AI Can Change the World and Still Disappoint Wall Street

Imagine opening a restaurant so popular that you need a second kitchen before the first month is over. Orders are pouring in. The food is good. Your accountant arrives, looks at the equipment invoices and asks when you intend to have any money.

You explain that you are transforming the future of dining.

Your accountant remains annoyingly interested in Thursday.

Something of that conversation ran through this week’s technology earnings. Investors questioned the cost of the AI build-out while companies reported rapid growth in the businesses selling the capacity. The market sell-off made an excellent backdrop for declaring the whole enterprise either doomed or misunderstood. Financial television does like a choice of two buttons.

I think a more interesting possibility deserves attention: AI could fulfil a great deal of its promise while delivering a thoroughly unpleasant experience for some of the people financing it.

That possibility has consequences far beyond a brokerage account.

Microsoft reported Azure revenue growth of 48% year on year for the quarter ending in September. On its earnings call, management said customer demand continued to exceed available supply.

Amazon reported AWS quarterly sales of $46 billion, up 39.3%. Across Amazon as a whole, operating cash flow reached $168.3 billion over the preceding twelve months. Yet free cash flow over that same period was negative $42.4 billion, with the company attributing the deterioration primarily to increased investment in property and equipment, chiefly for AI.

Those are different measures, scopes and periods. AWS growth is a quarterly business-segment figure; the cash-flow figures cover the entire company over a year. Negative free cash flow also does not mean Amazon made an equivalent accounting loss.

What the figures reveal is the scale of the wager. A business can attract more customers, generate more operating cash and still spend even more building for the future.

The queues outside the restaurant may be real. So are the invoices.

Cloud growth alone cannot establish how much profit the eventual users of AI are making, or whether every new data centre will earn an attractive return. It does, however, make the claim that demand has simply evaporated rather difficult to sustain.

The first clock measures capability: can the system complete a useful task reliably, at an acceptable cost?

The second measures adoption. Can an organisation connect it to the right information, give it appropriate authority, change its processes and persuade people to use it?

The third measures financial return. Who gets paid, how much capital did they commit, and how long must they wait?

A model can advance before a business has worked out how to use it. A business can become more productive before its supplier recovers the cost of the infrastructure. A supplier can eventually become profitable after an early investor has lost patience, money or both.

Share prices add another complication: expectations. A company may improve substantially and still disappoint investors who paid for something better.

The mistake is expecting all three clocks to strike together. Technology launches encourage that illusion. There is a demonstration, applause and a handsome slide suggesting that the remaining journey consists largely of updating the logo.

Then somebody asks whether it works with the accounts system.

This is where the argument becomes useful to people running ordinary businesses.

In his recent discussion of small-business agents, Farzad proposes a sensible test: how much time does the owner recover after setup, review and correction? He presents a way to assess the opportunity, rather than proof that a particular product already delivers it.

That distinction matters. An agent producing a proposal in thirty seconds is impressive. Whether the proposal contains the right prices, respects the delivery commitments and reaches the correct person determines whether it has helped.

Consider a hypothetical supplier whose quotations usually take two hours. An agent prepares one in five minutes. A person then spends twenty minutes checking it. That could be a substantial improvement.

If a mistaken specification subsequently consumes three hours of repair work, the stopwatch at the demonstration missed the expensive part.

Measure the complete job. Include checking, exceptions and the work created downstream. Otherwise the productivity gain can consist of moving a problem into someone else’s afternoon.

This is also where I remain optimistic. Useful improvement does not require a machine to run the entire company. Removing one recurring delay, reliably, can release capacity that a business has been unable to free for years. A modest task completed every day can be worth more than an astonishing task completed once for a camera.

Suppose AI really does make a service much cheaper to provide.

Initially, a business might retain the saving as additional profit. Competitors may then adopt similar tools. Customers may demand lower prices or faster delivery. Some of the benefit can pass to the buyer, even while the technology becomes more deeply embedded in the industry.

That would be successful technological diffusion with awkward consequences for anyone who assumed the savings belonged to them permanently.

The same question reaches the infrastructure layer. Strong usage is encouraging. Attractive returns also depend on pricing, utilisation, operating costs, financing and how long the equipment remains economically useful.

A system can be enormously valuable to its users without allowing every supplier to charge enormously valuable prices.

For business owners, this opens a more practical line of enquiry. Where could cheaper capability improve your service? Which parts of your advantage would remain if your competitors gained access to it too?

The answers might involve customer trust, specialist knowledge, distribution or an unusually well-run process. A subscription available to everyone requires some work before it becomes your competitive advantage.

The strongest objection to this argument is that it could become a convenient shelter for every failed investment: the technology is wonderful; the returns are merely arriving later.

That excuse deserves a short leash.

Persistent dependence on expensive supervision, weak customer retention, deteriorating economics or capacity that cannot earn its keep would weaken the case. A forecast cannot indefinitely substitute for a customer choosing to pay again.

Equally, evidence of reliable work, repeat use and improving economics should change a sceptic’s mind. Otherwise scepticism becomes another fan club, with less cheerful merchandise.

My reading of this week is that real demand and substantial financial risk are coexisting. The results justify investigating where the value is becoming durable. They do not justify treating the entire industry as one bet.

A falling share price cannot tell you that a technology is useless. A spectacular demonstration cannot tell you that its economics work.

For an investor, the challenge is understanding which business can retain enough value to justify its costs and valuation. For an operator, it is finding useful work that the technology can complete more reliably or economically, with the full cost counted.

Those decisions can lead in different directions. You can have reservations about the price of an AI company and still find an excellent use for its products on Monday morning.

I would keep the ambition, then inspect the receipts.

The future may be extraordinary. It still has to settle the bill.