Nvidia makes $10 billion from robotics AI, and China is its most eager customer

Nvidia is expanding its business in artificial intelligence for physical machines, and Chinese companies are now among its most important customers. This includes robots, vehicles, and drones—applications where models don’t generate text but control movement in the real world. This segment, known within the company as Physical AI, currently generates around $10 billion in annual revenue. For comparison, in the four quarters leading up to July, the company’s total revenue was $303 billion, so the robotics business accounts for about three percent of that. CEO Jensen Huang expects this revenue to increase tenfold within the next decade. To this end, Nvidia has spent the last few years building connections in the Chinese humanoid robotics industry. Local manufacturers there still rely on U.S. suppliers for computing chips and the associated software. This means an industry that Beijing has declared a strategic future technology is dependent on a supply chain subject to American export regulations. → Wall Street Journal

Synthszr Take: China’s humanoid manufacturers build the bodies; they buy the nervous systems from Santa Clara. Ten billion dollars is pocket change next to $303 billion in total revenue, but geopolitically, it’s the more interesting position: If Huang is right about the tenfold increase, in ten years Washington will have a say via export licenses in which robots perform which tasks in which factories. Beijing knows this and has been working on decoupling since Deepseek’s switch to Huawei silicon, but the robotics stack is tougher to replace than a language model: No one can replicate simulation environments, sensor drivers, and training data from millions of hours of grasping and driving in just two quarters. Europe doesn’t feature in this equation, even though a good portion of the industrial robot hardware is located here, while its control intelligence is being negotiated between California and Shenzhen. Thus, any factory automation planned for 2031 is also a foreign policy decision.

China’s automakers bet on humanoids: Xpeng raises over $900 million

Xpeng’s robotics unit has raised more than $900 million in a single funding round, at a post-money valuation of over $6.3 billion. The round was led by IDG Capital, with participation from Gaorong Ventures, Tencent, and Alibaba; according to the company, it is the largest single private funding round ever recorded in China’s Embodied AI industry. According to the Wall Street Journal, company founder He Xiaopeng and co-president Brian Gu invested around $100 million of their own money in the financing. The unit’s product is called Iron, a humanoid robot with a human-like form built for commercial use. Other Chinese manufacturers are following this path as well: Chery subsidiary AiMOGA is preparing for an IPO, according to Reuters, BYD introduced the humanoid Xiao Di, and Changan, GAC, Li Auto, SAIC, and Seres are also developing such machines. → Techpresso

Synthszr Take: Automakers are the natural incubator for humanoid robotics because they already have actuators, sensors, battery cells, and mass production capabilities in-house, whereas companies like Figure or Apptronik have to painstakingly piece their supply chains together. He Xiaopeng sees automotive margins shrinking and is shifting his focus so significantly that he and Brian Gu are investing $100 million of their personal wealth into their own robotics unit. Chery is preparing AiMOGA for an IPO, BYD is showcasing Xiao Di, and Hyundai is sending Atlas to its Georgia plant this year: The industry that accelerated the industrialization of electric cars is now repeating the pattern with bipedal machines. The software side remains an open question, and there, Tesla has a head start with years of driving data and training infrastructure that pure manufacturing expertise can’t automatically overcome.

Tencent releases new open-source flagship model

Tencent has released Hy4 preview, an open-source flagship model with 770 billion parameters, including 49 billion active parameters, and a context window of one million tokens. According to Tencent, the release specifically targets long-running software engineering tasks, document-heavy office work, game development, and scientific research. The company states that the model can convert scattered files into documents, spreadsheets, and presentations, including data analysis and financial models. In a blind internal comparison, 163 Tencent experts evaluated results for 203 engineering tasks: Hy4 preview achieved an average score of 2.99, GLM 5.3 scored 2.92, and Kimi K3 scored 2.94. Against Kimi K3, Tencent recorded 51.2 percent wins, 7.9 percent draws, and 40.9 percent losses. According to the company, the training data was built around the work of its own software engineers, game developers, financial analysts, and security experts, and the model is being developed in conjunction with the in-house tools CodeBuddy and WorkBuddy. → Techpresso

Synthszr Take: Tencent had 163 of its own experts blindly evaluate 203 engineering tasks, and Hy4 scored 2.99 compared to 2.94 for Kimi K3. The difference is statistical noise; the choice of benchmark is the real statement. A model trained on the work of its own engineers and financial analysts and developed alongside CodeBuddy and WorkBuddy is calibrated for precisely the tasks that cost salaries in-house.

New attack vector steals secret agent skills

In an arXiv-preprint, a research team led by Yu-Lin Tsai describes an attack called 'Daydreaming,' which can be used to replicate secret Agent Skills using only ordinary task queries. These skills bundle instructions, reference data, and executable helpers that a general agent uses to perform a specialized task; providers keep the files proprietary and only sell access to the results. The attack never asks the provider for the skill, nor does it have any reconstruction evaluated; instead, it strategically poses tasks whose results differentiate between hidden behaviors. The authors distinguish between three access levels (Differential, Trace, and Output) and focus on the weakest: The attacker only sees the final answer and the returned files. Across seven skills and four victim models, the authors claim the method achieves 86.8 percent of the original capability and outperforms the comparison method, SigLeak, by nearly fourfold. → Techpresso

Synthszr Take: A skill for which a provider has spent months compiling domain knowledge can be replicated for the equivalent cost of 32 API calls, and the customer doesn’t have to do anything that violates the terms of service. This hits any business model that sells secrecy as a moat: The lock protects the file, not the capability. 86.8 percent functional coverage is sufficient for the vast majority of customers, especially since the replica is installable and costs nothing to run afterward.

Open-source agents: DeepSeek Harness, OpenClaw, and Hermes compared

The MyClaw Newsletter compares three current open-source agent projects, categorizing them by their respective design goals. According to the article, DeepSeek Harness is the most infrastructure-oriented, providing developers with Agent-Harness building blocks to assemble their own reasoning, tool calls, context management, and workflow coordination. OpenClaw, in contrast, takes the approach of a perpetually present agent that remains accessible across desktop, smartphone, Telegram, Discord, and Slack, remembers past interactions, and executes scheduled tasks even when not actively in use. Hermes Agent is described as an autonomous executor: You provide a goal, and the agent plans, uses tools, manages and delegates subtasks, and works through multiple steps toward a result. The newsletter sums up the differences in a formula: DeepSeek provides the machine, OpenClaw gives it an identity, and Hermes gives it a task. → MyClaw Newsletter

Synthszr Take: Behind these three projects are three different assumptions about what is actually scarce when it comes to agents. DeepSeek says: the architecture, so you get building blocks and the operational load to go with them. OpenClaw says: the memory, hence the agent that lives on across Slack, Telegram, and desktop, and works through its tasks at night. Hermes says: the delegation, i.e., input a goal, then planning and execution without further questions, which in practice is precisely the point where a verification step is missing and trust erodes. Realistically, by 2027, all three bets will be in the same stack: building blocks from DeepSeek, persistence from OpenClaw, and execution logic from Hermes.

AI governance and open model weights are now a boardroom topic

PwC is observing that the choice between closed and open model weights is now being discussed at the board level, along with the issue of AI governance. Morning Brew interviewed Jenny Koehler, COO of Advisory at PwC, on this topic; the article is marked as sponsored content, so the assessment comes from the consulting firm itself. Koehler describes the shift as a market development: topics that were previously confined to specialized departments and research discussions are now appearing in leadership conversations. With closed-weight models, the trained parameters remain with the provider and are only made accessible via an interface, whereas with open weights, they can be downloaded, customized, and run on one’s own infrastructure. → Morning Brew

Synthszr Take: Koehler’s placement of this issue in the boardroom has an unromantic reason: model weights determine who is liable in the event of damages. A closed model means a contract, a service level, and a counterpart you can hold accountable; open weights in your own data center mean auditing duties, a patch cycle, and the burden of proof in-house. This is a budget and risk decision, and it can no longer be passed down to IT as soon as regulators start asking for documented accountability.

South Korea to build free AI for all citizens

South Korea has selected three consortia, led by SK Telecom, KT, and Kakao, to build a free basic AI service with no usage limits for every citizen. The government will provide the groups with a total of 512 Nvidia B200 GPUs over the course of 2026 and plans to cover a portion of the nationwide operating costs starting in 2027, while the companies must contribute their own funds and revenue sources. The three services are tailored differently: SK Telecom is focusing on phone and SMS in addition to app and web access for people facing digital hurdles; KT is integrating access into the Daum search engine and partner platforms; and Kakao is making KakaoTalk the main entry point and plans to develop specialized agents for senior care and taxes. Rules from July require that at least 50 percent of model usage must come from Korean foundation models, with at least 30 percent of that from domestic developers other than the consortium leader. Foreign models may cover limited functions but will not receive government funding for doing so. → Techpresso

Synthszr Take: A free service with no usage limit for an entire country, and the bill for it has yet to be written. The 18.4 million daily users figure is based on the assumption of four questions per person and a 500-token response length—roughly the behavior people exhibit in the first week and then never again. As soon as someone sends a government application with attachments through a Kakao agent that is supposed to make reservations, file applications, and process payments, every single one of these assumptions will be overturned.

Security researchers find three spyware implants in the firmware of Chinese routers

Security researchers have discovered three different backdoor-like implants in Chinese-made routers sold worldwide, reports Tom’s Hardware. According to the researchers, the implants are embedded in the devices' firmware, the layer that boots before the operating system and which a user never sees during normal operation. The affected devices are ones that have reached various markets through regular retail channels. The short report does not specify which manufacturers, models, or numbers of units are involved, nor who planted the implants. → Tom’s Hardware

Synthszr Take: Three different implants in a single piece of firmware sounds intentional, and they are located exactly where every encryption strategy fails: below the operating system, in the device’s factory state. For years, security budgets have been flowing into the software layer, into access rights and endpoint agents, while the piece of network hardware in the server room was procured through a tender where the unit price was the deciding factor. This has turned procurement into a security discipline, and questions about the manufacturing location, firmware origin, and signed updates belong in the tender documents, not in the follow-up meeting with IT.

OpenAI locks out Cursor: Model access ends November 12, Musk says 'I couldn’t care less'

On August 28, OpenAI announced that it would terminate its contract to supply the AI coding tool Cursor with its models; the company has set the shutdown date for November 12, 2026. The move was triggered by a change of ownership: SpaceX had acquired Cursor’s developer, Anysphere, on August 14 for $60 billion, after the corporation had already integrated X and xAI in February and gone public in June. As justification, OpenAI stated that it could not trust SpaceX to use the technology within its terms of service, based on experiences with contract breaches by Musk’s companies. It referenced the course of the Twitter acquisition as well as sworn testimony alleging that xAI violated the terms of service through the distillation of OpenAI models. Future models, including the anticipated Astra, will no longer be made available to Cursor during the phase-out period.

Cursor co-founder and CEO Michael Truell wrote on X that OpenAI models handle about five percent of Cursor’s user traffic and that they are in talks with the OpenAI team to resolve the matter. He said Cursor was one of OpenAI’s very first customers and had trusted the platform to be neutral infrastructure for its business. Musk commented on the incident with 'I couldn’t care less' and described Sam Altman and Greg Brockman as untrustworthy. It remains unclear how existing Cursor subscriptions will be affected and which specific models will be discontinued after the cutoff date.

Anthropic is filling the gap: Co-founder Tom Brown confirmed additional computing capacity to support Claude models on Cursor. According to data from Menlo Ventures, Anthropic now accounts for 40 percent of corporate spending on AI models, compared to 27 percent for OpenAI. In the second quarter of 2026, Anthropic’s revenue reached $11.5 billion, surpassing OpenAI’s $6.7 billion for the first time, with about $8 billion attributed to Claude Code. In parallel, the 'Claudeforce' partnership with Salesforce is making Claude the platform’s default reasoning engine; an IPO is targeted for October 2026.

The move comes at a time when OpenAI is narrowing its model policy anyway: the o3 model has been discontinued, and reinforcement learning training for Astra was paused on August 7 after a cybersecurity threshold classified as 'critical' was reached. The ongoing monitoring of Astra is reportedly tying up 20 percent of its inference compute power. → Business Today, CNBC, Forkast, StrictlyVC, Blockonomi

Synthszr Take: A five percent traffic share is the number that is saving Cursor right now: four years of partnership with OpenAI, and in the end, only one-twentieth of its usage depends on the provider that is now pulling the plug. This ratio is the result of routing through multiple providers, using its own models, and having a second pillar with Anthropic, which is now providing additional compute capacity. For most coding startups that show up with just an API key and a pretty editor, their product would come to a standstill on November 12. The termination was due to a $60 billion change of ownership—reasons like these arise in the capital markets, far removed from any roadmap. A second provider, productively integrated and tested weekly against your own eval suite, costs a few weeks of work and is cheaper than a three-month notice period with churning customers.