Sand.ai raises $200 million Series A+ for video AI | Magica
Sand.ai raises $200 million as Wangsu backs its video AI deployment
Editorial Team
••📖7 min read
Sand.ai says its $200 million Series A+ round lifts its valuation above $1 billion. Co-lead investor Wangsu plans inference and deployment cooperation, while its filing shows substantial losses behind the startup’s video AI ambitions.
Sand.ai announced a $200 million Series A+ round, its third financing of 2026.
Co-lead investor Wangsu has disclosed a planned RMB300 million investment and cooperation on inference and deployment.
Wangsu’s filing shows substantial group losses; reported video-generation costs still lack a reproducible competitor comparison.
Sand.ai, a Chinese developer of video-generation models and creator software, announced on November 5 that it had raised $200 million in Series A+ funding. The company says the round takes its valuation above $1 billion and brings its fundraising over the preceding 10 months to $300 million.
CoStone Capital, a Chinese investment firm, and Wangsu Science & Technology, a Shenzhen-listed IT infrastructure provider, jointly led the round. State-backed investors including Pudong Venture Capital and Shanghai Guofang joined, alongside other funds and returning shareholders. The company disclosed the financing totals and unicorn valuation in the funding announcement coverage; the valuation is a company-supplied figure.
The money is intended for video foundation-model research, training and inference infrastructure, product development and overseas expansion. Wangsu’s established content delivery network business distributes internet content to users. Its investment connects Sand.ai’s spending plan to deployment infrastructure, while offering Wangsu a route to expand its AI services business, as described in the analysis of the transaction.
Wangsu’s commitment predates the broader closing. Its September 29 exchange filing proposed an RMB300 million investment, approximately $44.5 million, for Series A+ preferred shares. As an interim arrangement, Wangsu would provide a convertible loan to Sand.ai’s Shanghai subsidiary and receive warrants for shares in its Cayman holding company. Wangsu put its prospective stake at about 4.4%, assuming all investors in the round completed their subscriptions. An October 19 update said the agreements had been signed on October 16, according to a .
Internet scans and leaked databases expose the spread of ARTEX, a security-testing framework used in cyberattacks. Researchers documented largely autonomous compromises, but the records also show failed cashouts and uncertain counts of independent deployments.
Wangsu said in the disclosures, as recounted in that report, that the companies intended to cooperate on audio and video inference, low-latency edge inference and enterprise deployment. Sand.ai's new models and related technologies would receive priority deployment through Wangsu's infrastructure, subject to the parties' agreements. These are disclosed cooperation terms; they do not establish that the partnership has already reduced serving costs or generated customer revenue.
A research pedigree and a paid application
Founded in January 2024, Sand.ai has offices in Beijing and Shanghai. Founder and company head Cao Yue is a Tsinghua computer-science PhD, a former principal researcher at Microsoft Research Asia and a co-author of the Swin Transformer computer-vision architecture. He also headed the Vision Model Center at the Beijing Academy of Artificial Intelligence, according to the financing analysis. He previously co-founded Light Year with Meituan co-founder Wang Huiwen in February 2023. Meituan acquired that AI company four months later for $281 million, according to the financing report. His record connects computer-vision research with experience building an AI startup.
Sand.ai's model line includes MAGI-1, which generates video in successive temporal chunks, and GAGA-1, which generates audio and video together. Its commercial application, VidMuse, is a paid creation service for music videos, product advertisements and animated content. Earlier investors included Wang’s family office Lollapalooza Capital, IDG Capital and Baidu Ventures. Existing shareholders including iMPC, formerly Matrix Partners China, and CAS Star invested again in the latest round, the financing report says, citing Sand.ai’s announcement.
The company says VidMuse exceeded $10 million in annual recurring revenue within two months of its January launch. That is an annualized revenue measure, not $10 million of receipts over those two months. The report on its commercial progress notes that the figure has not been independently audited; the available accounts do not establish customer retention or product margins.
There is a separate financial baseline. Wangsu’s September filing reports that Sand.ai’s Cayman holding-company group had revenue of RMB17.4 million and a net loss of RMB151.6 million in the first half of 2026. Both figures are unaudited. They cover the consolidated group over six months and cannot be equated with VidMuse’s annualized revenue claim. The group remained unprofitable over that period.
Sand.ai is pursuing model distribution and its own application together. The financing analysis recounts Cao’s argument that selling APIs alone is difficult in video models and that models and products need to develop together. VidMuse provides a direct route to paying creators and their feedback. It also means Sand.ai operates its own creator product while seeking to supply other content businesses, a qualification to the financing analysis’s description of it as a neutral supplier.
Company-reported, unaudited first-half group revenue and net loss, disclosed in Wangsu’s filing. Units are ten-thousand yuan. Chart: Magica. Source: Wangsu investment announcement.
What the efficiency claims establish
Sand.ai released and open-sourced MAGI-2 Preview in August. The analysis accompanying the financing describes a mixture-of-experts model with 114 billion total parameters. Sand.ai’s technical release specifies that about 6 billion are activated per token. The financing analysis also reports an inference cost of roughly RMB0.5 for a distilled model to generate a 10-second, 1080p video, characterizing that as about one-tenth of mainstream models' per-second cost.
That comparison needs qualification. The account does not provide a common hardware configuration or a named, reproducible set of competing models for the cost comparison. An inference-cost estimate also does not establish the price charged to customers or the margin earned on a finished video.
Sand.ai's own technical release supplies a further limit. Mixture-of-experts routing activates only part of a model for each token, but active parameter counts do not equal total computation, and attention still carries a cost. The company calls Preview an intermediate research release, with fuller scaling curves and controlled experiments still priorities.
Sand.ai says MAGI-2 did not directly carry over MAGI-1’s autoregressive chunking. Instead, it uses one Transformer backbone to process text, video and audio together. The release describes changes to architecture, training systems and data annotation, including retaining more complex motion, multi-person interactions and audio relationships. Sand.ai says some resulting capabilities are qualitative observations requiring controlled experiments.
The infrastructure burden remains substantial. Sand.ai describes communication between devices, uneven expert workloads and memory demands as potential offsets to sparse computation. Its training system distributes work across GPUs and uses different network arrangements within and between computing nodes. The claimed efficiency depends on this systems engineering as well as the model architecture.
Sand.ai’s company-reported architecture diagram contrasts conventional MoE with Multi-Head LatentMoE routing; it does not measure cost savings. Source: Sand.ai research release.
The full release must turn claims into comparisons
Sand.ai plans to release the full MAGI-2 in the near term, with a larger parameter count, according to the financing analysis. It faces continuing iteration from Kuaishou’s Kling and ByteDance’s Seedance, competing video-generation models. The contest encompasses controllability, inference efficiency and commercial uses, rather than parameter counts alone.
A rival has attracted substantially more capital: the same financing report says Kling's unit completed an approximately RMB19.8 billion financing in October while Kuaishou retained control. That shows the scale of capital available to a rival, though funding totals do not compare model quality or operating efficiency. Sand.ai's open releases and new capital alone do not settle that contest.
The next useful evidence would connect model quality to deployment economics: reproducible generation costs at a stated hardware configuration, comparable quality and latency tests, and evidence that creator revenue persists after launch. The financing gives Sand.ai resources and an infrastructure relationship to pursue those results. The disclosed figures do not yet show whether its claimed efficiency translates into a durable commercial advantage.
Meta's strategy chief and former CFO is joining Fireworks AI ahead of a reported fundraising push targeting at least a $30 billion valuation. The inference company announced a $17.5 billion valuation in July, making the hire part of a much larger expansion than its older funding figures suggest.
An advocacy group estimates Interior Secretary Doug Burgum’s AI holdings at $18.1 million–$64.4 million as his department advances data centers on public land. His ethics agreement allows retained investments with recusals, while pending permits will determine both the application of those safeguards and the public’s return. Interior denies his decisions are driven by personal investments.
Microsoft is evaluating fewer neocloud partners while expanding its own infrastructure. Delivery-dependent contracts underpin supplier financing, while IREN reports rising revenue alongside a $6 billion funding target for its broader expansion.
Physical Superintelligence says its AI system found a proof of a conjecture limiting quantum query advantages on random inputs. Public proof files enable scrutiny, but do not establish the private system’s cost or repeatability.
OpenAI is expanding its relationship with Hacktron, whose researchers previously reported an attack chain into employee accounts, to test third-party connections with greater access to cybersecurity models.
Google has acquired Crusoe’s southeast Wyoming data center site, where it plans a long-term $17 billion investment. Crusoe describes completed pre-development work; independent reporting points to customer setbacks. Power delivery and carbon capture remain key tests.
The team behind document-parsing service SoMark has released OXR-1.0 code and weights. A benchmark retest supports a strong result, but speed and accuracy were measured separately, and users above funding or revenue thresholds need a commercial license.
OpenAI is reportedly developing a Business Hub with daily mention and click charts, product updates and business agents. The unannounced feature could give verified businesses direct visibility into ChatGPT responses, without revealing prompts or competitor rankings.
Palantir and Nvidia have specified an integrated stack for running AI in customer-controlled environments. The architecture builds on existing software and infrastructure offerings, with published configurations spanning 32 to 640 GPUs, while pricing and workload economics remain unproven in the launch materials.
Democrats’ incoming congressional majorities put hearings and investigations within reach, but AI rules and crypto legislation still face bipartisan bargaining and a Republican president.
Microsoft reportedly declined to renew Oracle server rentals as older chips became less useful to it. The decision comes despite continuing capacity shortages, while Oracle says demand for earlier-generation GPUs remains strong.
An investigation puts Nvidia’s future commitments and guarantees at $530.5 billion. Its disclosures show rapidly expanding supplier commitments and phased support for customers whose infrastructure could generate further chip sales, with cash payments depending on timing, demand and contract terms.
TSMC is reportedly planning a possible move into 3D DRAM at Nvidia’s request. The foundry already helps integrate AI processors with memory, but a manufacturing expansion has no disclosed factory or production date.
Physicl has raised a $10 million seed round and introduced Studio, a beta app for generating targeted physical AI data. Its reported customer gains concern 3D reconstruction, leaving broader robot reliability and commercial economics to be demonstrated.
New cloud partners in Southeast Asia must enroll in Nvidia’s Fleet Intelligence system, according to a report, linking preferred chip access to monitoring while leaving data sharing and existing partners’ obligations unresolved.
Airbnb sold intellectual property to a private AI lab controlled by CEO Brian Chesky in a $7.1 million arrangement that also reimburses expenses. It obtained a license to certain lab technology and future investment rights, but disclosed no asset valuation or detailed license terms.
Chip shares recovered part of Thursday’s losses in early Friday trading. Fears about data-centre construction face a test against AI-chip supplier Nvidia’s strong reported sales, growth forecast and conditional infrastructure-financing plans.
SoftBank shares fell 10.8% as default-insurance costs reached a four-year high. Delayed listings and unfinished data centers complicate its AI investment outlook, while a $40 billion bridge facility has a March 2027 maturity and a repayment plan based on existing assets and other financing.
U.S. prosecutors allege a global fraud combined AI impersonation, sham confidential deals and cryptocurrency laundering. A blocked bank transfer shows why convincing an employee and moving the money are separate hurdles.
