AI Token and Power Costs Force Companies to Rethink Spending

A senior engineer at Netflix has built and open-sourced software that dramatically cuts the cost of using large language models, highlighting how quickly corporate enthusiasm for artificial intelligence has run into punishing bills. Tejas Chopra created Project Headroom to strip out redundant instructions, or tokens, before they reach models like Claude. He estimates up to 90 percent of tokens processed by many AI systems add nothing useful, yet companies continue to pay for them anyway.

Chopra’s own wake-up call came with a $287 bill for routine home coding tasks. What started as personal frustration has grown into a tool already adopted by several Netflix teams and external developers. In roughly five months the project has saved users an estimated $700,000 while freeing 200 billion tokens for other work. It has drawn 2,000 GitHub stars and more than 120 forks despite remaining at an early version.

The savings arrive as broader warnings mount about AI spending. Uber and Microsoft executives have already faced scrutiny after aggressive internal AI rollouts produced unexpectedly large invoices that threatened to erase savings from staff cuts. Industry analysts describe an earlier phase of “subsidized intelligence,” when providers kept prices low to lock in users. That era is ending. OpenAI and Anthropic are preparing to go public, and token prices are rising across the board.

The biggest cost driver is the shift from simple chatbots to AI agents that perform multi-step tasks such as booking meetings or writing code. Each agent workflow can trigger dozens of model calls, multiplying token consumption far beyond ordinary queries. Some development teams have coined the term “tokenmaxxing” for unchecked usage that now exceeds the cost of the employees doing the work.

At the same time, the infrastructure demands of this expansion are reshaping entire sectors. Electricity has become a strategic asset rather than a cheap commodity. Automakers like Ford have launched dedicated energy subsidiaries, while companies such as Bloom Energy and geothermal developer Fervo have seen stock prices surge on expectations that data-center demand will outstrip supply. Equipment orders at GE Vernova for data-center power systems already surpassed last year’s total in the first quarter alone.

Chopra’s open-source release offers one modest counterweight to these rising expenses. By making lossless context compression freely available, it lets smaller teams and individual developers avoid some of the token bloat that larger corporations have accepted as the price of staying competitive. Whether that grassroots fix can keep pace with pricing changes from the major AI providers remains to be seen, but the early results suggest many users were already looking for ways to push back against mounting bills.