Reported by |
Edited by | Wen Shuqi
The numbers are staggering: a fourfold surge in overseas paying users, a 400% jump in API revenue, and presence in over 200 countries and regions. These aren’t just stats; they are the latest report card for Moonshot AI’s flagship model, Kimi, as we head into 2026.
At the recent Amazon Web Services (AWS) China Summit, Huang Zhenxin, the B-side lead at NUPIAO (Moonshot AI), spilled the beans to the press. He revealed that Kimi’s enterprise business is picking up serious steam. Industries like internet, finance, manufacturing, education, and healthcare are now major clients, while international expansion is accelerating faster than ever.
For this star AI startup, which has only been around for three years, a new chapter is officially opening.
For years, NUPIAO’s brand was built on pure tech prowess: long context windows, mastering Scaling Laws, and pushing the boundaries of foundational models. This helped Kimi build massive industry influence almost overnight.
But even today, Huang Zhenxin keeps bringing the conversation back to the core: investing heavily in bottom-layer model innovation. The goal? To keep evolving model architectures and constantly test the limits of Scaling Law.
According to NUPIAO, we haven’t even hit the ceiling yet. Whether it’s pre-training, post-training, or architectural breakthroughs, resources are pouring in. Take “MuonClip,” for instance—it’s already being adopted by DeepSeek V4. New architectures like attention residuals are slated for the next-gen models.
However, as more companies rush to buy AI services, a tough question looms for every model maker: Are enterprises actually buying a model, or are they looking for a complete solution?
In the last six months, the game has changed. After countless companies started building apps around Agents, giants like ByteDance and Alibaba’s AI division shifted their focus to industry-specific solutions. Overseas, OpenAI and Anthropic are beefing up their enterprise teams. Especially Anthropic, whose “Forward Deployed Engineer” (FDE) model has become a cornerstone of their business strategy. It’s no longer just about providing an API; it’s about diving into client workflows and transforming them.
Tan Dai, President of Volcano Engine, recently noted that an enterprise’s moat boils down to two things: first, the model’s capability; second, how well you can integrate that model into the business. This means landing FDE strategies, understanding the industry inside out, deep partner ecosystems, and having a team that truly speaks “AI Solution.”
While the industry trend is getting heavier and heavier with delivery, NUPIAO chose a different path.
Huang Zhenxin believes the real bottleneck for enterprise AI isn’t on the model vendor’s side. It’s about how to effectively cut in and drive the enterprise through its AI transformation. “The approaches of the two overseas giants differ, and everyone is still figuring it out,” he noted.
When it comes to industry solutions, system integration, and that tricky “last mile” of implementation, NUPIAO prefers to let partners handle the heavy lifting. AWS is one of their most tightly integrated partners.
Here’s how the deal works: Kimi brings the model muscle, while AWS handles industry solutions, global customer access, and compliance frameworks. Currently, Kimi models are live on the Amazon Marketplace for API services. In the future, they aim for deep integration into Amazon Bedrock, where Kimi’s inference will run directly on AWS computing infrastructure. NUPIAO also plans towork hand-in-hand with AWSSolution Architects(SAs) to craft tailored industry solutions.
Huang shared that as the partnership deepens, they might explore collaboration even at the pre-training level, potentially running some training tasks on Trainium chips.
To Huang, making the model itself good enough is a massive challenge. That’s why NUPIAO wants to save its energy for the model. Even during their fastest growth phase for enterprise business, they’ve remained disciplined.
As of now, NUPIAO employs fewer than 300 people. Compared to the thousands in big tech AI teams, this startup crew is lean, and resource allocation remains laser-focused on model R&D.
“We ultimately want to push the ceiling of intelligence and aim to give those top three overseas model companies a real run for their money,” Huang stated.
With model capabilities continuously improving, NUPIAO is seeing better adaptation to complex environments. This reduces reliance on heavy external engineering frameworks, simplifying requirements for Harnesses. Internally, they’ve already started practicing “Loop Engineering”—a cleaner, newer approach compared to traditional Harnesses.
This year, nearly every model provider raised prices due to soaring compute costs. Whether in the US or China, compute resources can’t keep up with token demand, passing the cost pressure straight to the service layer.
Huang mentioned that while users are willing to pay a premium for high-performance tokens, model vendors are fighting back. Through optimizations like cache hit rates and inference improvements, they’re working to lower actual token costs. Right now, Kimi’s native service boasts a cache hit rate exceeding 90%.
For NUPIAO, what really determines the future battlefield is the model capability itself.
