Jevons Has Not Arrived. We Are Still Paying for the Machine.

Sonya Huang has given us this argument before.

In June, Luminity published Agentic AI: The Metric That Ate Itself, documenting the short life of tokenmaxxing as a management idea. We wrote:[3]

Sonya Huang, a partner at Sequoia Capital—one of the world’s largest venture firms, managing roughly $85 billion in assets and holding active stakes in OpenAI, xAI, and Anthropic simultaneously—had her Qu’ils mangent de la brioche moment, telling the Wall Street Journal: “We all should be tokenmaxxing.” The firm runs its own internal leaderboard and portfolio-wide office hours to drive usage up across its portfolio companies.[2][3]

Roughly four months after that declaration, Huang has returned with a broader version of the same consumption story. In an August 13 interview, she called Jevons Paradox a “freaking wonderful thing” and described an “everybody win situation”: AI usage rises, application-company gross margins improve, inference providers strengthen, and model-company revenues continue accelerating.[1]

Token consumption is no longer merely being offered as a proxy for organizational adoption. Now rising AI consumption is being presented as evidence that Jevons Paradox has arrived.

It is an attractive thesis. It may eventually be right.

It has not yet been demonstrated.

Tokenmaxxing was a spending instruction searching for a productivity result. The new Jevons story risks becoming the same mistake with an economic label attached.

A Rebound Signal Is Not Yet Jevons Paradox

The loose version of Jevons now circulating through AI commentary is simple: tokens became cheaper, usage rose, therefore Jevons Paradox has arrived.

That formulation removes the part that matters.

An efficiency gain can lower the effective cost of using a resource. Lower cost can stimulate additional demand. If that induced demand is strong enough, aggregate resource consumption can rise rather than fall. But rebound effects vary by market and by use case; they do not automatically erase efficiency gains. The International Energy Agency explicitly distinguishes measured rebound effects in some sectors from the much stronger proposition that efficiency broadly causes greater aggregate consumption.[7]

For AI, the evidentiary question is not whether token prices have fallen while token volumes have risen. Both are plainly occurring. The question is what caused the additional consumption, whether the associated work is economically durable, and whether efficiency-induced demand—not capital subsidies, migration, experimentation, or greater compute intensity per task—accounts for the aggregate increase.

That causal work has not been done.

What Huang Actually Observed

Huang’s claim rests on two views into Sequoia’s portfolio. She says application companies are increasing AI usage while their gross margins rise. She also says inference and model companies are showing improving customer cohorts despite price competition.[1]

Those observations matter. They are consistent with a rebound effect. They could become early evidence for Jevons Paradox.

But they do not establish it.

Selected venture portfolios are designed to contain unusually fast-growing companies. Their businesses operate amid investor financing, promotional pricing, capacity commitments, model subsidies, product experimentation, and a race to capture markets before their economics normalize. Improving gross margin at an application company can show that inference became cheaper relative to revenue. Stronger cohorts at an inference provider can show that customers expanded usage. Neither observation, alone or together, isolates how much new demand was caused by the efficiency improvement or establishes a durable economy-wide increase in resource consumption.

There is also an unavoidable perspective problem. Sequoia has investments across the AI stack and benefits when cheaper intelligence expands both application margins and infrastructure demand. That does not make Huang’s thesis wrong. It makes independent verification and careful causal language more important.

When venture capital subsidizes the meal, an empty plate is not proof of permanent demand.

We Are Still in the Capital-Formation Phase

The strongest evidence about the current stage comes from the companies building the capacity.

Alphabet reported $44.9 billion in capital expenditure during the second quarter of 2026, with the vast majority directed to technical infrastructure supporting AI. It raised expected full-year capital expenditure to between $195 billion and $205 billion. On the same earnings call, Sundar Pichai described enterprise adoption as “very, very early” and said companies were “barely scratching the early stages” of what is possible.[5]

Microsoft reported $115.9 billion in additions to property and equipment for its fiscal year ended June 30, 2026, up from $64.6 billion one year earlier. Its own risk disclosures still frame the cloud and AI strategy as substantial investment dependent on customer demand, technical development, competition, regulation, broad adoption, and sustainable revenue growth.[6]

Stanford’s 2026 AI Index reports that U.S. private AI investment reached $285.9 billion in 2025 while organizational adoption reached 88 percent.[4] Those are extraordinary numbers. They establish that capital and adoption are moving rapidly. They do not establish that the resulting consumption has normalized into stable production economics.

Taken together, these disclosures support a narrower and more defensible conclusion: AI is experiencing rapid adoption inside an infrastructure buildout of historic scale. Luminity’s interpretation is that this remains primarily a capital-formation and market-discovery phase. Demand is real, but its composition is not yet settled.

We are still financing the machine that may eventually produce the paradox.

Four Different Things Are Being Counted as Demand

Current AI consumption combines at least four economic phenomena.

Subsidized experimentation. Enterprises, model providers, and venture-backed companies are paying to discover what works. Consumption purchased for learning can be rational without representing a durable production workload.

Workload migration. Existing search, software development, support, analytics, and content work is being moved onto AI systems. This can create platform revenue without creating an equivalent amount of new final demand.

Rising compute intensity. Reasoning models, long contexts, tool calls, agent loops, and parallel agents can increase the tokens or compute required to complete one task. Sequoia itself argued in 2024 that inference-time compute opened a new scaling plane in which more computation could improve reasoning.[8] Higher consumption caused by a more compute-intensive production function is not the same mechanism as higher consumption caused by an efficiency-driven reduction in effective price.

New economically justified work. Falling costs can make previously uneconomic tasks worth performing. Continuous security testing, individualized analysis, new scientific search, and high-frequency decision support may expand because intelligence became cheaper. This is the category most relevant to a durable Jevons claim.

All four can drive token volume upward. Only careful attribution tells us which mechanism is operating. A token counter cannot.

Tokenmaxxing Already Taught This Lesson

In April 2026, Huang defended tokenmaxxing as an imperfect but useful way to push employees into an AI-first mindset. Sequoia used a leaderboard. Portfolio companies were encouraged to do the same.[2]

The metric survived roughly six weeks as a serious management idea before the backlash became explicit.

Meta’s informal, employee-created token leaderboard came down. Amazon employees admitted generating unnecessary activity to inflate their usage statistics. By May 28, Fortune was declaring tokenmaxxing over and reporting that Uber had exhausted its annual AI budget in four months as executives struggled to connect consumption with company-wide results.[3][9]

The structural error was straightforward: token consumption is an input measure; productivity is an output property. When consumption became a target, people optimized consumption. The metric stopped diagnosing adoption and started rewarding the appearance of it.

The present Jevons declaration risks making the same mistake at market scale.

Tokenmaxxing treated rising employee consumption as evidence of productive transformation. Premature Jevons treats rising market consumption as evidence that cheaper intelligence has already created durable new demand. Both promote an observable input signal into a conclusion about economic output.

First Huang told companies to maximize token consumption. Now she points to rising consumption as evidence that Jevons has arrived. The metric has changed costumes, but it is performing the same trick.

That creates a public analytical trail worth following. The individual claims will change; the recurring question is whether a new consumption signal actually demonstrates the value, productivity, or economic mechanism being claimed for it. Jevons is unlikely to be the last installment.

The first metric failed because it could not distinguish activity from value. The second claim remains unproven for the same reason.

What Would Establish the AI Jevons Case

A serious assessment would need longitudinal evidence across normalized operating conditions—not anecdotes from a selected portfolio during an investment boom.

At minimum, the analysis would need to show:

1. a measurable efficiency improvement that lowers the effective cost of completing a comparable unit of useful work;
2. an attributable increase in demand caused by that lower effective cost;
3. durable production use rather than temporary experimentation or incentive-driven consumption;
4. stable or improving value realization after infrastructure subsidies and promotional economics recede;
5. aggregate resource consumption that rises enough to offset the efficiency gain; and
6. controls for workload migration, capability changes, task complexity, model mix, and multi-agent amplification.

That analysis may ultimately confirm a direct rebound, a partial rebound, or a full Jevons-style backfire in particular markets. It may also reveal several different outcomes across coding, search, science, customer operations, and autonomous work.

“AI” is not one resource used in one market for one purpose. A single universal paradox may be the wrong level of abstraction.

The Analytical Claim

AI usage growth in 2026 is evidence of rapid adoption during a massive capital buildout. It is not, by itself, proof of Jevons Paradox.

Huang has identified a plausible mechanism: cheaper intelligence can improve application economics and stimulate additional consumption. But her portfolio observations do not separate efficiency-induced demand from subsidized experimentation, workload migration, increasing task intensity, and capability-led expansion. Calling the paradox already operative converts an investable hypothesis into a declared economic fact before the necessary evidence exists.

The distinction is not semantic. Enterprises deciding how much infrastructure to buy, which workflows to automate, and how to measure returns need to know whether consumption is compounding value or merely compounding activity.

The Architecture Still Has to Answer for the Economics

For the enterprise, Jevons is not a license to stop measuring.

Lower unit costs can justify broader deployment. They can also hide waste, encourage unconstrained agent loops, and make individually trivial actions material in aggregate. Whether total demand rises or falls, the governance requirement remains unchanged: connect resource consumption to authorized work, decision quality, business outcomes, and recoverable Audit Trails.

Token volume belongs in infrastructure telemetry. Cost per effective action belongs in operating economics. Decision fidelity, intervention rates, error rates, and outcome realization belong in the governance system.

If intelligence becomes cheap enough to trigger Jevons Paradox, architectural discipline becomes more important, not less. An enterprise must still be able to explain what the additional consumption did.

Jevons Is a Destination, Not Yet a Description

The AI rebound thesis is credible. Falling costs, improving capability, expanding access, and the emergence of new categories of machine work could eventually produce one of the most consequential demonstrations of Jevons Paradox in modern economic history.

But possibility is not arrival.

First, token consumption was mistaken for productivity. Now expanding market consumption is being mistaken for economic proof. The analytical habit is the same: observe the input, assume the outcome, and declare the mechanism settled.

We made tokens dramatically cheaper, so naturally we found increasingly elaborate ways to consume them. That may be an early rebound signal. It may also be the predictable behavior of an industry rewarded for maximizing activity while the bill is still being financed. A portfolio leaderboard can demonstrate consumption. It cannot demonstrate productivity—and it certainly cannot establish an economy-wide paradox.

Jevons may be coming. Today, we are still paying for the machine.