If You Build It, They Will Come ... or Will They?
Meta, Iridium, and the metric the AI build-out forgot to watch
On July 9, 2026, Mark Zuckerberg said something that should have made more headlines. Meta, he told Bloomberg, needs all the computing power it can get. And then, in almost the same breath, he said the company was exploring renting some of that computing power out to other companies; because the offers it receives to use its infrastructure are, in his words, “so high that it may make sense” to lease the capacity rather than use it internally.
Reading those two statements together, makes one scratch their head. A company racing to build one of the largest private ‘compute’ portfolios in history—a company whose CEO says it cannot get enough of the stuff—is at the same time shopping the surplus. That is not a contradiction. It is a tell.
Meta is doing what disciplined builders eventually do when the build runs ahead of confirmed internal demand: it goes looking for someone else to absorb the capacity. Nothing is wrong with that instinct. But it reveals the assumption the whole industry stands on—that the demand to justify this build-out is coming, at the prices builders need, and soon.
That assumption is now cracking. And the story of how it cracks is not a technology story. It is a decision-making story; about what you measure, and how you hold the short term against the long.
The Gap Nobody Wants to Name
The largest hyperscalers (Amazon, Google, Meta, Microsoft) are on track to spend somewhere near $700 billion on AI infrastructure in 2026. That is roughly double last year’s figure, deployed in a single twelve-month sprint.
Now set that against what the AI industry actually earns from end users, on the order of $75 billion; roughly a tenth of what is being spent to serve it. It is the technology version of the mantra from the movie The Field of Dreams: if you build it, they will come.
You don’t have to be a bear to find that gap alarming. The bluntest warning came from inside the industry. In a February 2026 interview, Anthropic’s CEO, Dario Amodei, walked through the math for his own company. Given their commitment to ~$1 trillion in compute, if revenue were even modestly short—$800 billion instead of a trillion, or 5x annual growth instead of 10x—then there is, in his words, “no hedge on Earth” that stops the company from going bankrupt. Being wrong by a single year could do it.
The bulls have a fair answer. Building ahead of demand is how you win a platform shift. Nobody laid the fiber after the traffic arrived.
Maybe. But the entire case rests on one critical load-bearing assumption. Every dollar of that $700 billion is a bet that buyers will keep paying premium prices for compute, indefinitely. And in the first half of 2026, the buyers started pushing back, hard.
The Assumption That Just Cracked: Demand Is Price-Elastic
Uber gave roughly 5,000 engineers access to agentic coding tools in December 2025. By April, it had burned through its entire 2026 AI budget; a full year’s allocation gone in just four months. It capped spending at $1,500 per tool per month, and its own president admitted the harder truth: rising token spend was getting difficult to tie to measurable product improvement. In June, the Financial Times reported that Amazon, Walmart, Cisco, Meta, and Uber were all curbing internal AI use, capping budgets, or steering staff toward cheaper models.
This exposes what is called the token-cost paradox: more adoption does not automatically equal more revenue at a price anyone can profit from. The industry assumes usage is value. Usage is just usage. When the meter runs and the output can’t be tied to a desired result, sophisticated buyers do exactly what sophisticated buyers always do—they cut spending.
And when they can’t cut, they substitute. When DeepSeek broke through in January 2025—wiping some $600 billion off Nvidia’s value in a single day—it was dismissed as a one-off scare. Eighteen months on, it looks less like a scare and more like a repricing:
The spread is brutal. Citi, the American bank, puts leading Chinese open-weight models at roughly 18 cents per million tokens against about $4 for the top Western models, a discount of more than 90%. The capability gap that once justified the premium has compressed from more than a year to, by some estimates, about four months.
Companies are moving, not just benchmarking. In June, the AI startup Lindy shifted 100% of its agent traffic off Anthropic’s Claude models to DeepSeek V4. Founder Flo Crivello said the switch saved millions and improved performance on core use cases. His summary is the whole thesis in seven words: “You don’t need God to write your email.” Airbnb and Cursor’s owner have both disclosed using Chinese open models.
The aggregate has flipped. On Vercel, DeepSeek’s share of token usage jumped from under 1% to 17% in a single month. On OpenRouter, the share of tokens U.S. companies run on Chinese models has held above 30% every week since February—against just 4.5% in the first half of 2025.
Note what this is not. It is not a collapse in AI usage; usage is exploding. It is a collapse in the assumption that exploding usage flows to premium-priced compute. The bet was never just “demand will grow”. It was “demand will grow and buyers will keep paying premium prices”. The first half is true. The second half is being proven wrong in real time, one routing decision at a time.
The Metric Problem: Counting Capacity Instead of Demand
Look at the numbers the industry celebrates: capex deployed, gigawatts under construction, chips shipped, tokens consumed. Every one of them is a supply-side metric. They tell you how much has been built and how heavily it is being used. Not one of them tells you whether that use clears a profit.
We have seen this before. In the late 1990s, the telecom industry convinced itself that internet traffic was doubling every hundred days; one seductive data-point, extrapolated into a building frenzy. Operators laid fiber across continents and measured progress in route-miles, not in traffic that paid. Much of that fiber was never “lit”, and the companies that spent the most, like Global Crossing, collapsed into some of the largest bankruptcies of the era. The capacity was real. The demand at the price the builders needed was not.
Falling in Love With the Solution: The Iridium Warning
The other error is subtler and more human: you can fall so deeply in love with a technical solution that you never stop to ask whether the market wants it at the price you’ll have to charge. No company illustrates this better than Motorola—because Motorola didn’t just lose money on the bet. It lost its market lead.
The idea was born in 1987, reportedly after the wife of a Motorola engineer, Bary Bertiger, grew frustrated that she couldn’t get a phone signal on a remote Caribbean beach. Bertiger and two colleagues sketched an audacious solution: a ring of satellites in low orbit that would blanket the entire planet, so a single handheld could place a call from anywhere on Earth. The original design called for 77 satellites, and because iridium is the 77th element on the periodic table, the project had its name.
It was breathtaking engineering, and breathtakingly expensive; roughly $5 billion and a full decade from concept to launch. When service went live on November 1, 1998, the ceremonial first call ran from Vice President Al Gore to the chairman of the National Geographic Society. Motorola had built a genuine marvel.
Almost nobody bought it. The handset was a $3,000 brick, the size of a shoe; airtime ran $3 to $7 a minute; and the phone needed a clear view of the sky, so it failed indoors, in cars, and in the very cities where business travelers actually lived. Meanwhile, in the decade it took to build, ordinary cellular had spread everywhere—cheaper, smaller, good enough. Iridium had spent ten years and $5 billion solving a problem the market was quietly solving on its own. A 1996 Gallup study had already flagged the business model as deeply flawed. Motorola launched anyway.
The forecast was 500,000 subscribers in year one. By mid-1999 Iridium had roughly 20,000. On August 13, 1999—nine months after that first call—Iridium filed for bankruptcy, among the largest in U.S. history at the time. The satellites were nearly de-orbited and burned up before a group of investors bought the entire system out of bankruptcy for about $25 million—roughly a penny on the dollar of what it cost to build.
Here is the part that matters most. While Motorola poured capital, engineering talent, and executive attention into its satellite moonshot, it was losing the race it had actually been winning. It had entered the 1990s as the undisputed king of the mobile phone—the iconic StarTAC had more than 30% of the global market—but it clung to profitable analog technology while a Finnish upstart went all-in on digital. In 1998, the very year Iridium launched, Nokia passed Motorola to become the world’s largest handset maker. Motorola never got the lead back. It took a roughly $2.5 billion write-off on Iridium; but the deeper cost was the years and the focus it burned building the wrong future while a rival built the right one.
The E.D.G.E. Framework: This is an ‘Establish’ Failure
Running this case through the E.D.G.E. framework reveals that the failure is largely in the first step itself.
Establish — Four Things the Builders Got Wrong
The Establish Your Foundation step asks four questions: What can you actually control? Are you aligning your goals with your actions? Are you spanning short-term and long-term goals? And what metrics are you tracking? The AI build-out is answering all four badly.
Control. Capacity built is controllable. Demand at a profitable price is not—nor is a buyer’s decision to route to a cheaper model, nor a community’s willingness to host your data center. The builders are pouring energy into the one variable they command, and treating the ones they don’t as settled.
Align. Watch what the spending is actually aligned to. Not to validated customer demand, but to rivals’ announcements; a land grab in which the trigger to spend is that someone else spent.
Span. They are not balancing the short term against the long. Racing to capture a genuine long-term platform shift is defensible. Overriding the near-term signal that demand is price-elastic, in the name of that long-term thesis, is not. Iridium is the cautionary extreme: a demand assumption set in 1987 and never revisited across a ten-year build. Balancing horizons means letting today’s feedback discipline tomorrow’s bet, not silence it.
Track. They are tracking the wrong metrics. Capex, gigawatts, chips, tokens: every headline number is a vanity metric dressed as a vital one. They measure the builders’ own enthusiasm. When you measure the thing you can control (spend, gigawatts, usage) instead of the thing that decides whether you succeed (durable demand at a profitable price), you feel intensely productive while walking toward a cliff.
Key question: Are you measuring what you can control, or what actually determines whether you survive?
Diagnose — Separate Signal From Noise
The number everyone cites—usage is exploding—is, for the survival question, noise. The real signal is unit economics: durable demand at a price that clears a healthy margin on a trillion dollars of committed capital. The 20% that drives 80% of this outcome is the price-elasticity of demand, not the adoption curve.
Key question: Is your headline metric a leading indicator of viability, or a lagging indicator of your own enthusiasm?
Go — Purposeful Action, Not Maximal Action
Purposeful is not the same as maximal. Purposeful action builds optionality; staged capex, exit ramps, assets that can be repurposed. Zuckerberg’s instinct to rent the surplus out is, read charitably, a Go move done well: it turns a fixed bet into a fungible one. The next tranche of capital should follow instrumented demand, at a real price.
Key question: What is the smallest commitment that keeps your options open without betting the balance sheet on an unproven demand curve?
Evolve — Build the Feedback Loop
The trap underneath all of this is extrapolation. Projecting limitless growth from a handful of early data-points; that is the narrative fallacy dressed up as a forecast. The antidote is a feedback loop that watches substitution and price sensitivity, not just usage. DeepSeek’s price and Uber’s cap are the feedback. Evolving means letting them inform your strategic response.
Key question: What signal would tell you your demand assumption is wrong—and are you watching for it, or watching the capex counter tick up?
The Discipline
You are almost certainly not deploying $700 billion. But the pattern scales down to any decision on your desk: falling in love with a capability before validating demand, spending because a rival spent, measuring effort instead of outcome, and overriding what this quarter is telling you because the five-year thesis is prettier.
None of this is pessimism about AI. The demand is real and growing. The discipline is narrower than optimism or doubt: refuse to confuse capacity with demand, and refuse to confuse growth with growth at a profitable price.
The builders may yet be proven right that the demand is coming. But the Establish question is the one Iridium never asked in time—and the one the meter at Uber just forced onto every finance team in the country:
At what price? And are you watching what your customers are choosing to do?
© The Uncertainty E.D.G.E. | Published every other Tuesday
If this resonated, I’d love to have you as a free subscriber — and forward this to a leader who needs to see around corners.
The Uncertainty E.D.G.E. is for leaders who are accountable for outcomes they can’t fully control — and want a clearer way to think when certainty won’t come.
Essays and conversations on decision-making under pressure, every other Tuesday. Join me at theuncertaintyedge.com.
If the human side of leadership is what draws you, I also write The Good Human Practice (thegoodhumanpractice.com) — on inner clarity and character for leaders called on to make those hard decisions under pressure.