Welcome to #90 of the AI edge.
The Middle East conflict keeps escalating, and this week, the second-order effects started hitting the supply chain.
The Strait of Hormuz is blocked. Most people see an oil story. The real risk is higher up the stack with chips, datacenter materials, and the supply chains AI is built on.
Qatar's helium production is offline, taking a third of global supply with it, and no real substitute exists in chip manufacturing. Sulphur disruptions are hitting metals processing that feeds into datacenter builds. TSMC is exposed across the board.
One waterway, and the entire hardware layer powering AI has a vulnerability nobody was pricing in.
BTC, somehow, doesn't care. It's been ranging between $65-70K all week like nothing's happening. Either the market has priced in chaos, or it's sleepwalking into it.
On the AI front, the headline wasn't a new model, but instead it was Karpathy demonstrating that agents can now do AI research autonomously. He pointed one at a neural network, left it running for two days, and it found real improvements across ~700 changes that a human hadn't caught.
We've gone from agents that write code to agents that improve AI itself.
Quick heads up: We have a small number of research partnership slots. If you’re a startup doing serious work in AI or robotics and want to work closely with us, fill out this form and I’ll personally follow up.
The Big Story: The Internet Just Trained a 72B Model
Every major language model in history has been built inside a data center. One cluster, one org, fast interconnects holding everything together. Templar (a Bittensor subnet) just proved that's no longer a requirement.
Their Covenant-72B run produced a 72-billion-parameter model on roughly 1.1 trillion tokens, coordinated across 70+ participants worldwide over the commodity internet, with no central infrastructure and no application process to join.
It now stands as the largest fully open decentralized pre-training effort ever completed, and it holds its own against peers like LLaMA-2-70B across standard benchmarks.
The Wedge
The fundamental problem with training over the internet is bandwidth. Inside a data center, you get 400+ Gb/s between machines. Templar's participants were running on household-grade connections (roughly 500 Mb/s down, 110 Mb/s up).
Their optimizer, SparseLoCo, addresses this by compressing the data communicated by over 146x. Participants train locally for 30 steps, then send heavily compressed updates instead of full gradients. The result is that even though the model is 7x larger than the previous record holder (INTELLECT-1), each sync round took just 70 seconds while also pushing compute utilization to 94.5%.
Trust (or lack thereof) is the big story here. Previous decentralized runs vetted participants upfront. Covenant was fully open. Peers could join or vanish mid-run, meaning the system had to catch freeloaders, garbage submissions, and adversarial poisoning automatically, every round. Their on-chain validation system, Gauntlet, scores each update against held-out data and normalizes contributions so no single node can hijack the model. Getting untrusted strangers to collaboratively train a model without anyone gaming the system is a problem that's been discussed for years. This is the first time it's actually worked at scale.
The Fine Print
"Open to anyone" comes with an asterisk. The minimum hardware to participate was 8x B200 GPUs, ~$250K in equipment! Not for the hobbyist. Technically permissionless, practically exclusive. This was more of a proof-of-concept for well-capitalized GPU operators.
The benchmark comparisons deserve context. LLaMA-2-70B shipped in mid-2023. Today's frontier models train on an order of magnitude more tokens with parameter counts in the trillions. Covenant is an infrastructure breakthrough, not a frontier model.
Making this accessible to smaller hardware setups is the next frontier. The team has published research on Heterogeneous SparseLoCo, which would enable resource-limited nodes to contribute via pipeline parallelism with compressed activations. But that's still on paper.
The bottleneck was never compute. It was getting strangers on the internet to train one model without cheating. Templar seems to have solved that. The question now is: can decentralized training actually catch up with the frontier? or does the gap keep widening?
VeryAI raised $10M (led by Polychain) to build “verifiable humanity” infrastructure with privacy-first palm biometrics + cryptographic ID to prove you’re real online without leaking sensitive data.
Ethereum’s dAI team + Virtuals published ERC-8183, a proposed standard for trustless agentic commerce that formalizes “Jobs” with on-chain escrow + evaluation, designed to plug into ERC-8004-style agent identity/reputation.
Nous Research’s new agent, the Hermes Agent, is now the #3 trending GitHub repo in Productivity, beating even OpenClaw, which is at #11.
Zeus Subnet kicked off its first pilot with a European energy trading firm, turning AI forecasts into renewable generation estimates to support real power-trading decisions.
General Tensor closed $5M across pre-seed + seed (with participation from funds like DCG), extending runway to keep building deeper into the Bittensor stack.
Modulr Robotics processed 100M+ testnet transactions in a month, a mainnet-scale throughput signal ahead of its upcoming launch.
Chutes AI partnered with a Harvard research team to develop prefix-caching techniques aimed at faster, cheaper inference across the Bittensor ecosystem.
Virtuals Protocol surpassed $3M in agent-to-agent service revenue (ex-fees) and is expanding Epoch 4 so non-tokenized agents can earn incentives too, pushing “aGDP” beyond just tokenized agents.
GEODNET launched GEO-SWARM on Kickstarter, a fully autonomous “drone-in-a-box” home security system that patrols, docks, and recharges on its own.
Openmesh unveiled XnodeOS, a NixOS-based operating system for DePIN nodes with reproducible builds and atomic updates to standardize infra deployments.
In case you missed it…
Reinforcement learning is driving a lot of the progress in AI. Right now, only a few giant labs can really do it at scale.
That may be starting to change.
If Gradient’s architecture holds up, this is not only about cheaper compute. It is about opening up one of the most important parts of modern AI.
This week, we published a deep dive on why that matters →
🔥 Our Weekly Top Tweets
#1 Fruit Fly “Brain Upload” Moment
Researchers simulated an adult fruit fly brain from its connectome and dropped it into a virtual body where behaviour emerged without training. A small but very real step toward brain emulation.
#2 The War in Replay Mode
An AI “agent swarm” scraped public OSINT in real time and built a minute-by-minute 4D reconstruction of the Iran strikes on a 3D globe, tracking airspace shutdowns, GPS jamming, satellite passes, no-fly zones, and shipping chaos.
#3 Alibaba’s AI Starts Mining Crypto
Alibaba’s experimental AI agent reportedly started mining cryptocurrency autonomously during a training run and repurposing compute for its own “side hustle.”
🔥 NEW: The Daily Chain
I've been building a daily intelligence briefing on the AI infrastructure value chain. That means the compute, memory, power, and supply chain layers underlying everything we cover here. Hyperscalers are projected to spend $3 trillion on AI by 2030, and that money has to flow somewhere.
Every morning: top signals and a short editorial on what's moving across the stack. Just takes 2 minutes to read.
You can read today’s edition on how ByteDance is exploiting the flaw in US chip controls, and what it means for NVIDIA. It’s free to subscribe.
Cheers,
Teng Yan & 0xAce
