For months, pundits have cast AI investment as a bubble on the verge of collapse, while techno-optimists see unbounded prosperity on the other side of datacenter buildouts. Both could be right. Railroads, canals, and broadband were all civilization-changing technologies that suffered speculative busts before becoming widespread. Electricity and cloud computing saw similarly frenzied investment in the technology’s nascent days, but scaled into massive revenue outcomes without an asset price collapse.

The fiber-optic cables laid during the 1990s broadband boom deliver massive economic value today, powering the modern internet, but they did not generate profit fast enough to sustain the bubble. By contrast, power plants built in the 1880s reached high utilization quickly enough to deliver phenomenal returns on capital. In today’s AI infrastructure buildout, the fundamental question is simple: what scale of revenue, on what timeline, needs to be generated for investment to be rational?

Take OpenAI as an example. The company has committed more than $1T to compute over the next decade. Translated into an annualized 2030 run rate, that implies a $295B compute bill. From there, the math is straightforward.

1. Shift from R&D to inference: To convert that compute into revenue, OpenAI needs most of its GPU usage to migrate from training (a cost center) to inference (a revenue engine).
‍
2. Apply margins: If we accept OpenAI’s projections of a 70% gross margin, that means compute costs should represent 30% of revenue.
‍
3. Solve for revenue: Revenue ≈ $295B ÷ 0.30 ≈ $1T per year. Put simply: if OpenAI’s 2030 compute bill is $295B, the business needs to generate roughly $1T in annual revenue for the economics to pencil out.

Source: Thomas Tunguz

And this is just one company. OpenAI only accounts for ~25% of global compute buildout, implying that the industry needs to cumulatively generate $4 trillion/year of AI revenue by 2030. If AI vendors capture only half the value they create, this translates into $8T in new annual economic output.

That number is historically staggering. Google currently generates $385B/year, after operating for 27 years. The entire knowledge work market, all the economic value generated by humans working in offices, is somewhere around $20 trillion/year. If AI was a country, it would have to rank #3 in GDP, only surpassed by the U.S. and China. Datacenter clusters will be the most expensive thing humans have ever built, more costly than the International Space Station and the Great Wall of China. All within the next five years.

Such a world may seem absurd, but it is not implausible. AI is poised to disrupt massive knowledge work industries like accounting ($676B/year), BPO ($300B/year), and call centers ($350B/year). Transportation alone is a $8.5T/year industry, of which ~30% is labor cost. Whether AI will generate revenue through labor displacement or augmentation, the market is large enough to accommodate it.

That’s the scale of revenue which must be achieved: $8T/year, but on what timeline? Unlike the fiber optic cables we use today, installed decades ago, the useful life of datacenter assets is much shorter. Roughly two thirds of datacenter CapEx is spent on the chips inside, which depreciate on a five year schedule (or much shorter, depending on who you ask). That means that a GPU installed today will be nearly worthless in 2030, due to the technology becoming obsolete. This is not an industry that can wait decades for revenue; value must materialize now.

In the present, continued investment is contingent on persistent faith that AI can generate value fast enough to meet that timeline. If that confidence falters, investment stalls and a bubble may pop. This could occur due to a multitude of reasons, including:
• Model performance stagnating.
• Margin compression causes datacenter P&Ls to look like airlines, not software.
• Changes in GPU depreciation schedules.
• Event-induced panics, e.g. the Deepseek scare of January 2025.
• Tightening private credit markets.
‍
2030 is a maximum deadline, not a guaranteed runway.

The concern of AI being a bubble fundamentally distills into one question: do you believe AI will usher in an age of unprecedented prosperity within the next few years? Will we be zooming around in self-driving cars, producing double our previous output at work, all while a robot does the chores at home? Or will AI end up stalling as a somewhat helpful chatbot generating pumpkin pie recipes?

Regardless, we have signed up for a world which looks vastly different in a few years, a world of untold wealth creation or devastating capital incineration. $8T/year by 2030 or bust.








Notes
‍† This framework is unique to infrastructure bubbles, which differ critically from asset bubbles and credit bubbles in that the investments are collateralized by productive physical assets rather than leveraged financial structures or expectations of resale value. Given that the vast majority of AI spend has been on datacenter buildout, mostly financed by hyperscaler balance sheets, it makes sense to focus on the infrastructure.

‍† Even if AI produces substantially less than $8T of value by 2030, asset prices may not collapse. Historically, hyperscalers overbuilt by 30–60% during cloud adoption waves. Overbuild is normal and not a bubble by itself if cost of capital is low.