In a conversation published on October 29, 2026, OpenAI CEO Sam Altman argued that pooling scarce DRAM in the cloud and running it at high utilization uses resources more efficiently than spreading it across personal PCs. He also said the DRAM installed in PCs is in use only "5% of the time," but he did not explain how that figure was measured or what it covers. His remarks come as demand for memory in AI servers surges, yet PCs and servers use memory differently and rely on different product configurations. To judge whether consolidating in the cloud really lets the same work run on fewer resources, you have to consider both the manufacturing constraints on memory and the conditions under which shared equipment pays off.
"5% of the time": a remark that came out of a discussion on memory supply
In the Cheeky Pint conversation, John Collison asked Altman how manufacturers would increase supply going forward, given rising prices for memory and GPUs. Altman replied that manufacturers, having lived through past booms, busts and overinvestment, are cautious about expanding capacity. His point was that even a sudden jump in demand does not immediately produce more production capacity.
From there, Altman turned to the idea that running open-source AI models on personal PCs could reduce reliance on massive data centers. If the DRAM in a PC is used only 5% of the time, he said, that is a "pretty terrible" allocation of resources worldwide. His argument is that, rather than running trained AI models on local devices (local inference), it is more efficient to concentrate limited memory in the cloud and share it among many users.
However, this "5%" is not a figure for memory capacity utilization. It does not mean that 95% of a PC's memory capacity sits idle, and it is unclear what Altman counted as memory being "in use." The types and number of PCs covered and the measurement method were not given, so the number cannot be treated as a statistic drawn from a survey of PC usage around the world.
The idea of reducing idle equipment in order to make good use of scarce resources is reasonable. But whether DRAM in personal PCs should be treated like equipment dedicated to AI is a separate question. PC memory is also used by the operating system, web browsers, image-editing software and many other non-AI tasks. Time spent not running local AI does not automatically count as time the memory is wasted.
Do AI servers and PCs use the same memory?
When considering the efficiency gains from consolidating in the cloud, the differences between the memory used in PCs and in AI servers also matter. In its third-quarter results announced on October 28, 2026, SK hynix reported a sharp increase in sales of server DDR5 and the start of full-scale HBM4 shipments. The company also plans to expand sales of DDR5 and LPDDR5-based products for AI servers. In other words, memory demand from AI servers extends beyond HBM to conventional DRAM products.
HBM stacks multiple DRAM chips vertically to move large amounts of data at high speed between the memory and the processor. AI servers also use DDR5 and LPDDR5-based products, for example as CPU-side main memory. These serve different purposes and use different packaging, and PC DDR memory cannot simply be used as HBM. Even so, different products are not unrelated when it comes to DRAM manufacturing resources.
In an explanation of HBM manufacturing, SK hynix says HBM requires more wafers than standard DRAM of the same capacity. Because HBM includes features such as through-silicon vias (TSVs) that connect the stacked chips, die area grows and fewer chips can be made from a single wafer. So even when the same amount of memory capacity is supplied, the amount of wafer needed differs by product. It is not as simple as adding up PC DRAM capacity and redirecting it to AI servers.
SK hynix forecasts that global memory demand will continue to exceed supply through 2030. It also says that long-term contracts with 14 major customers, including cloud providers, account for more than 50% of its total production capacity for 2027–2028. But this reflects SK hynix's own demand forecasts and contract position; it does not mean half of the world's memory production capacity has been secured. Nor does it mean OpenAI alone has locked up that volume.
These circumstances illustrate how hard it is to expand memory supply to meet surging AI demand. But these figures alone cannot show how much the allocation of production to HBM affects supply or prices of PC DRAM. How much retail prices for PC memory will rise from here would need to be examined separately.
Does cloud consolidation really save memory?
The benefit of pooling memory in the cloud is that multiple users can share the same computing resources. For example, compared with preparing a dedicated AI environment for each user, the same equipment can handle multiple requests if usage falls at different times. Users no longer need to secure their own memory and processors, and idle time for equipment may also fall. Altman's argument holds up best where demand from users is combined in this way, raising equipment utilization.
Still, concentrating memory in AI servers does not by itself maximize efficiency. NVIDIA's technical explanation of Dynamo describes a mechanism that manages the data needed for inference across GPU-side HBM, CPU-side DRAM and storage. By keeping data in the right place and optimizing cache reuse and request routing, the design balances memory capacity against responsiveness. Even in cloud environments, what matters is not simply installing large amounts of memory but deciding where to store which data and when to reuse it.
When many users use AI at the same time, enough compute and memory are needed to handle that demand. If response-time requirements are strict, processing requests one after another is not enough. The size of the model and the amount of cache kept during inference also determine how much memory is needed, and redundancy to guard against failures is essential. Even if sharing raises utilization, equipment cannot necessarily be run at close to 100%.
Local inference, on the other hand, can have resource-efficiency advantages under certain conditions. If AI runs on spare compute while sharing the memory of a PC the user already owns with other tasks, no additional memory purchase may be needed. Conversely, if someone adds PC memory only to run a large AI model and then barely uses it, sharing in the cloud could be more efficient. What matters is not the total installed memory capacity but how many additional resources are needed to carry out the same processing.
Therefore, comparing the efficiency of cloud and local setups requires running the same task at equivalent quality, with conditions such as the number of concurrent users and response time held constant. Even if cloud sharing reduces the memory capacity needed, that alone does not mean operating costs or power consumption will necessarily fall too. DRAM supply constraints raise the value of cloud sharing, but a single utilization figure cannot settle which approach is better.
OpenAI also offers local AI: how does it fit with cloud consolidation?
OpenAI itself offers options for running AI models in local environments. Of the open-weight gpt-oss models it released on August 5, 2025, gpt-oss-20b is described as running on edge devices with 16GB of memory, making it suitable for local inference. By releasing the model weights, the company made it possible for users to run the AI on their own devices. However, the 16GB figure is the operating requirement OpenAI gave; it does not mean everyone must buy an extra 16GB of memory to use AI on a PC.
At the same time, OpenAI is pushing ahead with a large-scale expansion of its AI infrastructure. In its Stargate collaboration with Samsung Electronics and SK, announced on October 1, 2025, the company planned to expand production of advanced memory and set a target of 900,000 DRAM wafer starts per month. That figure is a target for wafers entering the manufacturing process; it does not represent supply already secured or current production. Nor does it mean OpenAI will monopolize that share of the world's DRAM supply.
For OpenAI, whose main business is cloud AI, there are business reasons to raise data center utilization and secure the memory it needs on a stable basis. Yet while Altman backs concentrating memory in the cloud, it is also a fact that OpenAI itself publishes models intended for local execution. The question of how society as a whole should allocate scarce memory and the question of which way of running AI suits each individual user need to be considered separately.
For users, network conditions and data handling are also important factors. Processing locally can reduce dependence on connectivity, but network traffic can still occur when linking with external tools or services. Local processing is also not necessarily safer or cheaper. The cloud involves usage fees and service availability; local setups involve the cost of buying or upgrading hardware and processing performance. Each carries different conditions to weigh.
To test Altman's claim, one would need to compare the memory capacity, additional cost, power consumption and response time required in the cloud versus locally, assuming the same task and required quality. Adding conditions such as the number of concurrent users and whether data can be sent externally would show how much resource saving cloud consolidation actually delivers. Even with memory supply constrained, which of the two suits a given case depends on how it is used and what performance is required.