EAIDaily — July 9, 2026

Focus: AI Coding & Embodied Intelligence Sources: AI Agents Directory daily brief, Embodied Global, Reuters / The Information, Cyberpress, IT之家, GMA News, NVIDIA Developer Blog, Cainiao / Roland Berger white paper, AI HOT curated feed

1. General Intuition raises $320M at a $2.3B valuation to build gaming-trained “world foundation models” for robots

What happened: New York–based physical-AI startup General Intuition closed a $320M Series B at a reported $2.3B valuation on July 8, led by Coatue with participation from Jeff Bezos, former Google CEO Eric Schmidt, Vinod Khosla, and researchers from MIT and Google DeepMind. The company was spun out of Medal.tv, a gaming-clip platform that gave the team access to millions of hours of gameplay footage paired with controller telemetry (which button was pressed, and when). Using that data, General Intuition trains a “world foundation model” intended to encode spatio-temporal intuition, then fine-tunes a downstream policy on a very small real-world corpus. The team demonstrated a quadruped that, after just eight minutes of real-world fine-tuning, zero-shot navigated an office with dynamic obstacles using only a front camera. The company reportedly turned down an OpenAI acquisition offer to stay independent and is also building Nerve, a marketplace that pays gamers to do teleoperation and data labelling.

Why it matters: A $320M Series B for a pure-software, no-hardware robotics layer is a signal that top-tier Silicon Valley capital is willing to bet on generalist world-model software even as Chinese humanoid makers scale production and open VLA stacks (π0, GR00T N1.7, LingBot-VLA 2.0) industrialize imitation learning. The eight-minute fine-tuning claim, if replicated, would meaningfully lower the real-world data requirement that has constrained VLA deployment outside labs — the same bottleneck that justifies General Intuition’s “ChatGPT moment for robotics” pitch. The bet also reframes the data race: instead of competing for scarce robot demonstration data, the company is substituting abundant, cheap, action-labelled gameplay video.

Source: Embodied Global — General Intuition Raises $320M at $2.3B Valuation to Train Physical AI Foundation Models on Gaming Data

2. NVIDIA GR00T N1.7 + Isaac Teleop land in Hugging Face LeRobot under Apache 2.0, posting +61% on DROID-F6

What happened: NVIDIA and Hugging Face announced on July 7 that Isaac GR00T N1.7, NVIDIA’s flagship vision-language-action (VLA) foundation model for humanoids, and Isaac Teleop, a commercial-grade teleoperation framework, are now native components of the LeRobot open-source robotics library under Apache 2.0. GR00T N1.7 swaps in a Cosmos-Reason2-2B backbone built on Qwen3-VL, replaces the old Eagle vision encoder, and is pretrained on roughly 32,000 hours of real human demonstration plus ~8,000 hours of simulation from BEHAVIOR, RoboCasa, and Simulated GR-1. Versus N1.6, NVIDIA reports DROID-F0 +10% and DROID-F6 +61% (the long-horizon, multi-stage benchmark), with smaller gains on SimplerEnv-Bridge and SimplerEnv-Fractal. The stack includes ONNX/TensorRT export pipelines targeting Jetson Thor, the Isaac Lab-Arena batched evaluator, and an open dataset of more than 350,000 real + simulated trajectories and 57M grasps (15M+ downloads to date). NVIDIA Cosmos 3, the next-gen physical-AI world model, is slated to enter the same loop “in the near future.”

Why it matters: Three things rarely ship together in robotics: a permissive commercial license, a VLA with clear gains on hard multi-step tasks, and a full teleop → sim → train → eval → deploy toolchain wired into the de facto community library. The release effectively turns LeRobot into a one-stop shop for humanoid policy development and tightens NVIDIA’s grip on the embodied-AI reference stack ahead of the Jetson Thor ramp. For startups and academic labs it removes a year of glue work; for Chinese competitors it raises the bar on what “open-source VLA” means in 2026.

Source: Embodied Global / NVIDIA Developer Blog — GR00T N1.7 VLA and Isaac Teleop Land in Hugging Face LeRobot: Apache 2.0 Model With Cosmos-Reason2-2B Backbone Hits +61% DROID-F6

3. Cainiao + Roland Berger 2026 white paper: China holds 80%+ of global commercial service robot shipments and tops humanoid patents

What happened: Cainiao Network, Alibaba’s logistics arm, and Roland Berger released a joint 2026 Robotics White Paper on July 8 projecting the global robotics market to grow from ~RMB 380B in 2025 to ~RMB 2.12T in 2035, a 5× expansion over a decade. Headline findings: Chinese companies account for over 80% of global commercial service robot shipments, China is #1 worldwide in humanoid robot patent filings, and leading Chinese players in light-interaction commercial scenarios (delivery, cleaning, household vacuums) already generate >80% of revenue overseas while warehouse robotics leaders see >70% overseas. The report also flags a new moat: for cross-border robotics, the binding constraint is no longer product innovation but global supply-chain and compliance capability (Class 9 dangerous-goods logistics, oversized/overweight SKUs, exporter/importer registration, ESG and IP). The paper is candid that dexterous manipulation, high-takt industrial actions, and autonomous navigation in complex homes remain bottlenecked, which limits humanoid exports today to education and research scenarios.

Why it matters: The white paper is the first systematic China-side attempt to size the global market and frame the next competitive phase. It positions Chinese robotics not as a domestic subsidy story but as a globally dominant export industry whose moat is now logistics + compliance rather than pure technology. For foreign robot makers, the 80% service-robot share and 5× market projection are the new competitive baseline; for procurement teams, the compliance bullet list (battery shipping, customs, certifications, data privacy) becomes a checklist before any humanoid rollout.

Source: Embodied Global / The Paper — Cainiao and Roland Berger: China Holds 80%+ of Global Commercial Service Robot Shipments, Leads World in Humanoid Robot Patents

4. xAI and Cursor reportedly preparing a joint AI model launch, with internal memo pointing to a July 8 window

What happened: Reuters, citing The Information, reported on July 7 that SpaceXAI (xAI) and AI-coding-tool vendor Cursor plan to launch their first jointly developed AI model as soon as July 8. The companies had delayed an earlier launch to improve efficiency. The Information’s report is based on a staff memo; Reuters said it could not independently verify the details, and Cursor declined to comment. The model is being framed around speed and competitive performance with leading systems, but no benchmarks, pricing, or rate-limit information has been published. If the model ships inside Cursor, it would give the IDE vendor a first-party compute path that reduces dependence on OpenAI and Anthropic.

Why it matters: The story is the first concrete sign of a coding-tool vendor trying to internalize model distribution in a way that ties compute ownership (xAI’s GPU buildout), developer workflow distribution (Cursor’s IDE), and model competition into the same product surface. If the launch lands and benchmarks hold, it breaks the implicit neutrality of Cursor as a multi-model host and forces OpenAI/Anthropic into a more contested channel. The skepticism is warranted — Reuters could not verify the report, Cursor declined to comment, and benchmark claims need independent evaluation — but the structural shift it implies is real: coding IDEs are becoming model channels, not just clients.

Source: Let’s Data Science / Reuters / The Information — SpaceXAI and Cursor Launch Joint AI Model

5. OpenAI wins U.S. government clearance to broadly release GPT-5.6 after national security review

What happened: Axios reported on July 7 that the U.S. Department of Commerce gave OpenAI approval for a broad public rollout of GPT-5.6 after additional testing and meetings between the company and government officials. The clearance ends a period of regulatory uncertainty that had delayed the AI system’s wider release, just one week after the U.S. government lifted similar restrictions on Anthropic’s latest Fable and Mythos models (which had been grounded for less than three weeks). GPT-5.6 — released in three tiers (Sol flagship, Terra mid, Luna entry) — adds two new run modes not available in GPT-5.5: a “max” mode that trades latency for reasoning quality, and an “ultra” mode that fans out work to multiple sub-agents in parallel. Sol is priced at $5 / 1M input tokens and $30 / 1M output tokens, roughly half the price of the comparable Anthropic tier.

Why it matters: The clearance is the second front-end model from a U.S. frontier lab to be unlocked inside a month, confirming that the federal “compute review” process is now a recurring compliance step rather than a one-off delay. For OpenAI, the timing matters: Sol, Terra, and Luna have been in a holding pattern while Anthropic pushed Fable/Mythos into general release. For the agent-coding market specifically, the “ultra” sub-agent mode is a direct competitive response to Claude Mythos’s sub-agent and long-horizon tool-use story and will likely show up in coding-assistant UIs within weeks.

Source: Cyberpress / Axios / Let’s Data Science — GPT-5.6 Launch Cleared After OpenAI Completes U.S. Government Security Review

6. China industry regulator warns of a “security backdoor” in Anthropic’s Claude Code tool

What happened: A Chinese industry regulator issued a public warning on July 8 that versions of Anthropic’s Claude Code tool embed a “security backdoor,” raising the geopolitical stakes of Anthropic’s enterprise push into China. The warning follows earlier reporting (referenced in CBS News coverage) that versions of Claude Code shipped with hidden user-tracking telemetry — the same disclosure that triggered Alibaba’s full ban on Anthropic tooling effective July 10. Anthropic has not yet published a full post-mortem of the backdoor claim; the regulator’s statement was framed as a precaution for state-owned enterprises and critical-infrastructure operators.

Why it matters: The backdoor warning is the first formal regulatory action against an Anthropic developer tool in China, and it lands the same week as the U.S. Commerce Department cleared OpenAI’s GPT-5.6. Together the two stories draw a tightening perimeter: foreign AI tools now need to clear both a U.S. compute review and a Chinese security review to operate at scale, and the safe paths for cross-border AI coding procurement are narrowing. For enterprise buyers it accelerates the “sovereign stack” thesis (domestic models + domestic coding tools) that the LongCat-2.0 and Microsoft-MAI moves have been pushing.

Source: CBS News / The Decoder — China Warns of ‘Security Backdoor’ in Anthropic AI Coding Tool

7. Microsoft makes Sales Agent and Service Agent generally available inside Dynamics 365 and Microsoft 365 Copilot

What happened: Microsoft announced on July 8 that its Sales Agent and Service Agent are now generally available inside Dynamics 365 and Microsoft 365 Copilot. The two agents are designed to handle customer interactions and streamline business processes autonomously — qualifying leads, drafting outreach, summarizing case context, and routing tickets — and ship with the same conversational framework Microsoft has been rolling out across the M365 Copilot product line. The release comes alongside Satya Nadella’s broader push to migrate Copilot workloads off OpenAI and Anthropic onto in-house MAI models.

Why it matters: This is the first large-scale enterprise deployment of Microsoft’s “MAI-first” agent strategy outside of Office. Whereas the MAI-Excel and MAI-Outlook stories earlier in July were about cost, Sales Agent and Service Agent are about distribution: they sit inside the CRM that already runs a meaningful share of the world’s sales pipelines, which gives Microsoft a channel to ship agent behavior into enterprise workflows without requiring customers to procure a new tool. For AI-coding-tool vendors, it sets a competitive reference for “agentic CRM” that few startups can match on data integration or compliance.

Source: CX Today / AI Agents Directory — Microsoft Integrates Agentic AI into Sales and Service Conversations

8. Three Chinese embodied-AI companies close nine-figure rounds in a single 24-hour window

What happened: Three Chinese embodied-AI startups announced nine-figure rounds on July 7–8, signaling that the capital cycle for Chinese robotics is still accelerating even as valuations in the U.S. tighten. Zhongke Xinsong, a Siasun spinout building full-stack industrial embodied AI for semiconductor fabs, intelligent welding, and assembly, closed a nine-figure RMB Series A+ co-led by Hongtai Fund and CDH Baifu, with Chengtong Fund, Dazhong Juding, and Guangzhou Suyuan participating; the company reports 5,000+ deployments across Southeast Asia, Europe, and North America. Lingjing Zhiyuan, incubated by Shanghai AI Lab, raised over ¥100M across Angel and Angel+ rounds led by Matrix Partners, shipping a “Dvořák” brain-cerebellum heterogeneous compute architecture with its own MS-OS to 200+ robot makers covering 70% of China’s embodied-AI market. Bharat Robotics, a Bengaluru-based Indian humanoid maker, closed a ₹50 crore (~$6M) Series A led by Venture Catalyst Partners to scale its Bharat-1 humanoid for automotive and electronics assembly, priced at ₹45 lakh with pilots reporting 30% productivity gains in packing tasks.

Why it matters: The pattern across the three deals is striking: each company is selling a vertical, not a humanoid. Zhongke Xinsong sells into semiconductor and welding lines, Lingjing Zhiyuan sells the chip-OS full stack, and Bharat Robotics targets Indian assembly lines. Combined with the Cainiao/Roland Berger white paper (Item 3), this is the clearest sign yet that the embodied-AI market is splitting into a “vertical industrial” track and a “general-purpose humanoid” track — and that capital is rewarding the former while the latter still chases product-market fit. The Indian entry also marks the first time an Indian humanoid vendor has surfaced in the China-aligned funding ecosystem at this scale.

Source: Embodied Global — Zhongke Xinsong A+ round / Lingjing Zhiyuan Angel rounds / Bharat Robotics Series A

Quick takes

• KEENON Robotics showed its XMAN-R1 humanoid serving fully autonomous coffee at the LEAP East 2026 conference in Hong Kong (July 8–10), powered by KOM 2.0, billed as the world’s first VLA built specifically for the service industry, paired with a role-specific ProS vertical model.
• Anthropic expanded Claude Cowork to web and mobile, with cross-device sync, offline background execution, and the disclosure that >90% of Cowork usage is non-software — pointing to its growing role as an “administrative coworker” rather than a developer tool.
• Meta released Muse Image, a new generative system within Superintelligence Labs aimed at graphics creation and creative collaboration, continuing Meta’s push to ship standalone creative AI products.
• Google DeepMind launched a $10M fund to study safety in multi-agent AI systems, formalising multi-agent safety as a research category.
• AIsa raised $6.5M (Alibaba and Tribe Capital co-led) to build a transaction network for the AI-agent economy, and Norm Ai raised $120M for legal-focused AI agents — both signalling that “agent-to-agent payments” is now a real funding category.
• RoboDojo released a unified sim-and-real benchmark (42 Isaac Sim + 18 real tasks) across 30 leading VLA policies, finding the best policy scores just 8.80% sim / 12.8% real success vs. 76.03% / 100% for humans — a sobering data point for embodied-AI capability claims.
• Zeroth, a Chinese humanoid robotics startup, announced $74M in pre-Series A financing to expand its home embodied-AI business, joining a growing cohort of Chinese home-robot bets.

Trend lines

• The “world foundation model” race is now the data story, not the architecture story. General Intuition, GR00T N1.7, MIRA, RynnWorld-4D, and MECo-WAM all converged this week on the same bet: spatio-temporal prior learned from large, cheap, action-labelled data (gameplay, teleop, egocentric video) is the binding constraint for embodied AI, and architecture is now a second-order problem. Whoever owns the data flywheel owns the next five years.
• AI coding is becoming a sovereign-software problem. In the same 24 hours, the U.S. cleared OpenAI’s GPT-5.6, China warned of a backdoor in Anthropic’s Claude Code, and Microsoft pushed more of Copilot onto its MAI stack. The neutral “best tool for the job” era of agentic coding is closing; geography is becoming the primary filter.
• Vertical industrial embodied AI is splitting off from humanoids. Three of this week’s biggest Chinese/Indian rounds went to companies selling into specific verticals (semiconductor fabs, welding, assembly, embedded compute), not general-purpose humanoids. The Cainiao white paper ratifies the split. Expect 2026 H2 to be defined by which general-purpose humanoid vendors find a vertical beachhead before capital tightens further.

One-sentence synthesis

On July 9, 2026, the AI-coding story was about regulatory geography (U.S. clearing GPT-5.6, China warning on Claude Code, Microsoft deepening MAI), while the embodied-AI story was about the world-model data race and the vertical-industrial split — with General Intuition’s $320M bet, NVIDIA’s GR00T N1.7 / LeRobot open-source release, and three simultaneous Chinese industrial robotics rounds marking the clearest signal yet that 2026 H2 will reward data flywheels and vertical depth over general-purpose demos.