Argument
Architecture

The asymmetry is structural, not incidental. Training costs scale with how many models you build — a finite, episodic number. Inference costs scale with how many people use them, how often, and how much computation each request demands. These are continuous, unbounded variables. As the number of users, applications, and per-query compute all grow simultaneously, inference becomes the dominant and ever-increasing claim on the world’s AI compute infrastructure. Training is the factory construction. Inference is the assembly line that never stops.

This is not a one-time step change. It is a continuous escalation baked into the direction of AI research. Every improvement in model capability is being achieved by scaling computation at inference time — test-time compute. The models are not getting cheaper to run as they get smarter; they are getting more expensive per query because they are getting smarter. This escalating per-query cost acts as a hidden multiplier on every other demand driver: every new user, every new application, every new modality of consumption runs on a treadmill that is itself accelerating.

Chat’s growth potential as an inference driver is often underestimated because it is constrained by human speed — synchronous, a handful of exchanges per session. But the constraint applies to frequency, not to intensity. The compute cost per interaction is rising with each model generation as reasoning becomes the default, and the user base is measured in hundreds of millions. Even at human pace, hundreds of millions of users generating requests that each carry an inflating compute payload produce enormous aggregate demand. Chat is not the growth story — but it is the floor beneath every growth story, and that floor rises automatically with each model upgrade.

Three dynamics change simultaneously when developers integrate model access: volume becomes programmatic (millions of calls per day), context lengths expand (scaling compute super-linearly), and calls chain (multiplying per-interaction cost). Developers select the most capable models because capability is product quality, and the most capable models are the most compute-intensive. Smarter models attract harder tasks. Harder tasks consume more compute. Every SaaS platform, developer tool, and enterprise workflow adding an “AI feature” adds a new persistent stream of inference demand whose per-request intensity escalates with each model generation.

The agent paradigm is multiplicative across every dimension simultaneously: token volume per task (100–1,000× chat), reasoning at every step (compounding the most compute-intensive form of inference), machine speed (no human pauses), and escalating context (each step more expensive than the last). The arithmetic is explicit: a 2023 chat consumed X tokens; a 2025 chat with reasoning consumes 5–30X; an API pipeline consumes 10–100X of the new baseline; a single agent task consumes 100–1,000X of that inflated baseline. Agents do not merely add to inference demand. They multiply it by orders of magnitude — and they are just beginning to deploy at scale.

The skeptical view of cost declines is that they reduce total spend and shrink the infrastructure opportunity. The empirical record says the opposite. Every price reduction unlocks a new class of application — agents, background processing, always-on copilots — that generates 10–100× more tokens per user session than the workload it supplements. The volume effect overwhelms the price effect completely. Google’s case is dispositive: 78% cost reduction, 134× volume growth, 400% revenue growth, $175B+ in forward capex. The demand created by cheaper inference dwarfs the demand it replaces. Cost declines do not shrink the market; they detonate it.

Token throughput is the closest available proxy for aggregate inference demand because it directly measures computation performed. The numbers are not projections or forecasts — they are disclosed actuals from the companies with the largest visibility into AI compute consumption. A 134× increase in 18 months, with the rate of growth itself accelerating (doubling from May to July, then continuing), indicates the demand curve is not flattening. The fact that Google must double serving capacity every six months — and is still falling behind — confirms that supply is not catching up to demand despite record-breaking infrastructure expansion.

Inference demand scales with two variables: the number of endpoints generating requests, and the depth of each endpoint’s engagement. Consumer adoption provides breadth (billions of users). Developer adoption provides depth — and crucially, persistence. Every developer building an AI-powered application creates a workload that runs continuously and scales with their own user base, not just with their personal usage. This is the demand surface expanding in two dimensions simultaneously: more humans using AI directly, and more software systems embedding AI as infrastructure. The developer metric is the leading indicator because each developer-integrated application becomes a permanent, scaling inference consumer.

Enterprise workloads differ from consumer chat in three critical ways: they are persistent (running 24/7/365), high-volume (millions of transactions per day), and growing in complexity (from simple classification to multi-step agent workflows). The convergence of signals is striking — every frontier lab reports enterprise as the fastest-growing segment; every hyperscaler reports demand exceeding supply; every CEO frames the current moment as early-innings. The specific metric of 350 Google Cloud customers each processing 100 billion tokens in a single month demonstrates that enterprise inference is already operating at scale that dwarfs consumer usage per customer. And the leaders closest to the demand unanimously state the S-curve is inflecting, not topping.

Revenue is the hardest signal available. It is not survey data, adoption forecasts, or management optimism — it is customers paying money for inference compute. Anthropic’s case is particularly telling: with ~80% of revenue from API inference, its revenue curve is nearly a pure proxy for enterprise inference demand growth. The fact that these companies are profitable on inference (not subsidizing usage) and still supply-constrained (not demand-constrained) means the revenue figures understate the true demand. If Altman could double revenue with double the compute, then the observable revenue is bounded by supply, not by willingness to pay. Amodei’s analogy of rice on a chessboard captures it: we are on the 40th square, and the shocks from the first 39 combined are a fraction of what’s ahead.

The enterprise workload opportunity is structurally different from the consumer and API stories in three ways. First, the scale: millions of private applications, each a candidate for AI integration. Second, the amplification: interconnected systems mean the compute demand is proportional not to the number of applications but to the number of connections between them — a nonlinear relationship in large enterprises. Third, the durability: proprietary data creates unique value that makes each workload irreplaceable and permanent. These are not discretionary features that can be turned off in a downturn. They are operational improvements embedded in the way the enterprise runs — persistent, 24/7 inference consumers that exist for as long as the enterprise itself operates.

The productivity framing dramatically understates what is happening. AI coding tools create three distinct sources of inference demand. First, the direct demand: every line of code written, reviewed, or analyzed by an AI model consumes inference compute continuously, across millions of developers, indefinitely. Second, the recursive demand: AI makes developers faster, faster developers build more AI applications, those applications generate inference workloads, and the cycle accelerates. Third, the strategic demand: when developers use AI on production codebases, the models gain visibility into how enterprises actually operate — the workflows, data schemas, and business logic encoded in millions of lines of code. This is not a feature. It is a channel into the operational core of every enterprise on Earth. One Google principal engineer acknowledged that Claude Code reproduced a year of architectural work in one hour. The revenue trajectory ($500M to $5–6B in one year) confirms the market recognizes this.

Every major enterprise technology adoption in the last three decades — ERP, cloud, mobile — followed the same pattern: technology emerges, early adopters demonstrate results, consulting firms systematize the transition for the rest of the market. The consulting ecosystem does not merely measure adoption. It causes it. When McKinsey publishes that 88% of organizations use AI and BCG shows frontier adopters capturing 3.6× the shareholder return, a CEO who reads those findings does not have the option of waiting. The consulting flywheel is now spinning: measure adoption, publish findings, create urgency, staff implementations, measure ROI, publish again. Each rotation converts more enterprise workloads from evaluation into production — and from production into the persistent inference infrastructure that sustains them. The keyword analysis of 6,027 earnings calls confirms this is no longer abstract: CEOs are discussing agents, infrastructure, and GPUs in specific, deployment-oriented terms.