AI’s $1.5T Question

Updated math for H2 2026, reflecting H1 CapEx re-acceleration and $3T cumulative revenue requirement since ChatGPT launch

This is a repetitive blog post and so I will keep it brief. A lot of people have reached out to me for the updated math behind AI’s $600B question. It is now AI’s $1.5T question.

I have been tracking AI data center spending as compared with total ecosystem revenue for almost 3 years now. The rough trajectory has been as follows:

• September 2023: AI’s $200B question

• June 2024: AI’s $600B question

• June 2025: AI’s $840B question

• July 2026: AI’s $1.5T question

In the initial post, I noted that OpenAI was the lion’s share of total ecosystem revenue, and that it was holding everything up. Today, there is an obvious change which is that Anthropic has scaled phenomenally well and is now rumored to be at $60B in ARR. That’s quite a jump!

With the rise of AI coding, there is a more clear path to monetizing data center CapEx than there was when I first started publishing these analyses. When you see a company go $0 to $60B in ARR, we should all stand in awe of that performance.

The rough shape of the CapEx curve has always been of interest to me, as I think it reflects the overall level of exuberance in the ecosystem. What’s notable is that 2026 saw a material re-acceleration over 2025 (driven by long-horizon agentic capabilities) and that this is reflected now in stock prices.

As a reminder, the $1.5T estimate can be derived as follows: Take Nvidia’s projected Q4 run-rate data center revenue x 2 (to reflect total data center CapEx, including non-chip expenses) x 2 (to reflect a 50% margin across the hyperscaler and the AI product company). This analysis arrives at the lifetime end-customer revenue requirement for a single year of CapEx. These numbers are cumulative — if you wanted to know the total required revenue for today’s AI buildout (cumulatively) since ChatGPT, you’d add up the numbers and arrive at roughly $3T of lifetime required revenue thus far. ¹

You can also arrive at the same number by taking the expected hyperscaler data center CapEx in 2026 and multiplying it by two (reflecting the margin requirement). Most forecasts have 2026 hyperscaler data center CapEx in the $750B ballpark. Doubling that yields the same $1.5T.

Over the last three years, this napkin math has been consistent and predictive across both methodologies. However, with skyrocketing memory prices and with heterogenous compute, I expect the GPU-based math will increasingly underestimate the future AI revenue requirements, as it will undercount TPU / ASIC revenue requirements, and it will undercount the increasing cost burden of memory and other bottlenecked materials on data center buildouts (these factors increase the multiplier on chip costs). Recently, the required revenue per GW of CapEx has sharply increased due to these bottleneck dynamics and rising costs of construction.

Suffice it to say that when I first published this analysis, it was a conversation that very few people were talking about. Today, this is an incessant discussion in the market, and can be summarized as the “AI ROI Debate.” Lots of people have commented on both sides, and the conversation has become robust. No lengthy commentary is required outside of simply laying out the latest math.

1. To be precise, the cumulative revenue requirement would be based on LTM (last twelve months) numbers rather than run-rate numbers, but $3T is the right ballpark.