Happy Sunday and welcome to Investing in AI. Be sure to check out our AI in NYC podcast if you want to catch up on AI issues of the day. Today I also want to promote our new Tokenminning community, created to counteract the stupid tokenmaxxing movement. It was covered by the NYTimes recently. You can subscribe to our newsletter on the site to get tips and trick and tutorials about token optimization, which is my topic for today.
For two years, the Street has underwritten the AI trade on the wrong number. We tracked seat growth, “AI-attributed” ARR, hyperscaler capex, and the crude shorthand of cost per million tokens. None of those metrics tells you whether a public company is actually building a durable, high-margin business on top of large language models—or quietly importing a commodity cost line it cannot control.
A new framework gives us a better lens. In a June 2026 paper, NYU’s Quanyan Zhu formalizes what he calls AI tokenomics: the study of how tokens are generated, consumed, priced, allocated, and optimized across AI systems. The core argument is deceptively simple and, for anyone underwriting AI exposure in public equities, genuinely clarifying. The token is no longer a technical accounting artifact inside a transformer. It has become the microeconomic currency of the intelligence economy—the unit through which compute, memory, energy, latency, pricing, and ultimately gross margin are all denominated.
The investable thesis follows directly: token expenditure and economic value are distinct. A company can burn enormous token volume and create almost no value, or sip tokens and capture an outsized share of an enterprise workflow. If you want to find the winners in the next leg of this market, you stop counting API spend and start underwriting token productivity, marginal value, and structural tokenomics.
The token supply chain runs through your portfolio
Zhu’s first move is to link the token to physical reality. Every token processed carries a measurable cost in floating-point operations, key-value cache memory, energy (Joules per token), and latency (milliseconds per token). Output tokens are generated sequentially and therefore cost far more to produce than input tokens, which is why providers price output at roughly four to six times input. That asymmetry is not a billing quirk; it is the thermodynamics of inference showing up on an invoice.
For public-market investors, this is the chain that connects Nvidia and AMD silicon, hyperscaler data centers, and the grid operators and independent power producers now being repriced as AI utilities, straight through to the gross-margin line of every software company that resells intelligence. The infrastructure layer—Microsoft, Alphabet, Amazon, Meta, and the merchant chip and power names—does not win simply by owning the most compute or the largest datasets. It wins on unit economics: who can convert a raw kilowatt-hour and a GPU-second into a high-margin token at the lowest marginal cost.
The risk to watch at this layer is token dilution: model architectures that require excessive token generation to reach a simple answer. Reasoning models that “think” through tens of thousands of hidden tokens to answer a basic query inflate cost of goods sold and quietly compress the software margins layered on top. When you evaluate a model franchise, the question is not how capable it is in a demo. It is how many tokens it burns to deliver a unit of enterprise value—and who absorbs that cost.
Why a token spent is not value created
This is the heart of the framework, and the part the Street consistently misprices. Zhu argues that the economic value of a token depends on five dimensions, and each maps to a question you should be asking on every AI name you own.
Marginal productivity. How much revenue or cost savings does an additional token generate? A token that closes a sale or prevents a compliance failure is worth orders of magnitude more than one that summarizes a meeting nobody reads.
Workflow position. Tokens spent at critical, deterministic execution steps—diagnosing a condition, executing a trade, finalizing a contract—carry far more value than tokens spent on top-of-funnel summarization. The case studies in the paper make the point sharply: in a multi-agent workflow, the planning stage consumed under 5% of total tokens but had a greater downstream impact on final output quality than the execution stage that ate nearly half the budget. The most valuable tokens are rarely the most numerous.
Hidden reasoning activity. Modern reasoning models burn invisible internal tokens before they answer. In Zhu’s legal-analysis example, hidden reasoning accounted for roughly a third to nearly 60% of total token consumption, even though the visible output was a 1,000-token memo. For investors, this creates a measurement trap: a company’s public-facing token count can look modest while its true cost base is enormous. Whoever absorbs the hidden-token bill—the application vendor or the customer—determines who actually owns the margin risk.
Downstream propagation and risk. One bad token early in an agentic workflow can cascade, triggering thousands of wasted tokens and, worse, an erroneous action with real liability. As AI moves into safety-critical, revenue-critical positions, this propagation risk becomes a balance-sheet item, not a footnote.
The application-layer takeaway for a public-equity book: favor companies positioned at high-value, high-marginal-productivity workflow steps over those generating high-volume, low-value token throughput. Volume is vanity; position is margin.
Enterprises now run a token budget—and it will reshape SaaS
The most consequential shift for public software investors is happening on the buy side. Enterprises are starting to treat tokens like raw manufacturing inputs rather than SaaS seats, and they are building token budgets to govern them. Zhu formalizes this as a token production function: workflows are optimized mathematically, trading accuracy (more tokens, more reasoning) against speed and budget.
This is the quiet threat to the entire seat-based software model. For fifteen years, the SaaS premium multiple rested on 75–85% gross margins and predictable, near-zero marginal cost per user. Bolt a frontier model onto a flat-rate subscription and that economic logic inverts: every power user now carries a variable, volatile token COGS the vendor does not control. The companies that re-rate downward will be those that priced AI features flat and discovered their best customers were also their least profitable.
The companies that hold their margins—and their multiples—will be token-efficient: capable of delivering elite enterprise value while minimizing the token pull from underlying foundation APIs. That efficiency comes from caching, smaller fine-tuned models for routine steps, disciplined context management, and architecture that spends reasoning tokens only where marginal productivity justifies them. When you read a 10-K or sit in a management call, “AI-driven gross margin expansion” should prompt one follow-up: expansion relative to what token cost trajectory?
The emerging market: dynamic pricing and the ERP for AI
Pricing itself is mutating from static, flat per-token rates toward dynamic structures—congestion-based spot pricing tied to compute demand and energy costs, and outcome-based pricing that charges for the result rather than the token volume. Outcome-based pricing is both a threat to seat-based incumbents and an opportunity for vendors confident enough in their token productivity to put margin behind a guarantee.
That volatility creates a brand-new investable category: token instrumentation and middleware, or what Zhu’s framing implies is an “ERP for AI.” Software that measures hidden-token consumption, calibrates true cost, audits and hedges token volatility, and dynamically routes requests across model registries to optimize cost will become as standard as cloud cost management. A multi-model routing layer is no longer a feature—it is structural defense against single-provider pricing risk, and a screen for which application companies will survive a provider’s next price hike.
The public-market due diligence checklist
Run this at your next investment committee meeting on any AI-exposed name:
Token productivity. Is the company driving high business value per token, or just flooding context windows to demo capability?
Workflow position. Is its AI answering low-stakes queries, or executing high-value, high-risk, deterministic actions customers will pay to protect?
Hidden reasoning cost. If it relies on reasoning models, who absorbs the invisible-token bill—the vendor’s margin or the customer’s budget?
Margin architecture. Does pricing pass token volatility through to customers, or does the vendor eat it on a flat subscription?
Token volatility defense. Is there a multi-model routing layer, or is the business hostage to one provider’s pricing whims?
The macro outlook
The parallel is the early internet. Investors who kept analyzing raw server bandwidth missed the names that learned to monetize engagement and unit economics. The current AI era demands the same transition—from raw compute and headline API spend to genuine AI tokenomics. Scaling parameters is table stakes; the economic value will accrue to the companies, and the layers of the stack, that master token efficiency, position themselves at high-value workflow steps, and convert tokens into durable margin rather than commodity cost.
The token is now the unit of account. Underwrite the businesses that understand that better than the market does.
And of course if you need to maximize your tokenomics, I cofounded the world’s first token engineering platform at Neurometric.ai
Thanks for reading.