OpenAI CEO Sam Altman believes that keeping large amounts of memory inside personal computers is an inefficient way to use the world's increasingly constrained supply of DRAM. In his view, memory sitting idle in home PCs for much of the day could potentially be put to better use inside large data centres where the same resources can support many different users and AI workloads.
Altman discussed the issue during an appearance on the Cheeky Pint podcast, where the conversation covered the ongoing global memory shortage, rapidly growing AI infrastructure requirements and the increasing popularity of running artificial intelligence models directly on personal computers. His comments highlight a growing tension between cloud-based AI services and users who increasingly prefer running models locally on hardware they own.
Altman Says Most Consumer RAM Sits Underused
Altman's central argument is based on utilisation. A desktop computer might contain 32GB, 64GB or even more memory, but that capacity may spend much of the day unused or only partially occupied.
He argued that from a global resource-allocation perspective, this is inefficient when DRAM is already becoming increasingly difficult and expensive to obtain. According to Altman, it is undesirable to have large quantities of the world's limited memory supply distributed across individual computers when much of it may only be actively used a small percentage of the time.
His argument is essentially that scarce resources should be placed where they can remain productive for longer periods.
Cloud Data Centres Can Keep Memory Busier
Data centres operate under a very different utilisation model from personal computers. Instead of one person owning a large block of memory that may sit mostly idle overnight or while the machine is switched off, cloud infrastructure can continually allocate memory between many users and workloads.
One AI inference request may finish and release resources that can immediately be used by another customer. Large cloud platforms can also balance workloads across thousands of servers to maximise the amount of hardware actively doing useful work.
From this perspective, concentrating memory inside shared infrastructure can theoretically allow the same amount of DRAM to serve significantly more AI activity.
The Argument Makes Sense From An AI Infrastructure Perspective
Altman's position is easier to understand when viewed specifically through the requirements of large-scale artificial intelligence.
Modern AI models can require enormous quantities of memory, both during training and inference. Training clusters may contain thousands of accelerators, while inference infrastructure must simultaneously support requests from large numbers of users.
For companies operating AI services, memory that remains unused represents wasted capacity. Data-centre operators therefore spend significant effort ensuring expensive hardware maintains high utilisation.
Applying the same logic globally naturally makes consumer PCs containing large amounts of idle RAM look inefficient.
Local AI Users May See Things Very Differently
The argument becomes more controversial when viewed from the perspective of people who deliberately run AI models on their own computers.
Local AI inference has become increasingly popular because it allows users to run language models, image generators and other AI systems without depending entirely on external cloud services.
For these users, having substantial local memory is not necessarily wasteful. It provides the ability to run models privately, experiment without recurring subscription costs and continue using AI even without an internet connection.
What appears underutilised from a data-centre perspective may therefore still provide substantial value to the person who owns the machine.
Privacy Is One Reason People Choose Local AI
One of the strongest arguments for running models locally is privacy. Data sent to a local model can remain entirely on the user's computer rather than being transmitted to a remote service.
That can matter when working with confidential business information, personal documents or proprietary source code.
A local model may not always match the capability of the largest cloud-based systems, but some users are willing to accept that trade-off in exchange for greater control over their information.
From this perspective, local RAM is supporting something cloud infrastructure cannot always provide as easily: complete local control.
Local AI Also Avoids Recurring Cloud Costs
Cloud AI services frequently charge through subscriptions, API usage or other consumption-based models. For occasional use, those costs can be relatively small.
Heavy users can see the economics differently.
Someone who already owns a powerful PC with sufficient memory may prefer making a one-time hardware investment rather than continuously paying for cloud inference. Once the hardware has been purchased, running models locally may have relatively predictable costs based primarily on electricity.
That can make large amounts of consumer memory economically useful even when the system is not operating at maximum capacity every minute.
Users Also Gain Independence From Service Providers
Local computing provides another advantage that is easy to overlook: independence.
A locally installed application continues functioning regardless of whether a cloud provider changes its pricing, introduces usage limits or discontinues a particular model. Users can choose which models they run and when they upgrade them.
Cloud platforms provide convenience and access to much larger infrastructure, but they also create dependency on the company operating the service.
For enthusiasts and businesses that value control, maintaining their own hardware can therefore be a deliberate strategic choice rather than simple inefficiency.
Personal Computers Need RAM For Much More Than AI
Another limitation of the utilisation argument is that consumer RAM exists for considerably more than artificial intelligence workloads.
Windows and other operating systems rely heavily on memory for normal operation. Web browsers can consume several gigabytes, while games, creative applications, virtual machines, development tools and professional software can require substantially more.
Modern users also frequently run several applications simultaneously.
A PC equipped with 32GB or 64GB may not use every gigabyte continuously, but additional capacity allows the machine to handle demanding workloads without constantly swapping data to slower storage.
Gaming Continues To Demand More Memory
Modern games increasingly require larger amounts of system memory as environments, textures and background processes become more complicated.
A gaming PC with only enough RAM to satisfy today's minimum requirements may become restrictive within a few years. Users therefore often purchase additional capacity to extend the useful life of the machine.
That unused headroom is intentional.
The fact that a computer is not using all of its memory at one moment does not necessarily mean the additional capacity serves no purpose.
Creative Applications Can Consume Huge Amounts Of RAM
Video editing, photography, animation and 3D production are other workloads where memory capacity can matter significantly.
Editing high-resolution video may require large files and complex timelines to remain immediately accessible. 3D rendering software may need significant memory for detailed scenes and textures.
Professionals working with these applications generally do not want their system configured so tightly that every demanding project pushes the machine into heavy disk swapping.
Extra RAM therefore provides performance headroom even when it is not always fully occupied.
Software Developers Also Benefit From Large Memory Capacity
Development environments can consume large quantities of memory, particularly when several services are running simultaneously.
A developer may have an IDE, browser, databases, Docker containers and multiple virtual machines operating on the same workstation.
Local testing environments can easily use tens of gigabytes of RAM without involving artificial intelligence at all.
For these users, substantial memory capacity is simply part of having a capable professional workstation.
Altman's Comments Come During A Serious Memory Shortage
The debate becomes more important because global memory supply is under increasing pressure.
AI companies and cloud providers are purchasing enormous quantities of memory for new data centres, training clusters and inference infrastructure. This demand is competing with manufacturers of consumer PCs, smartphones and other electronics.
When supply cannot expand quickly enough, prices rise.
That means AI infrastructure spending can eventually affect consumers who have no interest in running AI models themselves.
AI Companies Are Buying Huge Amounts Of Memory
Large AI systems need memory at extraordinary scale.
Training clusters consist of thousands of accelerators, each requiring extremely fast memory. Servers also need conventional DRAM, while the surrounding infrastructure requires additional storage and networking hardware.
As AI companies build larger data centres, their component requirements continue increasing.
These buyers also tend to have extremely large capital budgets, allowing them to secure long-term supply agreements that smaller customers cannot easily match.
Long-Term Supply Agreements Could Tighten Availability Further
A previous forecast by Jefferies Equity Research suggested that AI companies and cloud providers had already secured approximately half of the world's memory supply through long-term agreements.
The estimate suggested that figure could eventually reach around 70% if additional contracts are signed.
If those projections prove accurate, a relatively small group of extremely large technology companies could effectively reserve the majority of available memory production.
That would leave consumer electronics manufacturers competing for the remaining capacity.
Memory Prices Could Continue Rising Until 2028
The same forecast suggested that memory prices may continue increasing through 2028.
This would represent a major shift from previous memory cycles, where oversupply periodically drove DRAM and NAND prices sharply lower.
AI infrastructure could potentially keep demand elevated for much longer because data-centre expansion plans span several years.
Consumers who have become accustomed to inexpensive RAM upgrades may therefore face a very different market.
ADATA Chairman Warned The Shortage Could Last Even Longer
ADATA chairman Simon Chen has offered an even more pessimistic outlook, suggesting that the memory shortage could potentially continue for another decade.
His concern centres on the widening gap between demand and available manufacturing capacity as AI adoption accelerates.
Building new semiconductor fabrication capacity is expensive and slow. Manufacturers cannot simply double memory output overnight when demand suddenly increases.
If AI infrastructure continues expanding faster than new production capacity comes online, shortages could persist for years.
Consumers And AI Companies Are Competing For The Same Industry Capacity
The situation creates an unusual competition between consumer electronics and enormous AI infrastructure projects.
A home PC may require only tens of gigabytes of memory, while data-centre deployments consume memory at vastly greater scale. Individually, consumer devices use far less, but billions of PCs, smartphones and other products collectively represent enormous demand.
Memory manufacturers must decide where production capacity generates the best returns.
When AI customers are willing to pay more, suppliers naturally have incentives to prioritise higher-margin data-centre products.
The Shortage Could Affect More Than Desktop RAM
Consumers may notice the consequences across many different categories.
Desktop DDR memory could become more expensive, but laptops, smartphones, tablets and other devices also rely on DRAM.
Manufacturers facing higher component prices may increase retail prices or reduce the amount of memory included in entry-level products.
A laptop that previously shipped with 16GB as standard could potentially remain at 8GB longer simply because upgrading the base specification becomes more expensive.
Smartphone Prices Could Also Be Affected
Modern smartphones increasingly include substantial amounts of memory, particularly flagship Android devices.
Features such as on-device AI may push those requirements even higher. Manufacturers trying to run larger language or image models locally need additional RAM to hold those models and provide enough space for ordinary applications.
Ironically, AI could therefore increase memory demand at both ends of the market.
Cloud AI requires massive data-centre capacity, while local AI simultaneously encourages device makers to install more memory in consumer hardware.
On-Device AI Complicates Altman's Argument
The technology industry's wider direction also creates an interesting contradiction.
Companies including Apple, Microsoft, Google and smartphone manufacturers increasingly promote on-device AI as a major feature. Running models locally can improve privacy, reduce latency and provide functionality without constant cloud connectivity.
Those capabilities require additional memory.
If the future of consumer technology includes more local AI, then allocating substantial RAM to personal devices may become increasingly necessary rather than wasteful.
Cloud AI And Local AI Solve Different Problems
The debate does not necessarily require choosing one model exclusively.
Cloud computing works extremely well when users need access to massive models requiring hardware far beyond anything practical inside a personal computer.
Local AI works better when privacy, low latency, offline capability or user control are more important.
Many future applications will probably combine both approaches.
Smaller tasks can run locally while more demanding requests are sent to cloud infrastructure.
Hybrid AI Could Become The Practical Middle Ground
A hybrid model allows devices to take advantage of local hardware without requiring every task to remain on the machine.
Sensitive or lightweight requests could be processed locally. Larger workloads could automatically move to cloud infrastructure when additional capability is necessary.
This could help balance memory utilisation while preserving some of the advantages users value about local computing.
The industry is already moving towards this architecture across PCs and smartphones.
Data Centre Efficiency Does Have A Strong Economic Argument
Altman's core point about utilisation still has merit.
A 64GB memory configuration used heavily for only a few hours each day cannot match the utilisation of cloud infrastructure running workloads continuously.
When resources are scarce, higher utilisation means more computational output from the same quantity of hardware.
From a purely economic and global-capacity perspective, centralised infrastructure can therefore be more efficient.
The disagreement is really about whether utilisation should be the only metric that matters.
Ownership And Control Have Value Too
Consumers do not purchase hardware solely to maximise utilisation.
A car spends much of its life parked, yet people still value owning one because it is available whenever needed. Personal computing works similarly.
RAM sitting idle is still providing capacity that can immediately be used when the owner needs it.
That availability has value even if a centralised system could technically keep the same hardware busier.
Efficiency Is Not The Same As User Benefit
This distinction is important.
A resource can be globally underutilised while still producing substantial value for the person who owns it. A home workstation may sit idle overnight but become extremely valuable during the working day.
Cloud infrastructure optimises efficiency across many users.
Personal hardware optimises availability and control for one user.
Neither objective is automatically superior in every situation.
Cloud Dependence Also Creates Infrastructure Risks
Moving more computing into centralised cloud platforms introduces its own problems.
Service outages can affect millions of users simultaneously. Internet connectivity becomes essential, and organisations may depend heavily on a small number of large providers.
Centralisation can improve utilisation but also concentrate risk.
Local computing provides resilience by allowing some workloads to continue even when external services become unavailable.
Subscription Fatigue Is Encouraging Local AI
Another reason local AI is becoming popular is growing subscription fatigue.
Users already pay recurring fees for streaming services, productivity software, cloud storage and other digital products. Adding several AI subscriptions can make monthly costs increasingly difficult to justify.
Running open models locally provides an alternative.
The initial hardware cost may be significant, but users avoid another recurring service fee.
Open-Source Models Are Improving Rapidly
Local AI is also becoming more practical because open and downloadable models continue improving.
Smaller models can increasingly perform useful writing, coding, summarisation and reasoning tasks without requiring enormous data-centre infrastructure.
Quantisation techniques also reduce the amount of memory needed to run them.
As these models become more capable, users have stronger incentives to invest in local hardware.
More RAM Gives Local Models More Room
For local AI users, memory capacity directly influences which models can run comfortably.
Larger models require more memory, particularly when running at higher precision. Users experimenting with several models may therefore intentionally build systems with 64GB, 128GB or more RAM.
From their perspective, the additional memory is not wasted.
It is enabling workloads that would otherwise require an external AI service.
The Memory Industry Has An Incentive To Expand Capacity
Persistently high prices will naturally encourage memory manufacturers to increase production.
Companies can expand existing facilities, build new fabrication plants and invest in newer manufacturing technologies.
However, semiconductor capacity takes years to plan and construct.
This delay explains why supply shortages can continue even when manufacturers clearly see strong demand.
AI Demand Could Eventually Produce More Supply
There is also a longer-term possibility that AI-driven investment ultimately benefits consumers.
If enormous demand encourages memory companies to build significantly more manufacturing capacity, the additional supply could eventually lower prices once the market stabilises.
Semiconductor industries historically move through cycles of shortage and oversupply.
The current AI boom may therefore create painful prices in the short term while driving much larger production capacity over the longer term.
Until Then, Consumers May Face Higher Hardware Costs
For now, the more immediate consequence appears likely to be higher prices.
Memory costs feed directly into the manufacturing cost of laptops, desktops, smartphones and servers.
Hardware companies can absorb some of those increases temporarily, but eventually part of the cost tends to reach customers.
Consumers may therefore end up paying more for systems with the same memory capacities they previously bought much more cheaply.
Final Thoughts
Sam Altman's argument that personal computers are an inefficient place to keep large quantities of DRAM makes sense when viewed through the economics of large-scale AI infrastructure. Memory sitting mostly idle inside one home PC cannot achieve the same utilisation as memory inside a data centre serving thousands of users continuously.
The problem is that utilisation is not the only reason people own computing hardware.
PCs need memory for gaming, professional applications, development and everyday multitasking. Local AI users also deliberately invest in larger memory capacities because they value privacy, offline access, independence from subscriptions and greater control over the models they run.
The debate is becoming more important because AI companies themselves are contributing enormously to the memory shortage. Forecasts suggest cloud providers and AI firms may already have secured around half of global memory production through long-term agreements, potentially increasing to 70%, while some analysts expect elevated prices to persist until 2028 or even considerably longer.
That means consumers may increasingly find themselves competing with billion-dollar AI infrastructure projects for the same underlying manufacturing capacity.
Cloud computing may indeed use memory more efficiently. But telling consumers their RAM is being wasted while the AI industry buys unprecedented quantities of the same scarce resource will understandably be a difficult argument to sell.
For many PC owners, unused memory is not waste. It is capacity they paid for, control they own and computing power that remains available whenever they choose to use it.
Comments 0