Sam Altman Says Computer RAM Is Better Used in AI Data Centres Than Personal Computers

Sam Altman Says Computer RAM Is Better Used in AI Data Centres Than Personal Computers

OpenAI CEO Sam Altman has sparked debate across the technology industry after arguing that the world's limited supply of computer memory should be used more intensively in data centres rather than distributed across millions of personal computers.

Speaking during an interview on the Cheeky Pint podcast with Stripe co-founder John Collison, Altman discussed the growing demand for computing resources, the global memory shortage and the increasing popularity of running artificial intelligence models directly on personal computers.

He questioned the idea that open-source AI models running locally on people's computers would significantly reduce the need for large data centres.

According to Altman, centralised cloud infrastructure offers a more efficient way to allocate scarce memory because multiple users and applications can share computing resources.

His comments have attracted attention because the AI industry's growing appetite for memory is already putting pressure on the supply chain, with potential consequences for the cost of computers, laptops and other electronic devices.

What Exactly Did Sam Altman Say?

During the podcast conversation, Altman discussed the idea that people would increasingly run open-source artificial intelligence models on their own computers instead of relying entirely on cloud-based AI services.

He questioned whether this approach would reduce demand for large data centres.

His argument was that memory installed in millions of personal computers is often used only occasionally, whereas memory in a centralised data centre can be allocated to different workloads throughout the day.

He described this as a poor use of scarce resources from a global allocation perspective.

Altman said:

"From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer and being used 5% of the time."

He added that greater utilisation of memory in cloud infrastructure would be preferable while supply remained constrained.

The remarks were made during a wider discussion about the availability of computing hardware, the economics of semiconductor manufacturing and the growing infrastructure requirements of artificial intelligence.

Important: The 5% figure was Altman's estimate during the conversation. It was not presented as an independently verified measurement of how much memory is used across personal computers.

What Is RAM and Why Is It So Important for AI?

RAM, or Random Access Memory, is the temporary working memory used by computers, smartphones and other electronic devices.

It allows a device to hold information that its processor needs to access quickly.

For example, when you open several browser tabs, edit a video, play a game or run a software application, your device uses RAM to keep the necessary information readily available.

Artificial intelligence systems also depend heavily on memory.

When an AI model runs, memory is needed to store model parameters, manage incoming information and support the calculations required to generate responses.

For larger models, memory requirements can become substantial.

AI data centres use specialised hardware equipped with large amounts of high-speed memory, including graphics processing units (GPUs) and high-bandwidth memory (HBM).

Although HBM and the conventional DRAM used in many computers are different types of memory, they belong to a connected semiconductor supply chain. Increasing demand for AI hardware can therefore place additional pressure on memory manufacturers and production capacity.

Why Does Sam Altman Prefer Data Centres?

Altman's argument is based mainly on resource utilisation.

A personal computer may have 16GB, 32GB or 64GB of RAM installed, but its user may not need all that memory continuously.

A data centre, on the other hand, can serve many customers and applications using shared infrastructure.

For example, a cloud provider can allocate computing resources to one customer during the morning, another during the afternoon and additional workloads overnight.

This allows expensive hardware to support multiple users rather than remaining dedicated to one person's computer.

From an infrastructure perspective, centralised computing can offer several advantages:

• Resource sharing: Computing resources can be distributed across multiple users and applications.
• Higher utilisation: Data centres can schedule workloads to make more consistent use of available hardware.
• Centralised management: Infrastructure can be monitored and maintained by specialised teams.
• Scalability: Computing capacity can be expanded or allocated according to demand.
• Hardware efficiency: Large facilities can coordinate workloads and manage infrastructure at scale.

Altman's argument is that when memory supplies are limited, shared infrastructure can make better use of available resources.

However, this is not the only way to measure computing efficiency. Data centres also consume substantial electricity and require cooling, networking and expensive infrastructure. The full environmental and economic cost must be considered when comparing centralised and local computing.

Why Are AI Companies Buying So Much Memory?

The rapid development of AI has increased demand for specialised computing infrastructure.

Companies developing large language models and AI agents need powerful systems to train models, process user requests and operate services at scale.

These systems require processors, memory, storage, networking equipment, electricity and physical data-centre capacity.

Companies including OpenAI and major cloud providers are investing billions of dollars in expanding their infrastructure.

Memory manufacturers such as Samsung Electronics and SK hynix are among the companies supplying components to this growing market.

The Verge reports that long-term supply agreements between memory manufacturers and major cloud operators are contributing to tighter availability of memory for other customers.

The problem is that semiconductor manufacturing capacity cannot be expanded instantly. Building and equipping advanced production facilities requires significant investment, specialist equipment and time.

As demand grows faster than available supply, prices can rise and manufacturers must make decisions about how to allocate their production capacity.

This is one reason the AI boom is increasingly affecting industries beyond software, including semiconductor manufacturing, electronics and computing hardware.

Could This Make Computers and Laptops More Expensive?

Potentially, yes.

When demand for memory chips increases, manufacturers may raise prices or prioritise customers willing to commit to large long-term purchases.

This can affect the cost of components used in consumer electronics.

Computers, laptops, smartphones and other devices depend on memory chips, meaning that supply constraints can eventually affect manufacturers and consumers.

The extent of the price increase depends on several factors, including:

• Memory production capacity.
• Demand from AI data centres.
• Orders from computer and smartphone manufacturers.
• Availability of alternative memory technologies.
• Manufacturing costs and investment in new facilities.
• Global economic and supply-chain conditions.

It is important not to assume that every increase in computer or smartphone prices is caused by OpenAI or AI data centres alone. Multiple market forces influence the final price of consumer electronics.

However, the continued expansion of AI infrastructure is becoming an important factor in the global semiconductor market.

Not Everyone Agrees That AI Should Run in Data Centres

Although Altman sees greater resource utilisation as an advantage of cloud computing, many developers and technology users have different priorities.

The popularity of local AI comes from several practical benefits.

1. Privacy and data control

Running an AI model locally can allow users to process sensitive information without sending every request to an external cloud service.

This can be particularly valuable for businesses handling confidential documents, financial information or private customer data.

2. Reduced dependence on internet connectivity

Local AI systems can continue working without a constant connection to an external AI provider, depending on the model and application.

This is useful in areas where internet connectivity is unreliable or expensive.

3. Greater control

Developers can choose which models to install, customise their applications and decide how their computing resources are used.

4. Reduced dependence on subscriptions

Some open-source and open-weight models can be downloaded and operated locally without paying a recurring subscription to a cloud AI provider, although users still bear hardware, electricity and maintenance costs.

5. Lower latency for some applications

Local processing can reduce delays associated with sending information to a remote server and waiting for a response.

This can be particularly important for applications that require quick responses or continuous operation.

These advantages explain why many developers, researchers and businesses continue investing in local AI even as large cloud companies expand their data centres.

Is Sam Altman Opposed to Open-Source AI?

His comments should not automatically be interpreted as a rejection of all open-source AI development.

The discussion focused on how the world's limited memory supply should be allocated during a period of constrained availability.

Altman questioned the efficiency of installing large amounts of memory in individual computers when that memory may not be used consistently.

His argument favours greater utilisation through shared cloud infrastructure.

However, the debate is also connected to an important difference between the two computing models.

Cloud-based AI gives providers the ability to manage large infrastructure investments and offer AI services to many customers.

Local AI gives individuals and organisations greater control over the hardware and models they use.

Both approaches have advantages, and the most appropriate option depends on the application, cost, privacy requirements, performance needs and available infrastructure.

What Does This Mean for Everyday Computer Users?

For ordinary consumers, the discussion raises an important question: will the increasing demand for AI infrastructure make it more difficult or expensive to buy powerful computers?

If memory supplies remain constrained, manufacturers may face higher component costs, which could eventually affect retail prices.

People building gaming computers, buying laptops or upgrading their devices may also face increased competition for certain memory products.

For developers interested in running AI locally, the cost and availability of higher-memory computers will remain important considerations.

For businesses, the decision between local and cloud AI may become increasingly dependent on the total cost of ownership, privacy, performance and availability of computing resources.

The debate is therefore not simply about whether RAM should be inside a computer or inside a data centre. It is about who controls computing resources, how those resources are allocated and how the benefits and costs of the AI boom are distributed.

The Bigger Question: Who Should Control the Future of Computing?

Altman's comments highlight a growing divide in the technology industry.

One vision is built around large centralised data centres operated by companies that provide AI services through the internet.

Another supports more distributed computing, where individuals, developers and businesses operate AI models on their own hardware.

Centralised infrastructure can make expensive computing resources available to millions of people without requiring everyone to purchase powerful machines.

However, greater reliance on cloud services can also create dependence on a smaller number of infrastructure providers.

Local computing offers more direct control, privacy and independence, but requires users to purchase and maintain capable hardware.

Neither model solves every problem.

As artificial intelligence becomes more important in education, business, software development and everyday life, the availability and affordability of computing hardware will become an increasingly important issue.

The debate sparked by Altman's remarks is likely to continue as technology companies expand their data centres and consumers look for more ways to run AI on personal devices.

Sources and further reading:

• The Verge: https://www.theverge.com/ai-artificial-intelligence/1023820/sam-altman-ram-shortage-cloud-servers
• Full Cheeky Pint podcast transcript: https://cheekypint.transistor.fm/39/transcript
• Tom's Hardware: https://www.tomshardware.com/tech-industry/data-centers/openai-ceo-says-its-quite-terrible-for-your-pc-to-have-ram-sam-altman-says-data-centers-offer-better-utilization
• Cheeky Pint podcast episode: https://podcasts.apple.com/us/podcast/why-sam-altman-is-against-a-single-superintelligence/id1821055332?i=1000798698948