Three Key Takeaways
– OpenAI CEO Sam Altman said PC memory sits idle most of the time and would do more work shared in the cloud.
– AI-focused HBM and PC or server DDR are not interchangeable, but the same makers split factories and investment between them.
– Sharing suits occasional use of huge AI models, but high utilization does not guarantee lower user costs or freedom to switch.
โผ Watch the Japanese-language video explainer
This video has Japanese audio and slides. To read English captions, open Settings โ Subtitles/CC โ Auto-translate โ English, where available. Automatic translations may contain errors. If the option is unavailable in the embedded player, watch on YouTube. The accompanying English article provides the written explanation.
Altman Says PC Memory Is Used More Efficiently in Data Centers
OpenAI CEO Sam Altman appeared on “Cheeky Pint,” an interview show published by the payments company Stripe on October 29, 2026. The host was Stripe co-founder John Collison. Altman said the memory people install in their PCs sits unused most of the time and would be used more efficiently if it were pooled in data centers.
The remark came as a rebuttal to the view that local inference, meaning AI running on people’s own devices, together with open models, would make data centers unnecessary. Altman said PC memory is in use roughly “5% of the time,” compared with “maximum utilization in the cloud.” He did not cite a measurement for the 5% figure, and he did not propose restricting what individuals can buy.
On the same show, Altman also said memory makers have become cautious about expanding production because of past overinvestment and the downturns that followed. From October 30 onward, technology outlets including Tom’s Hardware and TechRadar covered the comments, and they drew discussion on social media.
Related Articles
How AI Uses Memory and Why Sharing Saves
Why does AI need so much memory?
Local inference refers to running an AI model on your own device, such as a PC, instead of on a provider’s servers. Whether a model runs locally or in the cloud, it needs memory. DRAM is the chip that temporarily holds the data a computer is working on. When an AI model generates an answer, it keeps its weights (the numbers that define what the model knows) and its intermediate calculations in memory, so larger models need more of it.
The GPUs used for large-scale AI computing are paired with a fast type of memory called HBM. According to Micron’s product description, HBM stacks DRAM chips vertically and places them in the same package as the GPU. DDR, by contrast, is general-purpose memory used not only in PCs but also in servers and workstations. Both are forms of DRAM, but they are not interchangeable: you cannot plug HBM into a PC’s memory slot, or the other way around.
They still end up in the same conversation because the same manufacturers decide how to divide limited factory capacity, wafers (the circular silicon discs chips are made from), and investment among these products.
Local AI and cloud AI
When people run AI locally, they often use open models, whose weights and other components are publicly released. Each model’s license determines what it may be used for, and the device’s hardware determines whether it can actually run. Cloud AI services such as ChatGPT run in the provider’s data centers, and users access them under the provider’s terms and pricing.
What sharing makes more efficient
Altman’s argument rests on utilization, the share of time a piece of equipment is actually working. A personal PC has many hours when it is not running AI, for example while its owner is asleep. A data center can fill that idle time with other people’s work. Pooling equipment also spreads the cost of power, cooling, and management across many users, which tends to lower the cost per task. These are economies of scale.
The 5% figure, however, is Altman’s own estimate of time in use, given without a stated basis. A fair comparison between local and cloud computing also has to set some conditions: the same workload at the same response speed, how network costs and peak-hour congestion are counted, and whether the question is about equipment people already own or about where newly produced memory should go.
AI companies securing memory, and the makers’ situation
On October 1, 2025, OpenAI signed letters of intent with South Korea’s Samsung and SK Group. A letter of intent is a document that confirms a shared direction before a formal contract. The figure of 900,000 wafers per month that appears in these announcements means different things depending on who said it.
| Source | What the 900,000 figure refers to |
|---|---|
| Samsung | OpenAI’s projected demand, up to 900,000 DRAM wafers per month |
| SK Group | Supply volume and terms to be negotiated |
| OpenAI | A plan targeting expanded advanced-memory production and 900,000 wafer starts per month |
None of these describes a confirmed supply volume or actual shipments.
On the manufacturing side, Micron said in its fiscal fourth-quarter 2026 prepared remarks, dated September 30, 2026, that the DRAM market was tight and that its own average DRAM selling price had risen by a high-teens percentage from the previous quarter. Micron also noted that adding factory capacity does not immediately increase shipments. Altman’s comment that memory makers are cautious about expanding reflects his own reading of this boom-and-bust history.
Measuring Efficiency, and the Freedom to Choose Again
How strong is the case for sharing?
For someone who uses a very large AI model only occasionally, renting shared data center capacity makes more sense than buying a large amount of memory. Because factories cannot be expanded quickly, concentrating newly produced memory in data centers and sharing it among many users delivers more computing from the same output.
Equipment people already own is a different case. Its purchase price is already spent. What has to be weighed against the cloud is electricity, upkeep, wear on the hardware, and the value the owner could get by using or selling the machine some other way. The answer also differs between someone who runs a machine for long hours every day and someone who uses it occasionally.
Even when providers’ costs fall, how much of that saving reaches users’ bills depends on competition and pricing structures. Higher efficiency for the equipment and lower total costs for users do not necessarily move together.
Whether the benefits of sharing win out has long depended on conditions. When Oracle promoted its “network computer” concept in the mid-1990s, the company’s 1997 annual report set the promise of lower ownership costs alongside conditions for adoption, such as low prices and ease of use.
What local processing offers, and what it costs
The main advantage of processing data locally is that you keep control over where your data goes and on what terms you use the model. Your prompts do not have to leave your device, and a model stored locally will not change its contents without notice because the provider decided to update it. Local processing also limits the impact of unstable connections or of a remote service changing its prices or terms.
The trade-off is that the owner bears the cost of hardware, electricity, repairs, updates, and security, and very large models may be hard to run at all. Open models come with license terms too, and users have to check for themselves whether to accept new versions or changes in distribution terms. Network improvements may narrow the latency gap, but the question of who sets the terms remains. Ownership is less an end in itself than one way of keeping the ability to choose those terms.
Pooling equipment is not the same as depending on a few companies
Concentrating computing in one place and relying on a small number of companies are separate issues. Sun Microsystems’ 1999 technical overview of its Sun Ray terminals described organizations pooling CPU and memory on their own servers while setting their own operating rules.
When computing is entrusted to an outside cloud provider, the provider decides prices, terms, and whether to discontinue a service. If a provider suspends access under its terms or shuts a service down, users cannot keep working the same way. The more dependence concentrates on a few companies, the harder it becomes for users to influence those conditions.
Altman speaks about efficiency as an interested party: OpenAI is also a major buyer of memory. His intent cannot be inferred from the remarks. What users can reasonably ask is whether the gains from sharing reach them as lower prices or better service. The key factors are how easily data can be moved to another service, whether users can choose among several services and open models, and whether local and cloud computing can be combined. Companies and local governments in Japan face the same choice as they adopt AI.
What will shift the balance
Technological progress can favor either side. Apple’s 2025 technical report presents both an on-device model of about 3 billion parameters (a measure of model size) and a server-based model. Making models more efficient expands what devices can do while also lowering data center costs.
On the network side, the International Telecommunication Union (ITU) published technical requirements for 6G in March 2026, and standards work is still under way. Faster, more stable networks could make remote processing less of a hassle. Computing may well continue to concentrate in data centers. A separate question remains: whether users keep enough room to choose again where and how they use AI.
Japanese Reactions to Sam Altman’s Comments on PC Memory
The posts below are reactions on X, not a measure of Japanese public opinion. They include three individual users and one commentator listed on a public profile as a financial planner and editor.
@valuefp (financial planner and editor, per public profile):
ใใฝใณใณใฎใกใขใชใชใใฆ็ก้ง้ฃใใใใผใฟใปใณใฟใผใซ้็ดใใๆนใใใใใใใฏ็พ็ถใงใฏๆญฃใใใจ่จใใพใใ่ชๅฎถ็จ่ปใฎ็จผๅ็ใ้ฑๆซใฎๆฐๆ้ใ ใใชใฎใจๅใใงใใใใฝใณใณใงใใคใจใณใใฎAIใๅใใใใซใชใฃใใใพใ่ฉฑใฏใฒใฃใใ่ฟใใพใใ https://t.co/OWg7bjbWl4
— ไธญๅถใใใตใฟ๏ผ Yahoo!ใใฅใผในใซ้ ไฟกใใ็ทจ้้ท๐๏ธๅฐ้ๅฎถใฎๅท็ญๆๅฐโ๏ธ่ญฆๆใใชใใฆใใFP๐ฐ (@valuefp) October 31, 2026
Pooling wasteful PC memory in data centers is right for now, like cars used a few weekend hours. High-end AI on PCs would reverse that. (translated)
@thara1129:
ไบบใ ใฏใAI ไปฅๅคใซใใณใณใใฅใผใฟใผใไฝฟใฃใฆใใฆใใใใฏ RAM ใๅฟ ่ฆใจใใฆใใใ
— HRA!.๐ ผ๐๐ (@thara1129) October 31, 2026
ใใใ AI ใ็ก้งใซ RAM ใไฝฟใฃใฆใใใฎใ ใใใ่ฒดๆนใฎใจใใใฎ AI ใฎ RAMไฝฟ็จๅน็ใๆนๅใใๅทฅๅคซใ่ใใใฎใไธ็ชๅน็ใใใใฎใงใฏใชใใใ๏ผ๐คฃ https://t.co/21bSc2J96x
People use computers for things besides AI, which also need RAM. If anything, AI wastes RAM, so isn’t improving your AI’s RAM use most efficient? ๐คฃ (translated)
@masa_zx:
ๅฐใชใใจใ่ชๅใฏ็ๆAIใใใใใๅไบบ็ใซใชใฝใผในใๅฟ ่ฆใจใใฆใใใ https://t.co/rMpkyyaqfN
— ใใต@็ซใจๆ ใจ่ๅใๅฅฝใใงใจในใใซใน๐งใจ้ๅฒก๐ปใๆใใ่ช็ฑไบบ (@masa_zx) October 31, 2026
At least for me, generative AI is why I personally need the resources too. (translated)
@Jean_Coc_Teau:
ChatGPTใซ่ชฒ้ใใใฎใใใใใใชโฆ https://t.co/mfaDBo8UA0
— ใใใจใใฉใ๐๐ (@Jean_Coc_Teau) October 31, 2026
Maybe I should stop paying for ChatGPT. (translated)
Using Less Waste, Keeping the Choice
What the efficiency argument answers, and what it leaves open
Altman’s case for shared memory is strong under one condition: occasional use of very large AI models. Pooling limited memory lets the same resources deliver more computing. What his argument answers, though, is how to waste less. Whether users pay less, and whether they can benefit from sharing while keeping the freedom to choose a different provider, are questions measured by a different yardstick. Owning hardware is one way to keep that room to choose. Easy data portability and the ability to use several services play the same role.
As computing concentrates, the real question may be which parts of our AI use we keep at hand, which we entrust to others, and whether we stay free to decide that mix ourselves.
Frequently Asked Questions
Did Sam Altman say people should not have RAM in their PCs?
No. In Stripe’s “Cheeky Pint” interview published on October 29, 2026, Altman argued that PC memory is used only about 5% of the time and is used more fully in the cloud, but he cited no measurement and proposed no restrictions on buying PCs.
Can AI-focused HBM and the DDR memory in PCs be swapped for each other?
No. According to Micron’s product information, HBM is stacked and built into the same package as the GPU, so it cannot be plugged in where DDR modules go, or the other way around. A key link between the two is that the same makers divide factory capacity, wafers, and investment among these products.
How did Japanese social media react to Sam Altman’s comments on PC memory?
Posts on X ranged from a commentator who called pooling in data centers correct for now, to users noting that PCs are used for more than AI or that they personally need resources for generative AI, to one user wondering whether to stop paying for ChatGPT. These are individual posts, not a measure of public opinion.
Sekahan on YouTube
We publish video summaries of articles like this one, along with short clips built around Japanese reactions.
Reference Links
- Sam Altman on Cheeky Pint (transcript)๏ฝStripe
- Samsung and OpenAI Announce Strategic Partnership to Accelerate Advancements in Global AI Infrastructure๏ฝSamsung Newsroom
- SK Group Partners with OpenAI to Advance Global AI Infrastructure๏ฝSK
- Samsung and SK join Stargate๏ฝOpenAI
- Fiscal Q4 2026 Prepared Remarks๏ฝMicron Technology
- High Bandwidth Memory (HBM)๏ฝMicron
- OpenAI CEO says it’s quite terrible for your PC to have RAM๏ฝTom’s Hardware
- Sam Altman says it’s quite terrible that you’re using RAM in your PC๏ฝTechRadar
