AI Token and Power Costs Force Companies to Rethink Spending

Companies that raced to embed AI across operations now confront electricity and token expenses that can erase projected savings from staff reductions. Uber and Microsoft executives have publicly questioned returns on aggressive AI adoption, while Meta reversed an internal directive that once rewarded high token consumption as a productivity metric.

The core tension lies in whether efficiency tools and cheaper models can contain costs before data-center buildouts and energy contracts lock in losses. A senior engineer at Netflix developed Project Headroom, released in January and now at version 0.22, to compress redundant data before it reaches large language models. The open-source tool has processed 200 billion tokens for users and delivered roughly $700,000 in documented savings, according to its creator Tejas Chopra. It targets server logs, JSON outputs, and file metadata that account for up to 90 percent of tokens in some workflows, while preserving reversibility through markers stored on Redis or SQLite.

Broader energy demand has elevated power itself into a corporate asset class. Ford launched a $2 billion energy-storage subsidiary after identifying data-center needs. Bloom Energy shares rose more than 1,200 percent over the past year, and GE Vernova recorded $2.4 billion in data-center equipment orders in the first quarter alone. Yet Heatmap Pro data show a record number of data-center projects canceled in the first quarter, representing more than $40 billion in planned investment, amid local opposition over water, noise, and emissions.

Price signals are shifting as well. Providers that once subsidized usage to gain market share are raising rates, particularly for agent-driven tasks that consume dozens of times more tokens than simple queries. Some firms now route work to smaller open-source or specialized models that can cut costs from $15 per million tokens to as little as five cents. Analysts at J.Gold Associates and Enverso note that token expenses have in isolated cases exceeded monthly employee compensation within the first one or two months of heavy use.

Unresolved questions remain around long-term demand durability and the scale of stranded infrastructure if cancellations accelerate. Chopra’s tool and similar projects such as RTK also reduce energy draw by shrinking context windows, though the net effect depends on whether lower per-task costs spur greater overall usage.