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
Agentic AI is consuming so many tokens that it's draining AI budgets way faster than expected. Jevons Paradox rings true 161 years after it was coined.
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 | Tom's Hardware
Skip to main content
Unlock world-class roadmaps & trusted Bench data. See More
× Unparalleled insights. Industry analysis. Insider access. Tom's Hardware Premium equips you with world-class coverage and detailed insights into the evolving hardware landscape.
✓Full access to our trusted Bench database: Access granular performance data instantly. ✓Exclusive hardware roadmaps: Peer into the future of the hardware industry. ✓Daily news analysis: Dive deep into the biggest stories.
Subscribe to our annual plan for just $29
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Contact me with news and offers from other Future brands
Receive email from us on behalf of our trusted partners or sponsors
By submitting your information you agree to the Terms & Conditions and Privacy Policy and are aged 16 or over.
You are now subscribed Your newsletter sign-up was successful
An account already exists for this email address, please log in.
(Image credit: Getty Images)
Copy link
X
Share this article
14
Join the conversation
Follow us
Add us as a preferred source on Google
Newsletter
Subscribe to our newsletter
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 FromIt 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.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.Stay On the Cutting Edge: Get the Tom's Hardware NewsletterGet Tom's Hardware's best news and in-depth reviews, straight to your inbox.Contact me with news and offers from other Future brandsReceive email from us on behalf of our trusted partners or sponsors
TOPICS
See all comments (14)
Jowi MoralesContributing WriterJowi 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
Set stupid target and you get stupid results!
Reply
>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
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
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
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
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 w
📰Originally published at tomshardware.com
Staff Writer