The economics of artificial intelligence are increasingly being written in a currency most investors have never priced before: the token. As enterprises push generative AI and autonomous agents into production, the volume of tokens consumed, and the supply available to meet that demand, has become one of the clearest signals of where value, cost, and eventually returns will land across the AI stack.
THE PARADOX AT THE CENTER OF TOKEN ECONOMICS
Per-token prices have collapsed at a pace with few parallels in modern computing history. Research from Epoch AItracking state-of-the-art model performance across major benchmarks found that inference prices required to hit a given performance milestone have fallen between 9x and 900x per year, depending on the task, with the fastest declines emerging only in the past year. Andreessen Horowitz has coined the term “LLMflation” to describe this deflationary curve, likening it to a Moore’s Law for language models, and enterprise spending data compiled by Ramp shows the average cost per million tokens across major providers falling from roughly ten dollars to two and a half dollars in a single year.
Yet falling unit prices have not translated into falling bills. Total enterprise AI spending keeps climbing even as the cost of any individual token keeps shrinking, a dynamic often described as the token cost paradox. The explanation lies in how AI is actually being used. Multistep agentic workflows, in which a system reasons iteratively, calls tools, and verifies its own output, can trigger ten to twenty model calls to complete a single task that once required one. That volume effect, layered on top of retrieval-augmented pipelines that inflate context windows and always-on monitoring agents that run continuously, means usage growth has outpaced price declines by a wide margin.
WHAT IS ACTUALLY DRIVING SUPPLY AND DEMAND
Token supply, in practical terms, is a function of compute capacity: chip availability, data center buildout, and the throughput each new generation of hardware can deliver. On the demand side, analysis from Boston Consulting Groupidentifies the forces pushing consumption higher within individual organizations. Token cost grows as users move from simple chat interfaces into multistep research, code generation, and full workflow orchestration. Task intensity varies enormously, since a short query and a long-running autonomous agent session can look identical on the surface while carrying very different economics. Context and loops compound the effect further, as agents carry forward instructions, history, and retrieved documents across repeated cycles. And model selection matters just as much, since defaulting to the newest, most capable frontier model for every task is a design choice with direct billing consequences.
On the supply side, the buildout underway to meet this demand is enormous. The five largest US cloud and AI infrastructure providers, Microsoft, Alphabet, Amazon, Meta, and Oracle, have collectively committed to roughly $660 billion to $690 billion in capital expenditure for 2026, nearly double 2025 levels, according to research from Futurum Group. Notably, the hyperscalers describe their markets as supply-constrained rather than demand-constrained. Microsoft has disclosed an $80 billion backlog of cloud orders it cannot fulfill because of power limitations, and Goldman Sachs Research has cited projections that token consumption could rise 24-fold by 2030, largely on the back of enterprise agent adoption. That is why infrastructure owners are willing to spend years ahead of visible application revenue.
A MARKET SPLITTING IN TWO
The clearest sign that token supply is reshaping pricing power is the divergence now visible between commodity and frontier inference. Rohan Panjwani, an industry analyst cited in a recent Register report, has argued that the token market is effectively splitting into two tiers, with commodity inference heading toward zero cost even as frontier inference prices hold firm or rise. Open-weight models such as Kimi and GLM are approaching near-parity with premium frontier systems on many tasks at a fraction of the price, pushing routine workloads toward the cheapest viable option. At the same time, enterprises continue directing a large share of spending toward the most capable frontier models for complex reasoning and engineering work, where quality still commands a premium.
This bifurcation matters enormously for how investors size up AI-exposed companies. Businesses that control their own inference supply, through proprietary data centers, custom silicon, or long-term compute agreements, have far more control over their cost base than businesses that simply resell access to rented compute. That distinction is becoming a meaningful differentiator in how the market values infrastructure providers relative to the application-layer companies sitting on top of them.
WHY THIS MATTERS FOR INVESTOR RETURNS
Goldman Sachs Research has found that equity gains tied to the AI buildout have concentrated heavily in infrastructure names: semiconductors, hyperscalers, data center operators, hardware providers, and power companies. Its basket of AI infrastructure stocks returned 44 percent year to date against a much smaller rise in forward earnings estimates for the same group, a gap that has made analysts more cautious about how long capital spending can run ahead of visible monetization. That caution has already shown up in market behavior, with capital rotating away from infrastructure names where operating earnings growth is under pressure and where capex increasingly relies on debt rather than cash flow.
The unit economics inside individual companies tell a similar story. Boston Consulting Group frames the relevant measure as return on AI, a ratio weighing the economic value an AI deployment generates against the combined cost of human oversight and token consumption. Companies able to raise that ratio over time, meaning they extract more value per token rather than simply generating more activity, are better positioned to convert AI spending into durable margin expansion. That distinction has already surfaced in public disclosures: enterprise software companies embedding AI features into their core products have reported measurable gross margin compression tied directly to inference costs, a reminder that the near-zero marginal cost economics that defined traditional software no longer apply once every customer interaction runs through a paid model call.
THE OUTLOOK FOR TOKEN SUPPLY AND PRICING
The forces at work point toward a market that keeps getting cheaper at the unit level while getting larger and more capital-intensive overall. Analysts at Gartner have forecast a further 90 percent reduction in frontier model inference costs by 2030, even as total enterprise AI spending continues to climb alongside agentic adoption. For investors, the practical takeaway is that headline growth in AI usage or revenue tells only part of the story. The companies most likely to convert that growth into durable returns will be the ones with genuine control over their token supply chain, whether through owned infrastructure, efficient model routing, or disciplined engineering, rather than those simply riding the wave of falling prices and rising demand.
References
Bijlsma, J., Kleine, D., and Scognamiglio, F. “Return on AI: How CFOs and CIOs Can Manage the Token Meter.” Boston Consulting Group, July 1, 2026. https://www.bcg.com/publications/2026/managing-ai-token-costs
Cottier, B., Snodin, B., Owen, D., and Adamczewski, T. “LLM Inference Prices Have Fallen Rapidly But Unequally Across Tasks.” Epoch AI, 2025–2026. https://epoch.ai/data-insights/llm-inference-price-trends
“Is AI Really Getting Cheaper? The Token Cost Illusion.” Artefact, April 1, 2026. https://www.artefact.com/blog/is-ai-really-getting-cheaper-the-token-cost-illusion/
“AI Is Becoming a Bargain Hunter’s Market, With a Few Luxury Models on Top.” The Register, July 8, 2026. https://www.theregister.com/ai-and-ml/2026/07/08/ai-is-becoming-a-bargain-hunters-market-with-a-few-luxury-models-on-top/
“AI Capex 2026: The $690B Infrastructure Sprint.” Futurum Group, February 12, 2026. https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
“Why AI Companies May Invest More Than $500 Billion in 2026.” Goldman Sachs, December 18, 2025. https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026
