Three of China's most closely watched AI ventures — DeepSeek, Moonshot AI, and the agent platform Manus — have all pulled in significant capital, confirming that Chinese AI has moved past the "catching up" phase into one where it sets terms. The funding, the model releases, and the architecture papers all point in the same direction. What stays genuinely open is how chip export controls reshape the capability ceiling over the next two years, and whether Western companies will have durable, compliant access to these tools at scale.
In January 2025, DeepSeek dropped its R1 model on GitHub with a technical report attached. The training cost figure — under $6 million for the full run, by the company's own account — was the number that traveled fastest. Nvidia's stock fell sharply that week. Every Western lab started auditing its compute spend.
That single data point did something useful beyond the market reaction: it shifted the frame. The conventional Western read on Chinese AI had been "big spend, state-backed, catching up." DeepSeek broke that read on two counts. It is not state-backed in the obvious sense — it spins out of High-Flyer, a quantitative hedge fund in Hangzhou, not a Beijing ministry. And it was not catching up; DeepSeek-R1's reasoning benchmarks landed alongside or above GPT-4 class models on several tasks, according to the company's published evaluations and third-party testing circulated that month.
Now DeepSeek, Moonshot AI, and Manus are all back in the funding conversation. The question worth asking isn't "are they raising" — obviously they are — but what the money is for and whether the underlying capability story holds.
DeepSeek is not the company most Western professionals have a mental model for. It doesn't have a splashy consumer app. It doesn't have a familiar VC cap table. Liang Wenfeng, who runs both DeepSeek and High-Flyer, has given almost no Western press interviews. What it has is a research team, a codebase, and a series of models that have been relentlessly benchmarked by people who didn't receive a press kit.
DeepSeek-V3, released in December 2024, and DeepSeek-R1, released in January 2025, are both open-weight. You can download them. Developers in Berlin, São Paulo, and Singapore are running them locally. That open distribution strategy is not altruism — it is a land-grab for developer mindshare, the same move Meta made with Llama 2 and 3.
Specific external round sizes for DeepSeek are unconfirmed in public sources. High-Flyer has historically self-funded operations from trading profits, which makes fundraising harder to track than a conventional VC-backed startup. What is documented is the research output rate: in 2024 and 2025, DeepSeek published more significant architecture papers than most Western-funded labs of comparable headcount.
The hardware constraint is real and openly acknowledged. DeepSeek has stated in research communications that it trained on H800 chips — the export-compliant downgrade from H100 that Nvidia shipped to China before the October 2023 rule tightening. Building future generations under those restrictions requires either domestic chip substitutes or architecture work that extracts more from constrained hardware. DeepSeek has, so far, leaned toward the latter.
Moonshot AI operates from Beijing's Zhongguancun corridor. The company's product is Kimi, an AI assistant built around long-context capability. Where GPT-4 and Claude 3 debated document length in 2024, Kimi was shipping 200,000-token context windows in Chinese before either Western frontier lab matched it at equivalent quality in that language. That is not a minor distinction for the enterprise tasks Moonshot is targeting: legal review, contract analysis, regulatory filings in Chinese.
Moonshot reportedly raised over $1 billion in 2024 with Alibaba among the investors, according to Bloomberg and Reuters reporting. The reported valuation placed the company at approximately $3 billion — one of the highest for a private Chinese AI company outside Baidu's ecosystem.
The Western relevance is API access. Kimi's API is accessible from outside China with some configuration. The terms of service, data residency questions, and long-term API reliability are legitimate concerns — not paranoia, just due diligence that applies to any vendor operating under a different jurisdictional framework.
Manus (sometimes abbreviated as "Manu" in investor shorthand) launched in February 2025 with a demo video that circulated faster than most Chinese AI products ever have in Western tech communities. The demo showed the system executing multi-step tasks autonomously — booking, researching, writing, filing — without handholding at each step. Invite codes were scarce. GitHub discussions appeared within days.
The company behind Manus is Monica, based in Hangzhou. A funding round was reported in early 2025, but the size and lead investor remain unconfirmed in public sources. What is verifiable: the product exists, the waitlist was real, and the underlying model stack reportedly draws on frontier models including API calls to both Chinese and Western providers. That makes Manus less a foundation model company and more an agent orchestration layer on top of available models — which is a different kind of bet, more exposed to underlying model commoditisation and to policy risk around cross-jurisdiction API access.
That distinction matters for how you evaluate it. Moonshot and DeepSeek are building at the foundation layer. Manus is building above it.
| Company | Type | Key product | Backing | Access model | Western viability |
|---|---|---|---|---|---|
| DeepSeek | Foundation model lab | DeepSeek-R1, V3 | High-Flyer Capital (quant fund) | Open-weight (downloadable) | High — no API dependency required |
| Moonshot AI | Foundation model + app | Kimi assistant + API | Alibaba + institutional VC | API (some geo-config required) | Moderate — terms and data residency questions |
| Manus | Agent orchestration | Manus platform | Monica (round size unconfirmed) | Invite/waitlist at launch | Lower — access and jurisdiction still unclear |
Three things are operationally straightforward.
DeepSeek's open-weight models are a real cost option. If your application is not fine-tuned on proprietary restricted data and you're running on AWS, GCP, or self-managed infrastructure, deploying DeepSeek-R1 or V3 is a legitimate engineering choice. The relevant questions are performance-per-dollar on your specific task, your compliance requirements, and your jurisdiction. Those are normal vendor evaluation questions.
Moonshot AI's long-context capability in Chinese is a genuine competitive advantage for teams doing China-market work — legal, financial, market research. The alternative is paying OpenAI or Anthropic API prices for a language task where Kimi's architecture was explicitly designed to excel. Price out that trade-off.
Manus is interesting but premature for production use. The agent paradigm has real utility; the specific product is early, access is constrained, and the model dependency makes it a moving target. File it under "worth evaluating in six months."
Don't use Chinese AI APIs for data-sensitive work without a legal review. This is not a geopolitical position — it is the same advice that applies to any vendor operating under a different jurisdictional framework. GDPR, HIPAA, and sector-specific compliance requirements turn "where does my data go during processing" into a legal question, not a preference.
Don't assume benchmark scores transfer to your use case. DeepSeek-R1 performs competitively on coding and reasoning benchmarks. If your application is not a coding or reasoning benchmark, run your own evaluation. This caveat applies equally to Claude, GPT-4o, and Gemini.
Don't skip the shelf-life question. If your product depends on API access to a Chinese model, the policy environment in 12 to 18 months is genuinely uncertain. Design your LLM layer with substitutability in mind — abstract behind an interface so you can swap providers without a rewrite.
Don't mistake funding rounds for capability validation. Capital is evidence of investor confidence and burn rate. It is not evidence that a product works for your use case. Evaluate the tools, not the press releases.
The efficiency-first approach is now the default. DeepSeek showed that reasoning performance doesn't scale linearly with compute spend. Every lab — Western and Chinese — is now running experiments on mixture-of-experts, sparse activation, and similar approaches. The post-DeepSeek landscape is one where "what did it cost to train" is a legitimate question, and the answer actually tells you something.
Physical AI is the next frontier in China, and it is moving fast. The funding story in Chinese AI is not limited to software. Unitree Robotics, Fourier Intelligence, and a cluster of Shenzhen-based humanoid robotics companies are shipping hardware that would have been impressive three years ago. Shenzhen's manufacturing ecosystem — component supply chains, rapid prototyping, proximity to factory deployments — gives physical AI companies a development cycle advantage that Beijing-based language model labs simply don't have. The humanoid robot story and the LLM story are converging: the models that run inference on-device inside a robot body are the same models being optimised for efficiency in the labs.
Chinese AI tools will compete in the Global South first. Before they compete for Western enterprise contracts, they will win in Southeast Asia, the Middle East, and Africa — markets that are price-sensitive, less entangled with US compliance frameworks, and often Chinese-capital-adjacent in infrastructure decisions. Those deployments are early signals worth watching.
The agent layer is the next battleground. Manus is one entry in a larger category. ByteDance, Alibaba, and Baidu all have agent products at various stages. Chinese companies are willing to experiment at scale faster in certain industrial verticals — manufacturing, logistics, customer service — and may find product-market fit in those categories before agent tools cross over to general knowledge work. The hardware-software integration that Shenzhen enables makes physical-world agent applications particularly worth watching.
Open-weight models don't go away. Every year that performant open-weight Chinese models remain publicly distributed makes the export control logic harder to apply cleanly. Restricting hardware works differently from restricting weights that are already downloaded on a hundred thousand servers worldwide. That's a policy problem nobody has solved, and it has a long tail for anyone relying on hardware-restriction as the primary lever.
For context on how US-China technology relationships continue to operate despite restrictions, the Huawei-Qualcomm patent licensing agreement reached earlier this year is instructive: commercial relationships at the component layer continue even when the geopolitical frame suggests otherwise, and the picture is more granular than a clean decoupling.
Is DeepSeek actually as capable as OpenAI's models?
On specific benchmarks — particularly coding and reasoning — DeepSeek-R1 performs competitively with GPT-4 class models, according to the company's published evaluations and independent third-party tests published in early 2025. On multilingual tasks outside Chinese and English, the comparison is less favorable. On safety and alignment work, DeepSeek publishes considerably less, and Western frontier labs have invested more in that dimension. "As capable" is task-specific. Run your own evaluation with your own data.
Can Western companies actually use these tools legally?
Generally yes, with caveats. Open-weight models like DeepSeek can be downloaded and deployed without any Chinese-jurisdiction API call. Kimi's API is accessible but requires legal review for data-sensitive applications. As of this writing, the US government has not designated these specific products as restricted for general commercial use — but the regulatory environment moves, and legal review is not optional at enterprise scale.
How are Chinese labs training competitive models under chip export controls?
A combination of approaches: older-generation Nvidia hardware acquired before the October 2023 rules tightened (H800, A800); architectural efficiency that extracts more from constrained compute; and, increasingly, domestic alternatives like Huawei's Ascend 910B. DeepSeek has been explicit about its H800 cluster use. Whether the next generation of frontier Chinese models can be trained at competitive quality without H100 or H200-class chips is genuinely unanswered.
What is High-Flyer Capital's actual role in DeepSeek?
High-Flyer is a quantitative trading firm in Hangzhou, founded by Liang Wenfeng. It built significant GPU infrastructure for trading algorithms, and that same infrastructure became the foundation for DeepSeek. Liang runs both entities. There is no Big Tech parent with enterprise sales infrastructure, which is why DeepSeek's commercial trajectory is harder to predict than Baidu's or Alibaba's — and why it behaves differently from every other major Chinese AI lab.
What does "open-weight" actually mean for enterprise users?
Open-weight means the model weights are publicly downloadable. You run the model on your own infrastructure, fine-tune it on your own data, and never send a query to a Chinese server. This is distinct from "open source" (which implies training code and data too) and from API access (which routes queries through a vendor's infrastructure). For data-sensitivity concerns, open-weight is the relevant access mode — and it is the mode that makes DeepSeek's models directly usable by Western enterprises without the jurisdictional questions that API access raises.
Is Manus actually available, and is it production-ready?
At launch, Manus was invite-only. Access has expanded since the February 2025 debut but remains limited. A broader public launch has been signaled but not fully completed as of this writing. Treat it as early-access software. The underlying agent architecture is genuinely interesting; the product reliability at this stage is consistent with its age.
Should Chinese AI's rise concern Western AI companies?
"Concern" is a frame that produces bad decisions in both directions — either dismissing real competitive development or catastrophising it. The useful frame is competitive analysis. Chinese labs have produced genuinely good work at lower cost. That expands the options available to developers and enterprises globally. For Western AI companies, the relevant question is which capabilities they are building that are defensible, and which are now commoditised. The funding rounds are a signal of confidence. The architecture papers are the actual evidence. Read the papers.
Senator Elizabeth Warren's formal invitation to Nvidia CEO Jensen Huang to testify before a Senate committee is the sharpest escalation yet in a years-long argument about whether export-compliant American chips are...
DeepSeek and Zhipu AI are racing toward trillion-parameter scale using Huawei Ascend clusters and optimized mixture-of-experts architectures — precisely while US export controls have cut off access to Nvidia's H100...
In April 2025, Washington tightened export controls on Nvidia's H20 — the last chip Nvidia could legally sell in China at scale — and the financial hit was immediate. The Chinese AI compute market is now pivoting...
ByteDance and Alibaba are pulling humanlike AI companion features from their flagship products before formal rules arrive from China's Cyberspace Administration — a front-running pattern these companies have turned...
Moonshot AI — whose Chinese name, 月之暗面, translates directly to "The Dark Side of the Moon" — has released Kimi benchmark results that set off a pattern familiar to anyone who watched AI stocks in January 2025: sharp...
A Hangzhou hedge fund turned AI lab published a technical report in December 2024 claiming it trained a frontier language model for roughly $5.5 million — a figure that, if accurate, rewrites every assumption Western...
A US government website reportedly integrated Qwen — Alibaba's open-weight language model, built by a team in Hangzhou — for search functionality, even as the FBI was actively pursuing intellectual property theft...
Huawei and Qualcomm have reached a multi-year global patent licensing agreement spanning 5G and AI-related intellectual property — and it moves through a completely different legal channel than the US export controls...
BestAIFor