Happy friday {{first_name}} !
Welcome to #64 of the AI edge. Every week, we round up the coolest things we see in AI x Crypto, plus our quick reflections on the space. Our promise: bringing you the things that matter, before they matter. You’ll also notice the newsletter is getting a bit of a facelift. Sharper and more fun. Got feedback? Just hit reply.
There’s been a lot of friendly fire lately. A few voices are blasting crypto AI projects as hype-driven nonsense, and taking potshots at Bittensor.
This will piss some people off, but most crypto projects are hype-driven BS with zero real value.
The most "valuable" crypto AI project is called Bittensor ($3.4B market cap rn), which claimed to be 100x more valuable than OpenAI - because it's decentralized. But it has no real
— #Yuchen Jin (#@Yuchenj_UW)
4:53 PM • Sep 7, 2025
The Bittensor crowd didn’t take it lying down, pointing out that the critic — Hyperbolic’s founder — started as a “decentralized GPU network” before pivoting into a plain old cloud provider.
As usual, the truth sits in the messy middle. Yes, crypto has plenty of noise and scammy projects. But that doesn’t erase the fact that real teams are building serious infrastructure. Some of Bittensor’s subnets, for example, are genuinely interesting (we’ve got a deep dive on one of them dropping next week 👀).
Meanwhile, many crypto AI teams are still hesitant to launch tokens, worried it’ll make them look scammy. But here’s the paradox: waiting might actually be the bigger risk. Once the space gets “legit,” the teams that already understand token economics will be way ahead.
I see a lot of crypto-AI projects reluctant to launch their tokens because they’re trying to “boil the frog” — warm up a community to the idea that crypto is involved in their business model, but afraid of the impending stigma blowout of involving “crypto”.
I think this is a
— #const (#@const_reborn)
1:37 PM • Sep 8, 2025
AI startups are scaling faster than SaaS ever did. Without crypto, though, most of that value stays locked up in the hands of a few giants.
Tokens are the unlock. They hardwire monetization into the product, give users skin in the game, and spread upside across the network.
Be courageous.
The COT Meme of the Week
Meanwhile, early movers are eating the pie 🍰.
REI is blazing and up >50% since we published our 5,000 word deep dive on it. Total coincidence, of course.
Unibase (UB) will debut on Binance Alpha Sept 12, with trading live and an airdrop available via Alpha Points.
Zypher Network has introduced $POP, the token powering its Proof of Prompt (PoP) protocol and ecosystem, bringing auditable, secure, and trustworthy AI to scale.
Openledger has launched Datanet contributions for whitelisted users, allowing data additions to strengthen AI models, with broader access coming soon.
💻 Aethir Reports $39M Q3 Run Rate
Aethir just dropped some spicy numbers: $13M in revenue for both July and August, which puts Q3 on track for ~$39M. That’s a clean 20% lift from Q2. For a DePIN project, those are eyebrow-raising numbers.
The Wedge: Aethir
Aethir’s pitch: cut dependence on hyperscale clouds by pooling idle GPUs from data centers, telecoms, gaming studios, and crypto miners. Hardware owners earn from unused compute, while developers tap cheaper, on-demand capacity.
The network runs on three roles: Containers supply compute, Checkers verify quality, and Indexers route requests. The goal is enterprise-grade performance without vendor lock-in.
Features include NVIDIA H100 support, infra tuned for real-time rendering, prices up to 80% lower than centralized providers, and hardware ownership stays with contributors.
Gaming is the wedge. Cloud rendering lets players on low-end devices run high-end titles, especially in Asia. The same architecture also supports AI inference.
Payments, staking, and governance run through $ATH, the network’s token.
Source: docs.aethir.com
The Fine Print
Self-reported numbers: Revenue isn’t on-chain. Investors are taking Aethir’s word here.
Reliability: Quality control is hard when compute comes from a patchwork of providers. The Checker role helps, but scaling consistency is unproven.
Token friction: ATH adds incentives, but also volatility and regulatory risk.
Aethir is one of the few DePIN projects with both meaningful revenue and a clear wedge. But the big question here isn’t just “how fast can it grow?”
It’s “how reliable is the reported growth, and can the network keep quality high as scale explodes?”
📊 Vana Launches Playground
Most AI training data today comes from scraped public sources like tweets or Reddit threads, useful at scale, but lacking depth and context.
Vana’s recently launched Playground offers a different approach.
Introducing Vana Playground. A self-serve way to explore Vana's datasets.
From the beginning, we’ve been laser-focused on building valuable datasets and commercializing them through our networks.
This is the evolution: allowing anyone to see and use the data on Vana.
— #vana (#@vana)
12:57 PM • Sep 10, 2025
The Wedge: Vana Playground
A portal where developers can explore datasets sourced directly from communities. Think Telegram chats, Spotify histories, ChatGPT logs, data you can’t scrape from public APIs.
Playground lets you inspect dataset structures, test with synthetic samples, and benchmark model performance before working with the community that owns the data.
Instead of buying opaque datasets through brokers, contributors set the rules for how their data gets used.
The roadmap is to expand Playground into a full marketplace where people can query and trade data directly, with approvals handled by DAO governance.
Source: playground.vana.org
The Fine Print
Data quality: Community-owned doesn’t guarantee signal. Some datasets may be too noisy or inconsistent to be useful.
Governance friction: DAOs move slowly. Negotiating data rights at scale could bog down marketplace activity.
Vana is treating data as an asset that communities can manage and benefit from, rather than just byproduct from platforms.
Are you building something awesome in crypto × AI? Or spotted a startup or product that more people should know about? Fill in this form and share it with us. We’ll feature the sharpest picks in the newsletter each week.
⚙️ Infra & Protocols
Rally has launched on GenLayer, an AI-powered marketing protocol, rewarding real signal over noise with AI scoring, onchain settlement, and open access for projects and creators.
USD.AI just financed the first-ever RTX 5090 loan, a $946K borrow at 18% APR. The protocol targets operators of all sizes locked out of traditional debt. It also hit a $250M deposit cap in just minutes (!)
The Institute for Decentralized AI (IDAI), backed by the Cosmos Institute, has launched to set standards for decentralized AI, starting with research on oversight and security for safer, auditable agent networks.
Gaib.ai's "Final Spice" initiative has attracted $70 million in contributions in less than two days, demonstrating massive capital momentum for building the onchain AI infrastructure economy
OpenSea has introduced OpenSea Mobile, an AI-native trading app that pulls wallets, chains, tokens, and NFTs into one place, with built-in intelligence to help users trade smarter in real time.
Veronica is live on Virtuals Protocol, the first AI agent running end-to-end D2C commerce, from catalog and marketing to payments and logistics.
Giza launched Pulse with Pendle, an AI agent that optimizes PT portfolios, starting with ~13% APR on ETH-PT. It manages allocations, rebalancing, and rollovers in real time to bring intelligent fixed yield to DeFi.
The State of Bittensor (TAO) Report by Yuma Group is live
🤖 Agents & Apps in the Wild
Skynet is live on Virtuals Protocol, drag-and-drop agents built in minutes. $SKY is now trading, powering Skynet’s modular system for launching ACP-native agents without touching a line of code.
StableWatch just unveiled StableSim, a multi-agent Monte Carlo framework for stress-testing stablecoins with real data on stability and resilience. Delpho is the first adopter.
Coinbase is rolling out x402 Bazaar, an open discovery layer where AI agents and devs can permissionlessly plug into APIs, data feeds, and services, the scaffolding for the machine economy.
rainmaker unveiled Cloud9 (C9) Agentic Terminal, an autonomous AI agent hunting prediction market arbitrage across Polymarket and Kalshi. Beta drops next week.
Gaia rolled out Edenlayer, an AI agent discovery and collaboration platform, debuting with the New Eden Dreams game. AI Phone owners get exclusive access and $30 credits, highlighting the rise of agent-driven apps.
🌐 The Web2 Giants
Google rolled out big upgrades to Veo 3 and Veo 3 Fast: vertical 9:16 video, 1080p HD output, and lower pricing, now stable and production-ready in the Gemini API.
MoonshotAI open-sourced checkpoint-engine, a lightweight middleware that updates 1T-parameter models across thousands of GPUs in ~20 seconds.
Tencent has released HunyuanImage 2.1, its latest OS text-to-image model with native 2K generation, ultra-long prompt support, precise text rendering, and rich visual styles.
ByteDance has released Seedream 4.0, a new image model that unifies generation and editing, handling complex multimodal tasks with faster inference and 4K output, and rivalling nano-banana.
Source: menlovc.com
This new corner is for cool things we’ve noticed in AI beyond crypto that are worth your attention.
Most AI teams stay loyal to their provider, but not their model.
A report from Menlo Ventures found that once a team picks a model provider (like OpenAI or Anthropic), they usually don’t switch companies.
66% upgraded to a newer model from the same provider.
23% made no changes at all.
Only 11% switched to a different provider.
Performance is king. When Anthropic launched Claude 4 Sonnet, nearly half of users adopted it within a month. Usage of Sonnet 3.5 cratered from 83% to 16%.
So, most teams aren’t trying to save money by using older models. They upgrade fast to get better results, without changing vendors.
Our deep dive this week is on Manus AI, a startup chasing the holy grail: a real “worker agent” that actually delivers global productivity gains.
Having used Manus for a bit, I’m actually very impressed with the product. We tell you what Manus is, how it works, and where it might go.👇
🔥 Our Weekly Top 5
#1: Former DeepMind researcher just brought anime to life.
With Spellbrush, his team created an anime short from scratch in 48 hours.
A year ago, I left DeepMind to chase a crazy dream with Spellbrush: to bring anime characters to life.
Last week, my colleague and I put our in-house AI models to the test. We made this short anime from scratch in 48 hours.
1/3
— #Jerry Li (#@JerryLiJiaming)
11:29 PM • Sep 6, 2025
#2: ChatGPT just entered its “snitch friend” era.
What once felt like a private conversation now feels risky, people are starting to wonder if their chats could be exposed.
chatgpt is potentially leaking your private convos to the police
people use chatgpt because it feels like talking to a smart friend who won't judge you.
now, people are realizing it's more like talking to a smart friend who might snitch.
this is the same arc we saw in social
— #GREG ISENBERG (#@gregisenberg)
1:03 PM • Sep 7, 2025
#3: Consumer AI startups are breaking the old rules.
They’re making more money from each customer over time by charging based on usage and by expanding into workplaces with team plans.
In the AI era, consumer startups are on a new trajectory.
The best B2C companies are exceeding 100% net revenue retention - which was previously only possible in B2B.
At @a16z, we call this "The Great Expansion." How companies achieve it👇
— #Olivia Moore (#@omooretweets)
3:07 PM • Sep 10, 2025
#4: After raising $500K, the Nous Psyche run has been paused for the past few weeks
The Psyche training run is stuck at 5% and the team seems to have gone ghost.
hello @paradigm @matthuang @togethercompute @DistributedG @NorthIslandVC @Delphi_Digital @rajgokal wtf is going on with @NousResearch's distrusted training run???
— #0_skre (#@skre_0)
5:05 PM • Sep 11, 2025
#5: The path to AGI won’t come from LLMs (according to Yann LeCun).
He argues for joint-embedding, energy-based models, regularization, and model-predictive control, using RL only as a fallback.
Yann LeCun on architectures that could lead to AGI
LLMs can take us only so far.
"If you are interested in human-level AI, don’t work on LLM
Abandon generative models
in favor joint-embedding architecturesAbandon probabilistic model
in favor of energy-based modelsAbandon
— #Rohan Paul (#@rohanpaul_ai)
3:25 AM • Sep 10, 2025
That’s a wrap for this week! Got thoughts, feedback, or something cool to share? Just hit reply. We read it all.
Cheers,
Teng Yan & Issy
Did you like this week's edition?
This newsletter is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.
