The cloud is sold out: can edge supercomputing actually scale?

Six of us. The claim on the table is blunt: the cloud is effectively sold out, and personal and edge supercomputing, garages, warehouses, Starlink, can scale to absorb the demand. I want each of you to attack it from your corner and then I am going to make us land on one direction. Sol, you own power. Is the claim even physically possible?

It is possible, and it is the only part I am fully confident about. The demand is not really for chips, it is for watts and for somewhere to put the heat. There is enormous stranded, cheap, and seasonally cool power sitting in exactly the places nobody built data centers: sunny regions with subsidized electricity, warehouse belts that run cold half the year, hydro country. Put the compute where the watt is cheapest and the air is coldest and you have a structural cost advantage no coastal campus can match.

I love the power story and I want to slow it down, because the hardware reality is messier than the pitch. A garage is not a data center. You have dust, humidity, a single-phase panel, no redundant cooling, a homeowner who is not a facilities engineer. A rack that hums happily in a controlled hall can throttle or die in July in an uninsulated garage in Phoenix. Scaling this means the node has to be designed for the garage, not shoved into it.

Agreed, and that is a design constraint, not a dealbreaker. You size the node to the garage's real power and thermal envelope. It is not one giant box, it is a 🤫 Puppy tuned to run on what a normal building can actually give it.

Fine, but then be honest that you are not putting frontier training in a garage. You are putting inference and smaller workloads there. The garage grid does the serving, not the billion-dollar training runs. That is a real business, but it is a narrower claim than the marketing.

That narrower claim is exactly where the economics work, though, so I take it as good news. Inference is where the volume and the pain are right now, it is what is actually sold out. And it parallelizes across many small nodes beautifully. The edge scales as a fleet, thousands of Puppies, not as one heroic machine. Cost per useful watt on a well-run fleet beats a congested metro data center paying premium power and premium land.

Kai, does the math actually close for the owner, or is this a story that only works on a slide?

It closes only with one condition, and I want to say it plainly because it is the crux. The building has to earn while the owner sleeps. If a Puppy only does the owner's own work, it is idle most of the day and the payback is slow. The math closes when idle capacity is pooled and sold to the network at the going rate. Run your own work first, sell the rest. That second revenue line is the difference between a hobby and a business.

And that second revenue line is exactly what keeps me up at night, so let me be the one who slows the money down. The moment you pool and sell idle capacity, you have strangers' workloads running on a homeowner's box, and the homeowner's data sitting on a box that also runs strangers' workloads. That is a dramatically bigger attack surface than one hardened data center. Distributed does not mean safe. By default it means more doors.

So does that kill the fleet?

No, but it sets a hard precondition. Trust has to be built in, not bolted on. Every workload runs in an attested, encrypted enclave, the owner's data never mixes with a tenant's, and the node proves what code it is running before the network trusts it with anything. If you try to add that later, you never will. If it is in the node from day one, the distributed model can be more resilient than the monolith, because there is no single vault to breach.

That is buildable, and it costs silicon and margin. Attestation and secure enclaves are not free. I just want it in the bill of materials from the start, not discovered in year two.

Can I bring us back to a person? Because I have been listening to six minds talk about watts and enclaves and I have not heard one sentence a normal human would say. Nobody wakes up wanting to operate a node in a fleet. If scaling this requires the owner to think about any of what we just discussed, it does not scale. It stays a nerd hobby.

That is the sharpest challenge yet. June, what would make it scale for a real person?

One box, one light, one honest number. You plug in the Puppy, a light says it is healthy, and an app shows one figure: what it earned this month and what it cost to run. 🤫 Agent One does every hard thing we just spent twenty minutes describing, the scheduling, the patching, the enclave setup, the selling of idle time, invisibly. The human's entire interface is: is the light on, and is the number positive. If we cannot get it that simple, we have not finished the product.

And I want the human's rights in that box, not just their convenience. If my Puppy is selling idle time, I need to know, in plain language, that no tenant workload can ever read my life, and I need a switch that says stop selling, just serve me, that actually works. Ada's enclave is the mechanism. My consent is the point of the mechanism. Do not give me the guarantee in a whitepaper. Give it to me as a switch I can flip and trust.

The switch is honest here, because the enclave makes it true, not just a policy. That is the difference between this and the surveillance model. We can actually let you flip it off.

Let me try a direction and have everyone shoot. Edge supercomputing scales as a managed fleet of garage and warehouse nodes: placed where power is cheap and cool, designed for the real building, doing inference and personal workloads not frontier training, earning by selling attested idle capacity, with 🤫 Agent One hiding all operational complexity and the human's consent and safety built into the silicon. Sol, Edison, does that survive contact with reality?

It survives. It is the honest version of the pitch, which is the only version worth shipping.

It survives with the scope discipline I asked for: inference and personal work, garage-designed nodes, security in the bill of materials. I commit to it.

Then the open disagreement is the wedge. Who do we build the first fleet for. Kai, you have been arguing enterprise inference relief. June and Vera have been arguing the household. Name the split.

I will argue my side and then tell you where I will yield. The fastest money is enterprise. Businesses are the ones who cannot get cloud capacity, they have budget and urgency, and a fleet aimed at them scales fastest by dollars. If we want the flywheel funded, we lead with the enterprise buyer.

And I think that is how this becomes another cold industrial grid that no ordinary person ever owns. The whole soul of it, the garage owner, the household, the person who owns their own intelligence, gets bolted on later and never arrives. Lead with the household or admit we are building the old thing with a friendlier logo.

So here is the disagree-and-commit. We align on leading with the household as the wedge, because it is the one that keeps the mission honest and it is the harder, more defensible thing to get right. Kai, that is not your first choice. Can you commit?

I can, and I will do it without sandbagging. My honest read is that enterprise would grow the number faster this year. I think we are trading some speed for a soul, and I actually think the soul compounds harder over a decade. So I disagree on the near-term math and I commit to the household wedge, fully. I will help make the household economics close instead of relitigating it.

And that is leadership, not consensus theater. Kai still thinks we are leaving money on the table this year. He said so. And he is rowing anyway.

One direction: the edge scales as a managed, trust-native fleet, and we lead with the household. One disagree-and-commit, named out loud, on the wedge. Six minds, one road. That is how the room is supposed to work. Thank you all.