How much productivity is truly at stake? An economic framework for AI revenue

We shift from speed to scale. For today’s infrastructure spending to prove rational, the productivity AI unlocks must vastly exceed its cost. In this article, we leave GDP percentages behind and use a transparent three-step framework. We first size the total value of the productivity gain, then apply historically grounded capture rates to estimate the steady state revenue the winning vendors can ultimately earn.

 1. Sizing the addressable knowledge base

The first step is to accurately identify the target of AI's augmentation, namely the global knowledge worker. These are the roles from software engineers and financial analysts to lawyers and customer service agents that deal primarily with information, communication and complex problem-solving.

This is a significant economic pool: 

• Global corporate revenues: The total annual value of all corporate activity globally is estimated to be ~US$110–120 trillion.
• Knowledge work share: Based on McKinsey and other global employment data, knowledge-intensive roles account for ~25–35% of global corporate activity, representing over a billion workers. 

This places the economic activity generated by this workforce at a staggering:

 Global knowledge worker revenue base (estimated): ~US$34.5 trillion (at the midpoint)

US$34.5 trillion is the pool of value AI is directly augmenting. For context, this single segment of the global economy rivals the combined GDP of the US and China.

2. Estimating the productivity uplift

The core economic driver of the AI revolution is its ability to make this ~US$35 trillion workforce significantly more efficient. The key is that AI is not an incremental improvement like a faster search engine; it is a transformational automation tool.

Studies on generative AI's early impact reveal substantial and immediate productivity gains:

• Conservative enterprise-grade automation: most analyses suggest an immediate and sustainable ~10% uplift in overall productivity across an enterprise.
• Typical knowledge workflows: gains in high-impact areas like coding, documentation and data analysis generally sit between 20–30%.
• High-impact automation: customer service and back-office functions can see up to 50%+ increases in efficiency by automating first drafts and routine queries. 

These figures largely reflect a "Copilot" model meaning humans working with AI assistance. However, the technology is rapidly shifting toward Agentic AI, where autonomous systems actively plan and execute multi-step workflows (e.g., 'resolve this customer refund' rather than 'draft a refund email'). Early data suggests that while Copilots yield 10–20% efficiency gains, Agentic workflows in controlled environments (such as software testing or claims processing) can approach 50–70% labour displacement. As these agents mature from 'human-in-the-loop' to 'human-on-the-loop,' they provide a structural upside to our base case, potentially doubling the addressable economic prize over the next decade.

To build a prudent estimate, we anchor to a conservative case that allows for implementation, friction and integration time. Hence, we will assume a 15% productivity uplift on the knowledge base.

Productivity unlocked = knowledge base x productivity uplift
Productivity unlocked = US$35T x 15%

Annual global productivity unlocked (conservatively) : ~US$5.25 trillion

Calculating the vendor capture rate

The newly unlocked productivity does not all flow to the technology vendors; the vast majority is retained by the end-user enterprise as improved profit margins or reinvestment. The capture rate is the portion of that value the AI and software providers can monetise.

Historical trends offer a reliable proxy:

• PC era: vendors (hardware, software, OS) typically captured ~15–20% of the value generated by the widespread use of the personal computer.
• Cloud/SaaS era: vendors (AWS, Salesforce, etc.) generally captured ~10–15% of the efficiency and flexibility they provided.

Given AI is a software and service-based model, we will use a conservative ~12% capture rate.

Steady state AI revenue = productivity unlocked x capture rate

Steady state AI revenue = US$5.25T x 12%

Steady state AI revenue = ~US$630B

While we anchor on corporate revenue to capture the total value expansion (both cost savings and revenue acceleration), we must note the distinction between gross productivity and net economic impact. Even if we discount the base to reflect only the global wage bill of knowledge workers (approx. US$15–20 trillion) but apply a higher 'task-specific' automation rate, the resulting steady state revenue lands in a similar band of US$500–$700 billion. The math holds; the prize is large enough to support the infrastructure.

The big picture: plausibility and scale

This number, ~US$630 billion per year in steady-state AI revenue, is large but economically plausible. It aligns closely with the long-term total addressable market projections from the world's largest technology companies. However, the ultimate size of the prize depends heavily on the capture rate, which sits at the centre of a structural tug-of-war.

A strong argument exists that AI vendors could capture a significantly higher share of value than the 15–20% seen in the PC era. Unlike the fragmented PC market, the AI era is defined by centralisation. The "intelligence layer" is controlled by the hyperscalers and vertically integrated model providers. This structural concentration may see the capture rate increase towards 20% or 25%.

However, we must balance this upside against the deflationary pressure of open source models such as Meta’s Llama series, which provide a "free" alternative that competes directly with proprietary models. If "raw intelligence" becomes a commodity, it limits the pricing power of the hyperscalers and acts as a natural ceiling on margins.

In our final piece, which will be posted tomorrow, we bring the entire thesis together, addressing key investor questions and discussing the primary risks.

Article 1 - The glacial start of the PC era

Article 2 – How centralised capex heralded the speed of AI

Article 3 – Compressing decades of diffusion into single digit years

Article 4 – How much productivity is truly at stake. An economic framework

Article 5 – Can AI justify the current capex boom