DeepSeek, MiniMax and Kimi are three different businesses

Ask a room of Asian operators about the non-American models and you usually get one answer back: the Chinese ones are cheaper. That is true, and it is close to the least useful thing you can know about them.

Gartner expects half of enterprises worldwide to have a Chinese large language or multimodal model somewhere in their AI portfolio by 2027, up from 5 per cent in 2025, according to its top ten AI trends in China for 2026 (China Daily, 21 September 2026). If that forecast is even directionally right, the question stops being whether to look and becomes which one to build on.

The three names that come up most often in Singapore, Jakarta and Kuala Lumpur are DeepSeek, MiniMax and Moonshot's Kimi. They are not one category. They are three companies running three different businesses, funded three different ways, and each one can fail you differently.

DeepSeek sells tokens, and almost nothing else

DeepSeek publishes its price list and it is unsentimental. V4.1-Flash bills $0.15 per million input tokens off-peak and $0.30 at peak, with output at $0.60 and $1.20. The larger V4-Pro runs $0.66 to $1.32 on input and $1.98 to $3.96 on output. Both carry a one million token context window, and off-peak is simply half price, covering every weekend and Chinese public holiday (DeepSeek API docs).

Set that against Anthropic's published rates of $2 and $10 per million tokens for Claude Sonnet 5, and $4 and $20 for Opus 5.5 (Claude pricing). The V4 weights are also public: DeepSeek-V4-Pro at 1.6 trillion parameters and V4-Flash at roughly 291 billion sit on Hugging Face, with V4-Flash alone showing 1.49 million downloads (DeepSeek on Hugging Face).

Volume followed. OpenRouter's own data has DeepSeek moving from just under 10 per cent of weekly token flow in January 2026, down to 5 per cent in February and March, then up to nearly 20 per cent by late June, and it has been the single most-used model on the platform since mid-May. V4-Flash accounted for 70 per cent of DeepSeek's agentic token flow within a month of release. In the same window, Chinese models as a group passed American ones, going from roughly 25 per cent of tokens in 2025 to more than half in 2026 (OpenRouter, 30 June 2026).

The business underneath is thinner than the usage suggests. The Information put DeepSeek's annualised revenue at between $400 million and $500 million with gross margins of 70 to 80 per cent on V4 access, while the company was seeking roughly 50 billion yuan, about $7.4 billion, at a valuation near 500 billion yuan, about $74 billion. That is around 148 times annualised revenue, and the company has hired banks for a Shanghai STAR Market listing (PYMNTS, 15 July 2026). Bloomberg reported in August that the round had reopened at close to $8 billion at a similar valuation (Bloomberg, 6 August 2026).

So what: DeepSeek is a price floor with a research lab attached. If you are buying tokens, it is the cheapest credible frontier option and the weights are yours to self-host. If you are buying a vendor relationship, there is very little relationship on offer, and a company priced at 148 times revenue has every incentive to eventually stop selling at this price.

MiniMax sells applications, and most of the money comes from outside China

MiniMax listed on the Hong Kong exchange on 9 January 2026. Shares were offered at HK$165 and closed the first day at HK$345, up 109 per cent, raising HK$4.8 billion, about US$620 million. It was the second Chinese large language model developer to list, one day after Zhipu AI, which rose a more modest 13 per cent (CNBC, 9 January 2026). The Hong Kong public tranche was 1,837 times oversubscribed (TechNode, 9 January 2026).

Then the numbers arrived. For the year to 31 December 2025, MiniMax reported revenue of US$79.0 million, up 158.9 per cent, with gross profit of US$20.1 million and a gross margin of 25.4 per cent. Adjusted net loss was US$250.9 million, against US$244.2 million the year before. AI-native products contributed US$53.1 million and the open platform and enterprise services business US$26.0 million. The company counted more than 236 million cumulative users across over 200 countries and 214,000 enterprise customers and developers, with more than 70 per cent of revenue from international markets and US$1.05 billion of cash on hand (MiniMax, 2 March 2026).

Read that carefully. A 25.4 per cent gross margin is a consumer applications margin, not a model vendor's margin, and the loss is three times revenue. This is a company selling video, voice and agent products to individuals, with a developer platform bolted on.

Two Asian data points make the strategy concrete. Singtel's AI Pass, the bundle Singaporeans unlock by enrolling in an eligible SkillsFuture course, contains six tools, and three of them are MiniMax products: Hailuo AI, MiniMax Code and MiniMax Audio (Singtel). Separately, HUMAIN, the AI company backed by Saudi Arabia's Public Investment Fund, unveiled humain-m3 at LEAP in Riyadh on 3 September 2026. It is a 428 billion parameter mixture-of-experts model built on the MiniMax-M3 lineage, commissioned by HUMAIN and delivered by MiniMax, then further pre-trained on more than a trillion tokens of Arabic-native content (PR Newswire, 3 September 2026).

So what: MiniMax is the one of the three already selling into the region through channels your staff will actually touch, including a government-linked skills scheme. It is also the one whose unit economics you should study before you build anything load-bearing on its platform. A quarter of a billion dollars of annual loss on $79 million of revenue is a runway question, not a technology question, even with a billion dollars in the bank.

Moonshot sells a position, and is repricing it every quarter

Moonshot AI's valuation has moved faster than its product. It ended 2025 worth $4.3 billion, reached $10 billion in early 2026 after a $700 million raise, then took $2 billion at more than $20 billion in May, led by Meituan's venture arm with Tsinghua Capital, China Mobile and CPE Yuanfeng participating. That was $3.9 billion raised in six months. Annualised recurring revenue topped $200 million in April (TechCrunch, 7 May 2026).

It did not stop. Bloomberg reported talks on pre-IPO funds at a $50 billion valuation on 21 July, a closed round at $35 billion on 29 July, and on 4 September that Moonshot was seeking up to $5 billion in a Hong Kong listing this year (Bloomberg, 29 July 2026; Bloomberg, 4 September 2026).

The pricing tells you what Moonshot thinks it is now. Kimi K3 lists at $3.00 per million input tokens and $15.00 output, with a cache-hit input rate of $0.30 and a one million token context window. K2.6 and K2.7 Code sit at $0.95 and $4.00 (Kimi platform pricing).

That is the detail most Asian buyers have not absorbed. K3 at $3 and $15 is not a discount model. It is priced between Claude Sonnet 5 and Opus 5.5, and it costs ten times what DeepSeek V4-Pro charges at off-peak input. The cheap-Chinese-model assumption has already stopped being true at the top of the range.

So what: Moonshot has earned developer trust, with K2.6 the second most used model on OpenRouter earlier this year, and it is monetising that trust at Western prices ahead of a listing. Budget for Kimi as a premium line item, not as a saving, and expect its pricing to keep moving while the IPO is live.

What an Asian buyer should actually do with this

The useful split is not American versus Chinese. It is what the vendor sells, because that determines what it does to you when it needs money.

DeepSeek sells inference at near cost and will be judged on volume, so the open weights are the real product and self-hosting is the honest way to use it. MiniMax sells applications to consumers and is judged on that, so treat its API as a supplier relationship with a genuine solvency question attached. Moonshot sells a pre-IPO story and is judged on multiples, so its prices will follow its valuation rather than its costs.

Two more constraints belong in the same decision. First, several governments in the region restricted the DeepSeek consumer application on official devices, which says nothing about whether you can run the weights yourself inside your own boundary, and everything about whether you can hand staff the app. The distinction between a hosted service and an open weight file is now a procurement distinction, not a technical one. Second, the regional models you may end up standardising on are already built on Chinese foundations. Singapore's SEA-LION v4.5 is built on Qwen, and Malaysia has committed $494 million to a sovereign AI programme running on Huawei silicon (Foreign Policy, 21 September 2026). The choice has been partly made for you further up the stack.

The through line

The cheap-model story served its purpose and is now getting in the way. Price was the opening move for all three of these labs, and only DeepSeek is still playing it. MiniMax has moved to applications and is burning capital to hold user share. Moonshot has moved to premium and is repricing quarterly into a listing. If you are choosing a model this quarter, anchor on the business you are buying from rather than the benchmark you are buying, because the benchmark will change again before your contract renews and the business model will not.