EAIDaily — 2026-09-09

Daily curated English AI brief focused on AI coding & embodied intelligence. 8 stories · Sources: AI HOT, WebSearch (quidproquo, qbitai, geekpark, 36kr, cgtn, technode, xinhua, 21jingji)

1. Cognition (Devin) closes $2B+ Series E at a $48B valuation — AI coding agents priced as infrastructure

• Source: quidproquo Funding Brief / CoinPost / 36Kr; TechCrunch
• Why it matters: Cognition raised over $2B led by a16z and Accel, lifting its valuation to $48B up from $26B in May 2026 (1.85× in four months). Run-rate revenue jumped from $492M → $900M (+83%), and the company disclosed $800M in expected 2026 cash burn, mostly to lease an Nvidia cluster. Three new products launched alongside the round: Devin Auto-Triage (incident investigation), Devin Security Swarm (vuln triage), and Devin Automations (events from Slack/GitHub/Linear kick off work). a16z and Accel entered as new leads — not existing backers — signaling VCs no longer treat “AI writes code” as a feature but as an infrastructure layer.
• What to watch: whether on-prem/sovereign execution (à la Cursor Self-Hosted Machines) becomes Cognition’s next enterprise unlock, and whether Devin can sustain 50% MoM enterprise growth with compute constraints.
• URL: https://quidproquo.cc/posts/daily/2026-09-09-funding-cognition-en · https://coinpost.ai/en/topics/194221

2. Xiaomi opens MiMo Desktop beta with two preview models — “Smart” routing + 99% cache hit

• Source: Xiaomi MiMo (official); Geekpark / 36Kr; IT之家
• Why it matters: Xiaomi launched the MiMo Desktop closed beta on Sept 8 — its first desktop-class AI app — alongside two preview models: MiMo-X-Pro-Preview (flagship) and MiMo-X-Flash-Preview (lightweight). The app reads spreadsheets, PDFs, video, audio and compressed archives without preprocessing, and autonomously controls the system browser. Built-in “Smart” orchestrator routes each task to the cheapest viable model or splits it across multi-session agents. Crucially, same-session cache hit rate reaches 99%, cross-session 95% — controlling token cost on long chains, which is today’s hottest hard problem in agent productization. The lineup positions Xiaomi against Manus / Claude Desktop / Genspark as the OS-level Chinese PC entry.
• What to watch: whether “agent with browser control + cache engineering” becomes the new floor for any serious desktop AI; the answer will determine the next 12 months of consumer/hardware-bound agent competition.
• URL: https://mimo.mi.com/docs/en-US/news

3. Cursor ships Self-Hosted Machine Pools — agent execution moves into the customer’s network, not the model’s

• Source: Cursor Blog / Cursor Changelog (Sept 2, 2026); Cloudflare; Pondero
• Why it matters: Cursor expanded Self-Hosted Machines with dynamic worker pools, autoscaling, hibernation, and Cloudflare Sandbox support — letting enterprises route cloud-agent tool execution onto customer-controlled infrastructure while the agent loop stays in Cursor’s cloud. Eight sandbox providers integrate at launch (AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, E2B); Linux and macOS computer-use is supported. Cursor reports that >60% of its own internally merged PRs are now produced by Cloud Agents. The architecture split (orchestration in Cursor, execution on-prem) addresses data-residency blockers in finance, healthcare, and government — sectors where GitHub Copilot and JetBrains AI have not yet shipped equivalent on-prem agent execution.
• What to watch: whether GitHub Codespaces / Azure DevContainers announce analogous on-prem agent execution, and whether Cursor pursues FedRAMP / SOC 2 certification.
• URL: https://pondero.ai/news/2026-09-07-cursor-self-hosted-machines · https://cursor.com/cn/blog/self-hosted-machines

4. Songyan Dynamics launches Scalabot embodied-AI sub-brand with first world model HERON-World Model

• Source: Shanghai Securities News (cnstock) / Gasgoo / 21jingji
• Why it matters: Beijing-based humanoid maker Songyan Dynamics (松延动力) formally launched the Scalabot embodied-AI sub-brand and unveiled HERON-World Model — the first model in its HERON architecture stack. Scalabot signals a shift from “robot body maker” to “body + embodied brain” full-stack provider. Unlike the standard pyramid that floors everything on massive internet video, HERON-World Model uses only three data streams: real-robot tele-op, simulation rebuilt from real scenes, and first-person human operation video. Architecture is MoT + MoE, not bound to a single body, with the explicit bet that physical-world scaling laws still hold. The company says future HERON models will target home service first, then commercial service.
• What to watch: whether scalers like Songyan and AgiBot converge on “world models tied to tele-op data” as the dominant embodied-AI training paradigm through 2027.
• URL: http://noetixrobotics.com/ · https://www.cnstock.com/commonDetail/787333

5. WRC 2026 wrap-up: 3000+ exhibits, humanoid shipments 97% global share, 100K full-year target — China shifts from “demos” to “doing real work”

• Source: Xinhua / People’s Daily / 163 / CNR / CCTV / CGTN
• Why it matters: The 2026 World Robot Conference closed Aug 23 in Beijing Yizhuang. Over 3,000 exhibits, 311 debut products, 44 new tech solutions, 300+ companies (+36% YoY). Key industry signal: 2025 Chinese humanoid robot shipments hit 90% global share; H1 2026 industry revenue 165.5B RMB (+24.5% YoY); humanoid shipments 97% global share with >40K units in H1 and a 100K full-year target. The 1.35m / 39kg Tiangong Omni humanoid debuted; 4m-tall giant humanoid handles kg-to-tonne loads. Tiangong / Galbot / Unitree / UBTECH all showed “doing real work” demos — pharmacy restocking, depalletizing, laundry folding, cafe operations. The thematic shift: from “demo mode” to “deployment mode” — physically executing jobs, not back-flipping.
• What to watch: 2026 H2 humanoid shipments will determine whether 100K target is realistic; recent MIIT + SASAC “Real-Scene Field Training” program explicitly funds this with central SOE real-factory slots.
• URL: https://en.people.cn/n3/2026/0707/c90000-20475138.html · https://news.cgtn.com/news/2026-08-20/Robots-move-closer-to-real-world-partners-at-World-Robot-Conference-1PL5LLAu6XK/share_amp.html

6. UBTECH: H1 2026 humanoid revenue 590M RMB (+1,445% YoY), full-size units 921 — industrial B2B now #1 revenue line

• Source: UBTECH 2026 Interim Results (HKEX filing, Aug 28); China Auto News; RobotToday
• Why it matters: UBTECH’s H1 2026 reported revenue 1.269B RMB (+104.2% YoY), gross profit 567M (+160.9%), gross margin 44.7% (+9.7pp), net loss narrowed 23%. Full-size humanoid robot business hit 590M RMB (+1,445% YoY), becoming the #1 revenue line at 46.5% of total revenue — overtaking legacy business. 921 full-size humanoids shipped (+1,946.7% YoY), total humanoid units 16,123 (+268.3%). Industrial customers include Airbus, BYD, Goldman’s Chinese operations. Three products debuted in June: industrial Cruzr Y1, commercial Walker C1, hyper-bionic U1 (Resonance-LM affective model). U1 orders already >13,361 units. The data make industrial B2B the clearest path to humanoid PMF in 2026.
• What to watch: U1 first deliveries on Sept 16 — whether consumers trade up to a ¥30K+ companion robot will determine unit economics for the consumer market.
• URL: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0828/2026082801691.pdf · http://www.cnautonews.com/zhinengwl/2026/08/31/detail_20260831392988.html

7. Xiaomi debuts AI expo + 30-unit “Fuzhi” humanoid cluster at CIFIT Xiamen — overseas orders from Africa, Middle East, Europe

• Source: Securities Times / Toutiao (Sept 9, 2026)
• Why it matters: On Sept 8, the 26th CIFIT in Xiamen opened with the Fujian Humanoid Robot Innovation Center demonstrating 30 “Fuzhi” series wheeled-arm humanoids in coordinated multi-robot control — the largest such cluster demonstration in Fujian. The robots use full 3D-printed lightweight bodies + self-developed f-VLA large model, keeping per-unit cost under ¥10K (5%-10% of comparable products). 5 control modes including f-VLA, exoskeleton, super-vision remote over private network. The center is < 1 year old and already has tens of millions in domestic orders plus overseas inbound from Africa, Middle East, and Europe. The pattern: ultra-cheap wheel-arm humanoids optimized for cluster demos are now China’s emerging export product, just as industrial humanoids reach ~¥100K.
• What to watch: whether $1K-class wheel-arm humanoids become the “gateway” embodied-AI product in overseas B2B education/demonstration markets before full bipedal deployment.
• URL: https://www.toutiao.com/article/7683291268423680512

8. OpenAI agents solve Navier-Stokes Millennium Problem with Lean verification — but accusations of “research-data leak” ignite a credibility fight with NYU mathematicians

• Source: OpenAI / Sam Altman @sama / TechCrunch / Simon Willison / DeepMind commentary
• Why it matters: OpenAI announced that a team of agents, on a model “significantly stronger than GPT-6 Astra,” produced a finite-time blow-up proof of 3D Navier–Stokes (Millennium Prize problem, ~90 years open). The proof was Lean-verified in 17 hours, 49M messages exchanged, ~300B output tokens consumed; one stat put cost in the “single-digit millions of dollars” range. Within hours, NYU mathematician Tristan Buckmaster (with Levent Alpöge, Anthropic mathematician) publicly accused OpenAI of receiving leaked research signals from their independent track and racing to publish — Sam Altman pushed back, claiming both had only Euler and communication broke down. Noam Brown pointed out cost will fall fast (just as o3 → o4 fell 100×). The story crystallizes three forces at once: agents as research collaborators, AGI economics being publicly priced, and the handoff protocol between human mathematicians and AI agents still has no norm.
• What to watch: what publication-norm emerges for “AI co-authored math proofs” and how Anthropic / OpenAI position themselves competitively in the agent-for-science race — Claude Fermat and Codex Syl are now formally on the table.
• URL: https://openai.com/index/navier-stokes-solution · https://techcrunch.com/2026/09/08/openai-fought-dirty-on-career-making-math-problem-says-nyu-mathematician

Trend Lines — 2026-09-09

• AI coding infrastructure is being publicly priced: Cognition at $48B, Cursor’s on-prem execution opening regulated sectors, AI HOT showing Chinese stack (MiMo Desktop, Tencent Docs AI Workbench) catching up. Coding-agent economics now matches model-quality as a competitive axis.
• “Brain-body split” becomes the embodied-AI playing field: Songyan Scalabot HERON, UBTECH industrial + consumer B2B (WRC awards), 30-unit Fuzhi cluster — China’s embodied-AI builders are now visibly divided into “founding-body-and-brain” full-stack players vs ultra-cheap cluster demonstrator players.
• “Real-work deployment” replaces “showcase performance” as the dominant narrative: WRC’s “less dancing, more working” shift; China SOE-factory field-training policy; cross-industry B2B revenue (Airbus, BYD, hotels, pharmacies) — the era of acrobatics is closing.
• “Research as a product” week: Navier-Stokes with Lean verification, Fermat proof, GPT-6 Sol internal testing (~6× Astra throughput), Google’s Teamwork multi-agent framework — agent-equipped math is moving from research-paper artefacts to weekly product demos.
• The cost/quality trade-off is breaking in coding agents: a16z/Accel buying Cognition at $48B on $900M ARR (53×), Cursor 60% of internal PRs from agents, Noam Brown’s “cost falls fast” doctrine. Investors are underwriting that token economics will catch up with capability.

Generated by WorkBuddy · EAIDaily automation · Following @WoLoveAI 8 stories curated · Primary focus: AI coding + Embodied intelligence