Exclusive: Here’s Why the Chutes Team Didn’t Attend Proof of Talk

Proof of Talk wrapped in Paris, and one name conspicuously missing from the floor was Chutes. No booth, no panel, no team huddled in a corner pitching SN64 to passersby. For a subnet the number 1 subnet on Bittensor, the absence got noticed.

On June 15, Timon Agar from the Chutes engineering and product team published the reason, and it’s refreshingly blunt: they were too busy shipping to show up.

The team runs on a GSD philosophy. Get Shit Done. And the last stretch happened to be one of those all-hands-on-deck windows where everything in flight converges at once, leaving no room to fly to Paris.

So what was so important that it beat out a flagship industry event? A lot, as it turns out.

The migration that ate the calendar

The Chutes team was busy migrating its entire platform, and all active workloads now fully run inside Trusted Execution Environments (TEE).

TEE, built on AMD SEV-SNP, means every component of the production stack is now memory-encrypted and attested. That migration forced a consolidation: idle inventory got trimmed, image and voice models got bundled into “packs” that run several models on a single GPU instead of one GPU each, and the validator no longer needs any GPU infrastructure of its own. That last point alone is a serious expense off the books.

The validator itself got rebuilt for cost. Nodes now autoscale, either through standard HPA or custom KEDA metrics, so the system can collapse during off-peak hours and snap back the instant traffic spikes. They left a provider that tried to hike prices dramatically, built their own observability tooling, and in the process killed a hefty monthly Datadog bill.

Network policies were tightened, WAN traffic to internal services like AlloyDB and managed Redis was eliminated, and cross-region replication and high availability got far simpler. The payoff there is regional partitioning, the kind of thing an EU client demands when data legally cannot leave Europe.

There’s a hardware milestone too. Kyle, working with the miner Pierre, brought the first Blackwell B200/B300 nodes online inside TEE, with NVIDIA’s Nemotron-3 Ultra 550B as the first model on the new silicon. Crossing from Hopper and RTX Pro 6000s into flagship Blackwell unlocks native NVFP4 compute, higher concurrency, larger models, and the death of several KVCache bottlenecks.

Meanwhile cxmplex is pushing CPU-only TEE infrastructure forward, wrestling with genuinely thorny attack vectors, like the fact that sensitive data can leak into host-visible system logs even when memory is encrypted and every channel is secured.

And then there’s Parallax, whose draft tech report is now out. It enables training large sparse MoE models across heterogeneous, disconnected nodes efficiently, chipping away at the barriers that have hobbled decentralized training. Paired with a new data pipeline built to collect and synthetically augment coding and agentic workloads, because, as the team puts it, data has been and always will be king.

Each of these has a dollar figure attached. That’s the whole point. Migrating production infrastructure produces a precise, measurable number that directly reduces sell pressure on the token.

They’re not skipping the next one though

Worth noting is that the team plans to show up from here. Chutes sponsored the upcoming Exploit conference at the platinum tier and plans a strong in-person presence. So the door’s open.

With the full TEE migration in place, on board of Blackwell going on, and Parallax hitting new milestones, Chutes spent the time shipping. And they’re not stopping.