Moonshot takes the limelight as China gets political with AI
Highly-rated Kimi K3 model is winning rave reviews just as President Xi launches AI initiative for developing nations
Welcome back to Asia Tech Review, your curated digest to keep up to date with tech news across Asia.
We are back on summer schedule now after a very quiet week due to my travel. Going forward, that means that for July and August we will publish three times a week unless there’s earth-shattering news to share outside of that.
To keep up with our issues, follow us on LinkedIn or WhatsApp. Please do share Asia Tech Review with people you know who want to get smart about Asia tech.
Highly-rated new Kimi model gives Moonshot AI its moment in the sun
It’s been a major month for Moonshot AI, after it released a highly-acclaimed new version of its Kimi AI model and was linked with a major public listing.
Moonshot is one of China’s ‘Six Tigers,’ a group of AI startup pioneers building frontier models but it fell behind in the narrative as flashier peers MiniMax and Zhipu (Z.ai) went public and DeepSeek grabbed headlines for its rapid rise and competitive products.
But Kimi K3, a new 2.8 trillion-parameter open model with native vision and a 1 million-token context window, is bringing it attention with performance that ranks favourably to Western competitors. The new release is claimed to be the world’s largest open AI model by parameter count. Not only that, but with no guardrails like those of Claude and ChatGPT, some users have found it to be more useful, too.
Demand was so high during the first 48 hours that Moonshot paused all new subscriptions in order to keep up with a huge spike in activity. In response, it will also carve out separate plans for subscribers who want coding usage and those wanting more general workflows to help it handle the increased load.
This publicity comes at a good time, for sure, as Moonshot is reportedly looking to list in Hong Kong potentially by the end of the year and at a valuation of over $30 billion. According to Bloomberg, it has already sent shareholders a request for approval.
The incredible performance of Z.ai’s share price, which is up nearly 1,000% this year despite a 50% pullback over the last month, means all Chinese AI firms will have IPO plans. Kimi’s newfound credibility could make that journey smoother for Moonshot.
Indeed, its initial release sent AI and semiconductor stocks lower as investors questioned whether US AI capex and chip demand are harder to justify if Chinese models win more enterprise usage.
But, markets are a war of narrative. Despite America’s retail investors wondering whether yet another Chinese model is showing up its big spending companies, Transformer News argues its capabilities have been over-stated.
But take a step back and actually look at the evidence. While Kimi K3 is certainly a very good model, by the company’s own admission, it is not at the frontier. That means that unlike Claude Mythos or GPT-5.6 Sol, it likely does not have dangerous cyber capabilities.
Similarly, Chinese companies can’t keep releasing open source models forever. It helps get attention and close a lead, but once its top models reach parity with Claude and ChatGPT, there will need to be some guardrails. We’ve written before that Beijing is already said to have opened discussions around restricting models outside of China, for example. There are also legitimate concerns around distillation.
Still, the narrative is clear, and that is that Chinese models continue to gain credibility and undermine the free-spending nature of America’s best AI firms. For now, at least.
China puts AI into its geopolitical strategy
The Kimi launch was one of many announcements and developments from WAIC, the World Artificial Intelligence Conference, a Chinese event held in Shanghai that attracted 3,000 exhibits from more than 1,100 companies including many of China’s top names. The main takeaway, though, was politics.
We suggested before that China’s AI firms could be the next global technology land grab, and comments from Chinese Premier Xi Jinping at WAIC make that reality seem likely.
The themes were as you might expect, as Reuters reports:
In a speech to the opening ceremony of the World Artificial Intelligence Conference (WAIC) in Shanghai, Xi urged countries to seize the “historic opportunity” of open-source AI, and pledged to help developing nations build AI capabilities, warning against the emergence of “new historical injustices” from unequal access to the technology.
Xi’s speech, which is translated on SCMP, included a rallying call to encourage open source, openness and collaboration, as well as security and control. He strictly called out “placing one country’s security over that of others” and claimed that AI development should not be from one single country.
It’s, of course, easy to make these comments when you are behind, which China certainly is. But Xi placed a lot of emphasis on building a global community and development for all countries.
Xi announced the WAICO Alliance which, ostensibly an alliance to develop AI worldwide, looks like a political play. Its founding 29 member countries include Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia and Pakistan. That included a five-year pledge to open training and seminar opportunities (5,000 to be oddly specific) and AI application centres with regional bodies from Asia, Africa, Latin America and beyond.
This certainly looks like the next political-tech crossover. Now we wait to see how much of the nice-sounding rhetoric makes it into practical policy, and, if so, which Chinese companies stand to benefit.
In a similar way to how Huawei and ZTE networked much of the world and Xiaomi, Huawei again and others made smartphones affordable in scores of countries, Z.ai, Moonshot, DeepSeek and others seem poised to build usage in the developing world. They’re cheap and accessible. Plus China is building its position in response to the US government’s intervention with the rollout of Anthropic’s Mythos and Fable models, and later OpenAI’s newest models.
Elsewhere from WAIC, key news included:
Alibaba launched a preview of its Qwen3.8 Max model, claiming the system rivals top frontier AI models and trails only Anthropic’s Fable 5
BrainCo unveiled a brain-to-robot platform that lets users control robots through thought alone
StepFun showcased agents that run across a smartphone and its own operating system
Alibaba’s chip design unit T-Head will open-source SAIL, the software stack for its Zhenwu AI chips
ZTE unveiled the NaviX Ultra, billed as the world’s first agentic AI smartphone, with ByteDance’s Doubao in tow
Dongfang Suanxin (DFSX) unveiled an AI chip road map built around software-defined computing and 3D-stacked near-memory architecture as it tries to reduce dependence on Nvidia and export-restricted components
MiniMax showcased its M3 multimodal large model
Deals
DeepSeek’s first external funding round valued the Chinese AI lab at CNY350.88 billion ($51.82 billion), after a fund backed by consumer goods firm Anhui Korrun invested CNY2.9 billion ($428 million) for an indirect 0.8265% stake [Reuters]
Indian AI coding startup Emergent raised a $130 million Series C at a $1.5 billion valuation, helped by $120 million in annual run-rate revenue and more than 200,000 paying customers [TechCrunch]
Indian smart-ring maker Ultrahuman is reportedly in talks to raise about $60 million from Steadview Capital and Blume Ventures at a $360 million-$400 million post-money valuation [Economic Times]
Rize, a Singapore-based sustainable rice platform, raised $31 million led by BNP Paribas Asset Management Alts to take its total investment to $47 million, it emerged from the 100x100 fund we previewed last month [Rize]
Singapore-based pQCee raised $3.9m in seed funding co-led by SGInnovate and Lotus One Investment, with the post-quantum cybersecurity startup planning to expand across Asia, the US, Europe and the Middle East. [TechNode Global]
Hyundai will buy SoftBank’s roughly 10% stake in Boston Dynamics, giving it full control of the US robotics company before deploying Atlas humanoid robots at a Georgia car plant from 2028 [Reuters]
SBI Holdings completed its majority acquisition of Singapore crypto platform Coinhako after winning approval from MAS, extending the Japanese group’s Southeast Asia digital asset push [The Block]
Markets
TSMC pledged another $100 billion for Arizona after second-quarter profit jumped 77% to a record T$706.6 billion ($22 billion) on continued AI demand [Reuters]
Japanese memory chip maker Kioxia saw its share price fall by as much as 16% in Tokyo, leaving it down 52% from last month’s peak and wiping at least JPY30 trillion ($185 billion) from its market value [Bloomberg]
Related: A Texas federal jury ordered Kioxia to pay Viasat $229 million for infringing a flash-memory patent developed for satellite error-correction systems [Reuters]
Indonesian coffee chain Kopi Kenangan is said to have held talks with banks about a potential listing that could value the company at up to $1 billion, with Singapore among the venues under consideration [WSJ]
Zepto’s planned IPO is drawing investor interest at lower valuations, with foreign investors around $4.5 billion pre-money and some domestic institutions at $3 billion-$3.5 billion [Economic Times]
CXMT’s Shanghai IPO drew applications from more than 9.4 million investor accounts, with retail demand reaching 212 times the shares initially available online [SCMP]
Singapore’s markets regulator has reportedly held talks with investment companies about possible tax cuts for fund managers in response to Hong Kong proposing sweeping tax changes to lure high-paid portfolio managers [FT]
AI and Chips
Z.AI is said to be on pace for $1 billion in annual recurring revenue after hitting its full-year sales target in July [Bloomberg]
Nvidia CEO Jensen Huang’s Japan trip is expected to include support for Noetra, the SoftBank-led sovereign AI consortium backed by 44 Japanese companies and up to JPY1 trillion ($6.2 billion) in government funding [Nikkei Asia]
Taiwan is building TAIDE, its own AI model, to preserve traditional-script Chinese and counter mainland political framing in global AI systems trained mostly on simplified-script data [Nikkei Asia]
Thinking Machines Lab released Inkling, its first general-purpose AI model, using a DeepSeek-V3-based architecture and post-training data from Moonshot AI’s Kimi K2.5 [FT]
Nvidia and four Japanese industrial automation companies will expand robot development work through its Cosmos physical AI platform [Bloomberg]
Chinese tech giants including ByteDance, Alibaba and Tencent are turning AI tokens into an internal corporate currency as cheaper model pricing drives AI deeper into daily workflows [SCMP]
The US House China committee urged the Trump administration to bar US companies from buying memory chips from CXMT and YMTC over AI infrastructure and supply chain security risks [FT]
Tata Electronics is preparing to produce India’s first semiconductor wafers on 90nm process technology at its Dholera fab, stepping back from the 28nm process Tata Sons touted in March 2025 [Bloomberg]
Indonesia’s planned copyright rewrite would grant copyright privileges to people who use AI to generate content, setting up a potential fight with Google and other platforms over training data [Reuters]
DeepSeek is pushing China’s AI price war into enterprise adoption, with its V4 Flash model completing a standard intelligence task for about $0.02 versus more than $2 for Anthropic’s premium model [Bloomberg]
Japanese chipmaker Rapidus will work with US software company Cadence to offer customers agentic AI tools for chip design [Nikkei Asia]
The UK’s AI Security Institute said leading open-weight models GLM-5.2 and DeepSeek V4-Pro now perform on cyber tasks like frontier closed models released four to seven months earlier [UK AI Security Institute]
Shanghai Xingshu Tiansuan launched the first constellation for a planned 1,000-satellite space-computing network to process AI and remote-sensing data in orbit [Reuters]
Japan plans to buy 27,500 next-generation Nvidia Rubin chips for a JPY387.3 billion government-backed sovereign AI project led by newly formed Noetra [Bloomberg]
In other news
Hyundai Motor workers in South Korea staged a partial strike after the company showed its Atlas humanoid robot, with the union demanding a say before robots enter production lines [WSJ]
Sensitive files related to India’s Kudankulam nuclear power plant were exposed in a data breach and posted on the dark web [Reuters]
Cambodia’s Lim Heng Group leased casino-site buildings on the Thai-Cambodian border to a compound used for industrial-scale scams and human trafficking [Reuters]
South Korean authorities searched Chinese chipmaker Montage Technology’s local office over a potential competition law violation, with South Korea accounting for CNY2.93 billion ($432.63 million) of Montage’s fiscal 2025 sales [Reuters]
Balaji Srinivasan asked Malaysian Prime Minister Anwar Ibrahim’s office for a meeting after immigration officers raided his Network School in Johor’s Forest City [Bloomberg]
A South Korean fine on Coupang over a data leak has drawn criticism from Washington and forced Seoul’s US ambassador back to the capital for consultations [Reuters]
A Taiwan court sentenced BitShine ringleader Shih to 22 years in prison for unlicensed virtual asset services, fraud and money laundering that hit more than 1,500 victims for NT$1.27 billion ($39 million) [The Block]