# Personal agents are running into the compute bill URL: https://www.thedeepfeed.ai/posts/2026-11-07-personal-agents-hit-the-compute-wall/ Category: Products Published: 2026-11-07 Author: the-deep-feed Tags: agents, compute, openai, xai, meta, pricing Kind: deep > Within a few weeks OpenAI said Dots are compute-heavy, Musk said Grok Bot costs were unsustainably high relative to revenue, and Meta said it needs Muse to get 10x more efficient. Always-on agents cost real money. ## TL;DR - OpenAI's head of ChatGPT consumer, Adam Fry, called Dots **"on the compute-heavy side"**. Consumer access is on the **$100/month** Pro plan. - On Nov 3 OpenAI said it is adding **GPT-6.1 Terra** to Dots, so a dot can use different models for different parts of a task. - Musk said the cost of serving Grok Bot was **"unsustainably high relative to revenue"**. SpaceX reports **800,000** paid weekly users. - SpaceX's Q3 AI capex was **$16.4B**, or **84%** of total capex, according to Fortune. - Meta, per Fortune, expects to improve Muse's compute efficiency **tenfold** over the next year. Three of this year's biggest consumer agent launches ran into the same limit within a few weeks: compute cost. Each agent gets its own computer and, in OpenAI's words, works "24/7". That is a different cost structure from a chatbot. The responses so far are cheaper model routing, usage throttling and premium pricing, and those choices will decide who can afford a personal agent in 2027. ![Three responses to the compute bill: route to cheaper models, throttle usage, bet on efficiency](/post-images/2026-11-07-personal-agents-hit-the-compute-wall/compute-bill-three-responses.jpg) # OpenAI: route to cheaper models OpenAI launched [Dots](https://openai.com/index/introducing-dots/) on GPT-6 Astra, with each dot running on its own cloud computer. Its November 3 update said it is expanding Dots' models to include GPT-6.1 Terra: "Your dot can draw on different models for different parts of its work." The day before, Adam Fry, head of ChatGPT's consumer business, told reporters that Dots are "on the compute-heavy side," [Fortune reported](https://fortune.com/2026/11/04/will-compute-curtail-personal-ai-agent-push/). OpenAI is rolling Dots out to Pro and Business Premium users, with an Enterprise beta that workspace admins must enable; for consumers, Fortune notes, that means the $100-a-month Pro plan. Fry said he expects compute efficiency to improve enough for costs to fall quickly. [tbreak](https://tbreak.com/chatgpt-dots-automatic-model-selection/) notes that OpenAI hasn't said when automatic model selection reaches all users, or whether users can override it. We covered Terra's positioning as the [cheap tier for agents](/posts/2026-10-26-gpt-6-1-terra-cheap-tier-for-agents/). # SpaceX: throttle usage Musk replied on X that "The cost of serving Grok Bot was unsustainably high relative to revenue," [TechCrunch reported](https://techcrunch.com/2026/11/02/elon-musk-throttles-grok-bot-limits-due-to-unsustainably-high-costs/), as users noticed tighter limits. [The Verge](https://www.theverge.com/ai-artificial-intelligence/1024272/elon-musk-appears-to-have-throttled-grok-bot-for-being-unsustainably-expensive) said the same. Demand isn't the problem. According to [Fortune](https://fortune.com/2026/11/03/grok-bot-is-growing-fast-but-elon-musk-says-its-costs-are-unsustainably-high-compared-to-revenue/), SpaceX President Gwynne Shotwell said Grok Bot reached 800,000 paid weekly users, double September's figure. SpaceX spent $16.4 billion on AI capex in Q3, 84% of its $19.6 billion total. Much of that division's revenue growth came from leasing data centre capacity to Anthropic and Google. Musk said SpaceX aims to "split it 50-50" between its own AI products and renting compute to others, though Grok demand could change that. TechCrunch notes the "Super Intelligence" division still turned a $431 million profit. # Meta: bet on efficiency Fortune also reports that Meta, whose Muse agent has 10 million users since its September 8 launch, expects to improve Muse's compute efficiency tenfold over the next year. # Our read The common thread is that "agent minutes" have become the unit of cost, and none of the three has a subscription price that covers heavy use. OpenAI keeps Dots on premium plans and routes work to cheaper models. SpaceX cuts usage. Meta bets on efficiency. Fortune adds context: Anthropic's S-1 lists access to compute as a risk factor, and OpenAI's latest round reportedly came in at $11 billion against a $30 billion target. Expect more tiered limits and per-task model routing before personal agents become cheap. We tracked the supply side in [hyperscaler capex hits supply](/posts/2026-10-29-hyperscaler-capex-hits-supply/) and [2027 is already bought and financed](/posts/2026-11-01-2027-is-already-bought-and-financed/). ## Sources - [OpenAI — Introducing Dots (updated Nov 3, 2026)](https://openai.com/index/introducing-dots/) - [Fortune — Fortune Tech: Will compute curtail the personal AI agent push? (Nov 4, 2026)](https://fortune.com/2026/11/04/will-compute-curtail-personal-ai-agent-push/) - [TechCrunch — Elon Musk throttles Grok Bot limits due to 'unsustainably high' costs (Nov 2, 2026)](https://techcrunch.com/2026/11/02/elon-musk-throttles-grok-bot-limits-due-to-unsustainably-high-costs/) - [Fortune — Grok Bot is growing fast, but Elon Musk says its costs are 'unsustainably high' compared to revenue (Nov 3, 2026)](https://fortune.com/2026/11/03/grok-bot-is-growing-fast-but-elon-musk-says-its-costs-are-unsustainably-high-compared-to-revenue/) - [The Verge — Elon Musk appears to have throttled Grok Bot for being unsustainably expensive (Nov 2, 2026)](https://www.theverge.com/ai-artificial-intelligence/1024272/elon-musk-appears-to-have-throttled-grok-bot-for-being-unsustainably-expensive) - [tbreak — ChatGPT Dots will pick the AI model for each task (Nov 3, 2026)](https://tbreak.com/chatgpt-dots-automatic-model-selection/) --- Canonical: https://www.thedeepfeed.ai/posts/2026-11-07-personal-agents-hit-the-compute-wall/ Site: https://www.thedeepfeed.ai Full corpus: https://www.thedeepfeed.ai/llms-full.txt