AI demand is outrunning compute. Plan for higher model prices.

Revenue at the frontier labs is growing far faster than the compute that serves it, which gives the labs room to raise prices. Companies that move repeated work to specialist models, and measure cost per accepted outcome, keep more of their margin.

Demand is growing faster than compute.

Reports put Anthropic's 2025 revenue at almost twelve times its 2024 figure. The resources used to train and deploy AI are growing about 3.4x a year, according to Epoch AI. Put together, demand for model usage is rising about three and a half times faster than compute supply.

If that holds, intelligence becomes useful faster than compute becomes plentiful. That gap lets labs move capacity to inference, run their fleets harder and widen gross margin.

In most software markets, inputs like storage and bandwidth trend toward zero as hardware improves and competitors arrive. Access to the best models looks set to move the other way, toward scarcer and more valuable, and labs will price it that way. Products built on those models need a plan for that increase.

The tax on generality.

Prototypes need the smartest model and the shortest path to a working answer, and frontier models deliver both. Trouble starts when the work turns routine and the prototype architecture stays in place.

Production systems show their shape over time. The same tools, schemas, policies, corrections and failure modes keep returning, and reviewers define good work as the task narrows. Yet most companies still route every call to a generalist whose range the workload never touches. That is a tax on generality.

The lab gains twice. The customer pays the bill, and the lab learns from running its model across the market, while the customer often loses the experience that could make the next answer cheaper. As pricing power grows, handing over the surplus created by your own prompts, domain knowledge and volume gets harder to defend.

Give frontier models a different job.

Frontier models remain the quickest path to new capabilities and fit novel, hard or one-off work. They should stop being the default worker. Use them to explore new workloads, set baselines, help train candidate models and handle exceptions. Send repeated work with well-understood requirements to specialist open-weight models.

Measure cost per accepted outcome.

Price per million tokens is easy to compare and often misleading. A better measure:

Cost per accepted outcome = (inference cost + retries + review + operating cost) / successful task outputs

Latency, reliability and the share of traffic a route can safely carry belong in the comparison too. That requires your own evals. A public leaderboard says nothing about whether a model obeyed your rules, called your tools properly or satisfied your experts. Spending fewer tokens does not earn a model production traffic. Passing a test built for your workload does.

Move traffic only when a model passes.

Production work still generates learning: corrected tool calls, accepted answers, failed runs and preference signals. The loop is to capture, evaluate, train when the evidence supports it, deploy when benchmarks are met, and repeat.

Start with the cheapest fix that works. Sometimes that is a tighter prompt, schema or parser, or a current model that already clears the bar. Fine-tuning or dedicated serving makes sense once a task has the volume to justify it.

Routing should not be a fixed rule of small model for easy questions and large model for hard ones. A frontier model may start with almost all traffic. A cheaper route earns a slice after passing a held-out eval and gains more as it improves, while the frontier model keeps uncertain and new problems. Savings per accepted outcome pay for new use cases, extra usage generates evidence, and that evidence widens what the specialist covers.

Plan for higher model prices.

If lab revenue keeps growing at three or four times the rate of available compute, expect labs to defend their margins. Companies can respond by owning the data and models their operations produce, and by making providers compete for their frontier traffic.

Orchestra records how a current AI workload performs, builds held-out evals from that production evidence, trials lower-cost routes, and moves traffic to a specialist model once it clears the quality bar. Frontier models keep the hard, uncertain work.

Evaluate your workflows to find the repeated work that could move to a model you own.