Grok Bot and OpenAI Dots Face Compute Costs

Elon Musk said this week that running Grok Bot, the personal assistant SpaceX introduced in August, costs “unsustainably high relative to revenue.” He suggested SpaceX was reducing the computing power allocated to it, making users reach their usage limits faster. At a New York dinner with reporters, Adam Fry, who leads ChatGPT’s consumer business, said OpenAI’s Dots assistants are “on the compute heavy side.” Each Dot runs on its own cloud computer to carry out tasks using accounts and services a user has authorized.

Dots help users read and answer emails and find and book event tickets. Fry said a feature called background mode is intended to keep Dots available continuously. OpenAI’s ChatGPT voice mode has 250 million weekly active users, Fry said, while the company claims its search queries with search intent equal one-fifth of Google’s volume. Consumer devices designed by Jony Ive are set to debut this year.

Dots are currently limited to Pro subscribers, who pay $100 per month. Fry said he expects computing costs to fall rapidly as efficiency improves, even as model performance rises. OpenAI announced a task routing system that assigns work to different AI models, including GPT-6.1 Terra, which uses less computing power than GPT-6 Astra, the model that previously powered all Dots tasks. The company said the system will reduce Dots’ computing demands.

OpenAI this week began calling data centers “compute campuses” amid bipartisan opposition to data centers nationwide. The company said the campuses would ease pressure on local power grids and avoid taking electricity from local families; community resistance could slow permits or prevent facilities from being built. An updated Anthropic S-1 filing this week reported $895.4 billion in long- and short-term purchase agreements, “substantially all” related to compute and infrastructure, and listed difficulty securing more compute as a major risk.

OpenAI has not filed its S-1 or confirmed IPO timing. Fortune reported that sources expect it to face similar compute challenges; unlike Anthropic, whose infrastructure uses four chip types, OpenAI relies largely on Nvidia GPUs and is developing its Jalapeño chip, with production expected to scale in 2027. The company is still seeking funding and is unprofitable. The Wall Street Journal reported its latest funding round raised $11 billion, below a $30 billion target. The article says even $30 billion could be small relative to compute spending and that going public may be the only way to finance the personal agent push, without guaranteeing success.

Elon Musk said this week that running Grok Bot, the personal assistant SpaceX introduced in August, costs “unsustainably high relative to revenue.” He suggested SpaceX was reducing the computing power allocated to it, making users reach their usage limits faster. At a New York dinner with reporters, Adam Fry, who leads ChatGPT’s consumer business, said OpenAI’s Dots assistants are “on the compute heavy side.” Each Dot runs on its own cloud computer to carry out tasks using accounts and services a user has authorized.

Dots help users read and answer emails and find and book event tickets. Fry said a feature called background mode is intended to keep Dots available continuously. OpenAI’s ChatGPT voice mode has 250 million weekly active users, Fry said, while the company claims its search queries with search intent equal one-fifth of Google’s volume. Consumer devices designed by Jony Ive are set to debut this year.

Dots are currently limited to Pro subscribers, who pay $100 per month. Fry said he expects computing costs to fall rapidly as efficiency improves, even as model performance rises. OpenAI announced a task routing system that assigns work to different AI models, including GPT-6.1 Terra, which uses less computing power than GPT-6 Astra, the model that previously powered all Dots tasks. The company said the system will reduce Dots’ computing demands.

OpenAI this week began calling data centers “compute campuses” amid bipartisan opposition to data centers nationwide. The company said the campuses would ease pressure on local power grids and avoid taking electricity from local families; community resistance could slow permits or prevent facilities from being built. An updated Anthropic S-1 filing this week reported $895.4 billion in long- and short-term purchase agreements, “substantially all” related to compute and infrastructure, and listed difficulty securing more compute as a major risk.

OpenAI has not filed its S-1 or confirmed IPO timing. Fortune reported that sources expect it to face similar compute challenges; unlike Anthropic, whose infrastructure uses four chip types, OpenAI relies largely on Nvidia GPUs and is developing its Jalapeño chip, with production expected to scale in 2027. The company is still seeking funding and is unprofitable. The Wall Street Journal reported its latest funding round raised $11 billion, below a $30 billion target. The article says even $30 billion could be small relative to compute spending and that going public may be the only way to finance the personal agent push, without guaranteeing success.