Six things investors may be getting wrong about AI

AI is changing so quickly that becoming wedded to any single view seems particularly dangerous. Over the past month, several developments have changed the way I think about the technology. Some have made me more bullish about the scale of the opportunity, while others have made me more conscious of the risks.

Most importantly, I think the shape of the AI economy is becoming a little clearer. The debate has moved beyond whether the models will work. They clearly do, and their capabilities continue to improve remarkably quickly. The harder questions now concern economics. What happens as intelligence becomes cheaper? Where does value accrue as models proliferate? And how much infrastructure will we ultimately need? 

The answers are still evolving, but over the past month six ideas have made the shape of that future a little clearer.

1. Intelligence may bifurcate rather than simply commoditise

I used to think increasingly capable open models would steadily commoditise intelligence. I now think that view is too simplistic. A more nuanced view is that intelligence splits into distinct tiers.

At one end, increasingly capable open models could perform enormous quantities of relatively routine work at very low cost. At the other, frontier models from OpenAI, Anthropic, Google and others may remain materially more capable and command premium economics for the hardest problems. Open models may ultimately process the most tokens, while frontier models continue to capture a disproportionate share of economic value. Vercel AI Gateway data already offers an early glimpse of that possibility, with open-weight models rising from 28.4% of token share in late June to 62% by 22 August.

The key point is not that frontier models are going away. Rather, token volume and economic value may increasingly diverge. That is intuitive when you consider how an autonomous AI system might work. A frontier model does not need to complete every step itself. It can decide what needs to be done, break the problem into many smaller tasks, assign routine work to cheaper models, and then review or synthesise the results. In that sense, the architecture begins to resemble an organisation: you do not use your most capable person for every mundane task; their value comes from directing and evaluating the work of others.

Abundant cheap intelligence may actually increase the value of exceptional intelligence.

This also changes how I think about open source. If a company moves a workload from an expensive frontier model to a much cheaper open model, some of the saving may simply represent model-layer margin disappearing. The underlying computation hasn't necessarily disappeared, and the money saved can instead be spent doing vastly more work.

That potentially creates an interesting economic structure: cheap models dominate volume, frontier models retain premium economics, and both consume compute. This is one reason I like the strategic position hyperscalers occupy. Amazon (AWS), Microsoft (Azure) and Alphabet (Google Cloud) don't necessarily need to correctly predict which model wins. If frontier intelligence wins, it requires compute. If open models proliferate, they require compute. If applications increasingly route between dozens of models, they still require compute. The hyperscalers are effectively supplying infrastructure to several competing versions of the future simultaneously.

2. Falling AI costs may be bullish, but we need to know what would prove that wrong

A common objection to the AI infrastructure boom is that AI is becoming far more efficient. Chips and models are improving, and inference costs are falling. Doesn’t that eventually mean we need less infrastructure?

This echoes the early cloud-computing debate. Compute and storage costs kept falling, making it easy to question how businesses selling declining unit prices could ever earn attractive returns. What mattered, however, wasn't simply the price of a gigabyte of storage or a unit of compute. It was how much more customers consumed once those resources became cheap. Falling storage costs meant companies could retain datasets that previously would have been uneconomic to keep. Falling compute costs enabled workloads that previously would never have existed. Lower prices didn't destroy the cloud opportunity. They expanded the market.

AI may follow the same pattern. This is essentially a version of Jevons paradox: improvements in efficiency can increase aggregate consumption by making a resource economical for vastly more uses. But investors shouldn't simply assume that happens. The number that matters isn't a token's cost. It is the inference cost of completing a useful unit of work.

Suppose AI becomes ten times more efficient, but lower costs and better capabilities drive a hundredfold increase in useful AI work. In that case, total compute demand rises sharply. But if efficiency improves one hundredfold while demand grows only tenfold, total compute requirements fall.

Alternatively, useful intelligence may not become cheap enough to unlock the many new applications investors expect. In the first case, efficiency overwhelms demand; in the second, high costs prevent that demand from emerging.

So far, though, the evidence doesn’t suggest either failure mode is playing out. None of this proves today’s infrastructure investment will earn attractive returns. But it offers something more useful than another AI forecast: indicators you can test. Sustained declines in GPU rental prices, easier GPU availability, weaker memory pricing, and slower token growth would all make me more concerned. A good investment thesis shouldn't just explain why you might be right. It should tell you what to watch to discover that you're wrong.

3. The AI stack is becoming much clearer

If intelligence is becoming abundant, the investment question is increasingly about what remains scarce.

At the bottom of the AI stack are the physical constraints required to produce intelligence. In the middle, models are becoming increasingly competitive, although frontier intelligence may retain premium economics. Applications and workflows turn that intelligence into something useful, while the hyperscalers occupy an unusual position across the stack, providing the infrastructure on which almost every version of this future depends

At the top of the stack is something investors may still underestimate: context and control.

Consider what an autonomous AI agent operating inside a large company actually requires. Intelligence isn't enough. It needs to know who you are. It needs access to the appropriate documents and databases. It needs to understand what information is confidential. It needs permission to send an email, modify a spreadsheet, access a customer record or approve a transaction. And somebody needs to determine what it is allowed to do.

This is one reason Microsoft's position remains so interesting. Microsoft doesn't necessarily need to own the world's smartest model. It already owns significant pieces of the environment in which enterprise intelligence operates: Office, Teams, Outlook, Excel, GitHub, Azure, Fabric, security and enterprise data.

As intelligence becomes more abundant, permission to let that intelligence act may become increasingly valuable. In that sense, the most important scarcity in enterprise AI may eventually have surprisingly little to do with the model itself.

4. As AI becomes more capable, control becomes more valuable

The rise of increasingly autonomous AI has also changed the way I think about cybersecurity. But I increasingly think cybersecurity is only part of the story. The deeper issue is control.

Traditional software largely executes instructions. An AI agent can be given an objective, reason about how to achieve it and take actions along the way. As those agents become more capable, the question is no longer simply what they can do, but what they should be allowed to do.

One recent example made this distinction tangible. During a cybersecurity evaluation, an OpenAI agent was placed inside a restricted testing environment and given an objective. The environment was designed to prevent it from accessing the open internet. Rather than accepting that constraint, the model searched for another path, found a vulnerability in infrastructure accessible from inside the sandbox, and escaped the intended environment while trying to complete its task.

The important point isn't that the model became malicious. It didn't. It pursued an objective and found a pathway its designers hadn't anticipated. The more capable the intelligence becomes, the more valuable control over it becomes.

This has implications well beyond conventional cybersecurity. Enterprises will need to determine what an agent can see, which systems it can access, what actions it can take, when it needs human approval and how its behaviour can be monitored and audited. Identity, permissions, least-privilege access, sandboxing and governance therefore become part of the infrastructure required to deploy autonomous intelligence safely.

And that brings us back to the AI stack. Intelligence may become increasingly abundant. Permission to act will not.

5. Capability is moving much faster than economic diffusion

This may be the most important caution from recent months. AI models have improved at extraordinary speed, yet the broader economy still looks much the same.

Earlier this year, investors worried about a so-called “SaaSpocalypse”, with software shares falling as markets considered a world in which AI could write code, automate knowledge work and disrupt large categories of existing applications. The technology behind those concerns was real, but the economic impact has so far been much slower.

Businesses have not stopped using software. Accountants, lawyers, consultants, programmers and designers have not disappeared. Enterprise workflows have not been rebuilt overnight.

Perhaps that shouldn't surprise us.

Markets can price technological capability immediately. Economies absorb technological capability gradually.

Electricity was transformative, but it took decades for factories to reorganise themselves around it. The internet existed long before every company became an internet company. Cloud computing existed for years before large enterprises migrated meaningful portions of their infrastructure.

AI may diffuse faster, but adoption still requires organisational change. Data must be cleaned, permissions set, processes redesigned and employees trained. Regulators need confidence, management teams need to see clear economic returns, and in many cases companies must become comfortable letting increasingly autonomous software act on their behalf.

This creates an important paradox. AI can simultaneously be more important than sceptics believe and slower to disrupt existing businesses than markets fear.

6. Intelligence may eventually become like electricity

I increasingly wonder whether we'll eventually stop talking about AI at all. Companies today don't generally discuss their “electricity strategy”. They don't calculate their return on electricity. Electricity is simply an input into almost everything they do.

That wasn't always the case. Electricity was once a revolutionary new technology whose potential was debated intensely. But over time it became cheaper, more accessible and embedded throughout the economy. Eventually, the interesting question was no longer whether a business should use electricity, but what that business could do because electricity existed. Intelligence may ultimately follow a similar path creating a virtuous cycle. 

Today we obsess over AI return on investment, AI revenue, AI workloads and AI strategies because machine intelligence remains novel, relatively expensive and highly visible. But imagine intelligence becoming extraordinarily cheap and available on demand. Rather than a separate product or strategy, it could simply be embedded throughout the economy, quietly augmenting the software we use, the decisions we make, and the work we do. Every business process could draw on whatever level of intelligence it requires, whenever it requires it.

At that point, asking for the “return on intelligence” may sound as strange as asking for the return on electricity.

Intelligence becoming a commodity doesn't mean intelligence becomes worthless. It means intelligence becomes an input into almost everything else.

And that changes the investment question. It becomes less about who owns intelligence and increasingly about who owns the scarce resources required to produce, distribute, control and apply abundant intelligence. 

There will still be winners and losers. Frontier intelligence may retain enormous value. Open models may process most of the world's tokens. Applications may own specialised workflows and proprietary context. Identity and cybersecurity may become increasingly important. And hyperscalers may occupy the unusual position of supplying infrastructure across several competing versions of the future.

But beneath all of those questions sits a larger one. If intelligence becomes as accessible as electricity, how much intelligence will the world choose to consume? 

We cannot know the answer today. But computing history offers an intriguing clue. When the price of a powerful input collapses, we tend not to consume the same amount for less. We find entirely new things to do with it.