The strategy of offering services at no cost to capture market share is well-established: Google deployed it for search and productivity tools, Meta for social networking, and Cloudflare for web infrastructure. Today, the same playbook is being applied to artificial intelligence. Yet beneath the surface of "free" lies a fundamental economic problem that grows more acute by the day.

Describing a service as free obscures genuine operational costs that must be covered somewhere. The term itself acts as a powerful competitive barrier—rivals must match zero pricing to remain viable. History shows the model can work when subsidies eventually convert to revenue. Uber absorbed a USD 5B operating loss in three months during 2019, taking thirteen years to reach profitability with backing from SoftBank, Saudi Arabia's Public Investment Fund, and Google. DoorDash posted a USD 1.4B loss in 2022, with financial backers subsidizing every delivery. The expectation is that once established, these businesses turn profitable. But AI's infrastructure requirements operate at a different scale entirely.

Paying the AI bill

Large language models demand computational resources of staggering proportions. A modern AI data center typically requires around 60,000 advanced graphics processing units deployed at 72 GPUs per rack, totaling roughly 3,000 to 8,000 racks. These units connect via lossless fabric using 800G optics and high-density switches to high-speed storage systems.

Power consumption becomes the limiting factor. Each rack demands 150kW to 200kW of continuous delivery. The resulting heat necessitates liquid cooling systems throughout the facility, which itself requires substantial water sources. The entire installation needs a power supply infrastructure capable of sustaining up to one gigawatt—a load profile fundamentally mismatched to how electrical grids were designed. Unlike residential and commercial demand that varies across hours, weeks, and seasons, AI data centers run at constant maximum capacity. Solar and wind generation cannot fill this need. These facilities require dedicated substations, high-voltage connections, and significant grid-interconnection investment, plus additional power generation capacity.

Construction costs reflect these demands. A representative 200MW AI training campus currently costs approximately USD 8.2B, with roughly two-thirds covering IT equipment and one-third for real estate and power infrastructure. The known construction pipeline shows 183 GW of capacity being built by 2032, with another 118 GW planned afterward. This translates to USD 10.3 trillion in total infrastructure spending across the United States alone from 2025 to 2032—nearly 4% of annual US GDP, exceeding historical investments in rail, electrification, transportation, and telecommunications infrastructure.

This spending scale overwhelms the internal resources of the companies building it. Combined capital expenditures by Oracle, Microsoft, Amazon, Meta, and Alphabet grew from USD 97B in 2020 to over USD 400B in 2025, with projections exceeding USD 800B in 2026—surpassing their combined operating cash flow for the first time. Hyperscalers have turned to external financing channels: data center developers, infrastructure investment funds, private equity, bond markets, banks, private credit funds, and securitization vehicles. Investment-grade hyperscaler tenants make these structures financeable by supporting long-duration contractual cash flows.

Yet this financing architecture concentrates rather than disperses risk. Assets moved into separately financed vehicles can increase leverage on underlying infrastructure even when reported hyperscaler leverage appears low. Long-duration debt depends on uncertain factors:

  • Uncertain AI demand and revenue models
  • Rapid technological change
  • Timely access to power and hardware
  • The continued credit quality of a small number of tenants

The resulting capital structure makes exposures more layered, correlated, and difficult to observe. Not only is this infrastructure exposed to technological and operating risks, but the financing structure can transmit and amplify those risks. Long-duration debt and contractual claims are written against assets whose usage prospects, technological relevance, and residual value remain highly uncertain.

Technical obsolescence presents another threat. For six decades, the silicon chip industry has simultaneously increased computational capability while reducing costs and power requirements. But that trajectory may be ending. If chips stop improving or become more expensive to produce, AI will not become cheaper—it will become more expensive. Models may grow more efficient and some tasks cheaper per token, but total spending will accelerate because demand is growing faster still. Users want longer contexts, multimodality, agents, search, memory, and task execution capabilities.

Should silicon progress stall, advancing AI capabilities would require larger and more expensive computational assemblies and matching demand for bigger data centers with greater power needs. This scenario creates a vicious cycle: growing debt as AI companies amortize massive investments. Meta closed 2025 with USD 72B in capital expenditure and anticipated USD 115B to USD 135B for 2026. Anthropic announced a USD 30B financing round in February, valuing the company at USD 380B. Microsoft, NVIDIA, and Anthropic formed an alliance committing Anthropic to purchase USD 30B of Azure capacity and up to an additional gigawatt of compute.

Google undertook a USD 32B global bond issuance in 2026 across the US, Canada, Japan, Europe, and Australia to finance its AI data center buildout. Alphabet indicated capital expenditures could approach USD 175B to USD 185B in 2026. The company's 100-year bond signals a shift beyond traditional equity investors toward long-horizon institutional capital such as pension funds and insurers, reinforcing infrastructure-style financing as AI investment intensifies.

Chip designers, fabricators, cloud operators, model developers, data center operators, and infrastructure investors are interconnected through long-term contracts, strategic investments, and financing arrangements. A shock to one segment—weaker demand from model developers or the end of continuous silicon chip refinement—can propagate as falling service operator revenues, lease payments, asset values, and creditor recoveries. This circularity amplifies exposures when any single activity fails.

Making money with AI

Google pioneered a two-sided market model to fund its consumer services. Search is provided without direct cost to users, but Google assembles consumer profiles and sells them to advertisers. The more users search and the more comprehensive the profiles, the greater their value to advertisers. Economist Hal Varian, later Google's Chief Economist, argued in the late 1990s that spam represented a failure of consumer information. Better information about users meant better ability to convert advertisements into transactions. Google's advertising currently generates USD 300B annually—advertisers are effectively paying for search.

Advertisers did not increase budgets to add online advertising to existing channels. Instead, Google offered to increase advertising effectiveness while potentially decreasing total spending. What did advertisers cut? Newspapers and free-to-air television bore the brunt. The newspaper industry became an impoverished shadow of its former self. Some governments now force digital giants to return small percentages of forgone advertising revenue—Australia's News Media Bargaining Code exemplifies this approach. Free-to-air television lacked comparable bargaining power, so the content industry shifted to consumer subscription models. These discretionary costs are vulnerable to economic cycles. When interest rates rise, petrol costs spike, or aging populations shift to fixed retirement income, content subscriptions are often the first expenses cut.

AI's cost structure presents a different challenge. Capital expenditure alone reaches USD 750B in 2026. Beyond that lie power and cooling costs, tool refinement, and capital reinvestment. Total annual operating costs likely already exceed USD 2 to 4 trillion.

In previous eras, Moore's Law assumptions allowed companies to burn through startup funds gaining market share for free services, betting that operational costs would decline within years as silicon became cheaper. The mobile phone industry followed this path, adding features while maintaining retail prices and encouraging device replacement cycles. But if silicon progress has stalled, this strategy collapses. If computational capability cannot continue improving while power demands fall, users must pay more for AI services.

The cost may be packaged into employer services or bundled with products like office suites, evolving into an infrastructure cost largely invisible to end users. The advertising-funded model is also reemerging. OpenAI recently introduced unlimited access to Luna, a lightweight GPT-5.6 version, for free ChatGPT users and low-cost Go tier subscribers. The company announced plans to introduce advertising for Free and Go tier users to support continued low-cost and no-cost access.

Google is transforming its search service. In AI mode, the engine no longer returns resource pointers but uses AI synthesis to generate direct responses. Google reports that "Just one year after its debut, AI Mode has surpassed 1B monthly users, with queries more than doubling every quarter since launch". The real question is whether this change increased advertising revenue or represents last-ditch defense of existing advertising revenue against OpenAI and Anthropic incursion. The latter appears more likely.

Premium AI subscriptions will persist, offering better models, faster processing, more capable responses, and zero advertising. But these remain accessible only to those with discretionary income. Like streaming services, such revenue depends on broader economic well-being. When interest rates, petrol, food, and clothing costs rise, or when aging populations shift to fixed retirement income, discretionary subscriptions are often the first to be cut. AI operating costs are not necessarily based on query volume, so falling subscription revenues without accompanying cost reductions represent a very real prospect.

The digital advertising market is valued at approximately USD 700 to USD 800B annually. AI operating costs appear to be easily more than twice that amount, with rough estimates of USD 2 to 4 trillion per year. The math does not work.

Where will the money come from?

To fund today's search engines, the economy destroyed newspapers and free-to-air television. Funding AI requires identifying what part of economic activity will be sacrificed to generate the remaining capital.

The answer may already be visible: shedding people and their jobs. Employers laying off workers appear to be betting they can save sufficient money through reduced wages bills to fund premium corporate AI subscriptions performing the same functions. Many tech companies have experienced mass layoffs in recent years. Dario Amodei, CEO of Anthropic, has predicted that AI could wipe out half of all white-collar jobs and push unemployment to 20%. A labor market upheaval of this magnitude would cause enormous suffering for many households and pose significant challenges for policymakers, shifting costs to the public sector while retaining income and value in private hands.

Society has revised its definition of full employment to approximately 4% unemployment. At 20% job loss, the economy would leave the territory of economic downturn and enter full-blown economic meltdown.

A strong counter-voice challenges this dire prediction. Current figures fall far short of labor market collapse due to AI. While popular perception suggests AI will shrink the workforce rather than sustain growth, available data shows no clear signs of significant workforce shrinkage.

Another perspective holds that AI expenditure represents mass hysteria, with AI's actual ability to replace people far more limited than claimed. This view sees AI fitting within existing advertising revenues, with those failing to secure significant revenue streams facing failure.

Where is this all heading?

Meeting the total costs of constructing all specialized data centers, power systems, and IT plants appears unlikely. The capital requirements are simply too vast, and sufficient diversion into this sector cannot or will not occur to sustain everyone's borrowings. Some AI entities will likely fail financially, with survivors picking through the remains of failed enterprises for salvageable assets.

Which competitors are more likely to survive? Do incumbents hold an advantage through existing revenue and customer bases that sustain them through this period? Or can challengers capture sufficient market share and revenue through innovation and the absence of legacy business models that weigh down incumbents?

The current state of AI appears financially unsustainable, and changes will occur as the journey unfolds. Yet societies and their institutions demonstrate considerable resilience. Bitcoin was supposed to be finance-and-banking Armageddon. AI has been sold as an agent of shattering change in how people work and play. Perhaps that is overdramatic, and coming years will bring changes at a slower pace with more opportunity to adjust along the way.