MiHoYo burned 2 million yuan worth of Tokens overnight. Even senior executives from large companies are starting to question: Can burning Tokens create value? And who is getting rich from this?

Finally, a senior executive has stepped forward to directly question the cost of tokenmaxxing based on their own practical experience.

Andrew Macdonald, the head of operations at Uber, directly stated in a recent interview that it has become increasingly difficult for the company to justify the cost of AI tokenmaxxing.

Previously, Praveen Neppalli Naga, the chief technology officer of Uber, revealed in April that Uber had used up its 2026 budget for Claude Code ahead of schedule. This statement sparked intense online discussions. Macdonald described feeling as if his "head exploded" at that moment, and the company also initiated a discussion about token consumption and the trade - offs it brings, such as whether it would affect the number of employees.

Through communication with Uber's senior engineering management, Macdonald realized that a higher token usage did not translate into a proportional increase in useful consumer features.

"The connection hasn't been established yet, right?" he said. "I think perhaps on an implicit level, more things have been delivered, but it's difficult to draw a clear line between these data and 'Okay, now we've really produced 25% more useful consumer features'."

Macdonald said that due to the inability to establish a direct connection, it is even more difficult to rationalize the trade - off costs brought by AI. It should be noted that earlier this month, Uber CEO Dara Khosrowshahi said in an earnings call that the company is slowing down recruitment to hedge against AI investment.

Macdonald added that if you're just "a user sitting there thinking of various interesting use cases" and don't have to pay yourself, AI does seem free. But ultimately, the company foots the bill.

Think Twice Before Going for Tokenmaxxing: Do You Have Enough Money?

For large companies, even a single daily attempt without using "Tokenmaxxing" can be very costly.

On May 20th, Zheng Yinhe, the head of the AI NPC & Gameplay technology team of the "Honkai" series, inadvertently revealed the cost of an internal agent experiment at the 2026 Alibaba Cloud Summit: An employee built dozens of agents to collaborate for a project, and as a result, it burned through tokens worth 2 million RMB overnight.

"We accept that there are costs and learning fees when exploring AI, and this also helps to improve our agent platform." However, some netizens pointed out that the value of these 2 million RMB doesn't seem to be significantly reflected.

Not every company can afford such "generosity". As some netizens joked, "A single 648 - yuan purchase is only a little over half a gold, and two million is just the money spent by three full - 65 hardcore players to support the game."

It's worth noting that on May 15th, Liu Wei, the co - founder of Mihoyo, said at an external event that the company will invest up to 100 billion yuan in the AI field in the next three years. He frankly stated that "even if it doesn't succeed in the end, we'll accept it. It's like setting off a big firework."

"As a company with the slogan 'all in ai', one day after spending 100,000 yuan on tokens, we blocked everyone's API," a netizen commented.

Some Companies Have Started to Make Adjustments

While large technology companies are still vigorously promoting the so - called "tokenmaxxing", some companies have started to make adjustments in the opposite direction.

For example, Duolingo previously decided to include AI usage in performance evaluations. However, after employees questioned whether they were just using AI for the sake of using it, the company withdrew this practice.

Duolingo CEO Luis von Ahn said in a podcast interview in April:

"Duolingo did include AI usage in performance evaluations, but later we cancelled it. I sent an internal memo to the company stating that performance evaluations would include AI usage. As a result, I found that employees were wondering, 'Are we using AI just for the sake of using it?' In the end, we retracted this requirement because the most important thing in performance evaluation is to do the job to the best of one's ability. AI can be helpful in many cases, but there's no need to use it compulsorily. We don't want people to ignore actual work results just to meet the form, as AI is not applicable in some scenarios."

Companies like Shopify have also started to find ways to prevent out - of - control token consumption.

Shopify is one of the early companies to try the Token leaderboard. Its engineering head, Farhan Thawar, once said that the company would praise those who used the most tokens because they might be doing valuable work with AI. For example, a developer who spends $1000 on Cursor in a month might have built a team of agent employees behind the scenes.

However, Shopify later renamed the "Token leaderboard" to the "Usage dashboard". The reason is that the company doesn't want to encourage employees to compete for the top position. Token spending is still displayed on the internal profile page and the usage dashboard, but the focus has shifted from "ranking" to "understanding usage".

Shopify has also set up a "circuit - breaker mechanism". If an employee's personal spending suddenly surges in a single day, the company can immediately cut off access. If this surge is intentional or due to an out - of - control agent, the employee can apply for restoration.

Farhan said that this mechanism not only helps the company detect out - of - control agents but also exposes bugs in the infrastructure. More importantly, Shopify will pay attention to what high - token users are actually doing. High - spending employees will be asked about specific usage scenarios, and if someone is just mindlessly using tokens, they are likely to be exposed in this process.

Farhan also proposed a more valuable perspective: Instead of just looking at "who spends the most money", it's better to look at "whose generated tokens are the most expensive". Behind some high - cost tokens, there are often more in - depth and complex development tasks, which are more valuable for reference than a simple total - volume ranking.

The Winners in the Tokenmaxxing Trend

The essence of Tokenmaxxing is to change the internal AI usage in enterprises from "on - demand call" to "encouraged consumption". As long as enterprises regard token usage as an indicator of advancement, model manufacturers will have an almost perfect growth flywheel.

The most direct beneficiaries of Tokenmaxxing are basic model manufacturers such as OpenAI, Anthropic, Google, and xAI.

Because tokens are the billing unit for model calls. Once enterprises turn token usage into an internal leaderboard or a performance signal, employees will shift from "calling AI on demand" to "actively creating AI usage". This will directly increase model API calls, enterprise subscriptions, and inference revenues.

Meta's case best illustrates this. Meta employees consumed 60.2 trillion AI tokens in 30 days. According to the Anthropic API price, this cost could be as high as $9 billion; even if Meta gets a discount, the cost could still exceed $100 million.

This means that once Tokenmaxxing is scaled up within large enterprises, it is no longer an ordinary tool expense but an inference bill in the hundreds of millions of dollars.

Anthropic's revenue growth also reflects this trend. According to the latest Reuters report, Anthropic expects its second - quarter sales in 2026 to exceed $10.9 billion, higher than the $4.8 billion in the first quarter, and it is expected to achieve an operating profit of $559 million. Previously, it was reported that Anthropic's Claude Code was approaching an annualized revenue of $1 billion in the year of its launch.

That is to say, for model manufacturers, Tokenmaxxing essentially transforms internal corporate anxiety into API revenue. Of course, this doesn't mean that model companies don't have computing power pressure, but Tokenmaxxing at least gives them considerable revenue.

The second category of winners are AI programming tools and agent platforms such as Cursor, Claude Code, Windsurf, Replit, and Lovable. For example, Anysphere, the company behind Cursor, completed a new round of financing in November 2025, with a valuation approaching $30 billion. The company said that its annualized revenue has exceeded $1 billion, and its sales - driven revenue has increased 100 times since the beginning of 2025.

However, this doesn't mean that AI tool companies can be profitable. According to foreign media reports, companies like Cursor and Windsurf are growing rapidly and have high valuations, but many of them are not yet profitable. One reason is that they rely on basic models such as Anthropic, and the inference cost is very high. Some companies have started training their own models to reduce costs and enhance control.

Therefore, AI programming tools can only be said to be winners in terms of traffic and valuation in Tokenmaxxing, and not necessarily all winners in terms of profit.

Tokenmaxxing ultimately burns computing power, and the biggest winners are still NVIDIA, cloud providers, GPU cloud service providers, and AI infrastructure companies.

Take NVIDIA as a typical example. Its financial data already reflects this trend. NVIDIA's annual revenue in fiscal year 2026 (ending in January 2026) reached $215.9 billion, a year - on - year increase of 65%; its revenue in the fourth fiscal quarter was $68.1 billion, a year - on - year increase of 73%. Subsequently, in the first quarter of fiscal year 2027, which ended in April 2026, its revenue further reached $81.6 billion, a year - on - year increase of 85%. Among them, the data center business and network hardware are the core driving forces, and the network sales reached $14.8 billion, a year - on - year increase of about three times.

Cloud and inference infrastructure startups also benefit. The AI infrastructure startup Modal Labs reached a valuation of $4.65 billion after the latest round of financing, and its annualized revenue increased from $60 million in September last year to $300 million. Its business exactly hits two trends: the soaring demand for AI coding and the increasing scarcity of computing power resources.

In this wave, the biggest losers are still those enterprises that cannot convert token consumption into assets but blindly follow the trend.

Big Rich Companies Are Too "Obsessed" with Tokens

Looking back, a lot of tokens have been wasted in this Silicon Valley "token consumption war", and the ones that are more "obsessed" are still the big rich companies.

Many Meta engineers said that there is a lot of meaningless AI usage internally. Some developers run internal agents similar to OpenClaw, consuming a huge amount of tokens but producing almost no effective results. Some developers also mentioned that some SEV online accidents are presumably related to careless AI code generation: developers focus more on using AI to quickly produce a large amount of code rather than the quality of the product itself.

What's more controversial is that the leaderboard has turned token usage into a gamified competition. Employees at the top of the leaderboard may not necessarily have done more valuable work. Instead, they may have created a large amount of disposable code and prompt records. After internal staff checked the AI prompt word trajectories, they could clearly see that many of these works had no practical meaning.

After an online backlash on social media and the disclosure of relevant data by foreign media, Meta took down the leaderboard.

There are also similar token usage dashboards inside companies like Microsoft, OpenAI, and Anthropic. In the early days, this mechanism did promote AI tool experiments, but problems also emerged: when token usage is linked to performance evaluation, promotion signals, and AI native degree, it is no longer just an observation indicator but a goal that employees deliberately pursue.

A Microsoft engineer admitted that he would actively engage in Tokenmaxxing, not to top the leaderboard, but because he was worried about being considered "not using AI enough". He would ask the AI questions that were already written in the documents; let the AI generate function prototypes that he would never develop; even if he knew that writing by hand was faster, he would still use the agent by default and then watch it fail.

This engineer has not been with the company for long and is more sensitive about job security. Therefore, he chooses to consume more tokens to prove that he is an "AI - native enough" employee.

Salesforce's approach is even more radical. It directly allows employees to compare token consumption.

According to an internal engineer, the company launched two tools: one is a Mac widget that updates personal spending every 15 minutes and shows the minimum expected spending at the same time. For example, the weekly target for Claude Code was $100, and the target for Cursor was $70; the other is a web tool that allows you to view the token spending of any colleague.

Salesforce also set a maximum monthly spending limit of $250 for Claude Code and $170 for Cursor, but this limit can be removed with a simple click of a button. Some engineering organizations have even removed the maximum limit, citing the reason of "eliminating any friction in the development process".

This sends a clear signal to employees: They should use at least about $170 worth of tokens per month, otherwise they may be considered to be using AI insufficiently.

As a result, employees started to actively burn tokens to reach the minimum line or be slightly above the average. Some developers would let Claude or Cursor generate projects that are not related to work and will not be actually delivered; others would check their colleagues' spending, calculate a "safe range slightly above the average", and then increase their token usage to that level.

Tokenmaxxing Is Even Promoted as a Startup Methodology

However, this trend has not stopped. YC is turning "Tokenmaxxing" from a Silicon Valley buzzword into a startup methodology.

In the latest issue of "Startup School", partner Diana Hu told entrepreneurs that the key transformation in building an AI - native company is not to continue to expand the employee scale but to maximize AI token usage. She said bluntly, "Maximizing token usage, rather than maximizing the number of employees, will be the key transformation. The best companies will be those that practice tokenmaxxing."

Hu believes that AI is changing the cost structure of startups. In the past, the work that a pre - AI - era company needed a large engineering team to complete can now be done by a person who is proficient in using AI tools. She said that this means that the engineering, design, human resources, and administrative teams can all be more streamlined.

She further suggested that entrepreneurs should be willing to bear "an uncomfortably high API bill" because this expenditure replaces the more expensive and bloated human - resource cost in the past. This means that in YC's new startup concept, a high - token bill is not necessarily a waste but may be a signal that the company is using AI to replace organizational expansion.

However, is this set of suggestions really suitable for early - stage startups?

As mentioned before, even AI programming companies with a good user base are not yet profitable. How can new startups bear the cost after tokenmaxxing? Do they really have to "sell their equity cheaply" to Altman to get $2 million worth of tokens and burn them for a month and a half?

Unlimited, ungoverned, and non - transferable token consumption will drag down these startups.

Conclusion

To some extent, the essence of Tokenmaxxing is that enterprises, under the anxiety of AI transformation, mistake "consumption" for "productivity".

It packages token usage as the degree of AI nativeness and the degree of AI nativeness as organizational advancement. Without data verification on whether more tokens bring faster delivery, fewer bugs, lower accident rates, higher revenues, and more reusable capabilities, blindly encouraging endless token burning is more like paying an expensive "transformation anxiety tax".

Reference Links:

https://www.businessinsider.com/uber-coo-andrew-macdonald-ai-token-spending-harder-justify-2026-5

https://mp.weixin.qq.com/s/Xvoh9hGnqe7rJ_ns5tRwBQ

https://mp.weixin.qq.com/s/JaGOyQ20UOTKkrXkJ2AXBw

https://blog.pragmaticengineer.com/the-pulse-tokenmaxxing-as-a-weird-new-trend/?utm_source=chatgpt.com

This article is from the WeChat official account "AI Frontline" (ID: ai - front), written by Chu Xingjuan, and is published by 36Kr with authorization.