Bittensor Already Runs the Open Source Version of Every Product the Leading AI Labs Sell

Jensen Huang's first-ever tweet was a letter asking Washington to keep AI open source, co-signed by twenty five companies. The part that matters more than the letter itself is this: Bittensor already built what it asks for, and has been running it for months.

Every product category the closed labs sell - inference, training, coding agents, trusted compute - already has an open, revenue-generating counterpart live on Bittensor, backed by real usage and real emissions rather than a policy aspiration. That's why value will flow to the specific subnets that provide these AI products for a lower cost while maintaining the performance that closed labs provide.

What follows is the letter's actual ask, the subnet-by-subnet mapping against it, and the mechanism that routes value to the winners.

The Letter's Ask, Stripped Down

Nvidia, Microsoft, Meta, Palantir, Dell, and 20 others signed a letter arguing that open weights protect American AI leadership from concentrating inside a handful of closed labs. At the same time, OpenAI, Anthropic, and Google stayed off the list entirely. The letter also asks regulators not to treat distillation, training one model on another's outputs, as theft, since it's a technique the industry has relied on for years.

Huang put some numbers on the stakes back at CES 2026, months before the letter existed: one in every four tokens generated today already comes from an open model. That's the scale the letter is trying to protect, a share of the market too large to call niche and still growing.

The letter's underlying bet is that startups need open alternatives to survive without paying frontier-model prices for every task. That bet already has a track record on Bittensor across four categories worth walking through in order.

The Mapping

Where a closed lab sells inference through its own API, Chutes (SN64) serves Llama, DeepSeek, Qwen, and Mistral through a decentralized network of independent GPU operators, at roughly 85% lower cost than comparable AWS deployments. It has become the highest-emission subnet on Bittensor, processing billions of tokens a day through OpenRouter alone.

Templar (SN3) answers the same problem one layer up. Where a closed lab pretrains a frontier model behind its own walls, Templar already did the equivalent work in the open: Covenant-72B ran to 72 billion parameters across 1.1 trillion tokens, coordinated by more than 70 independent contributors running commodity GPUs with no centralized cluster behind them. The weights and checkpoints now sit on Hugging Face under an Apache license, and the resulting 67.1 MMLU score beats Meta's own Llama 2 70B.

For the developer paying $20 to $200 a month for Claude Code or Cursor, Ridges (SN62) runs a winner-take-all coding-agent competition that its own team benchmarks directly against both, chasing what one industry estimate puts at a $400 billion-a-year software engineering market. The top agent claims all of that round's emissions, and no single party owns the model that wins.

And where an enterprise pays AWS or Azure extra for confidential compute, Targon (SN4) delivers the same guarantee through Trusted Execution Environments running on more than $70 million in NVIDIA-certified hardware, already generating roughly $100,000 a month in revenue committed entirely to alpha token buybacks.

Four categories, and in each one a Bittensor subnet already provides the open substitute for what the letter's signatories, and the labs that refused to sign it, charge for today.

Where the Value Actually Lands

The mechanism worth being precise about is dTAO. Staking TAO into a subnet swaps it into that subnet's liquidity pool in exchange for the subnet's alpha token, so demand for the subnet's service, more Chutes inference bought, more Ridges agents licensed, pushes emissions and alpha value toward whoever staked into that specific pool rather than toward TAO holders generally.

That mechanism is why the mapping matters more than the letter itself. Every dollar spent on Chutes instead of a closed API becomes emissions to Chutes stakers, and every developer running Ridges instead of paying for Anthropic's tools turns into revenue inside Ridges' incentive structure rather than a SaaS subscription line. The open substitute captures the exact demand Nvidia's letter argues needs somewhere open to go, and it captures it subnet by subnet, not across the ecosystem as a whole.

The thesis has one honest crack worth naming. Rayon Labs, the team behind Chutes, also contributes to SN19 and SN56, and one industry estimate puts its combined share near a quarter of total Bittensor emissions across those three subnets, a real concentration risk even inside a network built to avoid it. What separates that concentration from a closed lab's is that Bittensor's stake-weighted ownership can move: Conviction, the governance upgrade live since May, ties subnet control to locked, on-chain commitment that any staker can contest, giving the network a way to correct course if one team's share grows unhealthy.

Meta has no equivalent mechanism. It shipped Muse Spark in April, its first closed-weight model since the Llama era began, without a single staker, developer, or outside party holding a vote in that decision. None of the four subnets above carries that same single point of failure, and that's the structural difference sitting underneath the mapping.

Huang Already Said the Quiet Part

Huang made this argument well before July. On the All-In Podcast on March 20, after Chamath Palihapitiya described Templar training a model with no data center and no company behind it, Huang called it a "modern version of folding@home." His stated position, on the same podcast: the industry needs proprietary models and open models, "A and B," not one or the other.

Four months later, he signed a letter making the policy version of the same case, built on the same infrastructure sitting behind every entry in the mapping above.

Where This Leaves You

You don't need Washington to pass anything for this thesis to work. The open version of every major AI product category is already live, already cheaper, and already routing revenue through TAO instead of a subscription plan or a corporate balance sheet.

Stake into whichever subnet owns the category you think grows fastest: Chutes if inference volume is the trade, Ridges if coding agents eat into the SaaS market before Anthropic or Cursor can respond, and Targon if startups lean into privacy and need confidential compute. The letter asked Washington for an ecosystem that keeps AI open. You already have one you can buy into today.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or trading advice. The information provided should not be interpreted as an endorsement of any digital asset, security, or investment strategy. Readers should conduct their own research and consult with a licensed financial professional before making any investment decisions. The publisher and its contributors are not responsible for any losses that may arise from reliance on the information presented.