Mark Jeffrey Sat Down With Jon Durbin of Chutes. Here’s What Was Said

On this episode of Hash Rate, Mark Jeffrey sat down with Jon Durbin of Chutes (Subnet 64) for one of the more technically dense Bittensor interviews of the year.

The conversation covered Chutes’ real numbers, the state of decentralized inference economics, and, most importantly, Parallax, Jon’s ambitious plan to solve one of the hardest problems in AI: training frontier-scale models on consumer hardware.

Here’s what was worth pulling out.

Chutes just crossed the Kraken finish line, first Bittensor subnet token listed

Congratulations first. Chutes was the first Bittensor subnet token to make it through a Kraken listing. Jon confirmed a handful of others are dribbling out behind them, but Chutes led the line.

Chutes (SN64) Goes Live on Kraken, the First Bittensor Subnet Token to Reach a Major Exchange

The real numbers behind Chutes

Chutes went live on Christmas Day of 2024, and the growth curve since has been genuinely unusual:

• Peak traffic: 160 billion tokens in a single day.
• Current run rate: ~43 billion tokens per day.
• Total lifetime: 34 trillion tokens served.
• Registered users: ~700K.
• Daily active users: hard to track precisely; a huge portion of traffic flows through aggregators like OpenRouter and Red Pill.
• Revenue run rate: ~$5M ARR, with daily revenue swinging between $10K and $22K depending on usage patterns.

For context, Mark checked against Venice. Venice claims ~50B tokens per day with peaks around 80B, so Chutes is comfortably in the same weight class on LLM inference, while also handling image models, text-to-speech, embeddings, and private deployments that don’t show up in token counts.

The Venice vs. Chutes economics comparison

Venice recently raised $65M, and one of their funders reportedly said publicly that the reason for the raise was because Venice is unprofitable. Eric was hitting the wall on inference network economics. That funder’s video was later pulled down.

Jon’s response: “If you show me a profitable AI company right now, I will show you a liar.”

Every AI company (Venice, OpenAI, Anthropic, everyone) is running at 3x–5x more capex than revenue. The Ed Citron report on OpenAI suggests they’d need a 15x revenue increase just to pay for compute they’ve already committed to. Accenture is publicly saying it’s hard to measure AI ROI. SpaceX capped their AI spend. Microsoft’s CEO said they can’t even plug in the GPUs they already own because there’s no power.

That’s the context against which everything else in the interview lands.

The TAO subsidy question, cleared up

Mark asked about the widely-circulated figure that Chutes gets $45–50M/year in TAO subsidies.

Jon pushed back with facts:

• Chutes had zero emissions for a long stretch on the liquidity pool.
• Of what does flow: 18% goes to subnet owners (and now much of that locks into Conviction).
• 41% goes to validators (mostly redistributed to stakers).
• Only the miner share (~41%) represents anything close to “subsidy”.

The summary is that the real subsidy number is much lower than headlines suggest.

Now the important part: what Parallax actually is

The rest of the interview was Jon walking through why Parallax exists and what it does.

The tl;dr version:

The problem: Every serious AI company is losing money because inference is too expensive. You can’t fix that at the edges. You have to attack the root, and make the models themselves radically more efficient.

What Parallax does: Takes the same intelligence a normal frontier model would have and compresses it into the smallest, most efficient form possible, a form that can run on consumer hardware while being trained across globally distributed nodes with none of the bandwidth catastrophes that killed previous decentralized training attempts.

Why the current state of decentralized training is broken

Jon’s take on where things stand:

• Templar’s 72B run was a real feat but hit a ceiling. Beyond ~72B parameters, sparse loco methods start falling apart. Each machine in that run was basically training the full model and comparing notes, which doesn’t scale.
• Bandwidth is the killer. For a 107B parameter model with naive collocated communication, you’d need ~550 Gbps per participant. Even aggressive optimizations (DILoCo H24, 4-bit quantization) still demand 5.7 Gbps constantly, unrealistic over normal internet links.
• Mixture-of-experts (MoE) models are where the industry is going: GLM, DeepSeek, Kimi, all sparse. But token-routed MoE breaks in decentralized settings because each “island” of compute picks different experts, and when you try to average them, you get “a mongy mess.”

The Parallax technical breakthroughs

This is where Jon gets into the weeds. Three innovations stacked together:

1. Surrogate experts. Instead of syncing every routed expert to every participant (the alltoall communication that kills bandwidth), each participant owns a slice. Everyone else gets a low-rank approximation, essentially a lightweight LoRA that guesses what the full routed expert would produce. Jon claims this maintains 99.98% forward accuracy while eliminating the bandwidth catastrophe.

2. Ternary routed experts. Backed by Microsoft’s BitNet distillation research and Prism ML’s Ternary Bonsai work, Jon’s team is compressing the routed expert weights from 16-bit floating point down to 1.58 bits per parameter (values of -1, 0, or 1). In a 1.6 trillion parameter model like DeepSeek V4 where 50B are active, ~90% of the model is routed experts, meaning almost the entire model can be ternary compressed with virtually no quality loss.

3. Hybrid attention architectures. Standard attention grows quadratically with context length. Every new token requires attending to all previous tokens. Parallax uses Gated DeltaNet 2 (GDN2) for the bulk of compute (fixed-size KV cache, constant decode time regardless of context length) mixed with normal attention for exact recall where it matters. Jon’s team believes they’re the first to build a unified framework for syncing recurrent attention models in decentralized training. Mamba 2, Mamba 3, GDN2, M²RNN; none of them synced properly before.

What Parallax can deliver

The concrete target:

“A 107 billion parameter model that would normally need two H200s should be able to run on a consumer RTX 5090 without any further quantization, lobotomizing, or tricks.”

That’s the pitch to enterprises.

Let’s lay it out exactly as Jon said it:

“What do they hate more than anything if you’re a CFO? Unpredictable costs. You’re trying to get them to adopt AI and you say ‘well, what’s our bill going to be?’ ‘It depends on what ChatGPT wants to charge that day, or Anthropic, or whatever. And you might lose access to the model. Maybe they deem your area against acceptable use.'”

Parallax’s solution: buy the training once, get a purpose-built 100B parameter model that fits your domain, run it on a $10K chip, own the whole stack. Fixed cost. No surprises. Your model, your data, your hardware.

The Bittensor-wide model idea

The most interesting long-term vision Jon floated is this: for really large runs (1T, 2T, 3T parameters), he wants to see all of Bittensor collaborate on it.

• Compute from multiple subnets
• Teutonic optimizing syncing algorithms
• Ready AI on data labeling
• Gradients for fine-tuning
• Affine running RL
• Chutes hosting and serving

He expressed it clearly:

“What is the north star of all of Bittensor right now? It’s kind of a fuzzy, ‘oh, let’s make AI’, but what does that actually mean?” A shared model built across the network could give the ecosystem an actual collective goal.

Privacy is baked in from the ground up

For enterprises worried about handing over proprietary data:

• Full Trusted Execution Environment (TEE) support with NVIDIA confidential compute
• Encrypted GPU traffic
• Encrypted storage volumes
• Miners can’t see workloads; it’s blind compute, take a job, run it
• Fully attestable via source-verifiable VMs

Why RTX Pro 6000 instead of B200s

Jon spent the last few weeks writing custom attention kernels for the NVIDIA RTX Pro 6000 because it’s:

• Much cheaper than a B200
• Actually available to rent (B200/B300 supply is brutal)
• TEE-supported for full confidential compute
• FP4 mixed-precision capable for higher training throughput

The timeline

Realistic expectations from Jon:

• First public model: a ~50–100B parameter coding model built for agentic use, with a 300,000-line data pipeline already in progress. Somewhere between 30 days and 4 months for the training run, depending on parameter targets.
• If that lands well, “all hands on deck, let’s make the biggest baddest model we could possibly do and show the world why Bittensor is the place to be.”
• Enterprise credit-card-swipe custom model service: goal is that customers own the hardware and pay training once, then run inference indefinitely at fixed cost.

The first model will be optimized for tools like Hermes and OpenClaw, running locally on a 5090 at 500+ tokens per second, tuned for the “million monkeys” iterative approach: fast, cheap tokens plus lots of them equals correct answers.

Two audience questions worth flagging

• Alpha holder discounts on inference? The code exists but isn’t enabled. Jon’s cautious about the SEC classification of the token. Chutes deliberately maintains the strictest definition of a utility token, which means being careful about any mechanic that could look like a security feature.
• Prefix caching? Already enabled by default on every LLM. What Chutes deliberately doesn’t do is store those caches on disk; everything stays inside the secure enclave. Never on disk, never readable, never exposed.

Wrapping up

Two things are worth walking away with:

On Chutes today: They’re a real business doing real revenue at scale, operating in the same weight class as venture-backed centralized inference companies while being profitable at the unit level (which basically nobody else is). The Kraken listing is the market catching up to that reality.

On Parallax: This is a serious swing at a real problem. If it works, even at the smaller model scale first, it changes what enterprise AI adoption looks like. If it works at the trillion-parameter scale with the Bittensor-wide collaboration Jon described, it changes what Bittensor itself looks like.

As Jon put it:

“Attention is all you need. Attention is all you need in business and in everything else too. For these models, attention is all you need. If you produce a model that drifts away in the wind, who cares?”

Full interview below: