Apple wants its users to use the new Siri AI on their devices all the time. Apple sells it as its grand return to AI. But behind closed doors, someone watched his employees’ AI spending excel (or numbers in his case) and began slicing a daily budget of tokens that Apple teams could use to improve their work.
This cut in the token budget for its employees is not just an Apple thing. Many companies that only a few months ago invited their employees to use AI in their work, have already received their first shipment of those tokens. The shock must have been so great that they stopped calling.
Apple scissors token. As published by BloombergLast month Apple quietly reduced some engineers’ daily token usage limit from $300 to $200. Now, the Turnus company is studying lowering that even further, leaving it at $100 for certain equipment.
Outside of engineering, some groups of employees will have $100 to $25 per day as tokens. Anyone who needs more can request an exception and its use will be studied. Apple employees have access to the cloud, ChatGPT and Gemini. Still, some teams are already asked to use cheaper models or models that work natively on devices.
Is the bill that expensive? This budget cut to use AI among employees draws attention from tech companies with the largest market capitalization. If Apple can’t pay for its employees to open bars in tokens, who can?
According to his calculations, the company spent 45,765 million on R&D in the last financial year, which is 32% higher than the previous year. Apple attributed the growth in part to its investment in AI. So why cut back? Employee cited by Bloomberg They maintain that “few workers use their entire $300 daily allowance,” so actual savings may be minimal. Other employees expressed their fear that these budget constraints could lead to more cuts or potential layoffs, as is happening at other tech companies.
Nvidia thinks just the opposite. Jensen Huang clears this up on the podcast All-in. An engineer with a salary of $500,000 will be very concerned if he spends less than $250,000 on AI tokens. Asked if Nvidia spends about $2 billion a year on tokens for its engineers, the CEO replied: “We’re trying.”
There is logic to Huang’s statement. Nvidia sells the chips that those models run on, so each token burned represents a little more investment in computing power and more sales for its chips. In Cupertino, which just pays the bills, the message is very different.
Apple is not the first. Known as event retardation Tokenmaxing It comes from months ago. There was an internal race to be the biggest token spender on the meta. On Amazon they stopped ranking similarly in May. Its subsequent memo discourages tokenmaxing, but the company says it has not set limits.
At Microsoft, a single employee burned through $28,000 in tokens in just one month. Microsoft responded. His vice president Jay Parikh wrote in an internal email to which he had access 404 medium That “tokenmaxing is not what we’re optimizing for.”
Change the Revolut model. With token bills in the spotlight, some companies are looking for other ways. Revolut has about 2,000 engineers programmed by default with GLM, a Chinese open weight model. Its parameters are public and can be run on its own server, which helps to significantly reduce the cost of tokens.
David Abitball, Revolut’s head of infrastructure, says they offer a tenth of the price of Frontier models. Additionally, the bill is fixed and does not increase with each request. Previously, the engineer for the cloud code was not $300 per month for a full month, now the query limit is defined by its own server.
In Xataka | From “Tokenmaxing” We’ve Moved to “Tokenwasting”: Waste in AI Reaches Unprecedented Levels
Picture | Unsplash (Wang Binghua)
