The founding document

The Third Door

An open proposal to Sam Altman

Sam —

On October 29, on Cheeky Pint, you said that from a global allocation perspective it's "quite terrible" to have the world's limited supply of DRAM sitting in everyone's own computer, used five percent of the time. The internet heard a man coming for their computers. I heard something else: the most honest sentence anyone has said about this industry all year. You're right. Scarce memory sitting idle in millions of boxes, while people who need compute queue for it, is a misallocation. Pooled, busy, professionally run infrastructure is the efficient answer to a supply crunch. My own company designs on that premise — our machines earn their keep around the clock, or they don't earn at all.

So let me first defend you against your critics, because their reading is lazy. You did not propose confiscating anyone's computer. You made an argument about utilization, and it happens to be correct. The "oligarch" framing is a way of not answering it.

But the fear underneath the framing is not lazy, and it deserves a precise name, because it is the part your answer doesn't reach.

Utilization is an argument about hardware. It settles nothing about policy. In the future as you've described it, the pooled machines arrive with your models on them — your refusals, your retention terms, your definition of acceptable use, privacy offered as a clause in a contract on infrastructure shared with everyone. The millions looking at local hardware aren't really shopping for RAM. They're shopping for a policy they control, and a privacy that is a property of the system rather than a promise in a terms document. That is why their intuition runs opposite to yours. And it is why the enclave roadmap, good engineering as it is, doesn't close the question: an enclave answers can the operator peek. It does not answer who decides what the model may say.

Those are two different questions, and the market currently forces everyone to answer both with a single purchase. Own the machine: idle silicon, your policy, your privacy. Or rent the cloud: busy silicon, their policy, their promise. Two doors, and a crowd pressing against both.

Here is our proposal: build the third door.

A dedicated machine — one customer per physical computer, no shared tenancy — rented by the month and professionally operated, so utilization stays high and your efficiency point is honored in full. Open weights on it, pinned and receipted, whose policy is set by the person paying for the machine, inside the law. Keys the tenant holds. A ledger the tenant can read. A certified wipe when the tenancy ends. Privacy that is architectural — a fact about the machine — instead of contractual. Sovereignty without ownership. Utilization without surrender.

We didn't assume this door was missing; we checked. This month our shop read the pricing pages and terms of some thirty providers across every category — GPU clouds, uncensored inference services, confidential-computing clouds, the labs' own private tiers. The results, with receipts, are published alongside this note. Refusal-free access exists only as shared, revocable API access. Dedicated control exists only as empty iron. Every architectural privacy tier from the major labs — including your own Private Inference — is still a preview or a paper. Nobody sells the combination. The middle of your argument with your critics is empty ground.

We're building on it, in the open, staging first, with the receipts published as we go. Not at your scale — we're a small shop in Newton, and we like it that way. But the size of the builder was never the point of a proposal. The shape is the point.

Because the disagreement was never really about memory, Sam. It's about whether pooled compute must mean pooled judgment. You have argued, correctly, that the memory should be pooled. Your critics have felt, correctly, that the judgment shouldn't be. Both things can be true at once — and a market that takes both seriously looks like the third door.

Consider it proposed.

— Frank Kurka
InferenceDevices, Inc.