Why Singapore Won’t Build Frontier AI | Generated by AI

Home 2026.06

Question: DeepSeek built V2 with ~100 people. Why can’t Singapore’s top 100–500 talents do that?

Answer:

They could technically. But Singapore chose not to—it’s a deliberate strategic decision, not a capability gap.

Here’s the asymmetry:

DeepSeek’s Actual Advantages

DeepSeek started inside High-Flyer, a hedge fund. During 2019–2022, High-Flyer spent ~$139M and stockpiled 10,000 Nvidia A100 GPUs before the US government imposed export controls on China in October 2023. That GPU inventory became DeepSeek’s moat—it could train frontier models after export controls hit, when most domestic Chinese peers couldn’t.

DeepSeek’s core team is <140 people. Outside that, it shares High-Flyer’s senior management, ops, HR, finance—basically everything except R&D. So it’s not 140 people doing everything; it’s 140 researchers + borrowed corporate infrastructure.

Key: DeepSeek’s culture emphasizes autonomy and bottom-up decision-making—”curiosity, passion, self-motivation”—similar to ByteDance’s approach. At ~200 people, this loose structure is its greatest competitive moat.

Why Singapore Won’t Do It

Singapore’s National AI Strategy explicitly states: “Our advantage does not lie in building the largest frontier models. It lies in deploying AI effectively, responsibly, and at speed.” This isn’t modesty—it’s deliberate policy.

Singapore’s homegrown models (SEA-LION, MERaLiON) prioritize training on regional Southeast Asian languages rather than frontier capability. They focus on deployment testing and application-level safety rather than model-level controls.

Singapore is building sectoral missions around advanced manufacturing, embodied AI, and partnering with NVIDIA’s new research lab—but within the context of deployment and testbeds (like Punggol Digital District), not frontier model development.

The Real Reasons Singapore Won’t Move

  1. Capital structure mismatch: DeepSeek had a hedge fund with $8B in assets. Singapore’s AI companies are venture-backed, needing exits in 7–10 years. Frontier models are 5–10 year bets with uncertain monetization. VCs won’t fund this.

  2. GPU access: High-Flyer acquired 10,000 A100s before export controls. Singapore has no historical GPU stockpile, and getting H100s is harder under current restrictions. China’s domestic supply chains matter.

  3. Regulatory comfort: Singapore positions itself as a trusted governance hub. Building a frontier model creates liability and geopolitical friction. Deployment + testing is cleaner politically.

  4. Small market: Singapore is 5.7M people. China is 1.4B. DeepSeek needed Chinese-language, Chinese-context models. Singapore’s regional focus (Southeast Asia) is weaker.

  5. Incentive misalignment: Singapore wants to be a hub where companies test and deploy AI faster than larger countries, not compete with OpenAI/DeepSeek directly. Building a frontier model doesn’t accelerate that mission.

The Hard Truth

Singapore has the talent (top universities, strong ML researchers) and capital (strong fintech ecosystem). But frontier model building requires:

Singapore’s positioning as a deployment hub is actually more defensible long-term. The frontier model race is zero-sum (China, US, maybe EU). But being the best place to safely deploy, test, and govern AI across Southeast Asia? That’s a durable moat.

The meta lesson: Don’t confuse “could do” with “should do.” Singapore could build frontier models. It chose to own a different, less crowded part of the value chain instead.

References:

Back Donate