Musk Calls Grok Bot Costs 'Unsustainably High' as AI Agents Hit Compute Wall

Elon Musk said this week that the cost of running Grok Bot — the personal AI assistant SpaceX introduced in August — is "unsustainably high relative to revenue." His admission, made in a post on x.ai, came with a warning that the company is throttling back compute allocation for the technology, causing users to hit their limits faster.

OpenAI is hitting the same wall. At an intimate dinner with reporters in New York City last night, Adam Fry, head of ChatGPT's consumer business, said the company's Dot assistants are "on the compute heavy side." Each Dot runs on its own cloud computer, which it uses to execute tasks on the user's behalf — reading and responding to emails, finding and booking event tickets, and a wide range of other assistant duties. A new "background mode" aims to make Dots an always-available presence.

Why the industry is betting anyway

Meta's Muse personal assistant has convinced the AI industry that proactive personal agents are the next logical step for the technology. The question is whether there's enough compute to power them on the timeline these companies are betting on.

OpenAI is stacking its pieces. ChatGPT's voice mode now has 250 million weekly active users, Fry said. OpenAI claims ChatGPT search volume — queries with search intent — is now one-fifth of Google's. And consumer devices designed by Jony Ive, the legendary iPhone architect, are set to debut this year.

What is OpenAI doing to cut compute costs?

Today, OpenAI announced a routing system that sends different Dot tasks to different underlying AI models, including the new GPT-6.1 Terra. Terra is smaller and draws less computing power than GPT-6 Astra, the larger model that originally powered all Dot tasks. The automatic task router reduces how compute-intensive Dots are.

Access remains the real constraint. Dots are limited to Pro subscribers paying $100 per month. Fry said he expects a breakthrough in compute efficiency will cause costs to fall rapidly even as model performance improves — but the more powerful and constantly running personal agents become, the more compute they consume.

Can the labs outrun the backlash?

The nationwide, bipartisan backlash against data centers has grown so strong that OpenAI kicked off a rebranding effort this week, renaming them "compute campuses." In its inaugural post for the new term, OpenAI promised to reduce strain on local electrical grids and not take limited power supply from local families. Community pushback could still slow permits or kill facilities outright if the local benefit on jobs and environmental protection isn't clear.

How much are the labs actually spending?

Anthropic's S-1 filing quantifies the bill: the company has entered into long-term and short-term purchase agreements totaling $895.4 billion, "substantially all of which is related to compute and infrastructure," according to an updated amendment filed this week. The filing lists inability to secure even more compute as a major risk factor that could cause Anthropic to be unsuccessful.

OpenAI hasn't filed its own S-1 or confirmed IPO timing, but sources say it almost certainly faces similar challenges — plus one unique to its fleet. Anthropic's infrastructure spans four chip types, while OpenAI depends largely on Nvidia GPUs. OpenAI is developing an in-house chip, called Jalapeno, set to scale up production in 2027.

Where does the money come from?

OpenAI, like Anthropic, doesn't yet make a profit and is still searching for more funding. Its latest private round fell short of its $30 billion target, landing at $11 billion, the Wall Street Journal reports. Even $30 billion would likely be a drop in the bucket against compute spending. The only option to fund the personal AI agent boom may be going public — and even that carries no guarantee of success.