Token prices are falling, AI bills are rising — the week the token became a board-level metric

Last week we covered how AI governance finally got a main stage in Geneva — and why the EU's August 2 deadline was the only date that actually mattered. This week the conversation shifts from what models can do to what they cost — and who's on the hook when the bill arrives. We're unpacking the token economy from the keyboard to the boardroom, looking at the tools helping teams stay ahead of it, and asking the question every CFO is starting to ask: are we actually getting value for what we're spending?

Three frontier labs cut model prices inside of eight days this month — Meta's Muse Spark 1.1 at $1.25 per million input tokens, OpenAI's GPT-5.6 Luna at $1 in and $6 out, and xAI's Grok 4.5 at $2 and $6. By the old SaaS logic, enterprise AI budgets should be exhaling. Instead, Forbes reported this week that agent spending keeps climbing — because the rate card is only one term in the bill.

The math is simple and brutal: total spend equals task volume, times attempts per task, times tokens per attempt, times token price — plus infrastructure. Only that last price term is falling. Everything else is exploding. Goldman Sachs forecasts token consumption will multiply 24x between 2026 and 2030, reaching 120 quadrillion tokens a month. That's the Jevons Paradox running at industrial scale — make a resource cheaper and consumption grows faster than the discount.

The bill is already forcing behavior change. Uber burned through its entire 2026 Claude Code budget in four months, and Gartner projects AI coding costs will rival developer payroll by 2028. The era of "tokenmaxxing" — throwing unlimited tokens at every problem and calling it velocity — is ending the way unlimited cloud spending ended a decade ago: with finance in the room.

I lived through this movie in the cloud era, and the plot is identical. The companies that won cloud didn't spend the least — they built FinOps muscle before their competitors did, then scaled with unit economics their rivals couldn't match. The business read: stop budgeting AI like a SaaS seat and start managing it like a metered commodity. Per-team spend caps, a daily dashboard, and one owner for the question that matters — are we getting more value per token than we did last quarter?

Read the full story →

The share of organizations that say they are actively tracking token spend in 2026 — up from just 3% in 2023. That 55-point jump in three years is the clearest single data point for why token economics has moved from an engineering footnote to a CFO line item. When nine out of ten companies weren't watching the meter, the bill was easy to ignore. Now that most are, the pressure to tie every token to a measurable outcome is only going to intensify.

Source: Lucidworks 2026 AI Benchmark Study

Caltech spinout PrismML released Bonsai 27B, a 27-billion-parameter multimodal model compressed to 3.9 GB that runs on an iPhone 17 Pro at 11 tokens per second, retaining over 90% of full-precision performance. It's Apache 2.0 licensed, and TechTimes reports Apple is evaluating the tech. The business read: every query that runs on-device is a query with zero marginal token cost — on-device AI is about to become a line item you can move spend to.

TSMC's Q2 profit surged 77% to a record on AI chip demand, with revenue up 36% year over year and the first revenue from its 2nm process, and the company announced an additional $100 billion investment in Arizona while raising full-year growth guidance past 40%. Every token you buy ultimately clears through this one company's fabs. For planners, TSMC's growth curve is the clearest signal that token supply — and pricing power — stays constrained into 2027.

After a 22-month regulatory wait, China's Cyberspace Administration approved Apple Intelligence for launch with Alibaba's Qwen models handling language features and Baidu powering search. Apple gets its AI story back in its second-biggest market; Alibaba gets the most valuable distribution deal in Chinese AI. The business read: the AI world is hardening into two model ecosystems, and multinationals will increasingly run different AI stacks — with different economics — in different markets.

In his first appearance at the World AI Conference, Xi Jinping called for a "symphony of global cooperation" and launched a new international AI cooperation organization, positioning China as the AI partner for countries squeezed by US export controls. NPR notes the pitch lands as US curbs tighten China's access to advanced chips. For global operators, the compliance map for AI procurement just got more complicated — and more strategic.

This week's Big Story makes the selection criteria obvious: the token economy is real, and the tools that help you manage it are no longer optional. This week's picks are built for teams that need to move fast without burning budget they can't justify — one for cutting the cost of the models you run, one for keeping your agents from going rogue on your invoice, and one for getting more done without touching a token at all.

Forbes talks to the CIOs and FinOps leaders building the discipline this week's Big Story demands — including one whose engineers each burn anywhere from a few thousand to $50,000 a year in tokens. The through-line: token spend is becoming a board-level metric, and "tokenmaxxing doesn't mean token usefulness." The perfect table-setter for a week where we unpack the token economy from the keyboard to the boardroom.