loading . . . The AI Compute Crunch Is Here (and It's Affecting the Entire Economy) Earlier this week, I wrote an article about startups that are spending money on AI compute (tokens on tools like Claude and OpenAI’s products) rather than hiring human employees. There are all sorts of ways this business strategy could fail, and we are beginning to see signs that one of the most obvious ones could be coming to pass: AI companies can’t endlessly subsidize their AI products by charging users less than it costs to actually run them.
This is the AI compute crunch, and the signs are all around us:
* GitHub announced it is pausing new signups for Copilot, tightening usage limits, and removing access to several more expensive AI models.
* Anthropic has tightened access to Claude Code, and tested removing access to Claude Code entirely in its $20 per month plan (keeping access in its $100 per month plan)
* As noted in The Verge, Anthropic restricted Claude access to users of OpenClaw because the heavy usage was unsustainable
* OpenAI’s CFO Sarah Friar has been talking endlessly about how the company does not have enough compute, which has manifested in decisions like deciding to shut down Sora
* Software that has AI tools embedded in them have increased between 20 and 37 percent according to some analysts; this has included increases in prices for Microsoft 365, Notion’s Business plan, Salesforce, and Google Workspace prices
* There is a general rationing of AI products and services
* Meta is laying off 10 percent of its workforce in part because it sounds like the company wants to spend some of the savings on AI infrastructure: The layoffs are “to allow us to offset the other investments we’re making,” the company told its remaining employees_._ Its main recent investments have been data centers and the tech to run data centers.
But it’s not just that AI companies are restricting access to their products, shutting down products altogether, and beginning to increase prices. The broader impact of the current unsustainability of AI can be seen across various sectors of the economy.
* RAM, graphics cards, and hard drive / solid state storage for consumers have skyrocketed in price and are sold out in many stores. The same 2TB external SSD I bought late last year cost me $159 at the time, cost $449 a month ago, and costs $575 today.
* Similarly, the general cost of consumer electronics is increasing as chip manufacturers and production lines shift their focus to building more AI capacity. The largest consumer electronics manufacturer in the world, Apple, says it is having trouble securing chipmaking capacity for upcoming iPhones.
* Home electric bill costs have skyrocketed in some states with high concentrations of AI data centers, leading in part to a widespread, concerted effort by some towns and states to reject and restrict new data centers entirely. There is a fear among experts that similar shortages and price increases could come for water supplies as well.
What this means is that the age of cheap, underpriced AI appears to be ending, or at least the compute crunch means the venture capitalists and investment firms funding OpenAI and Anthropic are going to have to be willing to burn even more cash in order to continue subsidizing their products.
On the podcast this week, I compared this situation to Uber (and any number of fast-scaling startups that sought to lock in customers then jack up prices). This comparison is only useful in that, like Uber, what AI companies are doing to this point is wildly unsustainable and is being subsidized by investors. For years, Uber’s investors subsidized the cost of individual Uber rides to keep prices for consumers artificially low in order to gain market share, crush competition, and destroy the taxi industry. Uber and its investors could only lose money on each ride for so long as it continued to burn cash. This eventually led to enshittification for both riders and drivers as Uber suddenly jacked up prices for consumers and sought to find ways to pay drivers less. The difference, as Ed Zitron has pointed out, is that Uber’s costs were extremely low because Uber is essentially an app that owns none of the infrastructure, and so jacking up the cost of its service went quite a bit further toward getting it to break even.
Some version of this is coming for AI companies, but the path toward sustainability is far more complicated because of the enormous infrastructure and societal costs of scaling AI even further. “Make Claude more expensive and limit its services” is a lever Anthropic can pull, but AI companies are also burning money trying to build new data centers, juggling the political backlash to those data centers, fending off various copyright and public safety lawsuits, and spending huge amounts of money trying to train the next frontier versions of their large language models. None of this is remotely sustainable as it currently stands.
This means that the startups that are using AI agents to scale their operations are doing so at a time when AI costs are unsustainably low and may wake up one day to find that their compute costs suddenly double, 10x, or that they simply aren’t able to access compute anymore.
The general, long-term hope for the AI industry seems to be one in which multiple things need to happen to avoid a broader AI bubble burst. There needs to be a widespread renewable energy revolution (which society and our environment desperately needs), vastly increased chip and component manufacturing, and models need to become more efficient. On top of that, AI needs to be widely adopted and prove to be enduringly useful and reliable across a bunch of different sectors and use cases, something the jury is still very much out on (and some studies have already shown AI use is creating more work for humans, not less). All of this must happen while AI continues to put pressures on these systems that are making the problem worse (AI is making energy more expensive in the short term; lots of data centers are powered by fossil fuels; AI is pushing up the costs of components, chips, and gadgets, etc).
Finally, all of this must happen while society juggles whatever potential mass unemployment / economic fallout comes from AI and the ensuing problems this causes for these employee-less companies who expect to sell their products to a populous that is struggling to find work. As many commenters pointed out in response to my last story: If companies begin replacing their employees with AI agents, who are they going to sell their products to? https://www.404media.co/the-ai-compute-crunch-is-here-and-its-affecting-the-entire-economy/
