“Free” has long been one of the most powerful and attractive economic tools for companies trying to break into a new market. For decades, consumer technology giants have deployed zero-price models to capture attention and build habits. Google famously used “free” to scale search, Gmail, Docs, and a suite of other consumer utilities. Meta leveraged the strategy to build its colossal Facebook social network, while Cloudflare applied it to web caching services. In a continuous cycle of technological innovation, offering a product or service at no initial charge acts as a vital mechanism to expose the market to new capabilities, hopefully cultivating a deep consumer dependency that can eventually be monetized down the road. Today, this exact playbook is being deployed on an unprecedented scale to introduce advanced artificial intelligence tools to the global public.
However, labeling a service as “free” often creates a misleading market signal, masking the very real, substantial operating costs that must be absorbed by the service provider. Describing technology as free can create the public perception that the underlying infrastructure can be delivered at a negligible cost, with expenses conveniently absorbed through indirect subsidies or alternative revenue streams. Furthermore, offering products for free acts as a formidable barrier to market competition, forcing rival firms to match zero-price offerings if they hope to successfully enter or remain viable in the industry.
Of course, the economics of disruption extend even further. A consumer service can be offered at a price point well below its actual cost to capture market share rapidly from competitors—a practice celebrated as positive “price leading” in optimistic corporate narratives or condemned as predatory “dumping” in regulatory discussions. The ride-hailing industry provides a stark historical parallel: Uber posted a staggering USD 5 billion operating loss in just three months back in 2019, and it ultimately took thirteen years for the company to become revenue positive. Financial powerhouses like SoftBank, the Saudi Arabian Public Investment Fund, and Google played critical multi-billion-dollar roles in subsidizing Uber through its protracted, cash-burning startup phase.
A similar trajectory was witnessed with food delivery giant DoorDash, which managed a USD 1.4 billion loss in 2022. During that period, every individual delivery was heavily subsidized by the startup’s financial backers. This dynamic is hardly novel; the underlying expectation is always that once a new market entrant establishes a dominant foothold, it can eventually pivot the business model, halt subsidization, and turn a profit. Venture-backed startups are essentially financed to buy rapid market share and exert intense financial pressure on their competitors.
When viewed through this rigorous economic lens, the current state of the artificial intelligence sector reveals striking parallels. By all accounts, the AI revolution is a remarkably expensive undertaking. The technology has been made widely available by a small, concentrated clique of hyper-scale enterprises racing fiercely against one another to achieve ultimate market dominance, establish ingrained consumer usage habits, and cultivate absolute user dependence.
Paying the AI Bill
Artificial Intelligence is profoundly expensive primarily due to the sheer, unprecedented scale of computation required to train and run the massive Large Language Models (LLMs) that fuel modern AI systems. Unlike software-centric startups of the past, this massive investment is grounded in physical plant and equipment on an industrial scale.
The architectural design process for contemporary AI data centers is remarkably straightforward in concept: take everything at the raw, bleeding edge of today’s advanced technology and integrate it into a single facility. A modern AI data center may require roughly 60,000 advanced Graphics Processing Units (GPUs), deployed at an extreme density of 72 GPUs per rack, resulting in a total deployment of 3,000 to 8,000 equipment racks. These GPUs must be densely mesh-connected to one another as well as to high-speed storage systems via a lossless connectivity fabric utilizing 800-gigabit optics and high-density switches, alongside immense mass storage capabilities.
Operating this vast web of hardware demands a continuous, monumental supply of electrical power, capable of delivering between 150 kilowatts and 200 kilowatts per equipment rack. Naturally, generating that much power creates an immense amount of heat. Consequently, every equipment rack must be equipped with specialized liquid cooling, and the entire facility must rely on a large-scale liquid cooling plant. Because traditional air cooling is entirely inefficient for such intense thermal loads, operators must tap into significant external water sources to cool the systems down.
These facility requirements demand power supply systems capable of sustaining up to a gigawatt of total energy for a single campus. That steady, heavy load profile does not align with the fluctuating load profiles that have historically dictated the construction of global electrical power infrastructure. Residential and commercial power demand naturally varies across a 24-hour cycle, throughout the week, and across different seasons. Nor does the AI load profile match the periodic generation cycles of renewable energy sources like wind and solar; operators cannot simply skim off excess production from the existing grid during lulls in aggregate public demand. Instead, these massive AI data centers necessitate dedicated electrical substations, high-voltage transmission connections, significant grid-interconnection investments, and the construction of entirely new power generation facilities.
The financial burden does not stop there. The rapid deployment of these facilities has severely strained global technology supply chains, forcing developers to pay heavy premium prices to secure hardware orders. Consequently, populating these data centers requires substantial, ongoing financial backing, alongside complex local community approvals—particularly when tech companies contemplate bold solutions like utilizing modular nuclear generators to meet their energy demands. Once all of that infrastructure is successfully established, operators must prepare to build an even larger, more advanced data center in roughly 18 months, as the capacity and performance parameters of the previous generation of hardware rapidly double.
Industry data indicates that a representative 200-megawatt AI training campus currently costs roughly USD 8.2 billion to construct. Roughly two-thirds of that figure reflects IT equipment costs, while the remaining third covers real estate and associated power infrastructure. Measuring future AI data centers solely by their power requirements reveals an extraordinary construction pipeline: known schedules indicate that 183 gigawatts of capacity are slated for construction by 2032, with an additional 118 gigawatts planned for development thereafter. This amounts to a staggering infrastructure spend totaling USD 10.3 trillion in the United States alone over the seven-year period from 2025 to 2032. This average investment level translates to nearly 4% of the United States Gross Domestic Product—surpassing previous historical infrastructure booms in rail, national electrification, transportation, and telecommunications.
The sheer scale of this capital expenditure clearly surpasses the internal cash resources of individual AI actors. Aggregate capital expenditures by major tech titans including Oracle, Microsoft, Amazon, Meta, and Alphabet surged from roughly USD 97 billion in 2020 to more than USD 400 billion in 2025, with projections indicating expenditures will exceed USD 800 billion in 2026. For the first time, these spending figures are outpacing the companies’ combined operating cash flows.
This reality has forced hyperscalers to dramatically broaden their financing channels. Data center developers, infrastructure investment funds, private equity firms, and bond markets are increasingly utilized to provide vital equity capital, while commercial banks, private credit funds, and securitization vehicles supply necessary debt. These funds cover not only physical real estate, power, and cooling plants, but increasingly complement vendor financing arrangements to cover expensive IT hardware. Investment-grade hyperscaler tenants make these complex financing structures viable by supporting long-duration contractual cash flows.
However, these financial arrangements frequently tend to conceal rather than eliminate risk. Shifting assets into separately financed special vehicles can elevate leverage on the underlying infrastructure, even while reported corporate hyperscaler leverage appears reassuringly low. Long-duration debt is supported by cash flows and collateral values that depend heavily on ongoing market demand, making financial exposures increasingly layered, correlated, and difficult to observe from the outside.
Consequently, extensive AI infrastructure investment is exposed not only to technological and operational risks, but its financing structure actively transmits and amplifies those vulnerabilities. Long-duration debt and contractual claims are being written against physical and digital assets whose long-term usage prospects, technological relevance, and residual market values remain highly uncertain. The severity of any adverse economic shock will therefore depend heavily on the underlying economics of AI market demand, as well as where financial leverage ultimately resides and how potential losses are allocated among tenants, asset owners, and creditors.
A parallel threat stems from rapid technical obsolescence. For over six decades, the silicon chip industry has fueled continuous tech evolution through relentless progress, simultaneously increasing computational capability while driving down the costs and power requirements of microchips. With ongoing access to faster, more capable silicon, the prevailing assumption has always been that computing power will become cheaper over time. However, this creates rapid depreciation in the value of existing IT hardware, maintaining a persistent need for continuous reinvestment just to track the bleeding edge of silicon capability and remain competitive.
Crucially, falling unit costs for manufacturing computational capability do not automatically mean that AI will become cheaper overall. The opposite is far more likely. While AI models may become more efficient, chips may improve, and individual tasks may cost less to process, total spending will continue to skyrocket because overall market demand is growing at an even faster pace. Users increasingly demand longer contexts, multi-modal capabilities, autonomous agents, advanced search, memory, and direct task execution.
Even that trajectory is facing growing uncertainty. The ability of the silicon industry to maintain its historic, predictable 18-month cycle of incremental evolution—increasing chip power while steadily reducing production costs—is encountering profound challenges. If upcoming generations of microchips fail to deliver significant performance or cost improvements, any future expansion of AI capabilities relying on larger, more powerful models will require exponentially larger assemblies of computational components, accompanied by a matching demand for massive data centers and unprecedented power generation.
The mounting burden of corporate debt is another growing concern as AI companies struggle to amortize their massive capital investments. Meta closed 2025 with USD 72 billion in capital expenditures, anticipating between USD 115 billion and USD 135 billion for 2026. Anthropic announced a massive USD 30 billion financing round, valuing the company at USD 380 billion, while forming a strategic alliance with Microsoft and NVIDIA that includes a USD 30 billion commitment to purchase Azure capacity and up to an additional gigawatt of compute power.
Similarly, Google undertook a USD 32 billion global bond issuance spree, tapping debt markets across the United States, Canada, Japan, Europe, and Australia to finance its aggressive AI data center infrastructure buildout. Alphabet indicated that capital expenditures could approach up to USD 185 billion, underscoring the massive computing resources required to support increasingly complex systems. Notably, Google’s issuance of a 100-year bond signals a strategic push beyond traditional equity markets, tapping long-horizon institutional capital such as pension funds and insurance companies whose liabilities favor ultra-long-term assets.
Ultimately, chip designers, fabricators, cloud operators, model developers, data center operators, and infrastructure investors are deeply interlinked through long-term contracts, strategic investments, and intricate financing arrangements. A shock to any single segment—such as weaker demand from model developers or a slowdown in silicon chip refinement—can rapidly propagate as a decline in service operator revenues, lease payments, asset values, and creditor recoveries. This circularity makes industry exposures far more correlated than they appear in isolation, acting as a powerful amplifier if any single link in the chain fails. Artificial intelligence no longer resembles an experimental technology project; it has rapidly transformed into a hyper-capital-intensive industry that must eventually demand profitability, or risk triggering a financial downturn of global proportions.
Making Money with AI
Having examined the staggering expenditures required for AI infrastructure, attention naturally turns to the revenue side of the ledger. The critical question is no longer whether AI will continue to be offered for free to everyday consumers, but rather who will ultimately subsidize these services and what economic assets will be traded in exchange.
Historically, Google mastered a classic two-sided market model to fund its consumer internet services. Search is provided without direct cost to the user; in exchange, Google compiles detailed behavioral profiles of consumers and sells targeted access to advertisers. The more users search, and the more accurate and comprehensive those consumer profiles become, the greater their value to advertisers. Hal Varian, who later served as Google’s Chief Economist, argued in the late 1990s that digital spam was fundamentally a failure of consumer information, implying that deeper user insights allow advertisements to be transformed into helpful suggestions, thereby driving transaction conversions. Google’s advertising ecosystem currently generates roughly USD 300 billion annually. To put it simply, advertisers are footing the bill for consumer search.
However, advertisers did not simply expand their overall marketing budgets to accommodate online advertising alongside existing channels. Instead, digital platforms increased advertising effectiveness while drawing marketing dollars away from legacy mediums. Newspapers and free-to-air television were the primary casualties of this structural shift. Today, the traditional newspaper industry exists as a diminished shadow of its former dominant self, prompting various governments to compel digital giants to financially compensate legacy publishers for lost advertising revenue, as seen with Australia’s News Media Bargaining Code.
Free-to-air television faced comparable pressures, forcing content creators to pivot heavily toward consumer subscription models. Yet these additional household costs remain acutely vulnerable to broader economic conditions. When households face rising interest rates, increasing mortgage payments, and higher fuel costs, discretionary subscriptions are often the first household expenses to be cut. Because the average consumer is not lavishly wealthy, any new spending commitment requires forgoing an existing one.
When evaluating the economics of AI, the capital expenditure component alone reaches an estimated USD 750 billion in 2026. Beyond capital costs, operators face steep expenses for electrical power and cooling, the continuous refinement of software tools, and ongoing infrastructure reinvestment. Consequently, the total annual cost of operating global AI services likely exceeds USD 2 trillion to USD 4 trillion.
In previous technological eras, sweeping assumptions regarding Moore’s Law and continuous silicon efficiency allowed computational power to halve in cost every couple of years, mirroring the mobile phone industry’s ability to pack in new features while maintaining stable retail prices. In an AI world, corporations could initially burn through startup capital to capture market share with free services, anticipating that rapid silicon advancements would eventually drive operational costs down to sustainable levels supported by modest consumer payments. However, if the industry is indeed reaching the limits of traditional silicon evolution, scaling capabilities will become increasingly difficult.
Superficially, the market solution appears simple: users will simply pay more for AI services. These costs may be packaged directly into employer software packages, bundled into corporate office suites, or quietly folded into consumer products until AI infrastructure costs become largely invisible to the end user.
Simultaneously, the familiar advertising-funded model is re-emerging across the landscape. OpenAI recently introduced unlimited access to a lightweight version of its models for free users alongside low-cost tiers, while announcing plans to introduce advertising to support sustained low-cost and no-cost access. Similarly, Google transformed its flagship search engine with its “AI mode,” which uses artificial intelligence to synthesize direct responses rather than merely returning traditional web links. Google reported that AI Mode surpassed one billion monthly users within a year of launch, with queries doubling every quarter. However, industry analysts debate whether this shift has generated incremental advertising revenue or primarily served as a defensive measure to protect existing advertising turf against aggressive incursions by competitors like OpenAI and Anthropic.
Undoubtedly, premium subscription tiers offering advanced models, faster processing, and zero advertising will persist for affluent users. Yet these revenue programs remain highly sensitive to broader economic well-being. As economic pressures mount—ranging from rising interest rates and living costs to aging populations shifting onto fixed incomes—discretionary subscriptions are frequently the first items cut. Because the operational costs of running massive AI models do not scale down simply because user subscription revenues dip, providers face persistent financial vulnerability.
This ultimately brings the discussion back to the concept of “free.” The global digital advertising market is valued at approximately USD 700 billion to USD 800 billion annually. Yet the estimated operating cost of running advanced AI infrastructure easily surpasses twice that amount, hovering between USD 2 trillion and USD 4 trillion per year.
Where Will the Money Come From?
To fund yesterday’s search engines, the digital economy largely disrupted newspapers and free-to-air television. The pressing question facing the modern economy is how the staggering financial weight of artificial intelligence will be funded, and what vital sectors of economic activity society is willing to sacrifice to foot the bill.
Early indicators suggest the answer may involve significant corporate restructuring and widespread workforce reductions. Many employers are betting that they can secure the capital required to fund premium enterprise AI subscriptions by aggressively trimming payroll expenses and replacing human labor with automated AI agents.
While the tech sector has experienced notable waves of mass layoffs over recent years, more extreme projections have circulated. Dario Amodei, CEO of Anthropic, publicly predicted that advanced AI could potentially eliminate up to half of all white-collar administrative jobs and push unemployment rates toward 20%. A labor market upheaval of that magnitude would impose severe financial strain on countless households, shifting substantial economic costs onto the public sector while centralizing income and productivity gains within private technology firms.
Revising societal definitions of full employment to hover around a 4% jobless rate underscores how disruptive a 20-percent workforce reduction would be, pushing nations past a standard economic downturn and toward a severe systemic crisis. Conversely, a strong counter-narrative argues that current labor market data falls far short of forecasting an AI-driven employment collapse. While popular perception often portrays AI as an immediate catalyst for workforce reduction, available labor statistics do not yet reflect widespread, systemic workforce shrinkage directly attributable to artificial intelligence.
Another school of thought suggests that a significant portion of current AI expenditure stems from market hype, and that the practical utility of AI in replacing human workers is far more limited than tech proponents claim. Under this view, AI operations will ultimately need to scale back to fit within realistic advertising and subscription revenues, leaving firms that fail to secure sustainable revenue streams vulnerable to financial failure.
Where Is This All Heading?
It is increasingly likely that the global economy will struggle to meet the astronomical capital costs required to construct every specialized data center, dedicated power grid, and physical IT plant currently on the drawing boards. The required financial investment is simply too massive, and the capacity to continually divert capital into this single sector is finite. Consequently, a number of over-leveraged AI entities will likely face financial distress, leaving surviving competitors to salvage what remains of their infrastructure.
The crucial question remains regarding which market participants are best positioned to survive this turbulent period of capital intensity. Do established tech giants hold an insurmountable advantage through existing revenue streams and massive customer bases, or will agile challengers capture sufficient market share through product innovation and the absence of legacy business models?
The current economic model underpinning the artificial intelligence boom remains unsustainable in its present form, and structural corrections are inevitable. Yet human economic systems have repeatedly demonstrated remarkable resilience. While technologies like Bitcoin were once predicted to trigger a complete armageddon for the traditional banking sector, and AI has been heralded as an absolute shatterer of modern work and leisure, the reality may prove less catastrophic. The changes ahead may unfold at a more measured pace, providing society with greater opportunity to adapt and absorb the transformation as it evolves.
The views expressed by the authors of this blog are their own and do not necessarily reflect the views of APNIC. Please note a Code of Conduct applies to this blog.
Leave a Reply