Structural or cyclical: What’s driving the token volume surge? 

In our view, the surge is structural rather than cyclical.

According to some estimates, monthly token volume has grown sharply, from about 9.7 trillion in May 2024 to roughly 3.2 quadrillion by May 2026—nearly a sevenfold increase in the past year alone.¹ On the application programming interface side, first-party models, or models built and served directly by the same company that developed them, were processing more than 10 billion tokens per minute by the fourth quarter of 2025, up from 7 billion the prior quarter.²

The token surge is being driven by a shift toward more computationally intensive AI workloads, for example:

• Reasoning models: AI models optimized for tasks that require deliberate, multistep problem solving, generating long chains of intermediate “thinking” tokens before producing a final answer.
• Agentic systems: An AI system that can pursue a goal through a sequence of actions—not just answer a single prompt—on a user’s behalf, using tools, reading documents, and iterating through many model calls per request.
• Multimodal inputs: Inputs that combine more than one type of data like images, audio, and video; they are converted into tokens at far higher rates than text.
• Embedded AI: AI capabilities built directly into everyday software such as search, productivity suites, and customer service tools.

We believe multimodal deserves special attention because it scales token volume in a way text alone never could. A minute of video runs to roughly 15,800 tokens, a minute of audio about 1,920, versus roughly 200 tokens for a minute of read text (Exhibit 1).³ In other words, one minute of video costs as much as dozens of pages of prose. So, as AI extends into images, audio, and especially video, reported token volumes can climb sharply even without a proportional rise in users or queries.

Are token prices really falling?

At a fixed level of capability, the cost of running a model has fallen dramatically and continues to fall.

Total spend, though, is a function of the price per token and the volume of tokens consumed, and the driver has been volume on a roughly flat price rather than a collapsing one. The blended price the market actually pays has held near $1 per million tokens over the past year.⁵ The Silicon Data LLM Token Expenditure Index, a usage-weighted gauge of what the market pays rather than list price, nearly doubled from its December 2025 inception to a May 2026 peak before falling back to about $1 by October 2026, roughly half off the peak, with the swings driven mainly by shifts in usage between cheaper open-weight and premium frontier models rather than by changes in list prices (Exhibit 2).⁶

How does the rise of cheap, open-weight models affect token pricing and what does that mean for closed-source/proprietary models?

Open-weight (or publicly published) competition is compounding an already steep price collapse for these lower-capability models. Because open-weight models can be hosted by any number of competing providers rather than a single vendor, they strip away the pricing power that comes from proprietary distribution, pushing prices toward the marginal cost of compute, or the incremental cost to serve one additional unit of AI output.

It is important to note, however, that while open-weight models reduce the subscription or consumption price per token, the infrastructure cost to serve an open-weight model is similar or in some cases even higher than the cost to serve a proprietary model. 

Moreover, falling unit prices don’t mean falling bills: Multimodal inputs multiply the tokens consumed per task—agentic coding alone uses on the order of 1,000 times more tokens than ordinary chat. This is the Jevons paradox at work: Cheaper tokens are expanding total consumption faster than they compress per-token revenue.

As a result, the economics of that expanding demand accrue to the infrastructure layer. Based on our forecasts, capital expenditure from US-based hyperscalers in 2027 is tracking to nearly $1.5 trillion, up about 60% from 2026, as the value accrues to the upstream and supply-constrained links of the chain where replacement cycles are long.⁷

For the established providers that sell access to frontier models on subscription or per-token terms, the effect is margin compression at the commodity end rather than an existential threat, and the outcome splits by workload. Tasks that a cheaper open-weight model can handle well drift toward the marginal cost of compute. Tasks that genuinely require frontier capability hold their pricing, which continues to support frontier economics. Where price does erode, providers are responding by moving up into the harness (the software layer around a model that enables it to execute multistep tasks as an agent) and application layer, embedding the model inside agentic workflows and enterprise data integration so they capture value at a level open-weight substitutes cannot easily replicate.

Why is this an EM supply chain story? And what does it mean for investors?

We view AI as an EM supply chain story: Many of the critical bottlenecks supporting AI infrastructure are concentrated in Taiwan, South Korea, and elsewhere across Asia. While US companies may capture a substantial share of the economics in cloud services, models, and accelerator systems, the underlying hardware stack—the chips, memory, packaging, and other physical components an AI system runs on—remains heavily dependent on Asian suppliers of advanced foundry capacity, high-bandwidth memory, advanced packaging, substrates, printed circuit boards, testing, materials, and specialized semiconductor equipment services.

These inputs are neither easily substituted nor quickly replicated. They require deep process expertise, lengthy customer-qualification cycles, and significant capital investment to expand capacity. Companies controlling hard-to-replicate portions of the hardware stack or the scarce, capacity-constrained component businesses through which every AI workload must pass may be positioned to benefit from sustained demand, pricing power, and rising capital intensity across the AI ecosystem.

In our view, differentiated EM opportunities are often found in the physical supply chain. These businesses can capture value regardless of whether the model layer is dominated by premium proprietary offerings or lower-cost open-weight alternatives.

Today the clearest profit pools sit in hardware. On a margin basis, the standout is memory: SK hynix earned an operating margin of roughly 72 to 76 percent through the first half of 2026 on high-bandwidth memory, an exceptional level for any hardware maker, as HBM sold out and pricing ran ahead of supply.⁸

As AI hardware grows more complex, is the supply chain narrowing—or widening?

In our view the supply chain is widening and that is central to the investment opportunity. Each new generation of AI hardware requires more components, not fewer. A single leading-edge AI server is typically built from thousands of individual chip components, and that silicon has the potential to account for the large majority of what a system costs to build. But it isn’t just about more chips: As these systems pack more computing power into less space, entirely new categories of supporting hardware are moving from pilot to volume production—advanced cooling systems to manage the heat these chips generate, new connection materials as traditional wiring reaches physical limits, and new equipment to power efficiently at much larger scale. Each of these is a new socket in the chain that has to be filled.⁹

A longer list of materials, combined with incumbents that are already capacity-constrained, is precisely the condition under which new suppliers get qualified by end customers like accelerator designers, server assemblers, and hyperscalers. When a category is new and demand outruns the installed base, customers actively second-source and bring on additional vendors ahead of the ramp rather than after it.

We believe that dynamic has been unusually favorable to emerging markets small caps historically. The manufacturing base for these categories already sits in Taiwan, South Korea, and increasingly elsewhere in Asia, so the credible new entrants are disproportionately Asian. These companies are small enough that a single qualification on a new component can move potential earnings and help improve a company’s quality materially.

This is why we frame AI investing less as a bet on a fixed roster of winners and more as gaining exposure to a lengthening chain. Each migration creates a new scarce input and, with it, a new set of specialized suppliers that can earn excess returns. That recurring creation of new, capacity-constrained pockets, filled disproportionately by under-followed Asian small caps, is the part of the AI trade we think is least reflected in consensus positioning.