Token Capital: Your Firm's New Asset Is Also Its New Bill

Token Capital: Your Firm's New Asset Is Also Its New Bill

Today we're going to do something a little more conceptual, because I think there's a single two-word phrase that has quietly become the most important idea in enterprise AI — and almost nobody has noticed that it means two completely different things at the same time. The phrase is "token capital." And the collision between its two meanings is, I'd argue, the whole ballgame for any leader trying to figure out their AI strategy right now.

This was inspired directly by an essay Satya Nadella dropped this week. It's called "A frontier without an ecosystem is not stable," he posted it on X, and depending on which tracker you believe it's been seen tens of millions of times. Now, it's the kind of unusually philosophical thing you don't often get from the sitting CEO of a three-trillion-dollar company, and VentureBeat's Michael Nuñez caught the detail that makes the whole essay worth a second read: the phrase "token capital" carries a double meaning. So that's what we're going to talk about today. First, token capital as an asset — Satya's argument. Then, token capital as a bill — the part the essay politely steps around. And then why those two things are actually the same story, and what it means for what you should be doing.

First, the asset.

Satya's core move is to give companies a new way to think about what they own. He says every firm is going to have to build two things: human capital — which he defines as the knowledge, judgment, relationships, and pattern recognition of its people — and token capital, which he calls the firm's AI capability it builds and owns. And his key claim, the one he's careful to italicize, is that these don't trade off. Human capital, he writes, doesn't become less valuable as token capital grows — it becomes more valuable. "Without human direction," he says, "you have compute running in circles."

Now, the part that's actually actionable is the architecture. Satya's argument is that the real opportunity is not picking the best model — it's building a learning loop on top of whatever model you use, where human and token capital compound. He prescribes three layers: private evals that measure whether a model is improving against your business outcomes, not just public benchmarks; private reinforcement learning environments that let models learn from the real traces of work inside your organization; and a knowledge base that makes institutional memory queryable. He calls the resulting thing a "hill climbing machine," and his line is that, unlike most assets, it compounds — every improved workflow generates better training signal, which accumulates tacit knowledge that's unique to you.

And here's the test he sets, which I think is the single most useful sentence in the piece: you should be able to swap out a generalist model without losing the company-veteran expertise built into your system. That's the sovereignty test. If switching from one frontier model to another wipes out your institutional intelligence, you never owned it in the first place — you were renting it.

Then the essay zooms out to the political economy, and the rhetoric gets sharp. "The last thing any of us want," he writes, "is a world where every company across every sector is ceding value to a few models that eat everything they see." He reaches for the globalization analogy: entire industrial economies hollowed out by outsourcing, GDP looking fine on the surface while the displacement was real. His warning is that if the AI industry doesn't distribute value broadly, the political system will step in to force it.

Now, I'd be doing you a disservice if I didn't point out the obvious. Microsoft sells the picks and shovels for exactly the world Satya is describing. A world where every enterprise builds a proprietary learning loop on top of commodity models is, conveniently, a world where Microsoft sells the cloud, the tooling, and the model layer to all of them. So yes, the philosophy and the self-interest are pointed in the same direction. And yet — and this matters — he's not the only one ringing this bell. Snowflake's Sridhar Ramaswamy has warned that big model makers want everyone else reduced to, in his words, a "dumb data pipe" feeding the big brain. Box's Aaron Levie keeps asking the same question: if everyone has access to the same expert intelligence, how does any one company differentiate? Three very different companies, one shared diagnosis. The idea has resonance beyond Redmond's balance sheet.

Okay. So that's token capital as the asset. Now let's talk about the bill.

Because here's what the essay glides right past: the learning loop Satya wants you to build is made of tokens. The literal kind. The ones you're metered and billed for. And in 2026, those tokens have become the fastest-growing line item in the enterprise.

Let me give you the ground truth, and it's worth sitting with. A team out of the Stanford Digital Economy Lab — Erik Brynjolfsson, Alex Pentland, Jiaxin Pei and co-authors — published a paper this spring called "How Do AI Agents Spend Your Money?" They ran eight frontier models across real coding tasks, and the headline finding is brutal: agentic tasks consume roughly a thousand times more tokens than a simple code-reasoning or chat interaction. Most of that is input tokens, not output — because an agent reads the task, gets a response, re-reads everything, acts, re-reads all of that plus the new response, and on and on. Pei calls it a context snowball. And it gets worse for planning purposes: run the exact same agent on the exact same task and the cost can vary by up to thirty times. The models systematically underestimate their own spend — the correlation between what they predict and what they burn tops out around 0.39. And more tokens doesn't even buy you more accuracy; accuracy tends to peak at intermediate cost and then flatten. You are, in other words, flying a budget you cannot forecast.

This is not theoretical. Uber burned through its entire 2026 AI coding budget in four months — after, and I love this, incentivizing adoption with an internal leaderboard ranking teams by usage — and then had to slap on a fifteen-hundred-dollar-per-month cap per employee, per tool. Meta had an employee build a leaderboard called "Claudeonomics." Amazon reportedly pushed people to "tokenmaxx." And in the most on-the-nose example imaginable, VentureBeat reported that Microsoft itself — the company whose CEO just told you to build token capital — is canceling the majority of internal Claude Code licenses in one division at the end of this month, after usage hit 84 to 95 percent and per-engineer costs ran five hundred to two thousand dollars a month, because they exhausted the budget. Nvidia's Bryan Catanzaro put the whole era in one line to Axios: "For my team, the cost of compute is far beyond the costs of the employees." Read that again. The compute now costs more than the people.

And zooming out to the macro, the FinOps Foundation's J.R. Storment lays out why this isn't a blip. Yes, the price per token at a fixed capability tier keeps drifting down. But that's no longer the relevant number, because consumption is exploding faster than price is falling. AT&T reported scaling from eight billion to twenty-seven billion tokens a day on multi-agent systems. Google reported processing roughly 1.3 quadrillion tokens a month — about a 130-fold jump in a year. Anthropic added something like twenty-one billion dollars of annualized revenue in six months, almost entirely on enterprise token consumption. The subsidy phase, where venture money priced tokens below cost to grab growth, is ending — OpenAI's own ChatGPT product lead said the quiet part out loud: an unlimited AI plan is like an unlimited electricity plan, and it doesn't work. Storment's summary is the line I keep coming back to: a token may get cheaper; tokens, in aggregate, are not.

So here's the synthesis, and it's the reason I wanted to do this post.

These are not two stories. They're one. The exact same tokens that make up Satya's compounding learning loop — the asset — are the tokens blowing up Uber's budget and Microsoft's own division — the bill. Token capital the asset and token capital the bill are the same substance viewed from two ends. And the implication is bigger than it looks: it means the thing your company actually needs is not an "AI strategy" in the Gartner-magic-quadrant, pick-the-right-vendor sense. What it needs is a discipline — and that discipline now has a name. The FinOps people call it token economics, or tokenomics: FinOps applied to AI, the practice of metering AI consumption and connecting it to business outcomes. It's serious enough that the Linux Foundation announced it's forming a Tokenomics Foundation to set open standards for AI cost management. Deloitte's framing in their "pivot to tokenomics" piece is that you have to govern AI like an economic system — the same rigor you'd apply to energy or capital allocation — because for some firms AI is already eating up to half the IT budget.

And the punchline that ties it back to Satya's sovereignty test: the team that can swap models without losing its expertise is usually the same team that knows its goodput, not just its throughput — that meters token quality, not just token quantity — that knows a reasoning-class workload can burn thirty times the tokens of a chat for the same-looking task. The cost muscle and the capability muscle are the same muscle. You can't compound a learning loop you can't afford to run, and you can't afford to run a loop you can't measure.

So if you take nothing else from this, take a short list. One: stop treating AI tooling as a SaaS line item — the seat fee is the floor, the metered consumption is the real bill, and the FinOps Foundation is right that the headline subscription price is no longer a reliable budget signal. Two: meter token quality, not just quantity — right-size models, cap context, don't pay premium-latency rates for work that tolerates Goldilocks-tier speed. Three: build the portable learning loop Satya describes, but instrument it from day one so your evals are tracking cost-per-outcome alongside accuracy. And four — this is the one I'd underline — resist the temptation to respond to runaway token costs with blunt, across-the-board spend caps. The Stanford data says you can't predict these costs yet anyway, and the most sophisticated organizations are responding not by spending less, but by spending legibly.

Your company doesn't need an AI strategy. It needs to get fluent in the economics of the token — because the token is simultaneously the most valuable asset and the most expensive liability on the AI-era balance sheet. And for now, that's going to do it.

RELATED RESOURCES

• Satya Nadella, "A frontier without an ecosystem is not stable" (sn scratchpad, Jun 14, 2026) — https://snscratchpad.com/posts/frontier-ecosystem/
• Michael Nuñez, "Satya Nadella warns that AI could hollow out entire industries…" (VentureBeat, Jun 15, 2026)
• Kamina Bashir, "Microsoft CEO Nadella Says Every Firm Must Build Token Capital" (Yahoo Finance, Jun 15, 2026)
• "How are AI agents spending your tokens?" (Stanford Digital Economy Lab, May 5, 2026) — https://digitaleconomy.stanford.edu/news/how-are-ai-agents-spending-your-tokens/
• Bai, Huang, Wang, Sun, Mihalcea, Brynjolfsson, Pentland, Pei, "How Do AI Agents Spend Your Money?…" (arXiv:2604.22750, Apr 2026)
• Merizzi, Smith, Mittal, Churiwala, Kearns-Manolatos, "AI tokens: How to navigate AI's new spend dynamics" (Deloitte, Jan 19, 2026)
• J.R. Storment, "Token Economics: The Atomic Unit of AI Value" (FinOps Foundation, May 10, 2026) — https://www.finops.org/insights/token-economics-the-atomic-unit-of-ai-value/