AI economics finally revealed… Microsoft shot first. GitHub Copilot is moving to usage-based billing on June 1. Pay-per-token replaces the flat subscription. Mario Rodriguez, Chief Product Officer, frames it as “alignment with usage”. Which makes sense: the flat-fee model was unsustainable. Frontier Labs have been absorbing $8 to $13 of cost for every dollar of revenue! Every question, clarification, silly picture or AI cat video has been subsidised by investor money, by a factor of ten. That’s why OpenAI decided to shutdown it’s AI video service Sora and cancelled its $1B licensing deal with Disney (who was caught by surprise in the process). AI pricing matters. As enterprises are pressured to adopt AI, they are building business cases and ROI models on subsidised economics that will not hold. Here is what is at stake: → Token Economics. Token consumption has scaled faster than price-per-token has fallen. Until now, asking the same question three times cost the same as asking it once, so variable and error-prone outputs were tolerated. Once each retry costs money, user tolerance for hallucinations will seriously diminish. Are the AI labs ready for the backlash? → Policy. In March, the US Department of War designated Anthropic a supply chain risk. The first time this designation has been applied to an American company. Defense contractors now have a countdown to remove Claude from covered workflows, regardless of technical fit. Anthropic is challenging it in court, but it sets a precedent. A sovereign model can become a prohibited dependency. → Enforcement. Last month, Anthropic's automated systems shut down 60 seats at a Latin American fintech (Belo) via a single email citing vague "Usage Policy" violations. No advance notice. The only escalation path was a Google Form. Access was restored days later as a false positive. For those days, all the company's AI-augmented workflows were down. Business continuity plans need to provision for these scenarios. Vendor concentration risk and the weak economics of frontier models are exposing companies to major disruptions. Diverging regulations only widen the uncertainty. Interestingly, China has been forced into a different posture. DeepSeek, Qwen, FP8 training, sparse activation (30 to 50% compute reduction at competitive performance). Constraint (restricted access to chips and energy) has forced architectural innovation. Narrow or Frugal AI is not a sustainability argument but an operational and economical one. Lighter models for narrow tasks. Edge inference where latency matters. Open weights where dependency becomes a business risk. For a deep dive on AI economics and how the US vs China are approaching the race to AI leadership, read my last piece on KoncentriK: link in comments. As the labs race to AGI, enterprises should build optionality as resilience and price AI projects on their real economics. Credits: Photo by Shamin Haky on Unsplash