AI is Finance's most valuable commodity, so why are returns still marginal?

In 2026, artificial intelligence has become the most aggressively accumulated capability in global finance. Hyperscalers, banks and markets are committing infrastructure-scale sums on data centres, GPUs and cloud capacity. AI is treated as a strategic commodity in global finance, yet the economic returns from this investment remain stubbornly narrow and uneven.

Aggregate capital expenditure at the largest US cloud and technology groups is projected to reach around USD 660 billion in 2026, roughly two-thirds higher than in 2025, with most of the increase driven by AI infrastructure rather than traditional IT. Their spending now resembles that of regulated utilities, with some investing close to half of annual revenue and peers such as Amazon, Alphabet and Meta also running extremely high capex.

Debt markets have adjusted quickly. Digital-infrastructure securitisation in the US - primarily asset-backed securities and commercial mortgage-backed deals - has grown sharply in under five years, with a rising share of issuance now backed by data centres. In 2025, data-centre ABS and CMBS tied to digital infrastructure totalled about USD 18.6 billion and are expected to exceed USD 35 billion this year, as operators refinance build-outs into long-dated, infrastructure-style structures. Infrastructure CLOs backed by data-centre loans have emerged as a small but growing product, offering investors extra spread over traditional corporate CLOs in exchange for underwriting long-dated, AI-linked cash flows.

Goldman Sachs reports that internal systems now draft roughly 95 per cent of an IPO prospectus in minutes, compressing work that previously took a six-person team about two weeks and freeing bankers to focus on judgment and client negotiation instead. Deloitte predicts that AI tools could help banks cut software-development costs by 20-40 per cent by 2028, primarily through higher engineering productivity, and several large firms are already reshaping their analyst intake as AI automates routine analytical and documentation work.

JPMorgan provides another signal. Its technology budget for 2026 is expected to rise by around 10 per cent compared with 2025, with management explicitly linking the increase to AI and broader technology investment. The bank reports operations-productivity gains of several percentage points per year, roughly double its historical trend, as AI is deployed across document processing, call-centre support, retail risk scoring and fraud detection. Exchanges and post-trade utilities show parallel patterns: Nasdaq is embedding AI into market-surveillance systems used by dozens of exchanges and regulators, while DTCC is deploying AI tools to support risk calculations and post-trade checks.

Despite massive investment, the economic returns from AI remain thin. A 2025 MIT study found about 95% of corporate generative AI pilots failed to deliver measurable financial gains, with only a small fraction scaling into revenue or cost-saving systems. Cloud providers have inflated demand through heavy incentives - Amazon alone pledged hundreds of millions of dollars in AWS credits - so much of today's AI use is still subsidised rather than self-sustaining.

Three structural frictions explain why AI looks like finance's most valuable commodity in theory but delivers marginal returns in practice.

The first is legacy infrastructure. The core plumbing of the financial system - real-time gross-settlement engines, securities depositories, central counterparties and main ledger systems - still runs on long-established mainframe and messaging stacks. It is relatively straightforward to layer AI around these systems, automating documentation, triaging alerts or supporting call-centre staff. It is far harder to redesign them so that AI sits in the execution path for payments, settlement or collateral movements, because those changes require multi-year programmes, complex cutovers and, often, coordinated industry action. As long as the core remains largely untouched, AI's impact will be powerful but peripheral.

The second friction is governance and liability. In regimes such as the UK's Senior Managers and Certification Regime, specific individuals are personally accountable for failures in critical functions, making it difficult to delegate binding, non-reviewed decisions on margin calls, default management or settlement finality to probabilistic systems. Global and international standard-setters highlight AI's capacity to amplify vulnerabilities through third-party concentration, correlated model failures and cyber-risk, and warn that dependence on a small group of providers could drive herding and procyclical behaviour in stressed markets. Until governance, explainability and assurance frameworks adapt, supervised institutions will keep AI at arm's length from their most critical control points.

The third friction is concentration and geopolitics. A handful of hyperscalers and frontier model developers dominates the AI stack in finance, creating operational and geopolitical single points of failure that supervisors are only starting to tackle. In the UK and EU, major cloud and AI providers are being treated as 'critical third parties', external to core market infrastructure, which invites tighter oversight but also reinforces caution about embedding them in systemic processes. For financial firms, concentration risk is now a live supervisory issue that shapes technology strategy. As a result, AI still sits as an overlay rather than part of finance's core machinery: it is changing how institutions work, but it has not yet been allowed to remake settlement, clearing or the transfer of systemic risk.

None of this is to say AI will not, in time, justify its treatment as a core commodity in finance. A small group of institutions already use it to cut the unit cost of core activities or expand capacity without matching headcount growth. For the wider industry, however, AI remains in transition. Regulators still treat AI as a potent but potentially destabilising tool, and most programmes struggle to show durable returns. If AI is indeed finance's most valuable emerging commodity, the priority now is integration, not additional capex: re-engineering legacy systems, building robust safeguards and explainability, and demonstrating, in earnings rather than narratives, that the investment is warranted. Until then, AI will sit uneasily between promise and proof.