Welcome to #92 of the AI edge.
New week, same old range for BTC. Grinding between $66-70K while the world burns around it. The more interesting story is what's happening upstream.
The Strait of Hormuz blockade is now threatening something bigger than oil prices: Taiwan's power grid.
Qatar supplies 1/3 of Taiwan's imported LNG, and over half of the island's electricity comes from natural gas. If this drags on, the semiconductor capital of the world could face energy shortages before it faces demand shortages.
And while TSMC is watching the grid, it may also need to watch Austin.
The big AI news this week was Elon’s announcement of Terafab: a semiconductor factory built jointly by Tesla, SpaceX, and xAI, targeting 1 terawatt of AI compute output per year. To put it simply, it’s like turning an entire country’s power grid into one giant GPU cluster. Logic, memory, and packaging under one roof with rapid design iteration.
It sounds absurd until you remember this is the guy who built a rocket company because launch costs were too high. If anyone is crazy enough to vertically integrate chip manufacturing from scratch, it's him.
We’ve got a great edition this week. Let’s dive in.
The Big Story: Trusted Compute on Untrusted Hardware
The big story is not that Targon found a new way to rent GPUs. It is that they are trying to solve the one thing that has kept serious AI teams away from decentralized compute: trust.
Data is easy to protect when it is sitting in storage. The hard part is protecting it while another machine is using it. That is the hole. If you run a model on hardware you do not own, the operator can, in theory, inspect memory, mess with execution, or pull out your weights.
Targon, Bittensor's Subnet 4, just co-authored a technical whitepaper with Intel ( yes, the $220 billion chipmaker) that directly tackles this problem. A chip company the size of Intel does not usually put its engineers on a paper unless the technical work is worth taking seriously!
The Wedge
At the core is the Targon Virtual Machine. It uses Intel and Nvidia security features to lock down a rented machine so the hardware owner cannot see what is happening inside it. That includes the data, the model, the GPU memory, and the files on disk. The machine only unlocks after it proves to Intel’s trust system that it started up correctly. If someone changed anything important during startup, it stays locked.
This check does not happen only once. About every 72 minutes, each machine must prove again that it is still clean by passing a new challenge from the network. The CPU and GPU checks are bundled into one proof. The whole setup assumes the hardware provider might try to cheat and is built to stop that.
Targon already runs one of the biggest compute subnets on Bittensor, with more than 1,500 H200 GPUs handling real inference traffic. This new setup is not for some future network on paper. It is a security upgrade for one that already exists. The research matters, but the rollout is the real test.
The Fine Print
Intel's involvement matters, but it is worth being precise about what that means. Two Intel engineers helped write the paper. That gives the technical design real credibility. It does not mean Intel is backing all of Bittensor. Still, this kind of engineering signal can do more than any press release.
There is also a cost to this extra security. Confidential computing adds overhead because the system has to encrypt memory and keep checking that the machine is still trustworthy.
The hardware requirements are also high. Providers need newer Intel Xeon chips and Nvidia Hopper or Blackwell GPUs. That means only well-funded operators can realistically join, at least for now.
Crypto networks love to talk about replacing cloud scale with incentives. Fine. But that only matters if they can also replace trust with something stronger than vibes. Targon’s paper is one of the first serious attempts I have seen to do that with hardware-backed proof instead of marketing copy.
One more thing: Targon runs 1,500 H200 GPUs on Bittensor. Here's what the supply picture looks like for those chips right now - it’s tight!
Our Compute Regime Score reads 74 out of 100. That’s deep in scarcity territory. Structured GPU contracts are priced 40% above spot, indicating long-term buyers are locking in supply while the rental market hasn't yet caught up.
NVIDIA tripled in 18 months because GPU demand outran supply. The signals were in the rental data and capex filings weeks before each leg up. That's what Tessara is built to surface. The supply chain shifts that move stocks, before they become consensus.
Opening a small private beta. Reply "beta" and I'll send you access.
io.net launched Agent Cloud, enabling AI agents to autonomously provision and manage compute resources and unlocking fully automated workflows without human intervention.
SurfAI introduced Surf Studio, a prompt-based platform that lets users build dashboards, portfolio trackers, and crypto apps instantly using multi-agent routing and domain-specific reasoning.
QuasarModels (Bittensor SN24) released Quasar Attention, a continuous-time linear attention mechanism supporting up to 5M-token context and outperforming GPT-OSS-20B on long-context benchmarks.
Virtuals Protocol announced the conclusion of its aGDP incentive program after driving $4M in agent-to-agent revenue and launching 32K+ agents.
Minara surpassed $1B in trading volume through its Copilot and Autopilot agents, highlighting growing adoption of autonomous trading infrastructure.
Degenclaw Arena (by Virtuals.io) launched a $300K USDC trading competition for AI agents on Hyperliquid perps, with top performers earning $100K backing and public copy-trading access.
Chutes AI cut token volume by 50% to eliminate unprofitable usage, boosting revenue efficiency by 37.7% per token and 44.9% per GPU, signaling a shift toward sustainable, pay-as-you-go growth.
🔥 Our Weekly Top Tweets
#1 Fresh Capital Flows Into Bittensor
Nearly 60K TAO (~$19M) in fresh capital has entered Bittensor over the past two weeks, with SN3 Templar capturing the largest share, signaling strong long-term conviction and renewed inflows across multiple subnets.
#2 OpenClaw’s First $175K Agent
FelixCraftAI, built by Nat Eliason, has generated $175K in revenue across four AI-native products, making it one of the most profitable OpenClaw agent to date and reinforcing the rise of autonomous, tokenized AI entrepreneurs.
#3 Hermes Hits 10K Stars
Nous Research’s Hermes Agent surpassed 10,000 GitHub stars, marking it as the team’s most widely adopted open-source project to date and signaling strong momentum in agentic development tools.
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.
