Azeem's analysis finds AI infrastructure demands ended a 16-year flatline in US power generation, driving 9 TWh monthly growth

6/ $2 trillion of CapEx is committed through 2026. This is the largest buildout in tech history and is increasingly funded by external debt rather than cash. As of Q4 2025, AI revenue only just covers the depreciation on that buildout and that's before every other cost.

13/ Our analysis suggests that AI demand is more revenue-validated than any prior platform shift. The investment case comes down to whether falling prices can move enough token volume to earn a return on CapEx.

FULL REPORT: https://www.exponentialview.co/p/the-state-of-the-ai-economy

Thanks to @alexolegimas @jaimesevillamol @shanumatthew93 for early feedback on the report.

So, the US is adding… just 12.3 GW of power/year? That's good, actually. China adds only something like 47, in pure 100% capacity factor terms. Of course, none of this actually matters. The US as well as Chyna can raise utilization of coal and get *hundreds* of extra TWh.

7/ Demand is so tight that even old, fully-depreciated GPUs still earn scarcity rents.

@alexolegimas @ShanuMathew93 @Jsevillamol Also covered in Bloomberg: https://www.bloomberg.com/news/articles/2026-06-25/ai-demand-begins-to-justify-massive-cost-of-data-center-buildout?srnd=undefined

5/ Compared with the wider economy, AI revenues remain tiny. It’s worth noting that all previous general-purpose technologies behaved similarly.

Electricity made light ~99.97% cheaper, but this gain wasn’t recorded in the GDP. When evaluated against the value consumers place on AI, we estimate that it tops monthly genAI revenues by 30%.

12/ Revenue is concentrated in infrastructure today, but value is visibly moving toward apps and models.

11/ The frontier is a treadmill. Last year’s best models commoditize into open weights within a year, and the most sophisticated users are already leaving.

10/ No one has yet cracked how to price the “intelligence” a token delivers.

Tokens are the billing metric of the AI economy, but not yet the unit of value we need to estimate the useful output of AI use.

We suggest that quality-adjusted tokens come closest to the unit of value we need. Over the past year, quality-adjusted tokens have kept pace with raw token growth.

9/ The efficiency gains drive lower token prices, but this is offset by higher demand. Industry-wide revenue per GW of data center capacity passed $7bn/GW.

8/ Token volumes top 30 trillion a month, growing 14x YoY.

@alexolegimas Thanks to @alexolegimas @ShanuMathew93 @Jsevillamol

@azeem @alexolegimas @ShanuMathew93 @Jsevillamol Unroll @threadreaderapp

@teortaxesTex America cannot "just raise coal utilization". Its coal plants are so old that they are struggling to keep current level of utilization.

@azeem Couple of clarification points:

- Our consumer surplus estimate is on top of whatever consumers may pay for these tools, so it should be 130% on top of revenues, not 30%.

- Revenues are mostly from enterprise use. Ideally would only compare with revenue from consumers.

@azeem @alexolegimas Unroll @threadreaderapp pls

@avi_collis @azeem We're planning to go into more depth here to work out how this best marries with your great research, but also let me know if it'd be helpful to have a chat through our data!

@teortaxesTex The 🇺🇸 can't just raise coal utilization for hundreds of extra TWh. Its coal fleet averages 40+ yrs old operating at declining capacity factors & being retired faster than upgraded. In 2025 alone, 🇺🇸 retired over 4 GW of coal capacity & that trend is accelerating.

@avi_collis @azeem Thanks Avi! Yes, we were deliberately conservative in our comparison for the reasons you state: particularly on 1) because of the consumer freemium model, and on 2) because of the prevalence of consumer subscriptions used in enterprises

@teortaxesTex Even if coal plants could produce more power, the grid can't move it where it's needed.

The 🇺🇸 transmission system is fragmented, with over 2,500 GW of projects stuck in interconnection queues for 5–7 yrs. Modernizing it will cost over $2 T.