations sensibly bought, and that simple economic reality created some of the most valuable companies on Earth. Build once, sell thousands of times, charge recurring licence fees. Awesome margins and great business if you can get it, right?

Except AI has started doing something rather awkward to the denominator in that equation. It's made software extraordinarily cheap to create. I have personally built entire ERPs in a fraction of the time it would have previously taken to do so, backed up by a control plane inspired by my prior life in the Microsoft low code ecosystem.

So what happens when building gets cheap enough that buying stops being the obvious answer? The productivity benefit doesn't disappear.

The money just stops going to the software vendor.

Now consider Microsoft

Microsoft is particularly interesting because it sits on both sides of this equation.

On one side it has some magnificent software franchises: Dynamics. Power Platform. Microsoft 365. SharePoint and the broader ecosystem surrounding all of them. I should know. I spent the majority of my career here, building and selling a company off the back of it.

On the other side sits Azure and an enormous investment in AI infrastructure. Along with it comes the straightforward bull case that justifies US$145 billion of capital expenditure in the last financial year alone, and counting.

Okay, but there is a teeny tiny problem with this thesis: what if the AI running in Azure helps Microsoft's customers build the things they previously bought from Microsoft?

Now Azure is not simply creating a new revenue stream, it is enabling the cannibalisation of an existing one. None of which will be news inside Microsoft. From my own interactions with senior people there, this tension is well understood. Nobody is naive about it.

It isn't a blind spot, it's a trade they have already made: cannibalise yourself in the hope of owning the layer underneath as a new moat, because the alternative is standing still while somebody else does the cannibalising for you. That is the classic innovator's dilemma with several billion dollars' worth of GPUs attached to it.

What I am less certain about is whether the trade works this time. The textbook version has the incumbent moving down the stack: give up the expensive product and sell the cheap thing underneath it at scale. Microsoft has done it before, and in this version, the playbook is trading licence revenue for the compute those systems run on.

But that assumes the cheap thing underneath still has to be bought from you. With compute, the layer you retreated to isn't just lower margin than the one you gave up, it is one almost anybody with a datacentre can sell.

Their moat is not our job

Don't get me wrong: Microsoft has moats. Identity is important, security is important, data infrastructure is important, and I'm not disputing any of them.

Those moats were never built on the difficulty of writing the software anyway. They rest on regulatory approval, contractual inertia and the sheer cost of switching, and none of that gets cheaper because a model can generate code.

I left an important one out though, one that Microsoft has been pitching hard: orchestration.

Orchestration is coordination logic: routing a request to the right model, holding context, retrying, evaluating, falling back. Genuinely useful, and precisely the kind of software AI has made cheap to write. A moat has to be hard for your customer to build, and orchestration gets easier every month.

Most of us are answerable to an organisation that is a consumer of services, and our remit is to deliver what it can afford, at a level of risk it can live with, without handing any single supplier the power to set our price later. We wouldn't run every database query on the world's most expensive computing infrastructure simply because it was the best. We match the resource to the workload, and we keep an alternative warm. This is not a new problem, in fact it's something that well-governed IT departments are pretty good at.

Apply that to AI. Why would we use the most expensive model available for every problem? Some inference happens locally on developer workstations, some goes to inexpensive external models, and some to open-weight models running on infrastructure we control. When we genuinely need frontier capability, we call a premium model: OpenAI, Anthropic, a Chinese frontier model, whoever happens to be best at that particular task at that particular time. The orchestration layer chooses, and vendor lock-in stops being something that happens to us.

An organisation could increase its AI usage fiftyfold while reducing its expenditure on premium frontier models.

Once we look at AI this way, something important changes.

Token consumption stops being the interesting number.

So what should we actually measure?

An organisation increasing its AI usage fiftyfold while spending less on frontier models is an uncomfortable possibility for anyone valuing model companies on the assumption that exploding token consumption necessarily means exploding frontier-model revenue. The assumption simply does not hold.

This line of enquiry led me to a metric I think is considerably more useful. Forget total tokens and divide organisational inference into three buckets:

• Internal inferenceModels running on infrastructure the organisation controls. A developer's workstation, a GPU or open-weight model in the organisation's data centre. The marginal cost is electricity and amortised hardware. Nobody sends an invoice.
• Commodity external inferencePaid API calls where the model is essentially interchangeable and the decision is made on price. Classification, extraction, summarisation, routing, and the enormous volume of unglamorous work most organisations actually have. Switching provider is a configuration change.
• Premium frontier inferenceCalls made specifically because the best available model is meaningfully better at that task, and the organisation is willing to pay a multiple to get it. This is the bucket that pays for the training runs.

Then watch the rate of change of the proportions. Call it frontier intensity - premium frontier inference as a percentage of total organisational inference. Which way that number moves over the next year or two tells us something much more interesting about the eventual economics of AI than total consumption ever will.

If premium frontier inference remains dominant, then the economics currently implied by some AI valuations become much easier to defend.

But suppose organisations become increasingly competent at AI.

• They invest in local infrastructure to avoid cost, lock-in and data sovereignty concerns.
• Open models improve, Chinese models create price competition, and orchestration matures.
• CIOs work out which workloads actually require expensive intelligence and which do not.

Premium frontier inference might eventually represent 5%, 10% or 20% of an organisation's total AI workload, while total AI consumption could have increased by orders of magnitude.

If this scenario plays out, that would be an astonishing technological success, yet a poor business outcome for some frontier model providers.

Which brings us to the circular financing

There is something undeniably circular about the current AI investment boom. Hyperscalers invest in model companies and model companies use enormous quantities of hyperscaler compute. Chip companies invest in AI companies. AI companies raise more money and buy infrastructure containing more chips.

Around and around we go.

Now, circularity by itself doesn't prove anything. Infrastructure has always required capital before the economic activity enabled by that infrastructure fully materialises, so I don't find the "look, the money is circular!" argument particularly compelling on its own. The much more interesting question is this:

Sooner or later somebody outside the circle has to pay for all of it. Who?

And that is where our software and orchestration problems come back, because AI isn't merely creating new things.

It is attacking existing profit pools. It has to.

AI-generated software competes with SaaS. Open models compete with proprietary models. Chinese models compete with American models. Local inference competes with cloud inference. Hyperscalers compete with each other.

Everyone increasingly orchestrates between all of them, simultaneously creating demand for AI while trying to reduce what they pay everybody else for it. It is the obvious thing to do as well when you think about it.

The pie might not be getting bigger in the way we think

This is where I have ended up.

I think AI is going to create enormous economic value. The productivity improvements I am seeing make it increasingly difficult for me to believe otherwise.

But I am not convinced that this necessarily translates into an equivalently enormous technology-sector profit pool.

Those are different statements.

And increasingly I suspect the somebody in question is the organisations adopting AI rather than the organisations supplying it.

• A manufacturer reduces administrative overhead.
• A professional services firm increases the amount of work each person can perform.
• A small company builds technology that previously required an enterprise software budget.
• A budget-constrained NGO now has several smaller partners to choose from for delivery of a bespoke system.
• A mining company optimises operations.
• A bank builds operational capability it would previously have licensed.
• Two small technology partners with deep domain expertise collaborate to offer an organisation a service that once required an enterprise vendor.

The productivity gain appears in those businesses. It might appear as higher margins, lower prices or competitive survival because everybody else has access to the same technology.

But it doesn't necessarily appear as another dollar of revenue for Microsoft, OpenAI, Anthropic or the long tail of hopefuls banking on compute.

Which gives us a rather strange conclusion

There is a bearish argument about AI that says the infrastructure investment cannot be justified because AI isn't creating enough economic value.

I am increasingly interested in almost the opposite argument.

What if AI creates so much value, and becomes so cheap and competitive, that nobody can capture enough of it?

AI works spectacularly well. Models could become dramatically better and token consumption could explode. Software creation becomes absurdly cheap. Businesses could become significantly more productive.

GDP could benefit, consumers could benefit, and yet some of today's AI investments could still produce disappointing returns.

There is no contradiction in that. In fact, the more AI starts behaving like a utility, the more plausible it becomes.

So perhaps the question we should be asking isn't whether AI has an ROI problem. Perhaps we should ask something slightly more uncomfortable:

If AI really does become abundant intelligence, why do we assume the companies producing that abundance will be the ones that capture most of its value?

I don't think we know the answer yet, but that's the question I'm watching.