Welcome to #103 of the AI edge.
BTC brushed $59k last week before clawing its way back above the $60-61k mark this week. The US-Iran conflict is winding down again, for what feels like the nth ceasefire of the year, and that took a bit of weight off risk assets.
Call it a mild tailwind for BTC. Though if the last few months are any guide, the calm tends not to stick around.
The bigger story this week was on the AI side. Barely a week into the long-awaited public launch of Claude Fable 5, the US government stepped in and ordered Anthropic to cut off access for every foreign national, citing national security.
One of the most capable models on the planet, yours on Monday, gone by the weekend if you happen to hold the wrong passport. If you were ever hunting for a clean case for decentralized AI, a frontier model getting geofenced overnight by a government memo is about as good as it gets.
And then there's Elon, who had quite the week. Days after SpaceX went public and made him the world's first trillionaire, the company turned around and dropped $60 billion to acquire Cursor.
The fit is almost too clean. Cursor brings millions of developers and no model to call its own. xAI brings a frontier model and data centers of compute sitting idle with nothing in coding to point it at. Put them under one roof and each side suddenly has the piece it was missing.
With that, let’s get into this week’s edition.
The Big Story: Bittensor's New Proposal Splits the Room
A proposal making its way through Bittensor has the community split this week, and the fight is a good one. It's called Root Reborn, and it goes after a potential design flaw that’s part of the network's economics.
Bittensor pays its safest stakers a base yield on "root," basically the network's risk-free savings rate. The catch is how that yield gets funded. Every block, the protocol automatically sells subnet tokens to mint the TAO it pays out. So the rate meant to anchor the whole network runs on permanently dumping the very assets that give it value.
Root Reborn wants to flip that. For now it's only a PR under review, nothing live on mainnet, but it's one of the bigger changes to root anyone's floated.
The Wedge
Instead of auto-selling, root validators would reinvest that yield into a basket of subnet tokens they choose, held under the validator and cashable back to TAO whenever a staker wants out. The roughly 1000 TAO of sell pressure opentensor flagged turns into buy pressure that supports subnet prices.
Validators stop being passive pipes and become active capital allocators. Delegators get a new thing to weigh too, since picking a validator turns into a bet on their allocation judgement rather than just their APY.
The argument for it leans on market discipline. Unlike the pre-dTAO days, when validators paid no price for backing weak subnets, a validator's yield now rides directly on the basket they pick. Choose badly and yield drops and stakers walk. Choose well and the staker base grows. Proponents say that feedback loop forces validators toward the subnets actually performing, with the market doing the policing.
The Fine Print
The argument against starts with moral hazard. Yuma's review warns validators can route capital into subnets they secretly own, or cut off-chain deals for favourable weights, the same conflict you get whenever a small group sets the rules and profits from them.
Critics also say this rebuilds what dTAO tore down. Letting validators direct emissions is what failed before dTAO, and opponents expect the same lobbying, political games, and capture by insiders.
And the other worry is about the passive staker, who only wanted plain TAO and now holds an opaque, actively managed basket of illiquid subnet tokens they didn't pick, with no clean way back to TAO and real bank-run risk if everyone redeems at once.
The design is clever, and both sides have a point. The flaw it targets is real, and so are the risks of fixing it this way. It's also still just a proposal with nothing live yet, which means Root Reborn is already doing its actual job: getting the people who run and hold the network arguing over how Bittensor should allocate capital.
Tessara Watch: The Glass Rod Inside The Fiber Shortage
This week the bottleneck in the AI buildout is optical fiber. The thin glass strands that wire GPUs together inside a data center, and right now there isn't enough of it to go around.
Inside a training cluster, tens of thousands of GPUs have to act like one machine, firing huge volumes of data at each other nonstop. Copper can't move that much past a few meters without overheating or dropping the signal. Light through glass can.
So the wiring between racks is optical fiber, the same hair-thin strand of glass behind home fiber internet, except a single AI rack eats about 36x more of it than an old CPU rack did. Demand is going vertical.
The tell is in who's paying upfront. Corning's optical revenue jumped 36% last quarter, and Meta, Nvidia and Amazon have each wired billions into multiyear supply deals just to guarantee they get glass at all. You don't pre-fund a vendor's factory for something you can just order.
Here's why it stays tight. Fiber gets drawn from a "preform," a glass rod about a meter long that's heated past 2,000 degrees and pulled into hundreds of kilometers of fiber. That rod is the choke.
You can bolt on draw towers fast, but a new preform line takes 18 to 24 months and a kind of glass chemistry only around 20 firms on earth have ever cracked, five of them sitting on 60% of supply. Money doesn't compress that. So the squeeze goes straight into price: standard fiber hit a seven-year high earlier this year, up roughly 75% in a single month.
And the one input that makes the rod work, germanium, is about 60% controlled by China and already behind an export license. The rod gates the fiber. Beijing gates the rod.
The full glass layer, and why Corning might be the cleanest way to play it, is in this week's Tessara Research newsletter
Quasar is preparing a 10 trillion-token decentralized training run on Bittensor SN24, beginning with a 5T-token phase to improve model quality through data, training direction, and continuous checkpoint refinement rather than simply scaling parameters.
GEODNET integrated with peaqOS, enabling robots to access centimetre-level positioning and autonomously pay for navigation services across chains.
Virtuals Protocol launched a $400K private inference program powered by Venice, giving builders free access to frontier models and agent infrastructure on Base.
vPay expanded its agentic banking network across the US, Hong Kong, and Singapore, adding new card options, bank accounts, and jurisdictional flexibility for privacy-focused users.
Shekel V3 is now live, allowing users to backtest AI trading agents on historical markets, simulate outcomes, and deploy the same strategies into live trading environments.
peaq launched peaqOS Stream, a trust layer that lets machines cryptographically sign, verify, and monetize their data, creating verifiable machine-generated datasets for the emerging robot economy.
🔥 Our Weekly Top Tweets
#1 An AI Agent Now Ships 15% of Block’s Code
Block revealed an internal AI system that coordinates agents across its entire codebase, handling 200k operations per day, merging 1,500 PRs per week, and contributing 15% of production code changes. What once took months can now be shipped in days.
#2 Humanoid Funding Is Exploding
Global venture funding for humanoid robotics has surged from $1.2B in 2022 to $9.8B by the end of 2025, an 8x increase in just three years as investors race to back the next generation of physical AI.
Cheers,
Teng Yan & Arvind
Quick recap: I launched The Chokepoint, a new weekly newsletter on the AI supply chain. Free, every Tuesday on Tessara Research.
