Welcome to #95 of the AI edge.
BTC pushed past 76k this week as the Iran and Strait of Hormuz situation finally showed signs of cooling off. March's range high is broken. Now we watch. Does it hold and run, or fade back to range lows? Next few days will tell.
On the AI front, the White House is preparing to give US agencies access to Anthropic's Claude Mythos model. The same government that called Anthropic a “supply chain risk” is now handing it the keys to federal infrastructure. Whatever this model can do, it's apparently powerful enough to make Trump change his mind.
Meanwhile, TSMC posted record Q1 revenue of $35.7B, up 35% year-on-year. Strait of Hormuz chaos since March? Didn't even dent the quarter.
But don't let the headline fool you. The cracks aren't in Q1. They're in what comes next. Earnings season starts in earnest next week, with several important AI companies set to report. That is where the market will start testing whether current optimism is justified. We break it all down in the latest edition of our AI infrastructure newsletter - one click to subscribe here.
With that, let's get into this week's edition.
The Big Story: The Subnet That Wouldn't Stay Down
A few editions ago, we covered Covenant-72B, the largest decentralized LLM training run ever completed on Bittensor. Jensen Huang name-dropped it on All-In. Chamath co-signed it. TAO ran 90% off the back of it. That's how big it was.
Last week, the team behind it walked. Sam Dare, Covenant AI's lead, published a blog accusing Bittensor co-founder Const of running a centralized operation behind a decentralized facade.
Then he liquidated over 37,000 TAO across his three subnet holdings. Roughly $10M out the door. TAO tanked 18% in a day. The team that brought the mainstream attention to the Bittensor ecosystem just torched it on the way out.
The exit exposed a structural hole. Nothing in the protocol stopped a subnet leader from dumping everything and disappearing overnight. Ownership was static. Commitment was assumed. And when that assumption broke, holders paid the price.
Bittensor's response came fast. Const pushed back on the accusations directly, disputed the core claims, took responsibility for what went wrong, and apologized to affected holders.
Then he shipped. Within days, he proposed Locked Stake, a new mechanism that ties subnet influence to time-weighted conviction. Instead of just staking tokens and claiming ownership, teams now lock their stake over a set period and accumulate a conviction score that builds the longer they stay committed. That score is verifiable onchain.
Sudden exits become structurally harder because locked tokens can't be dumped on a whim. Ownership becomes something you continuously earn.
On top of that, Const took Subnet 3 back. Templar has been renamed Teutonic, and a 1B parameter model is already training. 30B is targeted within a month. 1T is the endgame. The original Templar run took three months to reach 1B. Teutonic is doing it in a week. The codebase, the compute, the incentive design that powered Covenant-72B never left the network. Const wrote the original Templar codebase. The model still exists. The thesis survived even if the team didn't.
The Wedge
The Covenant blowup forced a governance upgrade that was probably overdue. Locked Stake directly closes the gap between Bittensor's decentralization pitch and what the protocol can actually enforce.
Testing starts on mature subnets with new ones getting temporary immunity. Rollout is gradual. That's the right call for something this foundational.
The Fine Print
Governance upgrades don't erase what happened. A flagship team publicly exited and dumped eight figures on holders while Grayscale was filing a TAO Trust S-1 days earlier. Institutional trust takes longer to rebuild than mechanisms take to ship.
Teutonic's training targets are aggressive. Compressing timelines that dramatically while switching leadership mid-stride needs runs on the board before it earns the hype.
Most protocols would've spent a month drafting a postmortem blog. Bittensor shipped a governance overhaul and started training a new model before the news cycle moved on.
Say what you want about the drama. The network is still standing and the code is still running.
Gensyn AI launched AXL (Agent eXchange Layer), a peer-to-peer encrypted communication layer for decentralized AI agents.
Zyfai launched its ATM on Virtuals, offering 7.9% APY via autonomous USDC rebalancing into the highest risk-adjusted yields.
Virtuals.io surpassed $4M+ in agent-to-agent commerce revenue, powered by 18K+ agents and 116K wallets, cementing measurable aGDP traction.
Amadeus Protocol acquired Bitte Protocol in a $1.7M $AMA deal, merging deterministic agent infra with Bitte’s 24K+ active user base.
PwC France partnered with Score (Bittensor SN44) to integrate decentralized “Eyeball Layer” vision AI into enterprise workflows.
Inference.net released Catalyst, enabling teams to train and deploy production LLMs for agents in minutes using existing app data.
Reppo launched a $1M Datanet Builders Fund to accelerate onchain data businesses and lower barriers for non-crypto-native builders.
🔥 Our Weekly Top Tweets
#1 Darkbloom Awakens Idle Macs
This Eigen Layer dev built Darkbloom, a private inference network that connects idle Apple Silicon Macs into an encrypted compute layer, claiming 230B-model capability at up to 32x lower cost than cloud, with hardware-level attestation so node operators can’t see your data.
#2 Agents Are Choosing Base
Recent x402 transaction data shows Base dominating agent payment volume, signaling that AI agents are increasingly settling commerce on Base over other chains.
Cheers,
Teng Yan & 0xAce
P.S. If you’re a startup doing serious work in AI or robotics and want to work closely with us, reply to this email and let’s chat.
