30 trillion knowledge work, 30 trillion consumer market -- Mastering the AI inference era

Morgan Stanley believes that the artificial intelligence industry is making a significant leap from the infrastructure construction phase into the "Age of Inference." This technological evolution will digitize and reshape the global knowledge work market worth $20-30 trillion and the consumer market worth nearly $30 trillion, driving an unprecedented sector rotation in the capital markets.

According to Wind Trader, on September 7, a research report led by Brian Nowak of Morgan Stanley, "The Morgan Stanley AI Guidebook: Navigating the Age of Inference," pointed out that as the capital expenditure growth rate of hyperscale cloud service providers' data centers is expected to peak in 2027 and slow down significantly in 2028, the flow of funds in the technology cycle is seeing a critical inflection point.

(By 2027/2028, hyperscale cloud providers' capital expenditures are expected to reach about $1.5 trillion/$1.6 trillion)

Investor funds are expected to marginally flow out of the semiconductor and hardware layers, with a massive rotation into the software layer and AI enablers. Technology giants with complete ecosystems and free cash flow such as Amazon, Google, Microsoft, and META will see new rounds of valuation expansion.

(The AI business cycle is expected to extend to software, services, and AI application enterprises)

The report emphasizes that enterprise and consumer-side AI applications are already displaying an accelerated explosive growth trend. As of the second quarter of 2026, about 25% of the S&P 500 companies have begun to quantify the financial benefits brought by generative AI.

At the same time, the AI infrastructure credit financing market is showing extremely strong support. Since the beginning of this year, global AI-related debt issuance has reached approximately $450 billion. The robust operating cash flows of hyperscale cloud service providers are fully capable of covering future debt requirements, alleviating market concerns over financing bottlenecks.

During this transition, market focus will fully shift toward AI investment returns, the competitive landscape of open-source models, and increasingly prominent power and regulatory constraints. How to balance surging computing power with physical limitations will determine the ultimate allocation of trillions of dollars in future inference expenditure.

Capital Flows at the Turning Point: Hardware Cools Down as Software and Enablers Rise

Morgan Stanley believes that the capital expenditure trajectory of hyperscale cloud service providers is the key indicator in determining sector rotation in the markets.

By 2027, data center capital expenditures are expected to increase by 60% year-over-year, reaching a total of $1.5 trillion. However, constrained by labor, materials, power, and the pre-positioning of some capacity for 2027-2029, the capital expenditure growth rate is projected to slow significantly to about 12% in 2028.

This deceleration, combined with the accelerated revenue growth at the software layer, marks the onset of a typical multi-year technological capital rotation. This mirrors the pattern in the mobile Internet era: after an initial explosion in hardware and semiconductors, value shifts toward application and software service layers.

(Relative stock performance during the mobile Internet era)

Although hardware stock prices are unlikely to experience a major plunge due to undemanding valuations, as earnings outperformance moves up the stack, software and enablers will enjoy greater opportunities for multiple expansion.

With capital expenditures being realized, the world's total computing capacity will grow exponentially. By 2028, total computing power is projected to surge from 35 GW in 2025 to about 145 GW.

(Morgan Stanley's projected path to 145 GW computing power by 2028)

Within this, the share of new computing power delivered by custom ASIC chips will leap from 34% in 2025 to 66% by 2028, with Google's TPU and Amazon's Trainium leading this structural shift.

(Over the next few years, the incremental capacity share of ASICs will continue to rise)

Penetrating a $60 Trillion Market: Tech and Finance Lead the Way

Entering the Age of Inference, the real test for generative AI is commercial implementation.

The market faces a $20-30 trillion global knowledge work digitalization opportunity, while the consumer market—including retail, tourism, autonomous driving, food delivery, and advertising—holds about $30 trillion in untapped potential.

(The research report estimates around $30 trillion in consumer spending awaiting further digitization)

Adoption at the application layer is accelerating. From a macroeconomic perspective, the technology sector is at the forefront of adopting and quantifying the benefits of generative AI, with finance, healthcare, and industrial sectors close behind.

(Application use cases cover the entire economic spectrum)

The ratio of tech sector earnings calls mentioning AI benefits has soared from 28% a year ago to 51% today.

Historical analogs suggest that AI's current penetration is likely being severely underestimated. In the second year of cloud computing adoption (2014), public cloud spending accounted for 4% of total IT budgets. If AI achieves a similar adoption rate, enterprise AI spending would reach about $800 billion by 2027.

(From an enterprise perspective, public cloud spending reached 4% of IT budgets in year 2, 2014)

Because AI does not require wholesale infrastructure migration and can deliver value in a shorter time, the actual adoption rate is expected to outpace cloud computing. On the consumer side, platforms with massive first-party data and distribution channels will have an absolute advantage.

(The development of generative AI should be even faster; even with a 4% penetration rate in its second year, 2027, enterprise generative AI spending would reach about $800 billion)

Computational Returns and Open-Source Models: Reshaping Business Models

Regarding the market's focus on capital returns, analysis shows that generative AI offers extremely attractive investment returns, with multiple monetization paths expected to deliver ROIC ranging from 25% to 50%.

The highest yields come from companies running model APIs on their own infrastructure (such as META, Google, SpaceX), which can achieve investment returns of about 46%. Even pure-play hyperscale GPU leasing (IaaS) operations can maintain an ROIC of around 30%.

(25%-50% return on invested capital for three generative AI frameworks)

At the model level, open-source models with lower service costs will not weaken compute demand, but rather become key drivers for AI's widespread adoption across the economy.

Open-source models will push down average token prices, triggering the Jevons Paradox—where lower prices lead to an explosive increase in inference demand.

This competitive dynamic will force leading AI labs to continually innovate to fight for inference spending, while further solidifying the core value of hyperscale cloud providers' "model orchestration layers" (such as AWS Bedrock, MSFT Foundry, GOOGL Vertex).

These cloud platforms maximize compute efficiency and monetization by matching the most cost-effective models to different tasks.

Overcoming Infrastructure Bottlenecks: Credit Expansion and the Power Supply Challenge

The foundation for this massive inference market is uninterrupted financing and infrastructure expansion.

Although AI-related debt issuance has surged to about $450 billion so far this year, high-quality hyperscale cloud providers still have sizable capacity for additional debt offerings. Market adjustment mechanisms will primarily manifest in widening credit spreads.

More importantly, the operating cash flows of Amazon, Google, META, and Microsoft are accelerating. By 2027 and 2028, these four giants are expected to generate $980 billion and $1.2 trillion, respectively, in operating cash flows, which is 7-8 times the $290 billion in debt needed during the same period—demonstrating exceptionally strong balance sheet resilience.

However, physical bottlenecks remain severe. State and local governments are increasing their political scrutiny of the impact of data centers on electricity costs, water consumption, and other issues.

To address grid connection delays, hyperscale data centers will increasingly adopt behind-the-meter on-site power generation, which is expected to add about $3 billion in capital expenditure per GW.

The rigid demand for power in exchange for time makes on-site generation companies and power shell asset providers direct beneficiaries with long-term investment appeal.

(Power shells and racks remain the biggest drivers of computing investment)