What AI investors can learn from a decades-old cellphone

On July 9, Mark Zuckerberg told Bloomberg that Meta needs all the compute power it can get.

In nearly the same breath, he said the company was exploring renting some of that power out to others, because the offers coming in are so high that leasing may make more sense than using the capacity internally.

A company racing to assemble one of the largest private compute portfolios in history is shopping the surplus.

That is what disciplined builders do when construction runs ahead of confirmed internal demand. They go looking for someone else to absorb the capacity.

What the instinct exposes is the assumption the whole artificial intelligence industry rests on: that the demand to justify this buildout is coming, at the prices the builders need, and soon.

In the first half of 2026, buyers began to challenge that assumption. For investors, the question is whether their portfolios are exposed to a capacity bet dressed up as a growth story.

Opinion: Why the bursting of the AI bubble would be a much more worrisome event than many people think

Start with the math. Amazon, Google, Meta and Microsoft are on track to spend nearly US$700-billion on AI infrastructure this year, according to Bloomberg and CNBC, double last year’s spend. That dwarfs the roughly US$75-billion the AI industry is earning from end users, leaving revenue at less than a tenth of capital investment.

In February, Anthropic chief executive Dario Amodei warned of the extreme financial risk involved in locking in massive infrastructure costs years in advance. He noted that if his company bet on investing roughly US$1-trillion in compute capacity but its annual revenue growth was cut in half, “there’s no force on Earth, there’s no hedge on Earth” that could prevent bankruptcy.

Pre-committing to such an enormous spend leaves zero margin for error. If the projected exponential revenue doesn’t arrive on schedule, the fixed costs would instantly outstrip the input with no way to recover.

Now consider the metrics the industry celebrates: capital deployed, gigawatts under construction, chips shipped, tokens consumed. All of these are supply-side measures. They say how much has been built and how heavily it is used. But none are saying whether that use clears a profit.

The evidence that it may not is accumulating. Uber gave about 5,000 engineers access to agentic coding tools in December, 2025, and had exhausted its 2026 AI budget by April. Its executives said rising token spend (the volume of text units processed by a large language model) was hard to tie to measurable productivity gains.

Opinion: Canada must step up to tackle AI’s catastrophic risks

In June, the Financial Times reported that Amazon, Walmart, Cisco, Meta and Uber were all reining in internal AI use, capping budgets or pushing staff toward cheaper models.

Call it the token-cost paradox. More adoption does not mean more revenue or profit for builders. The industry assumed usage was value; but usage is just usage. When the meter runs and the output cannot be tied to a desired result, sophisticated buyers cut spending. And when they cannot cut, they substitute.

The substitution is now measurable. Reuters, citing Citi, reported that leading Chinese open-weight (rather than proprietary) AI models cost as little as 18 US cents per million tokens, versus about US$4 for top Western models, while the capability gap has narrowed dramatically.

In June, CNBC reported that Lindy shifted all of its AI activity to the Chinese DeepSeek model, with founder Flo Crivello saying the move saved millions and improved performance. CNBC also reported that Chinese models have held above 30 per cent of U.S. companies’ token usage on OpenRouter – a platform that provides access to hundreds of AI models – every week since Feb. 8, versus 4.5 per cent in the first half of 2025.

Even though AI usage is exploding, we are seeing a collapse in the assumption that exploding usage flows to premium-priced compute.

The bulls would counter, correctly, that building ahead of demand is how platform shifts are won – nobody laid fibre after the traffic arrived.

Opinion: The AI boom isn’t a replay of the dot-com bust

But the analogy cuts both ways: Telecom operators in the late 1990s measured progress in route miles, the physical length of the fibre-optic cable, instead of paying traffic. Global Crossing, the biggest spender, collapsed into one of the era’s largest bankruptcies. The capacity was real. The demand at the price builders needed was not.

Motorola made a costlier version of the same error. In 1987, after an engineer’s wife could not get a signal on a remote Caribbean beach, the company planned a ring of 77 low-orbit satellites that would let a handset place a call from anywhere on Earth and named the project Iridium after the 77th element.

It took roughly US$5-billion and a decade to build. Almost nobody bought it.

The handset was a US$3,000 brick, airtime ran $3 to $7 a minute, and the phone needed a clear view of the sky, so it failed indoors and in cars. Ordinary cellular had spread everywhere in the meantime: cheaper, smaller, good enough.

Iridium, the Motorola subsidiary, forecast 500,000 subscribers in its first year but only had 20,000 by mid-1999, when it filed for bankruptcy.

And while Motorola poured capital and executive attention into the moonshot, it clung to profitable analog handsets as a Finnish upstart went all-in on digital. In 1998, the year Iridium launched, Nokia passed Motorola to become the world’s largest handset maker. Motorola never got the lead back.

Prosperity’s Path: AI data centres are the future. Canada must overcome the backlash

For investors, today’s practical work is less dramatic than picking a side in the AI debate.

Start with mapping exposure honestly. The bet on premium-priced compute now runs past the chipmakers into utilities, industrial suppliers, real estate and the private credit funds financing data centres, and the question for each holding is not whether AI will matter but what price of compute it needs to justify those valuations.

Then watch demand-side signals rather than supply-side headlines, since a capital expenditure announcement measures the builder’s enthusiasm, not the customer’s willingness to pay. Keep an eye on token pricing, substitution rates and enterprise renewal behaviour. The question worth asking before the next dollar goes in is: What evidence would show the demand assumption is wrong, and who is watching for it?

Read charitably, renting out the surplus is Mr. Zuckerberg’s most disciplined move. It turns a fixed bet fungible while the demand curve is still being argued about.

It is also an admission, not unlike one Toronto made a decade ago. The Union Pearson Express opened with fares at $27.50 a ride, riders balked, and the fare was cut by more than half within a year.

This loss was absorbed by taxpayers, not shareholders. AI’s builders have no such backstop.

The builders may yet be right that the demand is coming. But the question the meter at Uber has forced onto every investor is the one Motorola never asked in time, and that Toronto commuters answered before them: at what price?

As the late risk historian Peter Bernstein observed, the riskiest moment is when you are right.

Sam Sivarajan is a speaker, independent consultant and author of three books on investing and decision-making. He writes two free Substack publications on decision-making and the good life: theuncertaintyedge.com and thegoodhumanpractice.com