‘Free’ is a very attractive price when you are trying to break into a new market. ‘Free’ services have been used extensively by Google for search, Gmail, Docs, and many other consumer services. Meta has used ‘free’ for its Facebook platform, and Cloudflare for its web caching services. In a world of constant innovation, ‘free’ is a way of exposing a product or service to the market, hopefully creating a dependency which can be monetized later. These days, it’s used to introduce AI tools.
However, ‘free’ is often a misleading market signal, in that there are very real costs to the service operator that need to be covered. Describing a service as ‘free’ may create the perception that the underlying technology can be provided at negligible cost, with those costs absorbed through indirect subsidies. ‘Free’ also acts as a very effective barrier to competition, in that competitors need to match ‘free’ if they wish to enter or stay in the market.
‘Free’ is not the complete picture, of course. A consumer service can be offered at a price way below its cost to capture market share from competitors (often termed ‘price leading’ in a positive light, or ‘dumping’ in a negative one). Uber posted a USD 5B (yes, billion) operating loss in three months in 2019, and it took thirteen years to become revenue positive. SoftBank, the Saudi Arabian Public Investment Fund and Google all played a multi-billion-dollar funding role in subsidizing Uber through its protracted startup phase.
Then there was DoorDash, which managed a USD 1.4B loss in 2022. At the time, every delivery was subsidized by the startup’s financial backers. Of course, that’s not a novel situation, and the expectation is that once the new entrant becomes established, it can turn the business around and drop the subsidization payments. The startup is being funded to essentially buy rapid growth in the market and place pressure on its competitors.
Let’s use this lens of economic analysis to look at the current state of AI. By all accounts, it is a very costly undertaking. It has been made widely available by a small clique of enterprises who are racing each other to achieve a dominant market share, to establish consumer usage habits, and ultimately generate consumer dependence.
Paying the AI bill
AI is expensive because of the sheer scale of the computation required to generate the Large Language Models (LLMs) that fuel AI systems. The investment is not necessarily a software investment, but more conventional in physical plant and equipment.
The design process for AI data centres is quite simple: Take everything that is at the raw bleeding edge of today’s technology and push them together. These days, a modern AI data centre may require around 60,000 advanced Graphics Processing Units (GPUs), deployed at a density of 72 GPUs per rack, making a total of about 3,000 to 8,000 racks. The GPUs need to be mesh-connected to both each other and to high-speed storage in a lossless connectivity fabric using 800G optics and high-density switches. And then there is mass storage.
All this technology will need a massive amount of continuous power, capable of delivering 150kW to 200kW per equipment rack. A lot of power means a lot of generated heat. Every rack needs to be equipped with liquid cooling, and the centre needs to use a large-scale liquid cooling plant. Using air to cool this hot liquid is not efficient, so it makes sense to tap a large source of water to cool it down.
These plant requirements also call for a power supply system capable of sustaining up to a gigawatt of total power for the plant. That sustained load profile does not match the varying load profile that has dictated how we’ve built our power infrastructure so far. The residential and commercial power profiles that the grid was designed to serve vary across the 24-hour cycle, the week, and season to season. Nor does the AI load profile match the periodic generation profile of solar and wind renewable generators. We can’t just skim off excess production from the existing grid in times of lulls in aggregate demand. These AI data centres call for dedicated substations, high-voltage connections, and significant grid-interconnection investment, and call for additional power generation facilities.
It doesn’t stop there. Obviously, all this is not cheap, particularly as we have drained the world’s supply chains for technology items. You are going to pay a premium price to get your orders filled. That implies that to populate this data centre, you’re going to need substantial financial backing. You’ll also need many local community approvals, particularly if you want to be so bold as to try and use modular nuclear generators for power! And once you’ve done all that, then be prepared to build an even bigger data centre in about 18 months’ time, as most of the centre’s capacity and performance parameters will have doubled!
A representative 200MW AI training campus currently costs roughly USD 8.2B, about two-thirds of which reflects IT equipment costs, and one-third to fund the real estate and associated power infrastructure. If we measure AI data centres by their power requirements, then the currently known construction schedule has 183 GW being constructed by 2032, and a further 118 GW being planned for after 2032. That equates to a total infrastructure spend in the seven years from 2025 to 2032 of USD 10.3 trillion in the US alone. The average investment would be at a level of almost 4% of US Gross Domestic Product (GDP), larger than previous national landmark boom investments in rail, electrification, transportation and telecommunications infrastructure.
This scale of investment surpasses the internal resources of the AI actors. Aggregate capital expenditures by Oracle, Microsoft, Amazon, Meta, and Alphabet rose from about USD 97B in 2020 to more than USD 400B in 2025 and are projected to exceed USD 800B in 2026, surpassing their combined operating cash flow for the first time.
This has forced these hyperscalers to broaden their channels of finance. Data centre developers, infrastructure investment funds, private equity investors and bond markets are used to provide equity capital. Banks, private credit funds, and securitization vehicles supply debt. This is not used just for the buildings and power and cooling infrastructure, but now complements vendor financing to cover the IT equipment. Investment-grade hyperscaler tenants make these structures financeable by supporting long-duration contractual cash flows.
It’s no coincidence that these arrangements tend to conceal rather than eliminate risk. Moving assets into separately financed vehicles can raise leverage on the underlying infrastructure even when reported hyperscaler leverage remains low. Long-duration debt is then supported by cash flows and collateral values that depend on:
- 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 can therefore make exposures more layered, correlated, and difficult to observe
The result is that not only is this extensive AI infrastructure investment exposed to technological and operating risks, but this financing structure can transmit and amplify those risks. Long-duration debt and contractual claims are being written against assets whose prospects of usage, technological relevance, and residual value remain highly uncertain. The severity of any adverse shock will therefore depend not only on the economics of AI demand in the market, but also on where leverage resides and how losses are allocated across tenants, asset owners, and creditors.
There is also the risk of technical obsolescence. Fuelling this technical evolution over the past six decades or so has been the silicon chip industry. Its nonstop progress to simultaneously increase computational capability while at the same time reducing costs and power requirements of silicon chips has been little short of miraculous. With ongoing access to faster, more capable and lower-power silicon chips, it will always be cheaper soon, right? That implies a rapid depreciation of the value of the existing IT equipment, and an ongoing need for reinvestment to track the current extremes of silicon capability to remain competitive.
Just because the unit cost of manufacturing computational capability falls, it doesn’t mean that AI will be cheaper. The opposite is more likely to happen. Models may be more efficient, chips may be better, and some tasks will therefore cost less per token, but total spending will continue to skyrocket because demand is growing even faster. It is apparent that users want longer contexts, multimodality, agents, search, memory, and task execution.
But even that future is not so certain these days. The ability of the silicon industry to continue with incremental evolution to increase the computational power of its chips while simultaneously reducing its unit cost of production on a regular 18-month cycle is heading into a period of profound uncertainty. Maybe next year’s chips won’t be any better or cheaper than this year’s chips. That scenario means that any evolution in the AI capabilities that rely on larger and more powerful computational models will only be feasible with larger and more expensive assemblies of computational components — and a matching demand for larger data centres with larger power demands.
There is also the problem of a growing debt, as AI companies need to amortize their investments. Meta, for example, closed 2025 with USD 72B in capital expenditure, and anticipated between USD 115B and USD 135B for 2026. In February. Anthropic announced a USD 30B financing round (valuing it at USD 380B). Microsoft, NVIDIA and Anthropic have sealed an alliance that will see Anthropic commit to buy USD 30B of Azure capacity and up to an additional gigawatt of compute.
Google has undertaken a USD 32B global bond spree in 2026, tapping the bond markets in the US, Canada, Japan, Europe, and Australia to finance its aggressive AI data centre infrastructure buildout. Google’s parent company, Alphabet, has indicated capital expenditures could approach roughly USD 175B to USD 185B in 2026, underscoring the scale of computing resources required to support increasingly complex AI systems. Their 100-year bond also signals Alphabet’s push beyond traditional equity investors, tapping long-horizon institutional capital such as pension funds and insurers whose liabilities favour ultra-long assets, reinforcing a broader shift toward infrastructure-style financing as AI investment intensifies.
Chip designers, chip fabricators, cloud operators, model developers, data centre operators, and infrastructure investors are interlinked through long-term contracts, strategic investments, and financing arrangements. A shock to one segment, such as weaker demand from model developers or the end of the — to date — continuous process of silicon chip refinement, can propagate as a fall in service operator revenues, lease payments, asset values, and creditor recoveries. This circularity makes exposures more correlated than they may appear when each transaction is viewed in isolation. The interdependence can act as an amplifier when any single activity fails.
AI no longer looks anything like an experiment: It looks like a capital-intensive industry that sooner or later will demand profitability, or it will dictate the terms of a rather impressive financial crash of global proportion.
Making money with AI
Now we’ve investigated the expenditures for AI infrastructure; now let’s turn our attention to revenues. The question now is not whether AI will continue to be free for consumers to use, but who will subsidize it and the nature of the asset being traded in exchange for this subsidy.
Google used a classic two-sided market model to fund its retail services. For example, Google’s search is provided without direct cost to the consumer, but to fund this, Google assembles a profile of each consumer that it then sells to advertisers. The more users use Google’s search, and the more accurate and comprehensive the assembled profiles, the greater the value of the profile to the advertisers.
Economist Hal Varian (Later Google’s Chief Economist) argued in the late 1990s that spam was essentially a failure of information about the consumer. The implication was that the greater the level of information about the user, the better the ability to turn an ad into a helpful suggestion, and the higher the probability of converting a presented advertisement into a transaction. The greater the revenue available to support search, the better the ability to invest in assembling user profiles of search users, and the higher the value of Google as an advertising platform. Google’s advertising activity currently earns USD 300B per year. To put it simply, advertisers are paying for search.
But advertisers did not magically increase their advertising budgets to add online advertising to the portfolio of existing channels to market. In fact, the deal that Google offered was to increase the effectiveness of their advertising programs, and potentially to decrease their total advertising spend. What did advertisers pull back on?
The answer is obvious. Newspapers and free-to-air television were the major victims in this change of advertiser behaviour. These days the newspaper business is an impoverished shadow of its former dominant self. In some cases, governments are forcing these digital giants, who took legacy media’s historical advertising revenue, to pay back some small percentage of their forgone income (see the Australian News Media Bargaining Code as a good example of this).
Free-to-air television had little in the way of comparable bargaining capability, and the content industry turned to the consumer subscription market to fund their production models. Such additional costs to consumers are subject to changes in consumer spending. When times are tougher, when rising interest rates increase household mortgage payments, and when petrol costs rise, one of the first items of discretionary expenditure to go is content subscriptions. In most markets, consumers are not lavishly wealthy, and any new spending commitment is made by forgoing some other existing spend.
So now let’s turn our attention to AI and its costs. The capital expenditure component is USD 750B in 2026 alone. But the cost of that capital is not the only cost of AI. There is the cost of power (and cooling), the cost of continued refinement of the tools and services, and the cost of capital reinvestment. It would not surprise me if the total annual cost of operating AI services already exceeds USD 2 to 4 trillion.
Now, in a previous age, we could make some sweeping assumptions about Moore’s Law and the continued drive to more efficient silicon. The same level of computational capability would cost half as much in a couple of years, and this trend would continue in subsequent years. We could follow the path taken by the mobile phone industry, adding ever more features while maintaining roughly the same retail price and encouraging consumers to replace perfectly functional devices every few years.
In an AI world, we could chew through startup funds to gain market share for a free service, in the hope that in a few years, with the continuing functioning of Moore’s Law in silicon chips, the operational costs will come down to a level that is sustainable by consumer payments. But what if we are at the end of the silicon road? What if we can’t continue to improve the feature count and reduce the power demands of silicon chips?
Superficially, the answer is simple: Users will pay more for AI services. The cost may be packaged into services provided by employers or bundled with products such as office suites. AI is likely to evolve into an infrastructure cost, folded into other products and services until it becomes largely invisible to the end user.
The familiar advertising-funded model is also re-emerging. OpenAI recently introduced unlimited access to Luna, a lightweight version of its GPT-5.6 model, for free ChatGPT users and subscribers to its low-cost Go tier. It also announced plans to introduce advertising for users on the Free and Go tiers as a way to support continued low-cost and no-cost access to its AI services.
Google has been transforming its search service. In its ‘AI mode’, the search engine no longer returns pointers to resources that look relevant to your query. Instead, it uses AI synthesis of content from these resources to try and generate a direct response to the query. As Google asserts in a product promotion, “Just one year after its debut, AI Mode has surpassed 1B monthly users, with queries more than doubling every quarter since launch”. But the real question is whether this change in Google’s search product has managed to generate increases in advertising revenue. Or is it a more basic form of last-ditch defence of existing advertising revenue against the incursion into their market from OpenAI and Anthropic? The latter looks like the most likely answer.
No doubt the premium subscription services for AI will persist. These subscriptions will offer better models, faster processing, more capable responses, and of course zero advertising. But this is only accessible for those who can afford to pay, and only while they have a volume of discretionary income that can afford such a subscription.
As with the streaming models, such revenue programs are highly dependent on a broader environment of economic well-being. As we’ve already noted, when the costs of interest rates, petrol, food, and clothing rise, or when a greater proportion of an aging population shifts into a fixed retirement income, such discretionary subscriptions are often the first to be cut. The costs of running these AI services are not necessarily based on the volume of queries. The situation of falling subscription revenues without an accompanying fall in costs is a very real prospect.
We inevitably circle back to ‘free’. The digital advertising market is valued at approximately USD 700 to USD 800B per year. The operating cost of AI appears to be easily more than twice that amount, with a very rough estimate of USD 2 to 4 trillion per year.
Where will the money come from?
To fund today’s search engines, we destroyed newspapers and free-to-air television. How do we fund AI? What part of the spectrum of economic activity are we willing to chop off and forego in order to generate the rest of the money?
We might already be seeing the answer.
We are going to find the capital needed by AI by shedding people and the jobs they used to do. The employers of these laid-off workers appear to be betting that they can save sufficient money through a reduced wages bill to fund their premium corporate AI subscriptions to do the same function.
Many tech companies in the past few years have had experience with mass layoffs. As a more extreme illustration, Dario Amodei, CEO of Anthropic, has predicted that AI could wipe out half of all white-collar jobs and push unemployment to 20%. A labour market upheaval of this magnitude would cause enormous suffering for many households and pose a significant challenge for policymakers, shifting costs to the public sector while retaining income and value in private hands.
We’ve already revised our social definition of ‘full employment’ to equate to an unemployment rate of around 4%. At a 20% job loss, we would be leaving the territory of economic ‘downturn’ and entering a full-blown economic ‘meltdown!’
There is a strong counter voice to this dire prediction of widespread job losses due to AI. The figures so far fall way short of a collapse in the labour market due to AI. While there is a popular perception that AI will act as a catalyst for a shrinking workforce rather than sustaining growth, the available data so far does not show any clear signs of significant workforce shrinkage.
Another perspective is that all this AI expenditure has been an exercise in mass hysteria, and the reality of AI is far more limited in terms of its ability to replace people with AI agents. This perspective would see AI fitting within the confines of what can be done within the existing advertising revenues, and those who fail to secure a significant proportion of this revenue stream are doomed.
Where is this all heading?
It’s highly likely that we will be unable to meet the total costs of the construction of all these specialized data centres, power systems, and IT plants. It’s just too much money, and we can’t or won’t divert sufficient capital into this sector to sustain everyone’s borrowings. It’s likely that some of these AI entities will fail financially, and the others will pick over the leftovers of the failed enterprises to see if there is anything worth salvaging.
Who is more likely to remain standing at the end of this period of AI hype? Do the incumbents have the advantage of existing revenue and a customer base that will sustain them through this? Or are the challengers able to take sufficient market share and revenue through innovation and the absence of legacy business models that may weigh down the incumbents?
The current state of play in AI appears to be financially unsustainable, and changes will happen as we head along this journey. But there is a lot to be said for the resilience of our societies and their institutions. Bitcoin was meant to be the finance-and-banking industry’s Armageddon. AI has been sold as an agent of shattering change in the nature of the way we work and play. But maybe that’s just being overdramatic, and the changes we will go through in the coming years will occur at a slower pace and with more opportunity to adjust as we go along.
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.