Chutes (SN64) Last Week: 7 Models Cut, $294K/Trillion Tokens

Chutes (SN64) Last Week is the name of the new weekly recap series rolled out by the Chutes team, kicking off with coverage of the May 12 to May 18, 2026 period. The team plans to publish these summaries every Monday, packaging shipping news, catalog changes, partnerships, team location updates, and platform metrics into a single place for builders and ecosystem watchers.

The inaugural Chutes (SN64) Last Week edition spans four major threads: a tighter model catalog, a shift in platform economics, build week in Malaysia, and a fresh integration with Trishool (SN23).

Seven Models Removed in Catalog Consolidation

On May 15, Chutes removed seven older models from the network. The team pointed to the ongoing global GPU shortage as the driving factor. Supply of H100 and H200 cards remains tight, Blackwell units carry long lead times, and prices have stayed elevated across the board.

Because Chutes draws compute from a decentralized pool of GPU operators, that pool stays finite at any given moment. When two models cover similar use cases and route to the same nodes, both suffer in throughput. According to the team, the seven removed models all overlapped with stronger newer releases, so keeping them online was slowing down the models people actually use day to day.

Going forward, Chutes plans to focus on frontier open-source models and deprecate older versions faster as better ones land. In a related move, the team also closed the Base subscription tier last week. Existing subscribers retain access through the end of their current billing term, however no new signups or renewals will be processed.

Revenue Per Trillion Tokens Hits $294K

Two charts on chutes.ai/app/research illustrate how the consolidation is reshaping platform economics.

The first tracks revenue per trillion LLM tokens. A year ago, that number sat close to zero. As of May 18, the 7-day smoothed yield reached $286K per trillion tokens, with the 90-day average at $226K and the latest single-day reading hitting $294K.

The second chart shows platform revenue and token volume side by side. Raw token volume has come down from its 2025 highs as Chutes shed low-value traffic. Meanwhile, revenue climbed across the same period. Trailing 90-day platform revenue now sits above $1.41M, with $16,451 booked on the latest reporting day.

In short, the network runs less raw throughput than at peak yet earns more from each token processed. The catalog cleanup forms part of how that shift happened.

Inside the First Chutes (SN64) Last Week Build Sprint in Kuala Lumpur

Chutes contributor Fezicles_eth flew to Kuala Lumpur on Monday to kick off the team’s first build week. Two hackathons run in parallel through the week.

The first is the APU AI Marathon at Asia Pacific University, where Chutes is signed on as Main Sponsor. The second is Chutes Hack Malaysia 2026, organized together with Nyala Labs at Infinity8 Reserve in Sunway Square. Between the two events, over 1,000 attendees will pass through, including students, AI club members, and developers building on the Chutes API for the first time.

The team chose Malaysia for several reasons. Kuala Lumpur offers a strong technical scene, an English-speaking developer culture, and a regional position bridging Southeast Asia and South Asia. Costs also remain favorable for sustained build activity.

As for the hackathon format itself, week-long events generate dense feedback from developers actively shipping against the API. Much of the best product feedback Chutes received in the past six months came out of rooms like these. Throughout the week, team members will run workshops, deliver talks, and host conversations about decentralized inference.

Malaysia marks the first stop on this circuit. Additional regional events will follow through the rest of 2026.

Trishool Partnership Brings a Safety Layer to the Chutes API

Chutes also partnered with Trishool (SN23) last week. Trishool operates as a decentralized AI safety subnet on Bittensor, adversarially testing model behavior through a Petri-based evaluation framework.

The integration brings Trishool’s latest model, Halo Guard Alpha, into the Chutes ecosystem. Halo Guard Alpha works as a safety classifier that catches jailbreak attempts, and it will go live through both the Chutes API and Chutes Chat soon.

The partnership signals where Bittensor is heading as a whole. Subnets increasingly move from isolated operation toward composability, with compute, safety, training, and applications feeding one another. SN23 running alignment work on top of SN64 compute illustrates that pattern in practice. Additional crossovers should follow in coming months.

Further integrations with chains and clients remain in motion, although Chutes has held back on naming those partners. The miner side of the network also continues to grow, with new operators coming online to back the tighter catalog.

The next Chutes (SN64) Last Week edition lands Monday.

FAQ: