Why this matters to your business: The scale of capital being deployed into AI infrastructure is the single strongest signal that AI tool prices will continue falling over the next two to three years. $570 billion in AI-linked debt is not a market speculation bet — it is infrastructure investment that needs to generate returns through AI service revenues, which creates sustained pressure on vendors to grow their customer base by keeping prices accessible.
What Morgan Stanley is projecting
Morgan Stanley's 2026 forecast projects global AI-linked debt to nearly double from approximately $295 billion in 2025 to $570 billion in 2026. This category includes bonds, loans, and structured finance instruments where the proceeds are specifically earmarked for AI-related capital expenditure: data centres, GPU clusters, networking infrastructure, and the power systems that feed them.
The $36 billion Apollo-Blackstone structured debt facility funds Google's TPU (Tensor Processing Unit) manufacturing purchases, which are then used in the data centres running Anthropic's Claude models and Google's own Gemini models. Microsoft's $190 billion capex commitment — spread across Azure data centres globally — includes AI infrastructure for OpenAI's models and Microsoft Copilot services. Amazon's chip business, reportedly sold out through late 2026, reflects a similar infrastructure investment cycle on AWS. The common thread: every major AI model vendor is backed by infrastructure capital at unprecedented scale.
What this means for AI tool pricing
Infrastructure investment at this scale has a direct and predictable effect on the economics of AI tool pricing. When a data centre is financed at $10 billion in debt, the debt service requires ongoing revenue. That revenue comes from AI API calls, subscription fees, and enterprise contracts. The vendors have a strong incentive to grow their customer base — which means keeping prices accessible enough that adoption continues to expand.
Since March 2023, the cost of AI API access has fallen by roughly 97% on a cost-per-token basis — from approximately £24 per million tokens for GPT-4 to approximately £0.60 per million tokens for mid-tier models today. This did not happen because vendors became more generous. It happened because hardware costs fell, training efficiencies improved, and competition intensified as Anthropic, Google, and Meta entered the market. The $570 billion infrastructure investment accelerates all three of these mechanisms simultaneously. More infrastructure means lower marginal cost per inference. More competition means more pressure to pass savings on. The direction of travel is clear and the 2026 infrastructure investment makes it more durable, not less.
The caveat: energy costs are the floor
Not every cost in AI inference falls with scale. Energy costs — the electricity required to run the GPU clusters — are a real and growing constraint. The FERC grid fast-lane order reported this week is a US government attempt to solve this specific problem by expediting power grid connections for AI data centres. The UK faces similar energy infrastructure constraints, which affect the UK data centres serving UK customers.
For UK businesses, this means AI tool prices will continue falling, but not to near-zero. The floor is set by energy costs, and that floor is higher than it was two years ago. Build your AI business case on current pricing, not on an assumption that prices will halve again in twelve months. They may fall 30 to 40% over the next year — that is the realistic range given energy constraints — but the 97% compression of the last three years is not repeating.
What to do
Two strategic implications for UK SMBs
First: For AI platform commitments of two or more years — a specific tool, an API integration, a software stack — prefer vendors with owned infrastructure over pure resellers. The vendors backed by the $570 billion capital cycle (OpenAI via Microsoft Azure, Anthropic via Google/AWS, Google's own models) have the infrastructure security and vendor stability to support long-term commitments. A reseller dependent on a single upstream API has pricing and availability risk that infrastructure-backed vendors do not.
Second: Do not defer AI workflow implementation waiting for prices to fall further. The meaningful compression has already happened. Waiting another twelve months to save a potential 30 to 40% on per-token costs while competitors who started today are twelve months ahead in productivity compounding is a poor trade. The price advantage of waiting is real but small; the productivity advantage of starting now is large and compounding.
The Morgan Stanley $570 billion projection is ultimately a vote of confidence in AI as permanent, capital-grade infrastructure. Financial institutions do not finance $570 billion in AI-linked debt on speculative grounds. They do it because the revenue model is proven, the demand is established, and the infrastructure cost curves are clear. For UK small businesses, that confidence is worth more than the headline number: it means AI tools will still exist, will still improve, and will still get cheaper in two years' time. Build your AI strategy accordingly.