Sam Altman recently sat down for a conversation on John Collison’s podcast, Cheeky Pint. During the discussion, he addressed people who want to run AI on their own computers. He described the idea as “quite terrible.” His stated reason was that the RAM in a personal computer only gets used about 5% of the time, so that memory belongs in the cloud instead.
The clip spread quickly on X, drawing over 225,000 views and more than 365,000 quote posts within an hour. Many responses compared the comment to a Chipotle CEO calling cooking at home a waste. Others noted that a lawnmower used once a week does not mean the owner should rent one from OpenAI, and that a McLaren used only 5% of the time does not make private ownership a poor allocation of resources. The common thread in the reactions was that the shortage of memory was not caused by consumers.
Last fall, OpenAI lined up deals for up to 900,000 memory wafers a month. A couple of months later, Crucial stopped selling RAM to regular buyers. Micron’s decision to shut down Crucial for regular customers removed a brand that had supplied RAM for many gaming PCs. One technical note put the price of 32 GB of RAM at $82 before it rose to $310 in four months. The same company now connected to large memory supply was telling individuals that their unused capacity was a waste.
Putting the Argument in Plain Terms
Memory is scarce. Cloud systems keep chips running all day. Personal machines sit idle for long stretches. The suggested solution is to move the memory into data centers so it can be used more fully, then sell the resulting AI intelligence back one token at a time. The practical result is a meter, rate limits, and rules set by the provider.
On the same podcast, Altman noted that his own customers push back when pricing moves away from a per-token model. People want to know what they will pay. They do not trust the price to stay fair. That observation sits alongside the claim that personal machines waste resources by sitting unused 95% of the time.
A Pattern of Statements on Ownership and Control
The recent comment fits a longer series of positions. In 2016, Altman pushed basic income—free money from the government for everybody. He wrote that he was “fairly confident” it would appear at national scale. He also funded a study in which 1,000 people received $1,000 a month for three years. Those participants worked about 1.3 fewer hours a week.
In 2021 he published an essay titled “Moore’s Law for Everything.” It proposed a tax of 2.5% of a company’s value every year, paid in stock, plus 2.5% a year on all private land. In 2023 he told the Senate, “I would form a new agency that licenses any effort above a certain scale,” which would require a government license for AI work above a certain size. In 2024 he stated that “compute is going to be the currency of the future.” At a United Nations AI event that same year he said, “The whole structure of society itself will be up for some degree of debate and reconfiguration in the post-AI world.”
These statements share a common direction: ownership and decisions about compute and intelligence sit with centralized institutions rather than with the people or companies using the technology.
Why Local Inference Changes the Equation
When frontier models run only in someone else’s data center, the user loses control over how the system behaves for a specific business. Rate limits, pricing changes, and rule updates arrive from outside. Local inference keeps the model on hardware the user owns. That separation matters for teams that need predictable performance and no external metering.
Recent work on local setups shows what is already possible. Small systems built around chips such as the 7940 HS or the 70 ADM can deliver 8 to 10 tokens per second—enough for one agent or one developer, though the pace is modest. An Apple M5 Pro with 64 GB of RAM runs comfortably. Higher performance comes from GPUs. A 3090 or better is the practical starting point, even though GPU prices have risen. Pairing 64 GB of RAM with a 3090 and the added strata layer produces more than 60 tokens per second. That speed is enough for two or three developers. The same machine can run a 180-billion-parameter model (Qwen 3.8 Flash Next) at that rate. Systems with 5090 GPUs and 64 GB of RAM are also in use.
For machines with less memory, a Bonsai model—a derivative of a 27-billion-parameter Qwen 3.8 model—runs on a 3060 with 12 GB of VRAM at 20 to 30 tokens per second. Benchmarks on these smaller configurations have been strong. These results come from actual hardware tests rather than cloud accounts.
Making Local Agents Practical
OpenMonoAgent is an open-source project that is a few months old and has roughly 2,000 stars on GitHub. It supplies a harness so people can build and run their own AI coding agents. Setup instructions and the source code are available at no cost, and new features continue to ship. One recent addition expanded the range of hardware that can be used and added the strata layer that raised throughput on the 3090-plus-64-GB configuration.
Installation follows a short sequence. A curl script is run and option two is selected so the system can auto-detect hardware and dependencies. An OpenMonoagent tunnel is then created, verified with an email one-time password, and the relay configuration is copied. The architecture keeps the heavy GPU inference on the local machine and connects it to the coding agent through a secure relay. On the client side, the VS Code extension is installed, the relay is connected, and the local setup becomes available from anywhere. The entire process is designed to be straightforward.
Ownership Versus Rented Intelligence
The difference is not abstract. Local runs produce zero API costs and zero telemetry. The model and the data stay on hardware the user controls. Cloud inference, by contrast, meters usage and subjects the user to external rules and pricing. Teams that treat AI as infrastructure they own—rather than a service they rent—can integrate it into their own architecture instead of wrapping someone else’s API.
Startuphakk builds custom software, including AI systems that run in the customer’s environment rather than on a vendor’s cloud or pricing schedule. The approach treats AI as part of solid engineering rather than a separate trend. Fractional CTO work supplies the same architectural and delivery standards without full-time executive overhead. The open-source platform itself serves as a working example of terminal-native coding agents that run entirely on local models.
The debate that followed Altman’s remarks is less about idle RAM percentages and more about who decides how scarce compute is used and who sets the terms. Local hardware and open tools keep that decision with the people and companies that depend on the results.
Conclusion
Sam Altman’s view that local AI is a poor use of scarce memory puts control of intelligence in the hands of data centers and their rules. Running models on your own hardware keeps the meter, rate limits, and decisions with you. Tools like OpenMonoAgent make that practical today, with real speed on accessible GPUs and zero external costs or telemetry. The choice is straightforward: own the infrastructure that runs your AI, or rent it under someone else’s terms.
