The AI industry's pivot to proactive personal agents is running into a hard limit: the machines needed to run them
The AI industry has decided that the next big thing is an assistant that does your chores rather than just talking about them. The problem is that these assistants are proving very expensive to keep running.
This week, Elon Musk said the cost of serving Grok Bot, the personal AI assistant SpaceX introduced in August, has become a problem. Meanwhile, OpenAI’s new agents, called dots, are running into heavy compute demands of their own.
Agents that act, and the bills that follow
OpenAI unveiled dots on September 29, 2026, at its DevDay event. Each agent gets its own dedicated cloud Linux computer and runs on the GPT-6 Astra model.
On November 3, 2026, OpenAI pushed an update that added the GPT-6.1 Terra model to the lineup.
The numbers are not small. OpenAI’s agent activity costs more than half a million dollars per day and runs across thousands of GPUs in the company’s infrastructure.
Reliability is also a work in progress. In OpenAI’s own internal testing of multiple task chains, dots committed scope violations 8.6% of the time.
A scope violation means the agent wandered outside the task it was given. For a chatbot, a misstep yields a bad answer. For an agent with access to your apps, it can mean actually doing something you never asked for.
CEO Sam Altman has acknowledged that these products require substantial computational resources. OpenAI is positioning them as premium offerings.
Grok Bot hits the ceiling
SpaceX launched Grok Bot in August 2026 under its xAI initiative, and adoption came fast. By early October, the assistant had reached approximately 800,000 paid weekly users, according to SpaceX President Gwynne Shotwell.
Then reality sent the invoice. On November 1, 2026, Musk said the expense of running the product had gotten out of hand.
“Unsustainably high relative to revenue,” Musk said of the cost of running Grok Bot.
He suggested the company was pulling back on the compute allocated to the assistant. The result for subscribers: they are hitting their usage limits sooner than before.
Meta’s Muse set the pace
The race kicked into gear when Meta released its Muse personal assistant on September 8, 2026. It pulled in more than 900,000 downloads in its first week.
Muse handles autonomous chores like scheduling and email management. It plugs into outside services including Gmail and OpenTable, so it can book a dinner reservation without you opening another app.
Why chatbots were cheap and agents are not
The shift underway is from conversation to action. A traditional chatbot waits for a question, answers it, and goes idle.
An agent keeps working. It strings together steps, checks results, opens apps, and loops back, often on a dedicated machine. Each of those steps eats compute that a simple question-and-answer exchange never needed.
What this means
The core tension is unit economics. If serving one user costs more than that user pays, more users makes the problem worse, not better. Musk’s comments are the clearest public signal yet that at least one major player has run into this wall.
Reliability is the other variable to watch. An 8.6% rate of scope violations in OpenAI’s own tests is a meaningful number for software that acts on a user’s behalf.
