There are two ways the AI infrastructure story ends, and if you buy AI rather than build it, they lead to the same instruction.

Suppose the spending turns out to be justified. Capacity gets used, prices grind down, you buy compute from a healthy vendor.

Suppose it turns out to be far too much. Capacity strands, somebody loses an enormous amount of money, and you buy compute from a frightened vendor.

Either way the price of capacity falls and you come out ahead. What decides how far ahead is not whether you called the cycle. It is what you had built on the day the price moved.

That is why the bubble debate, entertaining as it is, is a fight between the people who put up the money. Your fight is a different one, and it is decidable this week.

What Cuban actually argued

His argument is getting flattened into a punchline, so it is worth reading straight.

Speaking on the All-In podcast in July, Mark Cuban said a lot of data centers are going to be turned into pickleball courts. He was not saying nobody will use AI. He was saying efficiency arrives before the payback does.

The fiber comparison carries that. Carriers laid long-haul networks against a demand forecast, then connection speeds improved and the bandwidth shortage dissolved on the technology side rather than the customer side. Cuban expects the same price-performance curve in compute.

On the financing he was blunter. In that episode he noted that the market leaders are borrowing hundreds of billions, that private credit already has a problem, and that these companies are spending all their cash flow on capex and then borrowing on top of it. He called that planning for perfection.

That phrase is the one to keep. It does not claim the plan is wrong. It claims the plan has no room for error. The Motley Fool picked the comparison back up on August 20, which is how it landed in my feed again a month later.

One piece of the fiber story Cuban leaves out is the part I find most useful. When the builders went under, the glass did not disappear. It sat in the ground, cheap, and the companies that arrived afterward built on it. The money was lost exactly once, by whoever laid the cable.

The money changed shape this year

The 2026 numbers give the warning some weight, and they are better read at the source than in the coverage.

Alphabet’s second quarter release reports $39.069 billion of operating cash flow and $44.924 billion in purchases of property and equipment. Subtract, and it is negative. First time. That happened in a quarter where revenue grew 24% and Google Cloud grew 82%, so this is not a business in trouble. It is the size of the bet.

Quarterly capex came in at exactly double the prior year. Full-year guidance moved to $195 to $205 billion.

In June, the same company raised $84.75 billion of equity, the largest equity capital raise by a listed US company, with $10 billion anchored by Berkshire Hathaway.

On July 9, S&P cut Oracle to BBB-, one notch above speculative grade, pointing at negative free cash flow and the concentration of its order book in OpenAI.

Cuban’s own position is less tidy than the headlines make it, and the messy part is the one worth keeping. When Alphabet raised guidance and the market got nervous, he defended the spending, comparing it to 1998, when he bought every server and every bit of bandwidth he could find because he needed more than anyone else. And he got it.

So he is not calling the leaders wrong. He is saying the field in aggregate builds more than gets used, and the ones who break are the ones who financed it with borrowed money.

Back in April I wrote that Google’s capex read as quiet conviction rather than fear, and I still think that was right for April. What changed is where the money comes from. Cash gave way to capital markets, and capital markets attach a clock.

What survives an 80% price drop

Run the question that matters for your company. If compute costs fall 80% next year, which of your current assets is worth more and which is worth less?

Worth less: any advantage resting on a negotiated rate, a volume commitment, or having picked the right vendor early. A competitor with a credit card matches all of that in an afternoon.

Worth more: a process you have already mapped, exceptions included, approvals included, the weird cases written down. That does not get cheaper when compute gets cheaper. It gets more valuable, because suddenly it can run at a scale that did not pencil out before.

This is the same case I made in the runtime is a commodity and the workflow is the moat and again in the harness is the moat. An infrastructure overbuild does not weaken that argument. It speeds it up.

There is a trap on the other side of the discount, and we looked at it in why a spending cap treats the symptom. When unit price falls, consumption rises faster than the savings if the workflow is undefined. Cheap capacity rewards the company with a finished process and punishes the one with only a budget.

What I would do about it this week

None of this argues for pausing an AI project. It argues for choosing where you anchor one.

If your plan for next year depends on the price per token continuing to fall, you are making the same bet Cuban is criticizing the builders for making, with less information and no balance sheet to absorb being wrong.

If your plan depends on three of your own processes being better understood than your competitor’s, the price of compute stops being a strategic variable. It falls and you win faster. It rises and you win slower.

That decision belongs to you, not to Alphabet. The AI Maestro discovery is built around it: two months to establish how the work actually moves, ending in a Go or No-Go before any tooling budget is committed.

The rest of it will be argued on television for the next two years. Those of us who watched the fiber cycle already know how that part ends. Somebody builds on top of it.

Let’s decide where your advantage lives before the price moves