AI Capex Is Justified:

I spent this week developing a model to estimate the impact on GDP. I mostly used Claude and ChatGPT for my analysis. It was a lot of nonstop prompting, and I ended up using Claude’s allowance daily and even exceeded the weekly usage limit. It was a lot of back-and-forth, but I am very pleased with the outcome. I will publish a detailed blog on this, but below is the summary -

The AI investment cycle is unlike anything we have seen. Microsoft, Amazon, Google, Meta, and the Stargate consortium are on track to spend roughly $4 trillion in cumulative infrastructure capital by 2036, rising toward $7 trillion by 2045. The natural question — one that the research community has approached from many different angles — is whether that capital is justified by the economic value AI will actually create.

PT-AIGE — the Primal Thesis AI GDP Estimator — is our attempt to answer it from the bottom up, with every assumption named, sourced, and open to challenge.

The Methodology:

The model is built from five multiplicative factors applied sector-by-sector across 21 NAICS industries, then extended with three annual layers.

The first factor — labor share — is the most consequential design choice. GDP is not the same as labor income. Real estate sector GDP is 88% imputed rent; Finance includes roughly 56% trading profits and capital returns. Since AI improves worker productivity rather than the return on capital, the model applies gains only to the worker-generated portion of each sector's gross value added, drawn from BEA 2023 compensation accounts.

The remaining factors follow a consistent discipline. AI coverage is drawn from Anthropic's March 2026 labor market study, which provides occupation-level theoretical exposure scores mapped to industry sectors via BLS OEWS 2023 employment weights. Productivity percentages are anchored to peer-reviewed evidence — Block CFO (+40%), MIT/GitHub Copilot (+26%), Goldman Sachs (+30%), and Brynjolfsson et al. (+14%) — with model assignments scaled from these anchors by sector characteristics. Adoption is adjusted for structural friction: Finance and Healthcare carry regulatory scores of 3 out of 3, indicating that their uptake curves are meaningfully delayed, while Education scores 0. And the capture rate accounts for the reality that low-skill productivity gains often deflate into lower prices rather than expanding measured output.

Three additional layers are computed annually. Demand expansion captures the volume effect of cheaper, faster AI outputs. Robotics gradually unlocks physical sectors, making them addressable from 2030, beginning with just 2% of the addressable pool in the base case. And electricity drag — incremental AI power operating cost — is subtracted from gross benefit every year. At $220 billion annually by 2036, this is a significant reduction that keeps the model honest about the true net economic gain.

The outcomes:

The model is run across four scenarios, each with a consistent set of assumptions about capture rate, friction residual, demand factors, and — in the Agentic case — sector-specific productivity multiples that reflect genuine workflow transformation rather than marginal task acceleration.

The base case produces a net annual benefit of $1,057 billion by 2036 — roughly 3.6% of current GDP. Three sectors drive the majority of this: Professional and Technical Services (+$460 billion), Information (+$327 billion), and Finance (+$219 billion). Together, they account for 85% of direct uplift — a result of their combination of high labor shares, high AI coverage, and strong productivity evidence from real deployments.

One result worth noting: the model extends beyond 2036. The base case rises from $1,057 billion in 2036 to $1,139 billion by 2045. Friction continues to ease across the horizon, demand effects reach full maturity around 2040, and the robotics contribution to physical sectors accumulates gradually. The model does not exhibit peaks or plateaus.

Is the investment justified?

The answer is yes — in every scenario. Cumulative AI infrastructure capex reaches roughly $3.9 trillion by 2036. The base case generates $7.7 trillion in cumulative net GDP by that year, yielding a 1.96x return. Even the most cautious scenario crosses payback in 2032 — six years from the model start date — before the bulk of productivity gains have fully diffused through the economy.

The ROI calculation uses cumulative net GDP after subtracting the electricity drag, with no discount rate applied. Capex scope covers Big 5 hyperscalers plus Stargate net-new construction—the infrastructure that underpins the entire AI ecosystem, regardless of which models or applications ultimately win market share.