And guess who is the oracle provider for XRP Ledger?

Eight months of ETF outflows, erased in five weeks. That's how fast this market turns. And every one of those positions is settled against a price, every second, whichever way it turns. That part never took the eight months off. Neither did we.

What was the first DeFi app you ever used? Bonus points if you remember which oracle it ran on.

DeFi has been declared dead more times than anyone can count. It's back. Again. The timeline's awake. Volumes are up, lending's busy, perps are running hot, and the builders who stuck around through the quiet part are shipping again. Here's what that means for us: every loan, every liquidation, every perp position is a price that has to be right. More DeFi means more of those, on more chains, all at once. Band prices DeFi. We never left

Rough day for the bears 🐻 Biggest day of ETF inflows in 11 months. ~$650M short liquidations in a single day. Every one of those liquidations came down to one price hitting one number. Getting that number right has been our whole job since 2018. Are we so back yet? 🔥

GM ☕️☀️ Are we so back?

Band Open Thoughts: TradingView shipped an official MCP server this week. You can now wire an AI agent straight into your account. Your watchlists, your alerts, your charts, all through conversation. It stops short of placing trades. Deliberately. Trust Wallet handed agents full trade execution to 220 million users back in March. So the line exists. Nobody agrees where it goes. Reading your charts? Fine. Managing your alerts? Probably. Sizing the position? Pressing the button? Where do you think the line should be drawn?

Never missed a departure ✈️

Two years ago this week, Band's price feeds went live on Flow mainnet. 54 symbols. Two years. Not one interruption. @flow_blockchain built something genuinely outstanding in that time: a chain the NBA, NFL, Disney and Ticketmaster actually ship on. Then turned around and built serious DeFi infrastructure on top. TVL up 600%. Most-built-on L1 at every ETHGlobal hackathon for six straight events. And the Forte upgrade has only made it better. Two years down. Here's to what gets built next. 🤝

First rate hike in three years. CLARITY Act rejected. Timeline in pieces. Here's what none of that actually did: it didn't un-ship a single thing. No block stopped being produced. No contract got deleted. Nothing that worked yesterday stopped working today. Band's still building in the background, every feed running perfectly. See you on the other side 👀⛰️

Open-Weight Wednesday: The four biggest US labs just agreed to slow down. Amodei called for it Saturday. Altman, Hassabis and Musk backed him within a day. The trigger was a swarm of agents breaking into Hugging Face on their own. To be fair, "pacing" isn't stopping. It means building in time for safety checks and outside evaluators. But here's the open-weight question nobody has a clean answer to: a slowdown is an agreement between companies. Open weights are already downloaded. Qwen, DeepSeek, Kimi and GLM weren't in that room, and nothing agreed on Saturday changes what's already sitting on people's hard drives. So does a frontier slowdown slow anything down? Or does it just change who's at the frontier?

Every few months people keep declaring the AI narrative dead. It keeps not being dead. Take a look at what's being built. Even in crypto: agent networks, decentralised compute. On-chain model training. Private LLMs. "The Data Layer That Trains AI Engines and Powers Blockchain Applications." You know what that means 😉

Model Monday: Qwen3.8 @Alibaba_Qwen 's biggest free model yet. Download it, run it on your own machine, no API key, no per-message bill. While US labs debate whether releasing models openly is too risky, Chinese labs have quietly made it their default. The useful part: it holds about a million tokens at once. That's an entire codebase, or a few novels, sitting in working memory without it losing the plot halfway through. Most models start forgetting the beginning long before that. Trade-offs worth knowing: it's text only, no images. And it always reasons before answering, so it's slower than a model built for quick replies. Anyone actually run it yet?

Band Open Thoughts: Someone on Reddit is running a 2.8 trillion parameter model on a gaming PC. That's the whole post. Ten years ago that sentence would have been science fiction. Now it's a Sunday.

Hide everything else.

Bigger screen. Same Unified Data Layer.

Open-Weight Wednesday: image gen edition. The most competitive corner in AI right now is image generation. Frontier closed models are clearly ahead on Chat. But on style fidelity benchmarks for image-gen, it’s even. OpenAI’s images 2.5 dropped today. Which open model do you think can beat it? FLUX, Qwen-Image, Krea 2, or something you're keeping quiet about?