This note began with a simple observation from our own workflow.

With the latest generation of reasoning models, our token budget did not fall as the models became more efficient. It went up. In our own work — and anecdotally among traders, engineers and researchers around us — usage has often more than doubled.

That sounds counterintuitive.

Better models produce substantially more value per token. Yet we are also willing to spend substantially more tokens per hour.

The reason is that AI demand is not a fixed quantity of work waiting to be served more cheaply.

As intelligence improves, the set of economically useful tasks expands.

A model that cannot reliably complete a task has almost zero demand for that task. Once it crosses a capability threshold, demand can appear very quickly. As it improves further, the task can be delegated, parallelized, embedded into an agent loop, and eventually automated.

This creates a different demand curve from ordinary software.

The central AI demand variable may not be price elasticity alone. It may be capability elasticity: how much new demand appears when a model becomes intelligent enough to cross a task threshold.

The public data does not yet prove that inference demand is mathematically exponential in model intelligence. But several independent datasets are already consistent with the thesis.

1. Token demand is already compounding

OpenRouter reported that weekly volume on its platform increased from 5 trillion to 25 trillion tokens in six months by May 2026.

That is not global AI demand, and OpenRouter is itself a fast-growing platform. But a fivefold increase in six months is difficult to reconcile with the idea that AI inference is settling into a mature, fixed pool of demand.

The composition of usage is changing as well.

A year ago, much AI usage looked like a conversation: one prompt, one answer.

Increasingly, the unit of consumption is an agentic workflow — long context, tool calls, retries, subagents, evaluations and parallel execution.

That changes the arithmetic.

A useful way to think about aggregate token demand is:

Token demand = users × tasks per user × model calls per task × tokens per call

Model efficiency can push the last term down.

Higher intelligence can push the other three terms up — sometimes dramatically.

2. Intelligence unlocks tasks, not just better answers

The strongest evidence for this thesis is not a benchmark score.

It is what users do when models become more capable.

OpenAI reported that between November 2025 and June 2026, median agent usage among active internal users increased:

OpenAI also reported that by May 2026, 70.2% of sampled individual Codex users had made at least one request estimated to represent more than one hour of human work, while 25.6% had made one estimated to exceed eight hours.

The important part is the direction of causality described in the study: as the agent became more capable, people shifted toward longer-horizon, more difficult and more cross-functional work.

This is what a capability threshold looks like in practice.

Before the threshold, the user does not spend five times as many tokens trying to force an incapable model to complete the job.

They usually do the job themselves.

After the threshold, the workload enters the addressable market.

METR's task-horizon work offers a second lens. Its measurements show an approximately exponential historical increase in the length of software-engineering, ML and cybersecurity tasks frontier agents can complete at a given reliability. METR's exact estimates are uncertain and sensitive to task composition, but the broad trend is difficult to miss: the frontier is moving from short tasks toward tasks measured in hours.

Every movement of that frontier potentially converts another class of human work into inference demand.

3. More valuable work appears to consume more compute

Anthropic's June 2026 Economic Index provides another useful piece of evidence.

Across work-related Claude conversations, Anthropic found a positive relationship between occupational wages and token consumption. Conversations mapped to top-wage-tercile occupations consumed roughly 2.07× as many tokens as those mapped to the bottom tercile.

More complex artifacts were also more compute-intensive: conversations about building applications used more than three times the tokens of the median conversation.

This matters because the economic ceiling on AI spending is not the cost of today's chatbot.

It is some fraction of the value of the work being produced.

If a model saves ten minutes on a low-value task, the willingness to pay is limited.

If a model materially improves a research process, discovers a trading hypothesis, resolves a production incident, finds a security vulnerability, or executes hours of engineering work, the economically rational inference budget can be orders of magnitude larger.

The relevant denominator becomes labor value and output value, not historical token price.

4. Price elasticity is real too

Capability is one side of the demand curve.

Price is the other.

In July and August 2026, OpenRouter ran large discounts on GPT-5.6 Terra and Luna. Compared with their pre-program daily averages, token usage rose 5.6× for Terra and 13.8× for Luna.

Sol, which initially remained at list price, rose only about 1.11× over the same period.

This is not a clean laboratory experiment: model launches, user mix and competing models all matter.

But the direction is striking.

When inference becomes cheaper, developers do not simply pocket the savings.

They often buy much more inference.

That is classic Jevons-paradox behavior.

The additional twist in AI is that price is falling at the same time capability is rising.

One expands the number of calls that are affordable.

The other expands the number of problems worth calling a model for in the first place.

Together, they can produce a much steeper demand response than either mechanism alone.

5. Reasoning changes the unit economics again

Modern reasoning models make the relationship between intelligence and compute explicit.

OpenAI's API exposes reasoning effort as a control across lower and higher settings. Higher reasoning effort allows the model to spend more reasoning tokens on planning, debugging, synthesis and multi-step tradeoffs, at the cost of more latency and compute.

This is important conceptually.

For many traditional software products, better quality means the same workload running on better software.

For reasoning AI, better output can literally mean buying more cognition at inference time.

At the same time, better models can often achieve the same quality with fewer tokens than older models.

Both things can be true:

  1. cost per unit of intelligence falls, and
  2. total spending on intelligence rises.

That is exactly what we should expect if the utility curve is expanding faster than efficiency improves.

6. Supply is growing fast — demand may be growing faster

Epoch AI estimated in May 2026 that global inference capacity at fixed model size and context length is growing at roughly 3.4× per year.

Its imperfect proxies for token demand were growing closer to 10× per year.

These estimates carry wide uncertainty and should not be read as precise forecasts.

But the mismatch is the important signal.

Even enormous improvements in chips, serving efficiency, quantization and model architecture can be absorbed by a workload whose surface area keeps expanding.

Agents intensify this effect because they can consume compute without waiting for a human to type the next prompt.

Parallel agents intensify it again.

At OpenAI, researchers were already using 3.1 agent-workdays of runtime for every human workday by mid-August 2026. The median researcher was consuming more than $600 per day of inference at API prices, while the 90th percentile exceeded $7,000 per day.

This is still an extreme frontier-user population.

That is precisely why it is interesting.

Frontier users often show what becomes ordinary later.

The investment implication: underexposure is a risk

The investment debate around AI is often framed as a single question:

Is too much capital being invested?

That is only half the problem.

There is another form of risk: being structurally underexposed to a technological transition whose addressable demand is still expanding.

We do not interpret that as "buy every AI stock."

The model layer can commoditize. Margins can collapse. Capacity can be overbuilt locally. Hardware cycles remain cyclical. Valuations matter. Competitive advantage matters.

But if the demand thesis above is broadly correct, then a portfolio with no meaningful exposure to the companies controlling scarce bottlenecks, distribution, infrastructure or durable software moats is making an implicit macro bet of its own.

In 2026, the scale of infrastructure commitment is already extraordinary.

Amazon has discussed approximately $200 billion of 2026 capex. Alphabet expects roughly $180–190 billion. Microsoft currently expects about $175 billion, and Meta $130–145 billion. The accounting definitions differ, so these figures should not be mechanically summed as if they were perfectly comparable.

The strategic message is clearer than the accounting comparison:

the largest and best-capitalized technology companies in the world are still behaving as if compute is scarce, not abundant.

Microsoft has said it expects capacity constraints through 2026. Amazon says a substantial portion of its AWS capex already has customer commitments behind it. NVIDIA's latest reported Data Center revenue rose 117% year over year.

This does not settle valuation.

It does tell us that infrastructure demand is not merely a narrative.

Where value can accrue

We currently think about AI exposure in three broad layers.

1. High-moat software

The best software exposures are not generic "AI wrappers."

They are products with existing distribution, proprietary data, workflow lock-in, trust, and a reason for AI to increase product depth rather than erase the moat.

Cybersecurity and observability are particularly interesting.

As AI generates more code, more agents, more identities, more tool calls and more autonomous actions, the attack surface expands. The number of machine decisions requiring monitoring, policy and verification expands with it.

The same force that improves developer productivity can increase the value of the security layer around it.

2. Hyperscalers

Hyperscalers sit at the intersection of compute demand, enterprise distribution, proprietary accelerators, networking, storage and data.

Model efficiency is not necessarily negative for them.

If a 10× reduction in inference cost causes a 20× increase in economically viable inference, aggregate consumption still rises.

The critical question is not cost per token.

It is total profitable compute consumed.

3. Hardware and physical infrastructure

Semiconductors, memory, networking, power delivery, cooling and data-center infrastructure remain the physical constraint underneath the entire stack.

This layer will remain cyclical. Expectations can overshoot fundamentals, and sentiment premiums can be violent in both directions.

That is why valuation discipline matters particularly here.

A secular growth thesis does not repeal the cycle.

The other cardinal error: sizing yourself into a forced sale

There is an important tension in saying that underexposure to AI is a risk.

The obvious bad interpretation is to respond by taking a position so large that ordinary volatility becomes existential.

Howard Marks provides the right counterweight.

In his 2023 memo Taking the Temperature, Marks explains that Oaktree generally does not sell attractive long-term holdings merely because it expects a market decline. He also argues that investors should normally operate around a sustainable risk posture rather than repeatedly attempting to time markets.

His sharpest formulation is that selling at the bottom is "the cardinal sin in investing."

The reason matters.

A drawdown is temporary if the underlying thesis survives and the holder can survive with it.

A forced sale converts that temporary fluctuation into a permanent loss of participation.

This produces a simple portfolio principle for secular themes:

A position that is too small to matter is a thesis problem. A position that is too large to survive is a risk-management problem.

Those two errors are not opposites.

They can coexist.

The goal is durable exposure: enough participation for the thesis to matter, with enough liquidity and balance-sheet resilience that a violent correction does not dictate the exit.

The strongest secular trend can still contain severe drawdowns in individual securities. A portfolio architecture that assumes otherwise is not expressing conviction; it is outsourcing the timing of its exit to the market.

The second strategic mistake: using AI shallowly

There is another risk that does not appear on a brokerage statement.

Not using AI deeply enough.

Casual AI usage gives people a view of what current products can already do.

Deep domain usage reveals what becomes possible next.

That distinction matters for both operating advantage and investment research.

A trader who uses AI only to summarize news sees a cheaper research assistant.

A trader who combines models with market data, event pipelines, simulation, execution logs, strategy evaluation and persistent research memory starts to see an entirely different product surface.

An engineer who asks a chatbot coding questions sees autocomplete.

An engineer who orchestrates parallel agents across a repository, tests, observability, deployment and post-mortems begins to see an automated engineering system.

A researcher who uses AI to summarize papers sees faster reading.

A researcher who uses agents to search, reproduce, challenge, synthesize and continuously evaluate hypotheses starts to change the feasible size of the research space itself.

That is where the most important demand discovery happens.

The frontier is no longer prompt engineering.

It is system engineering around intelligence.

From compression to extrapolation

The first phase of AI adoption mostly compresses existing work:

same task → less human time

The more interesting phase expands the work:

better intelligence → more hypotheses → more experiments → more parallel paths → larger search space

That is the transition from automation to extrapolation.

Instead of asking AI to reproduce the output a person already knew how to produce, we use it to search beyond the person's normal attention budget, disciplinary boundaries and cognitive bandwidth.

For investment and research organizations, this is potentially the larger opportunity.

The objective is not merely to employ fewer analyst-hours per report.

It is to build a semi-automated research system capable of exploring more markets, more data, more hypotheses and more adversarial counterarguments than the team could examine manually.

If successful, the output frontier moves.

What we are watching

The thesis is falsifiable.

We would become less constructive if capability gains stop translating into longer useful tasks, token volumes flatten despite falling prices, hyperscaler capacity shifts from constrained to persistently idle, or AI-generated economic value fails to keep pace with inference spending.

The metrics we consider most informative are:

The key metric is not any single benchmark.

It is whether each improvement in model capability creates more economically valuable work than the previous generation could address.

Conclusion

AI demand should not be modeled as a fixed pool of prompts.

The product being sold is increasingly machine intelligence applied to work.

As the quality of that intelligence rises, three things can happen simultaneously:

  1. existing tasks consume intelligence more intensively;
  2. previously uneconomic tasks become viable;
  3. agents create new machine-to-machine demand with little human latency in the loop.

That is why falling cost per token does not necessarily imply falling AI spend.

And it is why rising efficiency does not necessarily imply lower infrastructure demand.

We are still early enough that the demand curve itself is being discovered.

For investors, that creates two symmetrical responsibilities:

do not ignore a secular transition because the current spend looks extraordinary; and do not size the position so aggressively that a cyclical drawdown forces you out of it.

For operators, researchers and domain experts, there is a third:

use the technology deeply enough to discover the demand before it becomes obvious in the financial statements.

That may be one of the most valuable forms of research edge available today.

Sources & notes

  1. OpenRouter, OpenRouter Raises $113M Series B (May 28, 2026): weekly volume rose from 5T to 25T tokens in six months. https://openrouter.ai/blog/announcements/series-b/
  2. OpenRouter, GPT 5.6 Discounts & Jevons Paradox (Aug. 25, 2026): Terra +5.6× and Luna +13.8× daily token usage during the discount period; Sol initially served as a control. https://openrouter.ai/blog/insights/gpt-5-6-discounts-jevons-paradox/
  3. OpenAI, How agents are transforming work (Jun. 25, 2026): longer-horizon work, internal token-use growth by department, Codex adoption. https://openai.com/index/how-agents-are-transforming-work/
  4. OpenAI, Research acceleration: The view inside OpenAI (Sep. 6, 2026): researcher inference spend, concurrency and agent-workdays. https://openai.com/index/research-acceleration-view-inside-openai/
  5. OpenAI API documentation, Reasoning models: higher reasoning effort trades more reasoning tokens and latency for more deliberate processing. https://developers.openai.com/api/docs/guides/reasoning
  6. METR, Task-Completion Time Horizons of Frontier AI Models and Time Horizon 1.1: exponential task-horizon trend; estimates have important methodological limits. https://metr.org/time-horizons/ https://metr.org/blog/2026-1-29-time-horizon-1-1/
  7. Anthropic, Economic Index: Cadences (Jun. 26, 2026): token use rises with estimated work value; higher-wage work and complex artifacts consume more compute. https://www.anthropic.com/research/economic-index-june-2026-report
  8. Epoch AI, Is a compute crunch coming? (May 25, 2026): estimated inference capacity ~3.4×/year versus demand proxies ~10×/year, with substantial uncertainty. https://epoch.ai/gradient-updates/is-a-compute-crunch-coming
  9. Epoch AI, Inference economics of language models (Jun. 17, 2025): inference demand has grown rapidly alongside test-time compute and agentic workloads. https://epoch.ai/publications/inference-economics-of-language-models
  10. Amazon, 2025 Shareholder Letter / 2026 materials: approximately $200B of 2026 capex and substantial customer commitments behind AWS investment. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-2025-letter-to-shareholders
  11. Alphabet, June 2026 investor presentation: expected 2026 CapEx of $180–190B. https://blog.google/alphabet/investor-presentation-june-2026/
  12. Microsoft FY2026 Q4 earnings call: adjusted calendar-2026 CapEx expectation of approximately $175B and continued AI infrastructure investment. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4
  13. Meta Q2 2026 results: expected 2026 CapEx of $130–145B. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/
  14. NVIDIA Q2 FY2027 results: Data Center revenue of $89.0B, +117% year over year. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
  15. Howard Marks / Oaktree, Taking the Temperature (Jul. 10, 2023) and How to Think About Risk with Howard Marks (Sep. 12, 2024). https://www.oaktreecapital.com/insights/memo/taking-the-temperature https://www.oaktreecapital.com/insights/insight-video/education/how-to-think-about-risk-with-howard-marks

Perceptron Alpha research notes are for informational and educational purposes only and do not constitute investment advice, an offer, or a recommendation to buy or sell any security.