In February, delegations from more than 100 countries signed the Bangkok Declaration, committing to build and control their own artificial intelligence capacity rather than rent it permanently from Silicon Valley or Beijing. The declaration produced a wave of very literal follow-through: India now runs Sarvam AI, a 70-billion-parameter model trained across 22 Indian languages. Malaysia has ILMU. Indonesia has Sahabat AI. Singapore has SEA-Lion. South Korea has HyperCLOVA X Think. Japan has NTT's Sarashina. Taiwan has TAIDE. Every major economy in the Asia-Pacific now has a named, domestic large language model, built to work in local languages a global model tends to handle poorly.
The chip layer underneath all of that software is where Asia's dominance is least contested. Taiwan's TSMC posted $122 billion in revenue in 2025, 58 percent of it from high-performance computing and AI accelerators, and controls 71 to 72 percent of the global pure-play foundry market — meaning most of the world's advanced AI chips, regardless of whose logo is on the box, are physically manufactured on the island. South Korea's SK Hynix supplies the high-bandwidth memory that sits inside nearly every high-end Nvidia GPU, holding 53 to 57 percent of the global HBM market and posting record annual sales of $70.4 billion.
Every major Asia-Pacific economy now has its own named large language model. Almost none of them are trying to beat OpenAI at its own game.
Capital is following the same geography. Asian startups drew roughly $33.6 billion in AI venture funding in the first quarter of 2026 alone — the only region where deal volume rose rather than plateaued. Google committed $15 billion over five years to expand AI infrastructure in India in February, covering subsea cables, cloud capacity and data centres. Australia's AirTrunk followed with a separate $30 billion pledge to build five gigawatts of Indian data-centre capacity by 2030. Taiwan's government, for its part, announced a "Ten AI Initiatives" plan in January to expand domestic compute and talent rather than simply supply chips to whoever designs the models elsewhere.
What ties these efforts together is a rejection of the framing that AI is primarily an American race with everyone else as suppliers. Sovereign AI, as the field is now labelled, treats language models the way earlier generations of Asian governments treated steel or semiconductors: a capability too strategically important to outsource entirely, even to allies. A government that can't train a model in its own official languages, the logic runs, will always be a customer of someone else's AI rather than a maker of its own.
None of these efforts individually rival the scale of the largest American labs, and several of the newer sovereign models remain research projects more than commercial products. But the chips, the memory, and an increasing share of the capital all originate in the region regardless of where the best-known chatbots are built. The AI race was never only about who writes the smartest model. Increasingly, it is about who controls the compute the model runs on — and on that count, the balance of power already sits in Asia.