CONNECT WITH US

Tech

AI cost crisis hits tech giants as employee 'tokenmaxxing' backfires, sparking corporate pullback at Microsoft, Meta, and Amazon — agentic AI eats up to 1000x more tokens than standard AI

Published on

AI cost crisis hits tech giants as employee 'tokenmaxxing' backfires, sparking corporate pullback at Microsoft, Meta, and Amazon — agentic AI eats up to 1000x more tokens than standard AI

AI cost crisis hits tech giants as employee 'tokenmaxxing' backfires, sparking corporate pullback at Microsoft, Meta, and Amazon — agentic AI eats up to 1000x more tokens than standard AI

News
By Jowi Morales published

AI is getting too expensive.

Many tech companies are pushing their employees to use AI tools and increase their productivity, but it seems that this initiative has begun to backfire. According to The Verge, Microsoft has been reportedly pushing its people to switch to its own Copilot CLI rather than Claude Code because it wants to use an internal tool rather than a third-party one. However, sources say the primary reason is that the cost of using Claude Code has been steadily increasing as more people use the AI tool.

Microsoft is not alone in this, as Fortune reports that other companies are also pulling back on AI usage. While it’s true that the cost of training AI models is falling, making AI tokens more affordable, people have started using more tokens in their day-to-day tasks. This is particularly true for agentic AI, which can use a thousand times more tokens compared to querying an LLM, depending on the number of steps needed to accomplish your instructions. For example, OpenClaw creator Peter Steinberger claimed that his team spent more than $1.3 million in token costs in just a single month. Because of this, it’s now apparent that using AI is more expensive than hiring people, especially since it offers only limited productivity gains at the moment.

Decreasing token costs, paired with increased usage, reminds us of the Jevons Paradox, in which increased efficiency has led to more people using a particular tool or technology. There are many examples of this throughout history — the introduction of efficient steam engines during the Industrial Revolution led more firms to deploy these tools to increase productivity. This is also evident in the airline industry: as planes became more fuel-efficient, lower ticket prices led to higher demand, and air travel demand is now on track to double by 2050, according to IATA.

Latest Videos From

It seems that this is also true with AI tools, especially as many companies are deploying them in a bid to increase productivity. Nvidia CEO Jensen Huang famously said that its engineers should use AI tokens worth at least half their annual salary each year to be fully productive, even going so far as to say, “Are you insane?” to managers who discouraged AI use. This phenomenon, called “tokenmaxxing,” has led many employees to use AI for just about anything to hit internal targets. This was evident at Amazon, where some team members admitted to using the tool for unnecessary tasks to inflate internal usage scores, and it has also been reported at other companies, such as Microsoft and Meta. Incidentally, these companies are among the biggest spenders on AI development.

It’s unclear yet whether these companies will change their policies now that increased token use, which comes with associated costs, has become an issue. AI is indeed a useful tool, but some companies are using it to replace people in a bid to cut labor costs. If the number of tokens needed to accomplish tasks outpaces the speed at which these tokens become cheaper, then that move might just backfire.

Google Preferred Source

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Jowi Morales
Jowi Morales
Contributing Writer

Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.

14 Comments Comment from the forums
  • thisisaname
    Set stupid target and you get stupid results!
    Reply
  • JRStern
    >agentic AI eats up to 1000x more tokens than standard AI

    Hold on a second here, is vibe coding the same as "agentic AI"? I don't think so. This seems to be talking about vibe coding.

    It's been clear from the start this doesn't fly economically, even at fat discounts, they've been razzing it on YouTube. Might cost you $100k in tokens to fix a missing semicolon. SMH

    You know how you can have a discussion with an LLM and it saves it for you, so you can reload it and add a few more prompts? I asked ChatGPT about this months ago, "does it just save the text or does it save some kind of binary state?" Nope, just the text. Which means it's going to cost bigtime to reload a discussion. So if your "discussion" is 100 megabytes of your code project, it' gonna cost bigtime.
    Reply
  • DougMcC
    JRStern said:
    >agentic AI eats up to 1000x more tokens than standard AI

    Hold on a second here, is vibe coding the same as "agentic AI"? I don't think so. This seems to be talking about vibe coding.

    It's been clear from the start this doesn't fly economically, even at fat discounts, they've been razzing it on YouTube. Might cost you $100k in tokens to fix a missing semicolon. SMH

    You know how you can have a discussion with an LLM and it saves it for you, so you can reload it and add a few more prompts? I asked ChatGPT about this months ago, "does it just save the text or does it save some kind of binary state?" Nope, just the text. Which means it's going to cost bigtime to reload a discussion. So if your "discussion" is 100 megabytes of your code project, it' gonna cost bigtime.
    I've been working in a monolith that has >100k java files for months.

    My largest session hit 81MB. My 2nd largest is 31.

    It's not even possible to read that on opus-1m. Any task you would do with it would involve parsing it. But even if it were possible to insert that content as a simple turn, it would cost around $2 on my plan, which I assume is quite similar to most enterprise usage plans.

    I guess my point is: there is no realistic way to leverage such session transcripts in a way that burns massive amounts of money. It just isn't how such things are actually used.

    The notion of spending 100k to fix a trivial issue (fix a missing semicolon) even using an agentic implementation to do something so minor is highly unrealistic. It would take meaningful work to contrive a system so poorly designed as to spend even $10 on a minor fix. You'd really have to be embarrassingly bad at it.
    Reply
  • salgado18
    thisisaname said:
    Set stupid target and you get stupid results!
    https://i.pinimg.com/736x/44/13/66/441366100afea540c5d43efae5008f18--the-simpsons-ha-ha.jpg
    Reply
  • FoxtrotMichael-1
    DougMcC said:
    I've been working in a monolith that has >100k java files for months.

    My largest session hit 81MB. My 2nd largest is 31.

    It's not even possible to read that on opus-1m. Any task you would do with it would involve parsing it. But even if it were possible to insert that content as a simple turn, it would cost around $2 on my plan, which I assume is quite similar to most enterprise usage plans.

    I guess my point is: there is no realistic way to leverage such session transcripts in a way that burns massive amounts of money. It just isn't how such things are actually used.

    The notion of spending 100k to fix a trivial issue (fix a missing semicolon) even using an agentic implementation to do something so minor is highly unrealistic. It would take meaningful work to contrive a system so poorly designed as to spend even $10 on a minor fix. You'd really have to be embarrassingly bad at it.
    Hey, don’t let your experience and maturity actually, you know, using AI tools make you think you know more about these technologies than someone who clearly has never, ever used agentic AI to code.

    Claude Code writes the vast majority of code that I review and commit daily and I’m consuming maybe $35/day. I’m sure I’ll spend $100k to fix a semicolon issue any day now!
    Reply
  • JRStern
    DougMcC said:
    The notion of spending 100k to fix a trivial issue (fix a missing semicolon) even using an agentic implementation to do something so minor is highly unrealistic. It would take meaningful work to contrive a system so poorly designed as to spend even $10 on a minor fix. You'd really have to be embarrassingly bad at it.
    I'm not trying to do any of this myself but I've run across various discussions of it. When the organization encourages its use some people will push it to the limit, one report was a guy who used $150k/month in tokens at whatever their plan was, whatever their code mass was. They COULD consume the big mass and it could easily use exponentially more tokens with size. No doubt they were doing more than fixing semicolons.
    It might have been ten sessions a day for a month, or a hundred.
    Who knows, it might have been worth $150k and they can now fire nine other developers, but we just want to know what we're getting into.
    Reply
  • Plurality
    It's not a crisis because the companies involved don't care. The opinions of outsiders are irrelevant.
    Reply
  • King_V
    How the hell are these people who are trying to enforce more AI use, much to the harm of the company's bottom line, at all deemed intelligent enough to be put in charge?
    Reply
  • derekullo
    King_V said:
    How the hell are these people who are trying to enforce more AI use, much to the harm of the company's bottom line, at all deemed intelligent enough to be put in charge?
    Because they were ordered by management to promote AI at all costs.

    Management is operating under the assumption that if they don't force adoption now, while the tech is still messy, the company won't have the infrastructure or internal expertise when the inevitable industry-wide shift happens.

    They’re betting that the long-term competitive advantage of "learning by doing" outweighs the short-term cost of lost efficiency.
    Reply
  • 8086
    That headline had me thinking about:

    PdxJ-js52E4View: https://www.youtube.com/watch?v=PdxJ-js52E4
    Reply


Source link
Related Topics:#Tech

Disclaimer

We strive to uphold the highest ethical standards in all of our reporting and coverage. We StartupNews.fyi want to be transparent with our readers about any potential conflicts of interest that may arise in our work. It's possible that some of the investors we feature may have connections to other businesses, including competitors or companies we write about. However, we want to assure our readers that this will not have any impact on the integrity or impartiality of our reporting. We are committed to delivering accurate, unbiased news and information to our audience, and we will continue to uphold our ethics and principles in all of our work. Thank you for your trust and support.

Google Preferred Source