Casino Capitalism: Inside the financial architecture powering the AI boom

Casino Capitalism: Inside the financial architecture powering the AI boom

Capital is being committed faster than realised cash flows. Current valuations imply substantial revenue growth far into the future.

Published Sep 11, 2026 | 2:05 PM ⚊ Updated Sep 11, 2026 | 2:32 PM

Image used for representational purpose. (AI generated image)

Synopsis: The question is whether the financial architecture currently funding the AI boom is aligned with the speed at which economic value can realistically be created and realised. For India, the question is what form its participation in the AI economy should take.

Between 2022 and 2026, roughly $1.75 trillion was invested in artificial intelligence infrastructure. Over the next five years, another $4 to $8 trillion is likely to follow.

Data centres are being built at an unprecedented pace. Power-generation projects are being redesigned around future AI demand. The world’s largest technology companies are investing on a scale normally associated with national infrastructure programmes rather than private corporations.

The underlying assumption is that artificial intelligence is a general-purpose technology comparable to electricity, railways, or the internet. If that assumption is correct, today’s spending may eventually look prudent, even conservative.

But there is a second possibility.

History shows that transformative technologies and successful investments are not the same thing. Railways transformed nineteenth-century economies while destroying large amounts of investor capital. The internet changed the world while much of the first generation of dot-com investment disappeared. Telecommunications infrastructure eventually proved indispensable, yet many of the companies that built it failed.

The question, therefore, is not whether AI works.

This raises a question that would have been immediately familiar to both John Maynard Keynes and Hyman Minsky. Are we witnessing a period of productive capital formation that happens to look speculative because technology is new, or is it a boom in which genuine technological progress is becoming increasingly dependent on financial engineering or speculative finance?

The distinction matters not only for investors and technology firms. It matters for governments, power systems, public finances and countries such as India that are now deciding how deeply they should participate in the AI build-out.

The debate, ultimately, is not about whether AI is transformative. The debate is whether transformative technologies can be financed in ways that remain economically, financially and politically sustainable, both for advanced as well as developing economies like India.

Goldman Sachs frames the AI buildout as a multitrillion-dollar infrastructure problem, with widely discussed estimates of $4 trillion to $8 trillion of total capital investment over five years and a baseline model implying about $7.6 trillion of cumulative AI infrastructure capex between 2026 and 2031 across compute, data centres, and power.

Amazon, Microsoft, Alphabet, Meta and Oracle will spend $725–760 billion on AI capex in 2026 alone — a 77%+ jump in a single year (Figure 1 below).

With the current level of investment itself, the question is:

Can these firms monetise demand quickly enough, at sufficient margin, to justify assets with short economic lives?

Evidence of monetisation is emerging, but its scale remains modest relative to the pace of investment. A recent survey of nearly 6,000 executives across the United States, United Kingdom, Germany and Australia found widespread adoption of AI tools, yet more than 80% of firms reported no measurable impact on productivity or employment during the previous three years.

At the same time, these same firms anticipate significant gains in the years ahead. The gap between realised outcomes and expected future benefits illustrates a central feature of the current AI cycle: expectations are running well ahead of observed economic effects.

Another NBER paper suggests that AI is producing measurable improvements in specific tasks, such as email management and document preparation, while broader organisational changes remain slower and more difficult to realise.

Daron Acemoglu of the Massachusetts Institute of Technology reviewed the macroeconomic implications of AI and acknowledged non-trivial task-level gains, but estimates that aggregate TFP (Total Factor Productivity) gains may be modest, less than 0.53% to 0.66% over ten years under his assumptions – highlighting the possibility that infrastructure spending may run well ahead of realised economic impact.

AI may well prove transformative while still being overbuilt in the short run, because infrastructure investment is growing much faster than measured productivity gains (Figure 2 above).

Many other AI companies are still very small, though their run rate is high.

Figure 2. Revenue split: AI-specific vs. rest of business.

While the monetisation evidence is real, they are still not commensurate with the capex run rate or the expenditure by AI customers. Consequently, the companies are having to depend on external sources of finance, including debt.

All the hyper-scalers are raising or looking to raise external equity or debt. For example, Alphabet has raised USD 80 billion in equity along with a large amount of debt. Oracle announced that it would raise USD 45-50 billion during 2026. Meta is not far behind in building the war chest.

A large proportion of new equity is coming at peak valuation ratios, and debt is not being priced for potential risk and uncertainty associated with a disruptive technology that is at the early stages of delivering the required cash flows for large-scale investors.

A BIS (Bank for International Settlements) study concludes the following:

“While macroeconomic and financial stability risks from the AI boom appear moderate, the boom’s sustainability hinges on AI firms meeting high earnings expectations. The fact that equity prices have run far ahead of debt market pricing underscores this tension.”

Another BIS study highlights the fact that many AI infrastructure obligations are economically debt-like but sit outside reported corporate debt metrics via data centre partnerships, leases, and SPVs (Special Purpose Vehicles).

BIS argues that this creates new channels of financial contagion if AI revenues disappoint. It calls it shadow borrowing. In short, it is not only about risk arising from a correction in equity valuation, but we also have a situation where leverage and financial risk in the AI ecosystem are rising rapidly.

Moody’s, on the other hand, as quoted by CNBC suggests the possibility of a structural circularity within the AI boom.

Goldman Sachs research, cited earlier, argues that the economic useful life of AI silicon, the riskiest asset, is one of the most important variables shaping the scale of the AI build-out. Unlike data-centre buildings or power infrastructure, which may remain productive for decades, AI accelerators face rapid technological obsolescence. As a result, relatively small changes in replacement cycles can alter cumulative industry investment by hundreds of billions of dollars

Microsoft seems to have recognised the challenge in determining the economic life of some of its assets by bringing in an accounting change that has extended the life of its data centres and office buildings from 15 to 25 years. The policy change allows it to classify some of the financial leases as operating leases, lowering its reported capital expenditure level.

The question, therefore, is: Will AI monetisation, utilisation, and productivity diffusion arrive before depreciation, obsolescence, and financing pressure force a capex reset?

If valuations are to sustain, even the AI-firms need revenue to grow multiple times.

The New York Fed also warns that AI can create systemic vulnerabilities through expectation-driven asset valuations, adoption frictions, and model monocultures, especially if realised efficiency gains lag elevated asset prices.

The Federal Reserve’s May 2026 Financial Stability Report says broad equity valuations remain elevated, corporate spreads remain low by historical standards, and bond issuance among large investment-grade cloud-computing firms neared $100 billion in Q1 2026 with strong demand.

To make today’s valuations merely reasonable rather than extraordinary, OpenAI’s revenue would need to grow 4.3x and Nvidia’s 2.5x, even using a conservative 8x sales multiple [Fig. 3] — not the 20–34x multiples the market is currently paying.

Figure 3. Current revenue vs. revenue implied by an 8x “mature software” multiple

AI is a general-purpose technology, but the monetisable products built on top of it may have short product lives, weak defensibility, rapid price erosion, and high replication risk.

In the .com era, the internet was enduring, but many first-wave businesses were not. The same may happen in AI: model wrappers, copilots, agents, coding tools, search interfaces, vertical SaaS features, and workflow bots may be easy to replicate, bundle, or price down.

That creates a problem for infrastructure investors. If application-layer margins compress quickly, the ability to pay high compute rents weakens. The infrastructure owner then needs either massive volume growth, strong platform control, or bottleneck power to earn returns.

The AI capex model requires four things to be true at once:

However, power access and data-centre capacity can be scarce at a point in time, in a specific location, and under a particular permitting/grid constraint, but they are not structurally scarce in the way a natural monopoly railway route or urban electricity distribution franchise once was.

If AI demand disappoints or becomes more compute-efficient, power plants, data centres, substations, and GPU clusters are underutilised; returns fall quickly; debt and lease obligations remain; asset impairments appear on balance sheets; grid and environmental remediation costs may be shifted to ratepayers or the State, and local ecological costs, such as land use, water use, transmission corridors, and waste heat, become politically salient.

We will be facing the classic overbuild outcome, which will resemble telecom fibre in financial terms, but with a more visible physical and ecological footprint. Telecom overcapacity lowered bandwidth prices and hurt investors. AI-data-centre overcapacity could also leave behind stranded power contracts, underused grid upgrades, water-intensive cooling infrastructure, and debt-laden special-purpose vehicles.

If compute demand does explode, that is not costless either. It may produce higher local electricity prices; grid congestion; delayed interconnections for other industrial and household users; public subsidy pressure; accelerated fossil backup capacity; transmission and distribution costs socialised through tariffs; household backlash if data centres are perceived to crowd out ordinary consumers.

Consequently, the “bull case” also carries political economy risk. Even if AI demand validates the capex, the system may not be able to deliver that capacity at short notice without shifting costs to households and governments.

AI capex is exposed to a double commoditisation problem: the end products may have short monetisation windows, and the enabling infrastructure may lose scarcity faster than investors expect. Power and data centres are bottlenecks today, but bottlenecks are not the same as durable moats. If capacity overshoots, returns collapse. If demand explodes, costs are socialised through grids, tariffs, subsidies, land use, and ecological pressure. Either way, the State and households may absorb part of the adjustment.

Keynes would worry that AI infrastructure investment is increasingly being driven by expectations about future capital-market sentiment rather than demonstrated future yield. Capital is being deployed faster than realised operating cash flows and is being financed using external equity and debt, much ahead of uncertain demand.

He would have reminded us of his words, “Speculators may do no harm as bubbles on a steady stream of enterprise. But the position is serious when enterprise becomes the bubble on a whirlpool of speculation. When the capital development of a country becomes a by-product of the activities of a casino, the job is likely to be ill-done.”

It is, therefore, likely that Keynes would see the current situation to be a situation where markets are not serving productive economic activity, but productive activity is becoming secondary to financial trading. Capital allocation is being driven by market excitement, momentum, and gambling-like behaviour rather than by expected long-term returns.

Minsky argued that long periods of successful growth encourage firms, banks and investors to take progressively greater financial risks. The very success of the system alters behaviour in ways that make the system fragile. The recent experience of Situational Awareness is illustrative. A fund established to capitalise on the AI investment boom reportedly suffered severe losses after leveraged positions moved against it, forcing asset sales and highlighting how quickly confidence, leverage and market concentration can interact when expectations change.

Minsky’s question would therefore be simple: when investors begin demanding cash-flow-based returns rather than growth stories, will today’s financing structures still appear sustainable?

The significance of the AI boom extends beyond investors, technology firms and capital markets. A major technology cycle eventually is a public-policy question because infrastructure, power systems, labour markets and industrial capabilities are not allocated by markets alone.

Railways reshaped national development strategies. Electrification became a question of industrial policy and public investment. Telecommunications evolved from a private sector innovation into a strategic national asset. Artificial intelligence is expected to follow a similar trajectory.

If Keynes and Minsky are correct that periods of technological optimism often encourage capital to move faster than underlying economic returns, the consequences cannot be confined to shareholders. Decisions about where AI infrastructure is built, who pays for supporting energy systems, which countries capture the productive gains, and which countries absorb the risks become matters of national importance.

These questions are particularly relevant for countries that are not currently at the technological frontier but are being asked to participate in financing, hosting and supporting the physical infrastructure of the AI economy.

India provides an important example.

For India, the AI boom presents a different set of challenges from those facing the United States or China.

The United States hosts many of the world’s leading AI firms. China is investing across the entire technology stack, from chips and infrastructure to applications and industrial deployment.

India enters the AI era from a different position: it has a large digital economy, deep software development capabilities, growing infrastructure ambitions and significant development priorities competing for the same capital, energy and policy attention.

The question is not whether India should participate in the AI economy. It already has. The question is what form that participation should take.

These questions matter because the economics of AI infrastructure remain uncertain. If current projections prove correct, India could benefit from a substantial wave of investment, technological capability and digital infrastructure. If current assumptions prove overly optimistic, India risks absorbing part of the costs through subsidised power, water use, land allocation and stranded infrastructure.

It is against this backdrop that India’s recent AI commitments need to be assessed.

At February’s AI Impact Summit in New Delhi, Reliance Industries pledged $110 billion over seven years for AI and digital infrastructure; Adani Group committed $100 billion by 2035 to expand its data-centre platform from 2 to 5 gigawatts; the Indian government set a target of over $200 billion in near-term AI infrastructure investment; Microsoft, Google, OpenAI (via Tata) and Meta (via Reliance) all announced expansions. By one estimate, roughly $400 billion flowed into India’s AI ecosystem in the twelve months to mid-2026 [Fig. 4].

Figure 4. India’s AI infrastructure commitments, as announced through mid-2026.

That is real capital and real ambition, arriving in a country whose per-capita income — around $2,813 a year — sits well inside the global bottom half.

The government’s India AI Mission has assembled tens of thousands of subsidised GPUs, almost entirely Nvidia hardware sitting under US export jurisdiction — a dependency that stopped being theoretical in June 2026, when a US directive forced Anthropic to suspend access to its newest models for foreign nationals, India included.

These commitments reflect both confidence and urgency. Yet the uncertainties discussed earlier in this article mean that India should think carefully about the role it seeks to play in the emerging AI economy.

Not every country needs to compete across every layer of the AI stack. The United States and China are investing heavily in frontier model development, advanced semiconductors and hyperscale computing infrastructure. India’s comparative advantage may lie elsewhere.

The critical question is whether India gains more from replicating frontier-scale training infrastructure or from becoming the world’s leading platform for multilingual deployment, domain-specific applications and large-scale adoption in sectors such as agriculture, healthcare, education, logistics and public administration.

The choice is not simply technological. It is strategic. Every rupee committed to frontier compute is a rupee unavailable for data systems, digital public infrastructure, research institutions and human capital.

The economics of AI depend not only on GPUs and software but also on electricity, water, land, transmission networks and public infrastructure.

India therefore faces a difficult allocation problem. Should scarce reliable electricity be directed toward AI data centres before households, manufacturing, railways, hospitals, cold chains and small enterprises?

Are data-centre operators paying the full cost of grid upgrades and supporting infrastructure?

Are states assuming fiscal risks through subsidies, land concessions or preferential tariffs whose long-term costs remain unclear?

Much of the global AI debate focuses on model capability. India’s challenge is different.

The largest productivity gains will not come from access to more computing power alone but from improvements in data quality, administrative capacity, institutional redesign, procurement systems and trust. An AI model can assist a teacher, nurse, farmer, judge or local official only when it is embedded within effective processes and institutions.

The relevant question is therefore not simply how much compute India can acquire, but how effectively AI can be deployed to solve Indian problems.

Perhaps the most important question concerns contingent liabilities.

If AI demand grows more slowly than expected, who absorbs the losses from excess data-centre capacity, long-term power contracts and subsidised infrastructure? If demand grows faster than expected, who pays for the grid upgrades, transmission investments and system-level costs required to support it?

The answers matter because the AI boom is occurring at a moment when the economics of the industry remain uncertain. Public policy should therefore distinguish carefully between supporting innovation and socialising private investment risk.

India’s objective should not be to win a race defined elsewhere. It should be to ensure that AI contributes to productivity, capability and broad-based economic development while avoiding the costly mistakes that often accompany large technological booms.

Railways transformed economies while bankrupting investors. Telecommunications reshaped the world while many first-wave builders disappeared. The internet created extraordinary social value while much of the capital deployed during the dot-com era failed to earn satisfactory returns.

Artificial intelligence may ultimately follow a similar path.

The more difficult question is whether the financial commitments now being made are proportionate to the pace at which economic value can realistically emerge.

That question matters not only for investors. It matters for governments allocating public resources, for energy systems absorbing unprecedented demand, and for countries such as India deciding how they should participate in the next technological cycle.

The debate, ultimately, is not between believers and sceptics. It is between technological possibility and financial sustainability.

The first may prove extraordinary. The second remains an open question.

The consequences of AI may not be confined to financial markets. A second set of questions concerns employment, bargaining power, wage formation, and the distribution of AI-generated economic rents. These issues deserve separate treatment and will be examined in the next set of articles.

What AI can’t teach: The power of serendipity in learning

‘AI could kill us all by the end of the decade’: Anthropic researcher resigns with stark warning

Note: Generative AI tools were used in a developmental editing role during revision of this article. The authors remain responsible for all ideas, interpretations, examples, and final content.