EXPLORINEER
en

VN — VnExpress

Sam Altman: "It's a shame RAM sits in users' machines"

AI1 min read

OpenAI CEO Sam Altman said it is bad that scarce computer memory (RAM) sits in people's own devices instead of in data centers. His comments show a bigger fight over memory chips, which are getting more expensive.

What happened

In an interview with the YouTube channel Cheeky Pint, published on October 30, Sam Altman said many people want AI to run on their own devices. He said they believe the problem is solved when open-source models run on personal computers. Altman thinks RAM should be used in data centers instead.

What is local inference?

Local inference means an AI model runs directly on a personal device, such as a laptop or phone, instead of sending data to cloud servers. The device uses its own chips, like the CPU, GPU, or NPU, to do the work. Services like ChatGPT and Claude usually run in the cloud.

Key numbers and facts

  • Altman said RAM in a personal computer is used only about 5% of the time it is on.
  • Micron has signed 26 strategic deals, covering more than 35% of its expected revenue up to 2030. It also holds 32 billion US dollars in customer deposits.
  • Samsung is reported to have locked in long-term contracts for 75-80% of its expected memory chip output in 2027.
  • Analysts expect the memory chip shortage to last at least until 2028.

Why chip makers are acting this way

In past cycles, chip makers built more factories when demand rose. That often led to extra supply and falling prices later. This time, makers are choosing to raise prices many times over instead of building much new capacity. Supply stays tight, and high prices last longer.

Why it matters

Tom's Hardware said Altman is protecting OpenAI's business model. OpenAI mainly sells closed models that run in its data centers. If people can run AI on their own machines, they may pay less for them. The Verge's Steve Bonifiled said that the huge spending on RAM is making computers, phones, and other devices more expensive for everyone.

Glossary

RAM
Memory that a computer uses to hold data while it works. It is fast, but its supply is limited.
Local inference
Running a trained AI model on your own device instead of on a remote server.
NPU
A chip built to handle AI tasks quickly on a phone or computer.
OpenAI
An AI company that makes ChatGPT and other AI models.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

More in this topic

AIZDNet Korea

Samsung Executives Discuss AI Future Strategy with Sam Altman

Samsung held a briefing at its Yongin training center for presidents of major affiliates, including Samsung Electronics, with OpenAI CEO Sam Altman on the theme "The Next Decade Shaped by AI: Corporate Choices." The session, attended by about 50 executives, covered AI's impact on business decisions and organizational management, as part of Samsung's company-wide AI transformation (AX) push. Samsung said the aim is to raise executives' AI literacy and shape future management strategy.

AIThe Register

Amazon sends chills through its workforce with mixed message memo about AI use

Corey Quinn writes that an Amazon memo by VP and Distinguished Engineer Terk Terkowitz on the role of developers in the AI era has left many software engineers uneasy about their profession's future. The memo, first sent internally in September, emphasizes customer outcomes and engineering judgment and suggests checking whether teams with the highest AI token use actually increase delivery. Quinn notes that the memo lands amid Amazon's shutdown of the KiroRank token-usage leaderboard and reported new job cuts, though he does not claim these events are related.

AITechCrunch

Nvidia weighs playing the cosigner for Nscale’s AI data center in West Virginia

Nvidia is in talks to guarantee billions of dollars in debt financing for AI cloud provider Nscale's planned Anthropic data center in West Virginia, according to The Information, with Bloomberg later corroborating the report. The arrangement would be a residual value guarantee, under which Nvidia would step in if Nscale defaulted on construction payments, temporarily take over the facility, and help find a new buyer and tenant. The move underscores the financial pressure facing the AI buildout as investors grow wary of the large sums involved.

AIThe Register

Darkbloom, which allowed Mac users to rent out their GPUs for AI inference, warns of low payouts

Darkbloom, a distributed inference network from Seattle-based Eigen Labs, lets macOS users rent out their GPUs for AI model prompts routed through OpenRouter. Participants have reportedly earned only a few cents to a few dollars per day as token prices have fallen, with the median leaderboard earnings about $5.26 per day. Eigen Labs researcher Kydo said earnings have been low and are unlikely to recover soon, and the company is exploring other models and compute work.