Unable to sustain the burn, Meta calls a halt to the token consumption battle

Within just a few months, the "tokenmaxxing" model, once highly promoted by Silicon Valley giants, has gone through a process from rise to decline and finally to complete resistance.

According to media reports, Meta recently issued an internal memorandum, disclosing to about 6,000 employees that the company expects to spend billions of dollars on "internal AI use" alone in 2026. It also plans to officially implement a token management mechanism centered on budgets and quotas in 2027. This means that after strongly encouraging employees to use AI tools for several months, this tech giant is now turning to restricting internal token consumption.

Meta CEO Mark Zuckerberg admitted in the internal memorandum that the company made mistakes during the artificial intelligence transformation process. The company will arrange "meaningful new positions" for employees transferred to train artificial intelligence models.

In addition, Meta plans to guide employees to switch from third - party AI programming tools to the in - house developed programming assistant MetaCode to control token consumption costs. It is reported that the Applied AI Engineering department has arranged engineers to specifically improve the capabilities of MetaCode, training its programming response ability by repeatedly answering programming challenge questions.

The above developments have attracted market attention. Gary Marcus, a well - known AI researcher at New York University and the founder of the machine - learning company Geometric Intelligence, pointed out that "tokenmaxxing is giving way to tokenminimizing", and he expects this trend to cause the third - quarter revenues of Anthropic and OpenAI to be lower than their second - quarter performance.

Previously, there was a list called "Claudeonomics" within Meta to track employees' token consumption. This directly led to the chaos of "tokenmaxxing" - Employees competed to demonstrate their AI usage ability by increasing token consumption. Some even instructed AI agents to run multiple tasks in parallel, artificially inflating token consumption.

According to the list data, Meta employees consumed 60.2 trillion tokens in 30 days, with an overall cost of about $900 million. However, as AI costs continued to rise, the company did not generate real value internally. An insider bluntly said that what those at the top of the leaderboard produced was basically "one - time junk".

Many American companies have taken resistance actions. For example, Amazon recently clearly instructed employees "not to use AI just for the sake of using it" and switched to using "normalized deployment" indicators instead of token consumption. As of now, the "Kirorank" list used to track token consumption under its banner has stopped service.

Meanwhile, AI giants are gradually realizing the problem of the high cost of tokens. OpenAI CEO Sam Altman publicly questioned that "the increase in token support has not brought about actual productivity improvement". It is reported that the company is considering significantly reducing the token fees charged to users in order to snatch customers from its competitor Anthropic.

Guosheng Securities said that overseas large companies' views on token consumption are becoming more rational. Tightening token consumption has become a new trend among Silicon Valley giants. Research shows that for every dollar a company spends on tokens, about 80% is invisibly lost in bug fixing, code rewriting, and review delays. Individual efficiency improvement does not equal company revenue growth. Manufacturers are starting to shift from charging by tokens to charging by results, and the return to rationality is forcing the reconstruction of work processes.

This article is from the WeChat official account "Caixin Press AI daily", author: Zhang Zhen. Republished by 36Kr with permission.