Last Week in Chutes

July 14 to July 21, 2026Week ten. Jon pre-trained a 20B model for under $10 an hour of compute, in public, on rented VMs. Also this week: monetization hit an all-time high, Chutes is getting a podcast, dstack's private AI gateway now routes to Chutes, and we asked the timeline whether it wants TEE VM rentals.

Under $10 an hour to pre-train a 20B model

Jon Durbin publicly demonstrated pre-training a 20B MoE for under $10 an hour of compute. Read that again. Not on a reserved cluster. Eight rented single-L40S VMs scattered across two continents, plus a few consumer RTX cards, all holding roughly 6 seconds per step.Pre-training was supposed to be the part you couldn't do cheap. The part that belonged to whoever owns the biggest data center. If the training stack tolerates distance and mismatched hardware, the cost floor drops to what a small team can rent by the hour. This is Jon's north star, intelligence per dollar, applied to training itself.One thing we'll keep repeating: this was a public demonstration, not an official benchmark. It happened, which is the strongest kind of claim there is.

An all-time high: $340K per trillion tokens

The business side set a record of its own. As of 2026-07-20, Chutes revenue per trillion LLM tokens hit $340K, a new all-time high. The metric is exactly what it sounds like: take platform revenue, divide it by the tokens we serve, and you get what a trillion tokens of served intelligence earns.And it's not a one-day spike. The 90-day average sits at $282K and the line has been climbing since we started tracking it. Cheap intelligence only matters if the economics underneath it hold up. They're holding up, and then some.

DropZone: Chutes is getting a podcast

We've been sitting on this one. DropZone, the Chutes podcast, is on the way. First topic on the table: Parallax. Long-form and unhurried, straight from the people building it. More soon.

dstack's private AI gateway now routes to Chutes

Hang Yin of Phala posted that the dstack private-ai-gateway can now route to custom Chutes deployments. The part worth slowing down for is what happens before any prompt moves: the gateway verifies the target GPU TEE against its expected measurements, and only then sends prompts through, end-to-end encrypted the whole way.This is what we mean when we say the confidential-compute story is bigger than us. Other privacy stacks route through Chutes because the attestation is real and checkable, not because we asked nicely.

Would you rent a TEE-secured VM?

We put a genuine question to the timeline this week. We're exploring on-demand GPU and CPU VM rentals, running inside the same hardware-attested TEEs as our inference. Same attestation, same end-to-end encryption, but instead of API calls you get the whole machine.It's early, and we're reading the room before we build further. If that's something you'd use, reply and tell us what you'd run on it.

Dev Corner

The theme this week is trust you can see: clearer diagnostics when something fails, docs that let you fix your own deploy, and new partners building on the platform.Clearer GPU attestation for operators. When a confidential GPU server checks in, the platform now surfaces clearer attestation errors. If a check-in fails, operators can see exactly why instead of guessing, which makes running attested hardware a much less mysterious job.Deployment troubleshooting docs are back. We restored a Deployment Troubleshooting guide in the CLI docs, so developers can self-serve the common deploy issues instead of waiting on support. Small ship, big quality-of-life win.Other subnets can now build on Chutes, privately. Integrated subnet partners can run their chutes on Chutes while keeping their model and source code private, even on a publicly-runnable chute. ChronoSeek is already in. This is the ecosystem bet in action: the easier it is for other Bittensor subnets to build here, the stronger the whole network gets.

What we're working on next

Hardware-attested registry access. In the next TEE release, each confidential VM will present a hardware-attested identity to pull its container images. Image access gets bound to attested hardware, so the supply chain tightens another notch.Independently verifiable measurements. This one matters. We're building tooling that lets third parties build the confidential VM themselves and compute its expected measurements offline, independently verifying what's running instead of trusting our published values. Pinned kernel versions make the measured boot reproducible. It shrinks what you have to take on faith, and as a bonus it lets us onboard new hardware faster.A current OS underneath. We're continuing the move onto the newest Ubuntu LTS as the short-lived release before it reaches end of life. Unglamorous but necessary, and done right.

Next week

DropZone gets closer. The next TEE release keeps taking shape. And whatever the week surfaces.See you next week.→ chutes.ai/app→ chutes.ai/app/research