The global hunt for AI compute has SEA in its crosshairs

2026 was supposed to be all about agentic AI and enterprise adoption. But halfway into the year, another phrase has begun dominating conversations: compute shortages.

Spiking demand and limited production have led to a backlog of both graphic processing units (GPUs) and DRAM chips, which in turn has resulted in rising prices for cloud-based AI services. It’s a problem that falls disproportionately on the shoulders of small and medium-sized AI firms.

Image credit: Ulla

If you think this sounds like an issue that only concerns the US or China, think again. Increasingly, AI firms are on a global search for available – and hopefully affordable – GPUs, and many are looking at Southeast Asia.

Carmen Li is at the helm of two companies involved in hunting down additional compute. She is the CEO and founder of Silicon Data, a New York-based market intelligence firm she calls a “Bloomberg for compute,” and Compute Exchange, a Silicon Valley-based company that offers a marketplace where businesses can buy GPU capacity.

In short, one company tracks the prices while the other helps clients make a deal.

See also: SEA’s AI buildout is caught in a global memory squeeze

The shortage is real, Li tells Tech in Asia at the SuperAI conference in Singapore, adding that Compute Exchange could not fulfill half the requests for quotes that it has received.

Carmen Li, ceo of Silicon Data and Compute Exchange, speaks during the SuperAI conference in Singapore in June. / Photo Credit: SuperAI

GPUs and tokens have become the largest costs for many AI companies, Li says. It’s not a problem that can be easily fixed, as several factors including hoarding, export controls, a single dominant foundry, and government policies are pushing up prices.

The shortage is so severe that even rental prices for older chips have gone up. Colette Kress, chief financial officer of Nvidia, told investors in May that rental prices for the chipmaker’s H100 GPUs rose 20% in the first months of 2026. Prices for the older A100 GPUs climbed about 15%.

While the H100 is still the workhorse of most GPU cloud services, Nvidia has newer chips, including the H200 and the Blackwell microarchitecture.

But Li says that GPU compute is a far more fungible commodity than oil or electricity. A buyer short on local capacity can route work to servers in another region.

Why Southeast Asia

Hitting smaller players