The TatsuOS Whitepaper and six-month roadmap are live. TatsuOS is a new utility for the TATSU token — and a new answer to a problem the IoT industry has never properly solved: when you buy a connected device, the vendor owns its features, not you. TatsuOS changes that. Burn $TATSU, unlock a feature, own it forever — no subscription, no phone-home dependency, no vendor that can revoke it. The receipt is cryptographically signed and verified on the device itself. But the receipt mechanic is just the foundation. On top of it: a managed telemetry pipeline, fleet-wide dashboard, over-the-air firmware and model updates, on-platform AI model training, and AI agents that help you build models from your own device data — without standing up your own MLOps stack. One platform. One token. No separate vendors. The six-month roadmap: Month 1 — Hardening and Ethereum testnet deployment. Full burn → BLE → unlock flow on real ESP32 hardware. Production signing key in KMS. Contracts fuzz-tested and frozen. Month 2 — Closed pilot. Ten devices in five community members' hands. Real burns, real unlocks. Target: ≥90% receipt delivery success on first attempt. Month 3 — Multi-board support. Firmware runs on nRF52 and STM32 alongside ESP32. Telemetry SDK published as a standalone Zephyr module with templates for agriculture, industrial, and consumer use cases. Month 4 — Public beta. JS/TS SDK on npm. Self-service dashboard and docs go live — from clone to first unlock in under 30 minutes. Month 5 — Audits. Two independent third-party audits (contracts + firmware), full reports published, all critical findings remediated. Mainnet dry-runs complete. Month 6 — Ethereum mainnet launch. Contracts live under multi-sig. Firmware and SDK 1.0 released. First commercial OEM ships TatsuOS-enabled devices. First quarterly state-of-TatsuOS report: total devices, total unlocks, total TATSU burned. Read the full white paper: [tatsu.gitbook.io/tatsu-white…]

Sunday Brief – Oct 25, 2026 Brief Summary TatsuOS is live on testnet. On-chain payments in ethereum:0x92f419fb7a750aed295b0ddf536276bf5a40124f, USDC and USDT were tested end to end on Sepolia. This week we strengthened platform security, redesigned the console for desktop, tablet and phone, and added an internal admin dashboard. Chipforge is about to begin its next quest: building more advanced NPUs to run more AI models at higher speed. TatsuOS - The full TatsuOS stack is deployed on dedicated cloud infrastructure, with payment contracts live on Sepolia. Users can pay in ethereum:0x92f419fb7a750aed295b0ddf536276bf5a40124f (discounted), USDC or USDT. - The purchase flow is tested end to end: the wallet payment confirms on-chain and the plan or AI credits are applied automatically. - Security was strengthened: internal services are isolated, each device is restricted to its own channel, the core framework was upgraded, and payment confirmation was hardened. - Free, Pro and Enterprise plans are now enforced: device limits, how much history is kept, and data export. - The console was redesigned with light and dark themes, a fully responsive layout, clearer board pages and live readings across the fleet. New users get a guided first-run setup. - We added an internal, read-only admin dashboard protected by two-factor login. It shows user growth, plans, board activity, payments by currency and AI credit usage, and flags any payment that was never applied to an account. Chipforge Chip Design Team - The validation pipeline (benchmarking system) for Quest 2 is built and in its final stages of testing. In Quest 2, miners will design a complete AI chip, hardware plus software, and it will be tested on seven real-world AI models covering image classification, object detection and image segmentation, including a transformer model. - The test models range from about 2 to 26 million parameters. We kept them at this size so each validation stays fast. The chips that miners design are expected to scale to much larger models, up to several billion parameters. - Each design is judged on how fast it answers a single request, how much work it handles per second, and how much chip area it uses. - We tested the full validation on three different machines. Every validator reaches exactly the same result, and a full run now takes about an hour and a half instead of six. - Security was independently reviewed and every finding fixed. Anti-cheating checks are stronger, which minimizes exploits.

This is the part most people miss. NVIDIA’s moat was never just the GPU. It was CUDA. Co-designing hardware + compiler + runtime is how you actually win in AI acceleration.

Sunday Brief – Oct 18, 2026 Brief Summary - TatsuOS: The deployment plan to take the platform from prototype to production is complete, and all five repositories are now documented and checked for leaked secrets. - Chipforge: Quest 1 is closed and Quest 2, a full edge AI accelerator challenge, is being prepared. The subnet now runs on a tested, documented and secure production setup. TatsuOS Deployment Plan: From Prototype to Production With the prototype nearly complete, we finished the plan for running TatsuOS in production. - The platform is designed to scale to thousands of boards, each with multiple sensors, and its services grow automatically as load increases. - Credentials are kept out of code entirely and stored in a managed secrets service. - Rollout is split into clear, verifiable phases, ending with go-live checks. - All five repositories now have detailed READMEs, have been scanned for leaked secrets, and have had large build files removed. TatsuOS now has a clear, documented path to production. Chipforge Chip Design Team - Quest 1, the NPU v1 challenge, is complete and closed. - Quest 2 is being prepared: miners build a full edge AI accelerator, meaning the NPU hardware, its compiler and its runtime, designed and scored together. - Designs run seven widely used vision models (MobileNetV2, ResNet-50, YOLOX, DeiT, DeepLabV3 and more), covering classification, detection, segmentation and transformers. - Every evaluation reports industry-standard figures: FPS, latency, accuracy, TOPS, TOPS/W, area and clock speed. - The evaluator's models and test data have grown to several GB, so the evaluator now ships as a small signed package that fetches and rebuilds its large files at install. Every file is verified byte-for-byte, so every machine produces identical results. - End-to-end testing is under way with a dummy NPU. A full evaluation currently takes over 6 hours, and we're working to bring that down. Subnet Infrastructure Production, Security and Easier Setup The final part of this upgrade focused on production infrastructure, security, and making the subnet easier to operate. - The challenge server now runs in production with HTTPS, rate limiting, and caching. - Validators and miners get a one-command Docker setup, automatic restarts, and a settings check after updates. - The EDA server is locked down by default and automatically sizes itself according to the machine's available cores and memory. - All four repositories were scanned for leaked secrets during a security review, with none found. - READMEs were rewritten so new validators and miners can get started independently.

1,000× in under two weeks. Not a lab, not a megacorp, just engineers around the world competing on ChipForge. Decentralized chip design is real.

Smarter AI needs smarter hardware.

The AI chips the world needs most aren't in data centers. They're in cameras, robots and the devices in your pocket. That's the chip ChipForge is building: small, efficient, designed in the open by builders anywhere, scored on speed, power and area.

Sunday Brief – Oct 11, 2026 Brief Summary TatsuOS This week, we spent some time on R&D around BLE (Bluetooth Low Energy) and OTA (Over-the-Air) updates two areas that are important for making TatsuOS easier to deploy and maintain across supported devices. On the BLE side, we explored how devices can be discovered, connected, and configured more reliably with minimal user interaction. For OTA, we looked into how firmware updates can be delivered remotely in a secure and dependable way without requiring physical access to the device. There is still more work to be done in both areas, but this research gives us a clearer direction for how these features will fit into the broader TatsuOS ecosystem. Chipforge Chip Design Team - Our first real Neural Processing Unit (NPU) competition went live, and miners showed up in force. Within days we had more than 50 NPU designs from teams competing to build faster AI hardware. Several were hundreds of times faster than the starting point. - We caught people trying to game the system, and closed those gaps fast. Some miners found ways to post scores they hadn't earned. We shipped three validator updates in a week. Now only honest, working designs score, and cheating attempts are publicly marked down. - The next version is well underway. It moves from "build a fast accelerator" to "build a complete AI chip platform": hardware, compiler and software together, tested on larger, real-world AI models. - We're measuring what real chips care about. This week we built and validated a way to measure chip power accurately and quickly, plus a library of on-chip memories that miners can pick from, as in commercial chip design. - Our aim is to remove limits on miners' creativity. We reviewed every part of the competition for anything that holds miners back, and are fixing those so the best ideas can win. Subnet Infrastructure Fairer Winners, Steadier Validators - Winners are now decided when a batch closes: the highest-scoring submission that passes both gates wins if it beats the score to beat. Exact ties go to the earlier submission. - Scores are published only when the batch closes, preventing validators from seeing one another’s results early. - An optional minimum improvement margin can be configured, with logs explaining each decision. - Evaluation-side errors are no longer counted against miners. - The winner’s share of emissions can now be set centrally so every validator follows the same split. - Validator fixes prevent weights from flipping between the winner and burn, stop already-scored submissions from being re-sent, and verify downloads against the miner’s signed hash. Every validator now applies the same rules, and weights are steadier.

Chip startups spend years in stealth before anyone sees a benchmark. ChipForge put the scoreboard up on day one. Four edge AI models, exact answers only, scored on area and speed, every result reproducible, leader visible to all. SN108 is live. Take the top spot.

Everyone's fighting over data center GPUs. Meanwhile the AI chips that actually ship in volume go into phones, cameras, sensors and wearables. That's the market ChipForge is aiming at: an open NPU, built in public, proven on MLPerf Tiny. Live Oct 16, 9 AM EST. Go build.

Sunday Brief – Oct 4, 2026 This week, TatsuOS moved another step closer to its preview release, while ChipForge rolled out a redesigned live leaderboard and upgraded its challenge infrastructure. TatsuOS We’ve completed end-to-end testing of the TatsuOS ecosystem, and the team has now pushed each codebase as a service to GitHub. The next step is deploying the full infrastructure to our servers in an optimized and reliable way. Once that is complete, we’ll release a preview/beta version of TatsuOS. A few areas still need more work — mainly BLE (Bluetooth Low Energy) and OTA (Over-the-Air) updates. After the beta release, we’ll continue improving these features and expanding support for more boards across different hardware families. TatsuOS is getting closer to being ready for real-world use. Chipforge Chip Design Team - The NPU test cases were improved further. Designs are now tested on real, trained AI models from the industry benchmark, and every answer is checked against Google's reference software. - The right answers stay hidden while a design is being tested, so the only way to pass is to actually compute them. - A final security review found and closed two more gaps, and every result was re-checked after the fixes. Subnet Infrastructure A New Live Leaderboard ChipForge shipped a redesigned leaderboard and moved the challenge server back to subnet API. - Each submission now has its own row showing UID, result, score, score to beat, quality gates, and validator progress. - Live progress bars show submissions being evaluated, alongside a challenge countdown. - A scores-over-time chart and “who the chain paid” view show the winning design and how much validator stake agrees with the chain’s selection. - Submission pages explain results in plain language, while incomplete evaluations are clearly marked as Partially evaluated. - The homepage now shows the score to beat, current winner, and top designs in real time. - Miners can download their own evaluation logs, while protections prevent copies of winning designs from being submitted. Winning designs can also be revealed to registered neurons after a delay. Miners can now see where every submission stands and why.

TatsuOS Update We’ve completed end-to-end testing of the TatsuOS ecosystem, and the team has now pushed each codebase as a service to GitHub. The next step is deploying the full infrastructure to our servers in an optimized and reliable way. Once that is complete, we’ll release a preview/beta version of TatsuOS. A few areas still need more work — mainly BLE (Bluetooth Low Energy) and OTA (Over-the-Air) updates. After the beta release, we’ll continue improving these features and expanding support for more boards across different hardware families. TatsuOS is getting closer to being ready for real-world use.

Sunday Brief – Sep 27, 2026 Brief Summary This week, we made TatsuOS reach out. A dashboard is only useful if someone is watching it, so we built an alert system that finds the right person when something goes wrong. Using Twilio for SMS and voice calls, along with email for lower-urgency notifications, TatsuOS can now alert users and escalate critical issues when no one responds. Most importantly, users stay in control — they decide what triggers an alert, who gets notified, through which channel, and how aggressively it escalates. TatsuOS This week was about turning telemetry into action. Reliability Milestone: Alerts That Find the Right Person - Real notifications: TatsuOS can send SMS, place voice calls via Twilio, and send email alerts. - Configurable triggers: Users can create sensor threshold alerts, dead-man’s switches for silent devices, and model-driven alerts based on anomalies, predictions, or state changes. - Flexible delivery: Users choose contacts, channels (SMS / call / email), and severity levels. - Automatic escalation: If a critical alert isn’t acknowledged, TatsuOS moves through the configured contact tiers until someone responds. - Alert reliability: Cooldowns prevent notification storms, acknowledgements stop escalation, and crash recovery ensures unsent alerts aren’t lost. - Delivery visibility: Twilio and email callbacks feed back into TatsuOS, showing whether alerts were sent, delivered, or are still pending. The idea is simple: an alert nobody sees isn’t useful. TatsuOS now makes sure critical information reaches the people who need it. Chipforge We are actively improving: - The subnet codebase - The challenge server to provide a smoother experience for validators and miners - The website to make information more detailed and easier to understand We’ll be sharing more details about these improvements as they progress. Chip Design Team - The evaluation files for the first NPU challenge are ready. We put together the evaluator, the scoring setup and the reference results, then ran the whole thing on the EDA server with our test cases and a dummy NPU design. Everything came back matching what we expected. - We also made the evaluation faster. A single request now gets the whole server to itself, and when several requests come in together they share it sensibly instead of fighting for cores. If a run gets stuck, it is stopped after 45 minutes with a clear reason. - We spent time thinking like a miner and closed the gaps we found, so it is much harder to score well without actually building a better chip. The challenge will be launched once all internal testing on the challenge server is done.

Miners don't just submit the hardware. They submit the kernels to run on it. Plenty of accelerators look great on paper and die on software. Requiring both means designed hardware has to actually work, not just synthesize. Hardware and software, judged together.

Open models came first. Then open compute. Hardware was the layer everyone assumed would stay closed. That's why we built ChipForge. The next great AI accelerator shouldn't need a giant's balance sheet behind it. Just a better design.