I recently retired after a long career and began working on Maine Vision 2040, an initiative devoted to understanding Maine’s economic opportunities and preparing the state for the future.
I could not have undertaken this project at its current level without artificial intelligence. AI has helped me investigate industries, analyze technical presentations, build comprehensive and detailed financial models, develop policy ideas, and convert rough concepts into articles and presentations. Work that once would have required a research team, specialized consultants, and months of effort can now be explored in hours and days.
AI did not replace my job. I had already retired. It created a new avenue for contribution and expanded what I am capable of accomplishing independently.
That experience shapes how I view the debate over whether the AI infrastructure buildout is a bubble. The capital investment is enormous, and some projects will fail. But the bubble argument often assumes that AI must justify this investment through one narrow source of revenue, usually subscriptions paid directly to foundation-model companies. The opportunity is much broader.
The AI buildout does not depend on one company, one business model, or one path to profitability.
It can be supported by several reinforcing sources of value: model revenue, cloud infrastructure, enterprise productivity, scientific discovery, new-company formation, public-sector improvement, and products that do not yet exist.
A transformative technology can experience pockets of overinvestment without the entire economic expansion being a bubble. There may be too many data centers in particular locations, poorly financed projects, or model companies that fail. That does not make the underlying AI economy artificial.
The most specific and plausible framework I have encountered.
Brad Gerstner, CEO of Altimeter Capital, offered it during his presentation at the 2026 All-In Summit. He estimates that AI can address a $30 trillion global knowledge-work value pool. His financial model projects approximately $700 billion of AI-lab revenue in 2028, reaching a $1 trillion annualized run rate by year-end. Separately, he calculates that monetizing only 4% of the $30 trillion opportunity would produce $1.2 trillion in annual AI revenue, enough, in his view, to support the current infrastructure buildout. He defined the market this way:
| Knowledge-work category | Estimated value pool |
|---|---|
| Consumer services and advertising | $1T |
| Software development | $2T |
| White-collar workflows | $7T |
| Other knowledge work | $20T |
| Total knowledge-work value pool | $30T |
| Four-percent AI monetization | $1.2T |
Figures as presented by Brad Gerstner, Altimeter Capital, at the All-In Summit, September 2026. Top-down estimates, not forecasts.
The $30 trillion figure is a top-down estimate of the total addressable market that AI can influence. If the industry eventually captures 4%, or monetizes $40 of revenue for every $1,000 of knowledge-work value that AI helps perform, that level of demand could economically support the current buildout.
The $1.2 trillion in revenue does not have to be generated exclusively by the frontier laboratories, although they probably will account for a significant portion, at least in the short term. It can be distributed across the full AI economy. A customer may pay an application provider, which then uses part of that revenue to pay a model provider, cloud company, security vendor, and data platform. It remains one dollar of customer spending distributed across the stack, not separate revenue at every layer.
The strongest case against the bubble thesis is that the required revenue can emerge through many mutually reinforcing channels.
Not every path must reach its maximum potential.
01Proprietary foundation models
The most visible revenue will come from companies such as OpenAI, Anthropic, xAI, and Google charging for consumer subscriptions, enterprise licenses, API usage, coding agents, research agents, premium reasoning, voice, image, and video services.
These companies do not need to capture the entire AI economy. Their models can function as the intelligence layer beneath thousands of applications. A chatbot that answers a question has limited economic value. An agent that completes a software project, conducts financial analysis, or manages a business process can command substantially more revenue.
02Open models and managed hosting
Open-weight models from companies such as Meta, DeepSeek, and Moonshot AI will pressure proprietary-model prices, but they will not eliminate AI spending. An enterprise using an open model still needs computing capacity, hosting, security, access controls, data integration, performance monitoring, software updates, customization, backup, compliance, and technical support.
Many organizations will pay cloud providers, specialized AI clouds, and IT-service companies to manage those responsibilities. Open models may therefore accelerate adoption by lowering licensing costs and giving customers more control. Revenue shifts from model licensing toward hosting, security, integration, and management. The pattern resembles the commercial ecosystem surrounding Linux: the underlying software may be freely available, while cloud infrastructure, enterprise support, and managed services generate substantial revenue. Other companies will buy hardware and hire their own teams to maintain a competitive advantage in the marketplace.
03Cloud computing and AI infrastructure
Every meaningful AI application must run somewhere. Whether an enterprise chooses a proprietary or open model, it must purchase computing, storage, networking, and electrical power. This creates revenue for hyperscale clouds, specialized AI clouds, data-center operators, semiconductor manufacturers, networking companies, utilities, cooling suppliers, and electrical-equipment manufacturers.
The infrastructure opportunity is therefore broader than the success of any one model company. Even if the price per token declines, lower prices can stimulate much greater usage.
04Enterprise applications and digital agents
Most businesses will not interact with a raw foundation model. They will purchase AI through customer-management systems, accounting software, cybersecurity platforms, legal tools, healthcare systems, engineering applications, manufacturing platforms, and other industry-specific products.
The application layer converts general intelligence into a dependable workflow. A hospital does not simply need a language model; it needs a secure application that works with clinical records, follows medical procedures, and fits the daily routines of doctors and nurses. That specialized integration creates durable opportunities for existing software companies and new AI-native competitors.
05Corporate productivity and margin expansion
Companies will continue purchasing AI when the economic benefit exceeds its cost. Gerstner argues that successful AI deployment can increase annual corporate margin expansion from the historical rate of roughly 38 basis points toward 100 basis points, or one percentage point.
| Illustrative company economics | Before AI gain | After AI gain |
|---|---|---|
| Annual revenue | $10.0B | $10.0B |
| Net margin | 10% | 11% |
| Net income | $1.0B | $1.1B |
| Additional annual profit | — | $100M |
Illustrative arithmetic for one percentage point of margin, Maine Vision 2040. The 38 and 100 basis-point figures are Gerstner’s.
A company gaining $100 million in annual profit could spend $20 million, $30 million, or more on models, cloud infrastructure, applications, security, and integration while retaining most of the benefit. AI spending can therefore create a self-reinforcing cycle: productivity improvement raises margins and earnings, which supports additional AI investment.
The improvement need not come primarily from eliminating employees. It can result from producing more with the same workforce, developing products faster, improving sales conversion, reducing errors, responding to customers sooner, increasing equipment utilization, making better capital-allocation decisions and generating more revenue by expanded product offerings.
06Small businesses and new-company formation
AI gives individuals and small organizations access to market research, software development, financial modeling, advertising, contract analysis, customer service, data analysis, translation, and strategic planning that previously required large teams or expensive advisers.
Maine Vision 2040 is one example. I can investigate complex industries, examine competing arguments, and produce useful reports at a level that previously would have required a larger organization. This is not the replacement of existing economic activity. It is productive activity that might never have occurred.
Multiplied across millions of entrepreneurs, nonprofit organizations, researchers, and small businesses, this expansion may become one of AI’s largest contributions.
07Scientific discovery and industrial innovation
AI can identify drug candidates, predict the properties of materials, optimize electrical grids, improve weather modeling, discover manufacturing defects, design products, improve agriculture, and accelerate engineering simulations.
The value of discovering a successful drug several years sooner is not limited to reducing researchers’ hours. It includes earlier revenue, additional patent life, fewer failed experiments, and better health outcomes. Likewise, an AI-designed material, battery, or industrial process can create new revenue rather than simply reduce costs. These applications may support premium AI services because a successful result can have exceptional economic value.
08Healthcare, education and government
Some of the largest opportunities exist in sectors that have been difficult to automate. AI can assist with medical documentation, diagnostic support, personalized tutoring, benefits administration, permitting, regulatory compliance, public-safety analysis, infrastructure planning, translation, and accessibility.
These sectors may adopt AI more slowly because of regulation, security, and institutional complexity. Their enormous scale nevertheless means that gradual adoption can generate substantial long-term demand. Much of the benefit will appear as better quality and expanded access rather than reduced employment.
09Consumer services and new demand
Consumers will pay directly or indirectly for personal assistants, education, financial guidance, travel planning, health support, entertainment, shopping, and household management. Some services will use subscriptions; others will be supported by advertising, commerce, transactions, or inclusion in existing products.
As the cost of intelligence declines, services previously reserved for wealthy individuals or large businesses can become broadly available. Lower prices can expand the number of users and frequency of use, allowing total revenue to rise even while the price per token falls.
Better models make applications more useful. Better applications attract customers.
Greater usage creates demand for cloud infrastructure. Larger deployments and better hardware reduce unit costs. Lower costs allow AI to enter additional industries. Wider adoption creates more operational experience and more valuable applications.
This is why the buildout should not be evaluated solely by comparing current foundation-model revenue with current data-center spending. If proprietary-model revenue grows more slowly than projected, open-model hosting, enterprise applications, and cloud services may grow more quickly. If labor savings are limited, new products and increased output may create the value instead. If prices per token decline, the increase in usage may offset the lower unit price.
The market may be large enough. The timetable is still demanding.
Gerstner suggested that even if we focus only on foundation-lab revenue and ignore the other paths to value, the laboratories alone could eventually support the current infrastructure spending. Their growth has been extraordinary and unprecedented.
| Year | Hyperscaler capital spending | Estimated AI-lab revenue | Projected exit ARR |
|---|---|---|---|
| 2024 | $241B | $5B | $7B |
| 2025 | $422B | $20B | $30B |
| 2026 estimate | $788B | $105B | $180B |
| 2027 estimate | $1.06T | $300B | $450B |
| 2028 estimate | $1.195T | $700B | $1.0T |
| 2029 estimate | $1.30T | $1.0T | $1.5T |
Scenario figures as presented by Brad Gerstner at the All-In Summit, September 2026. ARR is annualized run-rate revenue at year-end. Projections, not forecasts.
These are scenarios, not established outcomes. Reaching a $1 trillion annualized run rate by the end of 2028 would require the leading laboratories and the surrounding ecosystem to convert rapidly improving capabilities into paid services. The buildout may still experience episodes of excess capacity, falling prices, consolidation, and failed investments.
The $30 trillion opportunity is economic activity AI can improve, augment, or expand.
AI can automate routine tasks, but it can also help existing employees accomplish more, accelerate scientific discovery, improve products, reduce errors, and make sophisticated services affordable to smaller companies and underserved communities.
Google’s workplace research indicates that most current AI use involves collaboration, ideation, information retrieval, and learning; full task automation remains uncommon. The immediate pattern is assistance and augmentation. Over time, some roles will change or disappear, but new products, companies, and categories of work can also emerge.
The automobile’s value was not limited to replacing expenditures on horses. It enabled suburbs, highways, tourism, logistics, and new industries. AI proponents believe inexpensive, widely available intelligence can produce a similar expansion.
A real technological revolution can coexist with overinvestment.
AI is already changing how people work and accelerating the development of new research, products, companies, jobs, industries, and sources of revenue. Demand for advanced computing capacity remains exceptionally strong. Leading AI systems and semiconductor production capacity are heavily committed, while new factories and data centers are under construction. Usage is expanding, models are improving, companies are reporting measurable benefits, and the addressable economic opportunity is enormous.
But a real technological revolution can coexist with overinvestment. Railroads, telecommunications networks, and the internet changed the economy while also producing failed companies and poor returns for some investors. AI will likely follow the same pattern.
The breadth of the opportunity matters. The buildout does not rely on one fragile assumption. It can be supported by proprietary models, open-model hosting, cloud services, enterprise applications, corporate-margin expansion, scientific discovery, public-sector improvement, small-business formation, and new consumer demand.
AI does not need to replace the people already producing economic value. It can justify the buildout by enabling millions of people and organizations to produce value that was previously impossible, unaffordable, or never attempted.
The most accurate conclusion is neither that AI is a bubble nor that a correction is impossible. AI is a genuine technological and economic transformation accompanied by an aggressive, forward-funded infrastructure cycle.
The $30 trillion knowledge-work opportunity suggests that the ultimate market is large enough. The unresolved question is whether revenue and productivity will arrive quickly enough, and accrue to enough participants, to support everything currently being built. That is not evidence that the AI economy is a bubble. It is an enormous execution test.
- Brad Gerstner, “No AI Bubble, Semis Eat the Nasdaq & AI’s Take Off Problem”, All-In Summit 2026, Los Angeles, September 13–15; presentation published September 17, 2026.
- NVIDIA, published AI-factory performance and energy-efficiency materials covering Hopper, Grace Blackwell, and Vera Rubin.
- SemiAnalysis InferenceX and AgentX benchmark materials covering large-model and agentic inference performance.
- Google, AI & Economy ATLAS, workplace adoption and task-use research, September 2026.
- OpenAI, The Next Era of Knowledge Work and related agentic-work research, June 2026.
At the Maine Data Center Advisory Council’s listening sessions, one concern came up in plain words: AI is a bubble, and Maine will be left with the empty building and the cleanup bill.
This essay is our longer answer to the first half of that sentence.
The second half deserves its own answer, and it does not depend on how the bubble question resolves. Every proposal should be stress-tested under multiple demand and closure scenarios, with enforceable decommissioning plans, financial assurance, and protection against stranded public costs. Our short answer, and the protections we would insist on, are in Straight Answers.