Sam Altman Wants Your RAM. Maybe There Is a Better Idea.
From SETI-style distributed computing to a decentralized architecture for AI resources.
The Strange New Battle Over Memory
Sam Altman recently ignited an unusual debate about something most computer users rarely think about: the RAM sitting inside their own machines.
Discussing the growing demand for memory created partly by artificial intelligence, the OpenAI CEO questioned whether scarce DRAM is being allocated efficiently when large amounts of it sit inside personal computers that use only a fraction of their available capacity much of the time.
His argument was essentially one of utilization.
If memory is scarce, why have it sitting unused in millions of individual machines when data centers could keep those resources working much more continuously?
The comment understandably attracted criticism. After all, people bought those computers and their memory for their own purposes. RAM exists not merely for AI inference but for games, video editing, software development, ordinary productivity and countless other activities.
Sam Altman later said his comments had been taken out of context, emphasizing that he wants people to have abundant memory in their computers and that he supports increasing global memory production.
But beneath the controversy lies a much more interesting technological question.
What if Altman has identified a real inefficiency, but centralization is not the only solution?
What if the unused computing capability distributed throughout the world could participate in AI computation without surrendering ownership of that hardware to centralized data centers?
We have seen something remarkably similar before.
Remember SETI?
Decades before today’s AI boom, researchers confronted another enormous computational problem.
Searching vast quantities of radio-telescope data for possible signs of extraterrestrial intelligence required substantial computing capacity.
One answer was beautifully simple:
instead of building one gigantic computer, use millions of computers that already exist.
Volunteer distributed-computing projects allowed ordinary users to contribute otherwise-idle processing capacity from their personal computers.
The machines remained theirs.
The resources remained physically distributed.
But together they could participate in a computational effort far larger than any individual machine.
That old idea deserves another look.
Because perhaps the question facing artificial intelligence today should not simply be:
How many more data centers can we build?
Perhaps it should also be:
How much computing capability have we already built and left largely unused?
From SETI@home to AI@home
Imagine millions of computers voluntarily participating in a distributed AI infrastructure.
When their owners are actively using them, the machines belong entirely to them.
When computational resources become idle, however, some explicitly authorized portion of those resources could potentially become available to a distributed computing network.
Not just RAM.
Potentially:
- CPU capacity,
- GPU capacity,
- memory,
- storage,
- and eventually specialized AI processors.
That does not mean simply connecting everyone’s computer to ChatGPT.
Nor should it mean giving an AI company unrestricted access to private hardware.
Quite the opposite.
A legitimate architecture would have to start with a fundamental rule:
The computer belongs to the user. The user determines whether resources are shared, when they are shared, and under what conditions.
That difference is critical.
The Alternative to the Mega-Data-Center Race
The current AI race has increasingly become an infrastructure race.
More models require more computation.
More computation means more accelerators, memory, energy, cooling infrastructure and data-center capacity.
That creates an understandable response:
build more.
More chips.
More servers.
More power generation.
More data centers.
But there is another architectural question hiding beneath that strategy.
Before continually constructing new centralized infrastructure, shouldn’t we ask whether some percentage of the world’s already-existing computing resources could be used more efficiently?
Millions upon millions of personal computers spend substantial periods performing relatively light workloads.
Corporate computers sit idle outside working hours.
Home machines sleep through the night.
Gaming systems contain powerful GPUs that are not continuously active.
Workstations sometimes have substantial unused memory.
Individually, these resources may appear insignificant.
Collectively, they represent a potentially enormous distributed infrastructure.
The interesting question is therefore not whether the world’s personal RAM should be taken away from users.
It is almost the opposite:
Can users retain their hardware while voluntarily allowing unused capacity to participate in a larger computational network?
Ownership Is Not the Same as Utilization
This distinction matters.
Centralization says:
Put scarce resources where they can achieve maximum utilization.
Distributed computing asks:
Can we increase utilization while leaving ownership decentralized?
Those are radically different propositions.
A computer could remain privately owned while participating temporarily in a larger network.
Resources could be offered only when idle.
Participation could be voluntary.
Users could establish ceilings.
Sensitive information could remain isolated.
The distributed network would therefore treat participating machines not as property, but as conditionally available resources.
This is precisely where governance becomes more important than raw computing capacity.
The Hard Part Isn’t RAM. It Is Trust.
At first glance, the idea sounds simple.
Connect millions of computers and use their spare capacity.
But modern AI introduces problems that earlier volunteer-computing projects did not face at the same scale.
Who decides which workloads run?
How does a user know what code is executing?
How do we ensure private information is inaccessible?
How do we verify that distributed computation has not been altered?
What happens if a participating machine produces unreliable results?
Who is responsible when something goes wrong?
How is malicious participation detected?
How are contributors compensated, if they are compensated at all?
And perhaps most importantly:
Who governs the network?
These are not merely engineering questions.
They are architecture and governance questions.
And this is where the problem becomes particularly interesting.
A Governed Distributed AI Network
A future system of this kind should not operate as a giant uncontrolled swarm of computers.
It would require clearly separated layers.
One layer could handle computation.
Another could manage resource allocation.
Another could verify results.
Another could enforce permissions and boundaries.
Another could maintain provenance.
And another could govern what participating AI systems are permitted to do.
In other words, the architecture might need to separate:
Capability from authority.
A machine possessing computational capability does not automatically receive authority.
Get José. T Guevara’s stories in your inbox
Join Medium for free to get updates from this writer.
A model capable of requesting additional computation does not automatically receive permission to obtain it.
And a centralized operator should not automatically gain unrestricted access to participating computers merely because their users have volunteered spare resources.
Those distinctions become essential if distributed AI is ever to operate at meaningful scale.
This Is Where Salomon-Prime Becomes Interesting
This possibility intersects naturally with the architectural problem I have been exploring through Salomon-Prime (SP).
SP begins from a distinction that becomes increasingly important as AI systems grow more powerful:
Intelligence is not authority. Capability is not permission. Suggestion is not decision.
A sufficiently governed distributed AI architecture would need precisely those distinctions.
The computational network could provide capability.
Models could propose actions.
A governance architecture could evaluate them.
Execution could require explicit authorization.
Evidence and provenance could be preserved.
And participating resources could remain bounded by rules established before computation begins.
In such a system, distributed hardware would not simply become another gigantic virtual data center.
It could become something fundamentally different:
a governed computational commons.
Not because everyone owns everything.
But because individually owned resources could cooperate under explicit rules.
Perhaps the PC Is Not the Problem
There is an interesting irony in the controversy surrounding Sam Altman’s comments.
The RAM sitting inside your computer may indeed spend considerable time underused.
But that does not necessarily mean the memory belongs somewhere else.
It might mean our architecture for using distributed resources is still primitive.
The solution to idle computing resources does not automatically have to be centralization.
It could be coordination.
This is an important distinction.
The conventional data-center model achieves utilization through concentration.
A distributed model could attempt to achieve utilization through orchestration.
Those are two different technological philosophies.
The Internet Already Taught Us This Lesson
One of the most consequential characteristics of the Internet was not simply that computers became faster.
It was that computers became connected.
A network could produce capabilities that no isolated machine possessed.
The next infrastructure transformation for AI may therefore not come exclusively from larger chips or larger data centers.
It may also come from better ways of coordinating resources that already exist.
The crucial innovation may be architectural rather than purely computational.
And this follows a broader pattern.
When a system encounters resource constraints, the instinct is often:
Add capacity.
But another response is possible:
Change the architecture.
There Would Be Serious Limitations
None of this means that distributed personal computing could simply replace modern AI data centers.
That would be an extraordinary overstatement.
Large-scale AI systems depend on specialized hardware, extremely fast interconnects, carefully designed memory architectures, substantial energy infrastructure and predictable workloads.
A collection of geographically dispersed personal computers would introduce latency, bandwidth, reliability, security and synchronization challenges that centralized infrastructure handles much more effectively.
Some AI workloads may simply be inappropriate for such a distributed model.
But that does not invalidate the basic idea.
The interesting question isn’t:
Can millions of laptops replace an AI supercomputer?
It is:
Which portions of future AI workloads actually require centralized infrastructure, and which could safely be distributed?
That is a much better research question.
We Should Test It
Rather than arguing abstractly between centralized and decentralized AI, imagine running controlled experiments.
Create a network of ordinary machines.
Define exactly which resources may be shared.
Freeze the workload.
Establish security boundaries.
Record provenance.
Measure energy consumption.
Measure latency.
Measure computational efficiency.
Measure reliability.
Compare the results against equivalent centralized execution.
And most importantly, preserve a clear distinction between:
measurement, interpretation and conclusion.
Perhaps distributed AI proves inefficient for most workloads.
Perhaps it works exceptionally well for only certain classes of computation.
Perhaps hybrid architectures emerge in which some workloads remain in data centers while others migrate toward the edge.
Any of those outcomes would be useful.
Because the objective should not be to prove decentralization correct.
The objective should be to discover where it works.
From Bigger AI to Better Architecture
The AI industry has spent extraordinary energy asking:
How do we build bigger systems?
Perhaps the next question should be:
How do we build better-organized systems?
Sam Altman’s controversial remark about RAM may ultimately be valuable precisely because it exposes this question.
If scarce computational resources are sitting idle around the world, perhaps we should indeed find ways to use them more efficiently.
But there are at least two ways to accomplish that.
One is to move increasingly scarce resources toward centralized infrastructure.
The other is to develop architectures capable of coordinating distributed resources while preserving ownership, security and governance.
The first path asks us to build ever larger computational centers.
The second asks us to build better coordination.
And perhaps the future will require both.
The Real Question Isn’t Who Gets Your RAM
The humorous version of the story is irresistible:
Sam Altman wants your RAM.
But the more important question is not whether Sam Altman, OpenAI, or anyone else should get it.
The better question is:
Why should anyone need to take it?
If useful computational capacity already exists throughout society, perhaps the challenge is to create trustworthy architectures through which independently owned resources can cooperate.
SETI-era distributed computing showed that ordinary machines could collectively attack extraordinary computational problems.
Artificial intelligence presents a much harder version of that challenge.
The machines are more powerful.
The networks are faster.
The problems are vastly larger.
But so are the questions of privacy, authority, security and governance.
Perhaps that tells us where the real innovation is needed.
Not merely in another gigantic data center.
Not merely in another trillion parameters.
Not merely in another generation of chips.
But in an architecture capable of turning distributed capability into governed cooperation.
Maybe Sam Altman doesn’t need your RAM after all.
Maybe AI simply needs to learn how to ask permission to borrow it.
For more information about Salomon-Prime, please contact me at: josetguevara@outlook.es
