Count the Tokens. Then Follow the Revenue.

A chart showing billions of AI tokens moving between models can look like a scoreboard for the entire industry. It is not. It records activity within a defined measurement system. Turning that activity into an investment thesis requires several additional steps.

TaylorVentureLab's research view is straightforward: track token usage, but keep adoption, spending, and profitability in separate columns. A model can gain usage without its creator capturing corresponding revenue. A provider can lose share while its business grows. Neither possibility is a contradiction.

Start with what the evidence actually measures

OpenRouter and a16z's December 2025 study analyzed a dataset of 100 trillion tokens. Most model-level and pricing analysis covered November 3, 2024, through November 30, 2025. The study reported that open-weight models reached approximately one-third of usage on OpenRouter by late 2025. Its token totals combined input and output, including reasoning tokens within output. [1]

That is useful historical evidence of adoption on one platform. It is not an October 2026 global market-share estimate. The report describes an observational dataset shaped by availability, prices, and user preferences; activity outside OpenRouter is not fully captured. [1]

Our interpretation: use this evidence to generate questions, not to declare a winner. Which workloads moved? Did customers keep paying after experimentation? Did the change reflect better economics, different task requirements, or a temporary promotion? A rising line alone cannot answer those questions.

Count comparable things

OpenRouter's usage-accounting documentation distinguishes prompt and completion tokens, reasoning tokens where applicable, and cached tokens where available. It also reports charges and uses each model's native tokenizer. [2]

Those distinctions matter to analysis. A total that mixes cached input, fresh input, and generated output is not a uniform unit of commercial value. Different tokenizers also complicate comparisons of identical underlying text. More tokens need not mean more customers, more completed jobs, or more useful output.

Consider an agent that revisits a long document repeatedly. Its activity could increase without a proportional increase in customer value. Conversely, a system that completes the same work with fewer tokens might improve customer economics while looking weaker on a volume chart. These are illustrative possibilities, not reported outcomes.

The practical rule: compare the same period, the same platform coverage, and the same token categories. Preserve the original metric definition alongside every chart.

Convert volume into money, carefully

Here is a deliberately simplified hypothetical, not a provider forecast.

Suppose a service doubles its billed token volume while its average realized revenue per token falls by 40%. Revenue becomes 2 × 0.60, or 1.20 times the starting level: a 20% increase, not a doubling.

If the realized rate instead falls by 60%, the same doubled volume produces 2 × 0.40, or 0.80 times starting revenue: a 20% decline.

This calculation assumes the volume and realized rate cover the same business and period. A real analysis must reconcile discounts, free usage, subscriptions, service fees, and the mix of token types. Public list prices are not necessarily realized prices.

Even a reliable revenue estimate would leave profitability unresolved. Delivery costs, utilization, research spending, and other expenses still matter. Do not label customer spending as provider profit, or use token growth as a substitute for disclosed financial performance.

Follow the entity that captures the payment

Open-weight adoption and model-company revenue are different questions. A researcher should distinguish the organization that creates weights from the operator that hosts inference and the application that bills the end customer.

Terminology deserves equal care. The Open Source Initiative's definition requires more than accessible weights: it includes freedoms to use, study, modify, and share, plus specified data information, code, and parameters. [3] This article therefore uses “open-weight” without assuming every model satisfies that definition.

Our thesis to test is that broader access to capable models may shift bargaining power toward customers and applications. It could also create opportunities for hosting businesses. These are scenarios, not established investment conclusions. Lower prices could pressure one business while expanding demand for another.

Nor should “open-weight” and “frontier” be treated as opposites. One describes access to weights; the other is a claim about capability that needs its own task-specific evidence.

Build a research record that can disprove the thesis

A useful monthly research sheet would record five things:

• Coverage: platform, observation dates, model versions, and known omissions.
• Activity: comparable token categories and absolute volume alongside share.
• Monetization: observed spending, free versus paid usage, and effective realized rates where available.
• Persistence: repeat paid use and retention within comparable customer groups.
• Economics: disclosed revenue and margins, with missing information explicitly marked unknown.

This is a proposed research method, not a claim that TaylorVentureLab has collected those private operating metrics.

Evidence supporting competitive displacement would include sustained migration of comparable paid workloads, not just a release-week traffic spike. Evidence against a simple displacement story could include a premium provider expanding paid usage and margins despite declining token share. Without those observations, the thesis should remain provisional.

The investment lesson is not to ignore token data. It is to respect what the data can establish. Count activity, reconcile spending, identify who receives the money, and test whether the economics endure. The most useful AI chart is the one whose limitations remain visible.

This article is educational research commentary, not a recommendation to buy or sell any security.