Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Ant Group's Robbyant has released code and six-billion-parameter weights for LingBot-VLA 2.0 after pretraining on data from 20 robot configurations. The release broadens its whole-body control interface, but Robbyant's published evaluations cover four platforms, require careful reading of partial-progress scores, and have not been independently reproduced.
Ant Group's robotics unit has made a substantial robot-control model available for others to test. What it has not yet established is that one downloadable foundation can be deployed economically and reliably across the 20 kinds of hardware represented in its data.
Robbyant announced LingBot-VLA 2.0 on July 8 as an upgrade to the model it opened in January. A release-day brief relayed the company's headline: 60,000 hours of pretraining data spanning more than 20 robot configurations and 17 brands, with added support for heads, waists, end effectors and mobile bases.
The composition matters. The team's technical paper separates the total into 50,000 hours of robot trajectories across 20 configurations and 10,000 hours of egocentric human video. A report reproducing Robbyant's detailed account says those sets were filtered from raw pools of 90,000 robot hours and 20,000 human-video hours. The filtering strengthens the data-quality claim, but the combined 60,000-hour figure should not be read as 60,000 hours of robot operation.
The more revealing comparison is with Robbyant's own first version. Its predecessor paper described about 20,000 hours of real-world data from nine dual-arm configurations. Version 2.0 therefore raises the robot-data total by about 2.5 times, while the addition of human video makes the overall headline three times larger. It also moves beyond a dual-arm-only action space.
Robbyant maps heterogeneous hardware into a shared 55-dimensional state-and-action vector: 14 dimensions each for arm joints and end-effector poses, 12 for hand joints, four for the waist, three for mobility, two each for grippers and the head, plus four reserves. Its project description pairs that representation with sparse mixture-of-experts layers and a future-prediction objective distilled from depth and video models.
That design is a mechanism for sharing a foundation, not evidence of zero-effort portability. The released workflow requires developers to map a robot's features into the common representation, compute normalization statistics and post-train on a downstream dataset. Training also depends on separate Qwen3-VL, depth and DINO-Video checkpoints. The repository documents the work; it does not eliminate it.
For the GM-100 bimanual benchmark, Robbyant says all nine tasks were jointly trained in a generalist policy rather than handled by task-specific specialist policies. Its released benchmark tables report two measures—task-progress score and full success:
| Platform | LingBot-VLA 2.0 | π0.5 | GR00T N1.7 | LingBot-VLA 1.0 |
|---|---|---|---|---|
| AgileX Cobot Magic | 66.2 / 34.4 | 59.1 / 32.2 | 36.3 / 17.8 | 58.2 / 30.0 |
| Galaxea R1 Pro | 34.6 / 15.6 | 27.4 / 8.9 | 16.4 / 5.6 | 32.7 / 15.6 |
The lead is real within Robbyant's table, but its size depends on the metric. LingBot's AgileX success rate is only 2.2 points above π0.5. On Galaxea, version 2.0 improves its predecessor's progress score but ties its 15.6 success rate.
In the mobile evaluations, progress is not a simple count of completed trials. Robbyant's account says long operations are divided into weighted substeps, with credit for stages such as moving, grasping, opening, placing and cleaning. A high progress score can therefore coexist with many failed end-to-end trials.
That gap is most visible in the mobile tests. On Astribot S1 refrigerator sorting, LingBot fell from 77.1 progress / 60.0 success in-domain to 37.0 / 13.3 out of domain. On the Cobot Magic-ARX X5 stove-cleaning task, it moved from 84.3 / 66.7 to 67.5 / 40.0. LingBot exceeded π0.5 in all four comparisons reported by the team, including out of domain, but the absolute completion rates still limit the company's “universal brain” framing.
These are company-run evaluations, not a neutral bake-off across the 20 training configurations. A separate analysis of the release found no published third-party reproduction and likewise highlighted the progress-to-success gap. Open weights make reproduction possible; they do not supply it.
Apache-licensed code and weights let a lab or robot maker inspect and adapt LingBot without negotiating private model access. That is useful distribution leverage for Robbyant as it tries to persuade hardware makers and data partners to adopt its representation.
It is not a unique advantage. Nvidia released GR00T 1.7 under Apache 2.0 on July 7 and describes it as a three-billion-parameter, cross-embodiment VLA model within an open modular workflow that runs from teleoperation and simulation to evaluation and deployment. Nvidia says in its technical overview that GR00T used about 32,000 hours of real demonstrations and egocentric human data plus 8,000 hours of simulated rollouts and demonstrations.
Those totals cannot be ranked cleanly against LingBot's 50,000 robot hours plus 10,000 human-video hours, and Nvidia reports results on different benchmark suites. Robbyant's own GM-100 test favors LingBot over GR00T N1.7; Nvidia's integrated data, simulation and Jetson deployment stack offers a different form of control over the developer pipeline.
Robbyant has Ant's resources behind it. A report on the unit's open-source program says Ant reported record research spending of $5.17 billion for 2025. No retained source allocates that sum to robotics, so it indicates the parent's capacity, not LingBot's budget or a financing round.
Robbyant says in its announcement that pilots with Leju, Ti5 Robot, GuoDa Drugstore and Longsheng Technology cover retail sorting, logistics and industrial automation. It has not published pilot outcomes, pricing or task-level reliability.
The only concrete inference measure in the release is about 130 milliseconds for one call on an Nvidia GeForce RTX 4090D with 10 denoising steps. That is a reproducible engineering target, not deployment economics. The current record does not disclose accelerator memory use, power draw, hardware cost at fleet scale, the amount of task-specific post-training data, or the labor needed to integrate a new body.
The next decision belongs to robot makers and operating customers: compare the released checkpoint with π0.5, GR00T and specialist policies on the same unseen tasks, cameras and hardware; report full-trial completion as well as partial progress; and publish the cost of adaptation and on-device operation. Until those results exist, LingBot-VLA 2.0 is best understood as a broader, testable foundation—not proof of one universal robot brain.
Anthropic’s 20-year lease gives TeraWulf a customer for 401 megawatts of future AI infrastructure, but rent starts only after delivery and the proposed utility deal passes power and grid costs to TeraWulf, leaving financing and project margin unresolved.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Twenty-nine countries signed an agreement creating WAICO as an independent intergovernmental organization, while China paired the launch with capacity-building offers that are not yet confirmed as WAICO programs.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
Nebius has arranged its first senior secured facility against an operating GPU deployment and one customer’s cash flows. The deal adds project-level debt to its expansion toolkit, but an unnamed customer and a larger rival financing limit what it proves about the rest of Nebius’s backlog.
Huawei publicly displayed a 16-cabinet Atlas 950 configuration rated at 1 EFLOPS in FP8, providing tangible evidence of its system-scale AI strategy while leaving price, power use and sustained workload performance undisclosed ahead of the full system's planned fourth-quarter release.
UK testing places leading open-weight models four to seven months behind selected closed-model cyber results, yet longer attack chains, U.S. benchmarks and mixed cost comparisons show why that interval is a warning signal rather than a universal capability clock.
Moonshot AI's Kimi K3 added to a global technology selloff with near-frontier performance, but unreleased weights, mixed cost comparisons and a recommended 64-accelerator deployment leave its effect on chip demand unresolved.
Two binding EU decisions require Google to give rival AI services comparable access to 11 Android features and offer eligible search competitors a restricted, anonymized dataset, but phased deadlines, certification, pricing and privacy safeguards leave the competitive effect unproven.
Databricks has signed a term sheet for a Coatue-led financing at a $188 billion valuation, while unidentified sources put the round at $3 billion. The proposed capital would deepen its push into AI governance, data agents and operational databases, but the transaction remains open and the company supplied no new operating figures.
OpenAI says a handful of GPT-5.6 Sol file-deletion reports most commonly involved Full Access without sandboxing or Auto-review. Its own evaluations show a more complicated risk picture, and the company has not yet published an incident rate or evidence that its promised safeguards stop the failure.
Netflix says generative AI workflows were used on roughly 300 titles in 2026, but its only quantified example tied to that disclosure covers 17 minutes and does not establish how much finished material, spending or labor changed across the slate.
A rogue browser extension can trigger Claude for Chrome’s fixed Gmail, Docs, Calendar and business workflows; the attack is constrained by default approvals but can run silently for users who enabled unattended action.
Gemini 3.5 Pro missed its expected June rollout. An anonymous-source account says a late-June data change fell short of Google's coding goals, but Google has confirmed only partner testing—not the reported cause, a new date, or public results and pricing.
Gold Eagle has begun collecting and prioritizing AI-discovered software vulnerabilities, but the voluntary federal clearinghouse has not disclosed results, operating rules or the resources that would turn findings into deployed fixes.
Google is bringing Gemini Omni editing and a reusable face-and-voice avatar to Vids, but the sharper distinction is account-level identity across Vids and Gemini rather than a new category of AI video; Vids already offered Veo generation and customizable avatars, while specialist rivals already sell digital presenters.