$1,000 a Day in AI Costs. Three Engineers. No Writing Code. No Code Review. But More Output.

A token-based economy is rapidly replacing “developer time” as the scarce resource in software—reshaping engineering work, enterprise budgets, and entire career ladders. Instead of paying for instructions executed by machines, organizations increasingly buy “units of purchased intelligence” measured in tokens. That shift matters because it turns intelligence into a variable, budgetable input: companies can dial up output by spending more tokens, but they also need new capabilities to aim that spending at measurable business value.

The cost side is falling fast. Inference costs per token are dropping at rates described as roughly 10x to 200x per year depending on benchmarks. Concrete examples include GPT-4–equivalent performance moving from about $20 per million tokens in late 2022 to roughly $40 “today,” while Claude 4.5 Sonnet is cited at about $3 per million input tokens, with expectations that prices could fall into the “cents” range within a year or two. The consumption side rises even faster: when intelligence gets cheaper, organizations use far more of it—an effect framed as Jevons’ paradox. The result is a new “physics of compute” built on hyperscaler infrastructure, where AI spending grows not because companies want waste, but because the economics make far more work viable.

That budget reality shows up in reported spending spikes and revenue-to-cloud-cost ratios. Strong DM CTO Justin McCarthy described a three-person team targeting roughly $1,000 a day in token spend with no handwritten code. Journalist Ed Zitron reported that Curser’s AWS costs jumped from about $6 million to over $12 million between May and June 2025, coinciding with Anthropic’s launch of priority service tiers. Zitron also cited Anthropic’s AWS spend of about $2.66 billion through September 2025 against estimated cumulative revenue of about $2.5 billion over the same period—before accounting for Google Cloud spend—implying cloud costs consuming more than 100% of topline revenue. Perplexity was also reported to have spent well over 100% of its 2024 revenue across AWS, Anthropic, and OpenAI combined.

The core organizational change is “token management” (or intelligence operations): the bottleneck moves from hiring and headcount to converting token spend into outcomes. Enterprises are building internal routing and platform layers that match tasks to the right model at the right cost, then measuring whether the purchased intelligence actually produces value. The speaker argues this is why token spend is increasingly treated as a lever for ROI rather than a cost to minimize—driving custom API agreements, consumption floors, and volume pricing with hyperscalers.

But token economics can also break businesses overnight when downstream providers raise prices. Cursor is used as a cautionary case: heavy reliance on Anthropic API costs reportedly made costs “uncontrollable” after priority pricing changes, forcing plan changes and triggering user backlash. The response included building its own model to regain control.

Career implications follow from the same premise. Software work is splitting into three tracks: (1) orchestrators who specify outcomes, manage agent workflows, and optimize token budgets; (2) systems builders who construct the infrastructure, evaluation pipelines, and routing layers; and (3) domain translators who combine technical fluency with deep market expertise to decide which problems are worth solving. The most exposed are those whose value is primarily generic application code. The most resilient are those who can manage intelligence throughput—whether inside large enterprises reorganizing around token-based productivity or in smaller, niche-focused startups where distribution and domain trust can outweigh raw compute scale.