Budget shock is already happening. Multiple major players have pulled back on AI features or subscriptions due to unexpectedly high token costs. Amazon removed its token leaderboard and Microsoft cancelled Claude Code subscriptions. These are early signals that the deploy-everywhere approach is hitting hard financial limits, not just theoretical ones.
Physical infrastructure is the binding constraint. Compute capacity, power consumption, cooling systems, memory bandwidth, and inference budgets are not soft or future constraints. They are active bottlenecks right now, determining what you can run, when, and at what price. Teams that plan AI deployments without factoring these in will be forced to reckon with them under pressure.
The real cost is price multiplied by volume, not unit price alone. Per-token pricing can look manageable on a spreadsheet until you multiply it by the actual volume generated by long-context windows, multi-step agent chains, retries, and verbose prompts. Model total inference spend, not just the listed rate, before declaring a workflow economically viable.
How Pricing Signals Should Shape Deployment Decisions
Higher costs are doing useful filtering work. Rising inference costs are pushing teams away from speculative or low-return AI deployments and toward use cases where the output clearly justifies the spend. This is a healthy market signal, not just a headwind.
Falling token prices don’t mean lower infrastructure demand. When the per-token cost drops, total usage often increases because more teams can afford broader deployment. The mix shifts toward cheaper, more efficient models rather than the market shrinking overall. Lower unit prices and growing infrastructure investment are not contradictions.
Demand elasticity is the key metric for evaluating use cases. Elasticity here means: how much does usage or output grow when costs fall? Where AI reduces costs for tasks with elastic demand, output can expand enough to actually increase demand for complementary human labor. Where demand is inelastic, productivity gains stay narrower. Think through this before committing resources to a new use case.
Reading regularly? Consider becoming a paid supporter 🙏
Building Token-Efficient AI Systems
Token-efficient design is a genuine competitive advantage. Prompt engineering, shorter context windows, retrieval-augmented generation (pulling in only the relevant data rather than feeding long documents in full), caching repeated responses, and batching requests are not just engineering hygiene. They directly shape unit economics and determine whether a product can scale profitably.
Model routing and fallback should be built into the architecture from day one. As cost sensitivity rises, teams will increasingly need to route different tasks to different models based on complexity and value. Architectures that make model substitution, fallback, and comparison easy will be far more resilient than those built around a single provider or tier.
Task-level operating metrics beat model benchmarks. Benchmark scores tell you what a model can do in a lab. Cost per resolved support ticket, cost per generated test, cost per completed analysis, or cost per workflow outcome tell you whether your deployment makes business sense. Track the latter.
Low-return AI experiments should be pruned on a portfolio basis. Treat AI deployments like a product portfolio and cut workflows that cannot demonstrate clear productivity, revenue, quality, or risk-reduction benefits. Speculative projects should compete for compute budget the same way they compete for engineering time.
Agentic Workflows: Tighter Economics Required
Autonomous agents need explicit cost guardrails, not just behavioral ones. Agentic systems (AI that takes sequences of actions with minimal human oversight) consume far more tokens and infrastructure than simple single-step assistants. Design them with clear stopping conditions, scoped permissions, per-run cost caps, and measurable value benchmarks before scaling them into production.
Complex workflows must justify their inference spend. Before building out a multi-step agentic system, compare the marginal productivity gain against the marginal compute cost. Many tasks that look like agent candidates are better served by a focused, single-step model call or a human-in-the-loop workflow at a fraction of the cost.
Human-in-the-loop design may scale better than full autonomy. Workflows that keep a person in the decision chain can often deliver strong results with fewer tokens, fewer compounding failures, and clearer accountability than fully autonomous systems. For most organizations, this is also the more defensible compliance and risk posture.
AI as a complement to human workers is the proven path. The most durable productivity gains come from pairing AI with people rather than replacing whole workflows: developers using coding assistants to write and debug faster, support teams using copilots to close tickets more quickly, and knowledge workers using models to compress research, drafting, and translation tasks.
Focused, bounded use cases scale better than grand automation visions. Search compression, drafting, testing, debugging, documentation, and support assistance are practical because they are bounded, measurable, and token-efficient. Prioritize use cases where you can clearly define the before-and-after outcome and verify performance at the task level.
The Frontier vs. Everyday AI Split
Two distinct AI tracks are forming. Heavy, inference-intensive frontier AI is becoming a specialty resource for organizations with deep capital and high-value problem domains. Simpler, cheaper models are the more practical productivity tool for the broader market. Most organizations will find their ROI on the everyday track and should plan their infrastructure and vendor relationships accordingly.
Frontier AI will concentrate among fewer, better-resourced players. Expect the most compute-intensive AI work to consolidate in firms that can absorb the infrastructure cost and operate in domains where solving genuinely hard problems at scale delivers outsized returns. For everyone else, simpler models are likely the right default until physical constraints ease materially.
The decline in LLM expenditure indexes is a substitution signal, not a demand collapse. Recent drops in token expenditure benchmarks reflect users shifting to cheaper, more efficient models rather than pulling back on AI overall. This is rational market behavior responding to cost pressure, and it is worth tracking as a leading indicator of where broader adoption is heading.
Governance and Business Planning
Cost governance is now a formal AI discipline, not a finance afterthought. Approval flows, usage monitoring, model-tier policies, and budget caps belong in production AI systems alongside safety and data access controls. Organizations that treat spending discipline purely as an engineering concern rather than a governance one will consistently be caught off guard by their token bills.
Budgeting should assume volatility, not stability. Token prices, capacity availability, and model performance will keep changing. Locking strategy or business cases around today’s specific prices and model costs is risky. Build in flexibility to adjust model choices, workflow designs, and spending thresholds as the economics shift.
Plan for uneven, slower diffusion rather than frictionless rollout. AI adoption will spread unevenly: concentrated at the frontier where returns justify the compute cost, and shaped by cost and capacity limits everywhere else. Planning assumptions built on gradual and selective deployment are more defensible than those built on ubiquitous, immediate adoption.
Long-term constructive, near-term cost-conscious. AI’s potential as a productivity-enhancing technology remains intact over a longer horizon, but the path there is more selective and cost-sensitive than markets have typically assumed. Build for a longer, more uneven adoption curve rather than pricing in a smooth upward trajectory.
