Its starting ingredients include sodium sulfate and ferrous sulfate, two chemicals produced as by-products of other industrial processes. Ferrous sulfate is an iron salt already used to treat drinking water and wastewater. It is also left over from making titanium dioxide, the bright white pigment used in paint. Pigment manufacturers remove iron from their ore, leaving an iron-rich by-product that can become a battery ingredient. Zoolnasm uses these ingredients to make sodium iron sulfate, known as NFS, for a battery’s positive electrode, the cathode.

Sodium batteries generally store less energy for their weight than lithium-ion batteries but perform better in the cold. Their raw materials are abundant and could lower costs, although making a finished cell cheaply is still hard. Those traits suit short-range vehicles, backup power and stationary storage, where a battery’s weight counts for less than in a long-range car.

The company sells this cathode material to other battery makers and also offers its own batteries. A cell is the individual unit that stores energy; multiple cells are assembled into a battery pack. Zoolnasm makes some cells itself and works with established cell manufacturers for larger-scale production. Its finished products include e-bike packs, backup batteries and storage systems. It lists 0.5 GWh of cell capacity against a 5 GWh target, with cell-line expansion the first stated use of the new financing.

The latest investors and valuation are undisclosed. Zoolnasm says it has raised more than ¥1B (about $149M) cumulatively. Earlier backers include energy-storage system supplier Pylontech (派能科技, 688063.SH), which agreed to invest ¥30M (about $4.5M) in 2023 at a ¥1.5B (about $224M) pre-money valuation.

Whether batteries made from industrial leftovers pay depends on what they cost to make and whether buyers will pay more for longer life and cold-weather performance. Zoolnasm is pursuing longer-lived e-bike packs and cold-weather backup power, while competing with cheap lead-acid batteries and lithium batteries whose raw-material price has fallen.

Why this chemistry?

Sodium batteries can be made with different cathode materials. Zoolnasm’s NFS uses sulfate, while the more common NFPP uses phosphate. Metals data provider SMM estimates that NFPP accounted for 81% of China’s sodium cathode output in December 2025. Cathode supplier JNION (珈钠能源), which specializes in NFPP, said in May that its annual capacity was already in the tens of thousands of tonnes. Zoolnasm lists 13,000 tonnes of NFS capacity, with a target of 30,000.

Read more

Founded in January 2024, Sand.ai makes models that generate video with audio from a text prompt, or from a prompt plus a still image. As of writing, its MAGI-2 Preview ranks 10th overall and second among open-weight models on Artificial Analysis’s image-to-video leaderboard. MiniMax H3 leads the open-weight models. Chinese developers including Alibaba, ByteDance and MiniMax feature prominently near the top of the video-model rankings.

Sand.ai founder Cao Yue attributes China’s strength to timing and its short-video market. In video, he argues, Chinese teams started at roughly the same time as their overseas rivals, so they had less catching up to do than in language models. He also credits China’s more active short-video ecosystem with getting the models into use faster.

Alongside its research, Sand.ai runs VidMuse, a paid video-creation service with tools for music videos, product ads and anime dramas. The service offers access to competing companies’ models. Wangsu’s stock-exchange announcement of the proposed investment includes Sand.ai’s financial results.

Ten Seconds, Eight GPUs

Sand.ai’s profile of Cao says he is an author of Swin Transformer, the Microsoft Research Asia paper that won the best-paper award at the 2021 International Conference on Computer Vision.

The MAGI-2 research preview produces ten-second clips with sound. Sand.ai has released its weights and generation code under Apache 2.0, so developers can download the model and run it on their own infrastructure.

The published setup requires eight NVIDIA Hopper GPUs, and the model’s weights total roughly 307 GB.

As of writing, MAGI-2 ranks 21st on Artificial Analysis’s text-to-video leaderboard. Both rankings cover video with audio and rank models by pairwise human preference votes on generated clips.

Sand.ai has also changed the approach we covered in June. MAGI-1 generated video in successive chunks, using earlier footage to guide what came next and extend the video. MAGI-2 did not directly carry that method over. Sand.ai says retaining it would complicate the combination of joint audio-video generation and distributed training at the 100-billion-parameter scale.

Joint audio-video generation is already shared by models including ByteDance’s Seedance 2.0 and MiniMax H3. H3 also runs video and audio through a single-stream architecture. H3 and MAGI-2 take different approaches to scaling: H3 uses a dense 33-billion-parameter model, while MAGI-2 uses a mixture of experts with 114 billion parameters, activating about 6 billion per token. This lets Sand.ai increase model capacity while limiting how much of it is used for each calculation. The company says the design explores an efficient path to scaling video generation.

Why Offer Other Companies’ Models?

In June, Yicai identified VidMuse as Sand.ai’s core revenue product.

VidMuse’s Pro plan costs $33 a month when billed annually and includes 4,000 monthly credits for generating and analyzing media. How far those credits go depends on the model, video length, resolution and settings; subscribers can buy more.

Read more

Z-TRAK’s proposed rocket, Fission-A (裂变号-A型), is designed to carry one or more satellites with a combined weight of up to 450 kg into low Earth orbit. The first-stage booster powers the initial ascent, then separates; the upper stage carries the satellites the rest of the way into orbit. Z-TRAK plans to land the booster and fly it again, spreading the expense of building it across multiple launches.

A satellite owner can buy space alongside other customers on a shared launch, or book a whole launch around its own mission. Sharing spreads the cost among customers, but the satellite has to fit the flight’s orbit and timetable. A dedicated launch gives the customer more control over both, while leaving it to cover the launch bill. Z-TRAK says rideshare customers wait an average of 6.8 months; it wants to make booking a whole launch a more affordable alternative.

The amount a rocket can carry depends on the orbit it needs to reach. Fission-A’s planned maximum is 450 kg to low Earth orbit, falling to 300 kg for a sun-synchronous orbit. That is a type of low Earth orbit used by imaging satellites to revisit a location under consistent sunlight, making changes on the ground easier to compare.

For a whole launch carrying up to 450 kg to low Earth orbit, Z-TRAK targets a cost below ¥15M ($2.24M). SpaceX’s published rideshare tariff charges ¥14.1M ($2.1M) for a customer to send 300 kg to sun-synchronous orbit alongside other customers’ satellites.

SpaceX already reuses Falcon 9 boosters on both dedicated and shared launches; the upper stage is used once. Z-TRAK is trying to bring booster reuse to a much smaller rocket, so a small-satellite customer could afford the whole launch rather than a share of a larger one.

Zhang says he grew up building rocket models and wrote the business plan with classmates in a Shenzhen co-working space in October 2024. He says he is still studying aerospace engineering at Hong Kong Polytechnic University. The venture has since moved to Zhuzhou, Hunan, through a local government program that uses investment to attract businesses.

On September 29, 2026, Z-TRAK announced a “¥100M-level” (~$14.9M) angel-plus round, its latest round of early-stage funding. Investors bought about 10% of its Hunan operating company at a reported valuation of ¥1B ($149M) after the investment, up from ¥300M ($44.7M) in May. Returning investor Zhuzhou Beidou Group (北斗集团), a local state-owned investment company, is helping establish Z-TRAK’s development and manufacturing base. The team has tested an engine on the ground; the rocket has yet to fly.

Batteries behind the burn

Z-TRAK uses batteries to power the pumps that feed methane and liquid oxygen into its engine’s combustion chamber. The methane burns to produce thrust; battery-powered electric motors drive the pumps, replacing the gas turbine used in many liquid-fueled rocket engines. This simplifies the machinery and makes thrust easier to adjust, which matters when slowing the booster for landing. Zhang said in a July 2026 interview that the team built its first engine prototype for under ¥500,000 ($74,600).

Read more

This was its sixth financing since incorporation in December 2025. Its fundraising adviser says investors have put more than ¥1B (~$149M) into the company in total. Its valuation rose from ¥3B (~$446M) on June 11 to ¥5B on July 31 and ¥6B in the latest round, doubling in about three months.

Mind over motor

The round was announced on September 17, the day LiberAI’s Liber-0 Preview model took first place on RoboDojo, which tests AI models controlling robot arms on 42 simulated tasks. A day earlier, RoboDojo’s team had added OpenAI’s GPT-6 Astra to the board. Tested without additional task-specific training, Astra achieved a 22.5% average success rate, putting it above every public entry at the time. Liber-0 Preview reached 25.5%.

Read more

Chen Tianrun (陈天润), born in 2000, is a Zhejiang University PhD student under Pan Yunhe (潘云鹤), a senior figure in Chinese AI research. He registered KOKONI in 2021 and the next year began selling a consumer 3D printer with an app that turned photos into printable models; kokoni.ltd still opens on the printer, its copyright notice stopped at 2024, and nobody has touched the old business since.

Then investors began pouring money into world models, AI that predicts how a scene will change, and Chen had something most model labs did not: a library of 3D objects and the tools to make more of them quickly. By his account KOKONI used that library as its training data: assemble the objects into scenes that move and you have the raw material for a world model, produced in house rather than scraped. The company pivoted to spatial intelligence in 2024. Customers so far, by the same account, are a few embodied-AI companies, digital factories, and surveyors who now turn three photographs of a site into a 3D map in minutes, work that used to take days with dedicated sensors.

Read more

The low altitude economy

China’s 低空经济 (”low-altitude economy”) is the government’s umbrella for all civilian use of low-altitude airspace, from eVTOL passenger aircraft and delivery drones to power line inspection, agricultural spraying, and emergency response.

Chinese megacities are running out of…

Why build a company around one joint? Because the hand is where a humanoid does its job. Legs get the robot to the work; the fingers do the work, and packing real dexterity into a human-sized hand is one of the hardest parts of the whole machine. Every joint needs a reducer (the small gearbox that turns a motor’s fast spin into a joint’s slow, controlled movement), and every reducer costs space, money and power.

One that is smaller, cheaper and self-locking is much more than a spec-sheet nicety. Without it, a joint either burns motor power to hold a grip, which drains the battery over a shift, or needs a separate brake bolted on, another part, more weight, one more thing to fail. A self-locking joint holds that grip for free, whether it is clamped on a tool for a minute of fine assembly work or carrying a box across the floor, and it lets a hand pack more independent fingers into the same space. The early buyers point to where that lands first: the dexterous-hand makers themselves, and electronics assemblers that want robots doing fine-motor work on the line.

Global humanoid robot shipments reached about 13,000 units in 2025, per market researcher Omdia, led by Chinese manufacturers, and Morgan Stanley expects China alone to ship 50,000 this year. Each one needs dozens of reducers. The dominant technology is the harmonic reducer, a coaxial gear (input and output share the same axis) that sits inline between motor and joint. One listed Chinese player, Leaderdrive (绿的谐波, 688017), shipped 425,000 harmonic reducers in 2025 and, per a 2026 disclosure, its share of the harmonic reducers going into Chinese humanoid robots had risen to 80-90%, up from around 70% in 2025.

The coaxial design’s structural constraint means power moves in a straight line, not around a corner. For joints that need a 90° turn, a wrist rotating perpendicular to the forearm or a finger bending off the palm, a harmonic reducer cannot do the job. Among reducer-based designs, the current workaround, the second…