The AI capex train isn't stopping, and Brazil pays the fare

The 2026.33 edition of This Week in Stratechery, by Ben Thompson, crystallizes a thesis that has been building over the course of the year: the AI problem has stopped being just about compute and energy, and has become about capital. And the piece's central item, the analysis Nvidia's Risky Business, points to a shift that changes the calculus for anyone building a technology business, including in Brazil.

What Nvidia actually announced

What the headline describes as a "new financing mechanism" is, in Thompson's reading, financial engineering: Nvidia is creating ways for its own customers to raise long-term capital to buy more GPUs. Add to that the move by Google, which according to the piece "is leading the way in raising via equity" to fund infrastructure construction, and the picture becomes clear.

The question Thompson poses is uncomfortable:

Everyone knows we're short on AI compute. Everyone knows we could soon run out of energy. But what happens if we run out of capital? If AI is as valuable as it seems, it should pay for itself, but that hasn't happened yet.

>

-- Ben Thompson, Stratechery

This is the point that separates the analysis from the hype. The dominant narrative treats every multibillion-dollar capex round as proof of confidence. Thompson reads it the other way: if AI already paid for itself, it wouldn't need structured financing to keep up the pace of investment. Financial engineering is the bridge between today's spending and the sustainable revenue that hasn't arrived yet.

Why this keeps the cost of compute rising

Here is the mechanism that matters to anyone deciding a budget. As long as long-term money can be channeled into buying GPUs, demand for compute stays pushed upward, even if the buying companies' AI revenue still doesn't justify the spend. Financing keeps demand artificially high, and high demand keeps prices high.

Thompson is explicit about the cost of this arrangement: it serves Nvidia's threatened margins and, at the same time, "expands the blast radius of a bubble." Translating that into operational terms: the more the buildout depends on third-party debt and equity, the more people get dragged down if the promised revenue fails to materialize.

For the founder, the practical implication is direct. The price of training and running models isn't falling on the strength of technological progress alone; it's locked into a system where the biggest buyers have the incentive, and now the financial instruments, to keep buying. Don't count on a structural drop in compute costs in the short term out of market generosity.

What changes for those building in Brazil

The Brazilian angle is where the analysis needs to go beyond translation. A founder here doesn't raise long-term capital in dollars on the same terms as an American hyperscaler, and doesn't have an Nvidia designing a financing mechanism for their data center. The cost of compute arrives in Brazil already inflated by this global dynamic, on top of currency exchange and a cloud infrastructure whose GPU prices track international scarcity.

This reinforces three strategic decisions:

• Don't compete at the foundation model layer. Training your own model from scratch means playing on a board whose entry price is set by those with access to capital that the Brazilian founder doesn't have. The source makes clear that even this abundant capital isn't making model economics trivially profitable.
• Build at the application layer, where the margin exists. The value for the Brazilian business lies in using third-party models with predictable costs, not in owning the infrastructure. If compute is going to stay expensive, the differentiator is how much real problem gets solved per token spent.
• Treat inference cost as a business risk, not a fixed line item. A product whose margin depends on a compute price anchored to a possible bubble needs unit economics that survive a correction scenario, whether up or down.

The counterpoint: what if revenue catches up with spending?

The pessimistic thesis has an honest flank, and Thompson doesn't hide it. If AI is indeed as valuable as it seems, Nvidia's financial engineering and Google's equity raise are exactly the rational bridge: you front-load capital to build capacity today because tomorrow's revenue will justify it. In that scenario, whoever waited for the cost to drop missed the window, and the structured financing was vision, not desperation.

There's precedent on both sides. Debt-financed infrastructure buildouts have created lasting value (railroads, telecom) and have also produced spectacular craters (fiber optics in 2000). The difference lies in how fast revenue demand catches up with installed capacity. What the source signals is that, so far, it hasn't, and that's the data point, not the opinion.

For the Brazilian founder, the counterpoint doesn't change the conduct, it reinforces it. In both scenarios, betting the company on owning compute infrastructure is the worst place in the chain: if it's a bubble, you burst along with it; if it's real, you compete on capital with those who have instruments you don't. The defensible position is the same under both hypotheses.

What remains open

Thompson's piece points to a systemic risk without pinning down a timeline, and it does well not to fake a precision it doesn't have. What the founder should monitor isn't the size of the next announced capex round, but two signals: whether the big buyers' AI revenue starts to close the gap with spending, and whether the inference cost reaching the clouds used in Brazil shows signs of decoupling from global scarcity.

The business reading the source allows is sober: the capex train keeps rolling not because AI has already proven its return, but because there now exists financial engineering to keep it moving while that proof is still pending. Building on the assumption that this makes compute cheaper in Brazil in the short term is betting against the very mechanism Nvidia just put together.