Silicon Valley didn’t see this coming: a little-known Chinese AI startup, DeepSeek, has burst onto the scene to challenge the industry’s giants. By developing cutting-edge models at a fraction of the cost, DeepSeek is forcing a rethink of how AI innovation can and should be approached. 

Previously we wrote about the headline and the opportunities it opens up for the AI industry, but this article digs deeper into the models themselves and implications for AI core startups.

A Quiet Revolution Begins

The story starts in Hangzhou, China, where founder Liang Wenfeng, a former quantitative finance expert, reimagined AI development. Rejecting the conventional reliance on massive computational power and billion-dollar budgets, DeepSeek opted for a lean, innovative strategy.

Born within High-Flyer, an AI-powered hedge fund, the startup found a nurturing environment for experimentation. Free from the pressures of Silicon Valley, DeepSeek quietly laid the groundwork for a revolutionary approach to AI.

The Key to Success: A Three-Pronged Strategy

1. Efficiency as a Game-Changer
   At the core of DeepSeek’s success is its groundbreaking Mixture-of-Experts (MoE) architecture. Their flagship model, DeepSeek-V3, boasts 671 billion parameters but activates just 5.5% (37 billion) for any given task. This selective activation slashes computational requirements while enhancing scalability and efficiency.
2. A Balance of Openness and Innovation
   Inspired by open-source pioneers like Meta’s LLaMA, DeepSeek combines transparency with unique advancements. By incorporating community-driven input into their distinct Mixture-of-Experts system, the company has created a hybrid model of openness and proprietary innovation.
3. Resourceful Foresight
   DeepSeek’s early acquisition of Nvidia A100 chips, ahead of U.S. export restrictions, secured the critical hardware needed to fuel their development. This strategic move allowed them to outpace competitors struggling with supply chain bottlenecks.

Breaking Barriers in Cost and Capability

DeepSeek’s models are as cost-effective as they are powerful. DeepSeek-V3 was developed for under $6 million, a stark contrast to the billions spent by rivals. This remarkable efficiency challenges the industry norm, proving that innovation doesn’t require exorbitant budgets.

A New Era in AI Reasoning

DeepSeek’s breakthroughs culminated in two key releases:

• DeepSeek-V3 (December 2024): A foundation model showcasing advanced efficiency with 671 billion parameters.
• DeepSeek-R1 (January 2025): A reasoning powerhouse that matches or surpasses OpenAI’s o1 model in tasks like mathematics and coding. Its top-ranking position on Apple’s App Store globally further validates its real-world impact.

Redefining Multimodal AI

In January 2025, DeepSeek unveiled Janus-Pro-7B, an innovative multimodal AI model. Unlike competitors, Janus-Pro integrates content understanding and generation into a unified framework, excelling in tasks involving text and image translation. With just 7 billion parameters, it outperformed OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion XL across benchmarks. Released under an MIT license, Janus-Pro invites global collaboration, marking another leap forward in AI development.

Implications for the Global AI Landscape

DeepSeek’s rise disrupts the conventional narrative that groundbreaking AI requires massive resources. Instead, it highlights the power of strategic innovation and efficiency. The company’s ascent has sent shockwaves through financial markets, shaking up shares of U.S. AI giants and sparking broader conversations about the future of the industry.

Looking Ahead

DeepSeek’s success poses critical questions:

• Could their efficient model become the industry standard?
• How will established tech leaders respond?
• What role will open-source contributions play in shaping AI’s future?

One thing is certain: DeepSeek is rewriting the rules of AI development. By prioritizing resourceful innovation over brute force, the startup is redefining what’s possible in artificial intelligence—and the world is taking notice.

Takeaways from AI core startups in Southeast Asia

DeepSeek’s success challenges the idea that AI innovation belongs only to the companies with billion-dollar budgets. Instead, they’ve proven that:

• Efficiency > Scale: Lean, smart architectures can compete with massive compute-heavy models. Optimizing model efficiency is just as important as raw compute power. AI startups must think beyond brute force scaling and invest in architectural efficiency to remain cost-competitive.
• Strategic Execution Matters: AI startups must plan years ahead for infrastructure, talent, and funding. When it comes to infrastructure in particular, the AI infrastructure bottlenecks are real. If you’re an AI startup in Southeast Asia or other emerging markets, securing compute power and partnerships early is key to long-term survival.
• Open-Source is an Opportunity: The best AI companies will leverage open innovation while building their own unique moat. That said, find the right balance between openness and defensibility. Open-source can be a powerful GTM strategy, but building proprietary differentiators is essential for long-term business sustainability.

For AI founders in Southeast Asia, this presents a massive opportunity—but also a challenge: Can regional AI startups carve out an edge through localized data, specialized models, and efficient architectures?

AI isn’t just for the giants anymore.