Moonshot AI vs. DeepSeek
Jan-Erik Asplund
TL;DR: While most Chinese AI labs have focused on small, efficient models usable by chip-constrained mainland businesses, Moonshot AI bet on scale, releasing the world’s largest open-weight model with Kimi K3 & finding customers among American AI companies building their own coding & legal agents. Sacra estimates Moonshot AI reached $1B in annualized revenue in August 2026, up from $300M in June. For more, check out our full report and dataset on Moonshot AI.
Key points via Sacra AI:
- ChatGPT’s launch in November 2022 kicked off a race to build the ChatGPT for China, with Moonshot AI (2023) differentiating from Baidu & Alibaba’s models through its focus on long context, echoing Anthropic’s strategy against OpenAI, with its flagship Kimi (October 2023) processing 200K Chinese characters versus roughly 80K for Claude 2 & 25K for GPT-4-32K, with the aim of winning businesses that need to analyze & do RAG over entire legal contracts or financial filings. Moonshot AI began charging roughly $5/month for priority access in May 2024 versus $20 for ChatGPT Plus or Claude Pro, while its 128K-context API charged approximately $8.30 per million input or output tokens, versus GPT-4 Turbo’s $10 input/$30 output—making GPT-4 Turbo 1.2x as expensive for input & 3.6x as expensive for output.
- After spending roughly $100M on advertising in 2024 on its consumer push, Moonshot AI grew Kimi to more than 20M monthly active users, second only to ByteDance’s Doubao among Chinese AI apps, before DeepSeek released its viral R1 reasoning model on January 20, 2025—the same day as Moonshot AI’s reasoning model k1.5—and seized the spotlight & users, hitting 35M DAUs in February while Kimi went into decline. Kimi’s monthly active users fell 55% to roughly 9M by year-end 2025 as DeepSeek seized the lead, while Kimi’s original long-context advantage was eroded by Alibaba launching a model with support for 10M Chinese characters & Baidu launching a 2.5M context window model.
- Post-DeepSeek, Moonshot AI cut ads & shifted to distributing its models through businesses, releasing Kimi K2 (July 2025) as open weights just as AI companies like Cursor, Cognition & Harvey were beginning to train their own specialized models to reduce dependence on OpenAI & Anthropic, control inference costs and improve performance through their own agent harnesses. Moonshot AI enters revenue-sharing agreements with inference providers like Fireworks AI ($800M annualized revenue in May 2026), Together AI ($1B annualized revenue in Feb 2026) & AWS Bedrock, which contract directly with customers to serve standard or fine-tuned versions of Kimi and pay Moonshot AI up to 30% of the resulting revenue, allowing Moonshot AI to monetize demand from US companies like Cursor & Harvey without directly handling their data or billing relationships.
- The launch of Kimi K3 (July 2026), the largest open-weight model in the world, gave Moonshot AI its own DeepSeek-style high-profile success, with the model ranking behind only Anthropic’s Claude Fable 5 & OpenAI’s GPT-5.6 on overall benchmarks & coming in at 2.8T total parameters, versus 763B for the subsequently released DeepSeek V4.1 Flash & roughly 743B for GLM-5.3. Due to US GPU export restrictions, Chinese labs have generally leaned towards creating smaller models like DeepSeek V4 Flash (1/10th the size of Kimi K3) that need less GPU memory & fewer chips to run, but Moonshot AI has bucked that trend & in the process won with American customers that are looking for open weights models with frontier-level agentic & reasoning performance.
- Sacra estimates Moonshot AI reached $1B in annualized revenue in August 2026, up from $300M in June & about $15M at the end of 2025, last valued at approximately $50B in October for a 50x annualized revenue multiple, ahead of their planned Hong Kong IPO. Compare to DeepSeek at $1B in annualized revenue in September 2026, up from $12M at year-end 2025, currently raising at a $75B valuation for a 75x multiple, and publicly traded MiniMax (HKEX: 0100) at $800M in annualized revenue in August, up 8x from $100M at year-end 2025, with an $8.2B market capitalization as of October 16 for a 10x multiple.
- Squeezed between DeepSeek’s rock-bottom API prices for businesses & free consumer AI products from ByteDance, Alibaba & Tencent at home, Moonshot AI is increasingly a global AI company, with overseas revenue overtaking domestic revenue in February 2026, overseas paying users reportedly growing 400% by June & more than 70% of new users coming from outside China post-K3. In addition to the U.S, Moonshot AI is pushing into markets like South Korea & Brazil, hiring local-language influencer marketers as they go up against sovereign LLMs like Naver’s HyperCLOVA X & Maritaca’s Sabiá, looking to win on the proposition of getting frontier intelligence at lower prices while local players tout domestic computing infrastructure & security.
- With Kimi K3’s July benchmark lead already overtaken by newer models from Xiaomi, Alibaba, DeepSeek & Z.ai, Moonshot AI’s next test is whether K4 can restore an advantage that attracts more enterprise API spending, or whether its applications can build recurring customer relationships even as buyers switch between competing Chinese models. Its growing suite spans Kimi Chat (ChatGPT), Kimi Slides (Gamma), Kimi Docs (Google Docs/Gemini), Kimi Sheets (Excel Copilot), Kimi Deep Research (ChatGPT Deep Research), Kimi OK Computer (Manus), Kimi Agent Swarm (Genspark), Kimi Claw (OpenClaw), Kimi Work (Claude Cowork), and Kimi Code (Claude Code/Codex), with international plans from $19–$199/month competing on breadth & price against OpenAI’s & Anthropic’s roughly $20 standard plans & $200 heavy-use tiers.
For more, check out this other research from our platform:
- Moonshot AI (dataset)
- DeepSeek (dataset)
- MiniMax (dataset)
- Fireworks AI (dataset)
- Together AI (dataset)
- OpenRouter (dataset)
- OpenRouter growing 29% MoM at $140M/year
- $400M/year Napoleon of sovereign AI
- $100B sovereign AI boom
- Fireworks AI customer at Hebbia on serving state-of-the-art models with unified APIs
- Together AI: the $44M/year Vercel of generative AI
