Nvidia’s AI boom is looking more Asia-heavy than ever.
According to data compiled by Bloomberg, Asian suppliers now account for roughly 90% of Nvidia’s production costs. That is a big jump from around 65% just a year earlier, and it shows how deeply Nvidia’s current AI empire depends on the region’s chip, memory, and server manufacturing network.
For Malaysia and SEA readers, this matters because Nvidia is not just some distant Silicon Valley company anymore. Its GPUs power the cloud services, AI tools, gaming laptops, workstations, data centres, robotics projects, and even future automotive systems that local businesses and consumers will eventually use. When Nvidia’s supply chain gets tighter, pricing and availability can ripple all the way down to us.
The Asia stack behind Nvidia’s AI machine
The main suppliers are familiar names in the semiconductor world. TSMC handles advanced chip fabrication, SK hynix and Samsung supply high-bandwidth memory and other memory components, while companies like Foxconn and Quanta assemble servers.
That setup already supports Nvidia’s huge data centre GPU business. But now, Nvidia is pushing harder into what it calls physical AI — robots, autonomous machines, and automotive platforms that need serious onboard computing.
One key example is Jetson Thor, Nvidia’s robotics platform released last August. It uses the Blackwell GPU architecture and is made on TSMC’s 3nm process. The high-end T5000 module delivers 2,070 FP4 TFLOPS and comes with 128GB of LPDDR5X memory. Nvidia also introduced a cheaper T4000 version at CES 2026, offering 1,200 FP4 TFLOPS and 64GB of memory at US$1,999 per unit in volume.
Both modules use Arm Neoverse-V3AE CPU cores and LPDDR5X memory from Samsung or SK hynix. In other words, even Nvidia’s robotics push is leaning on the same Asian suppliers already feeding the data centre GPU boom.
Robotics and cars are joining the queue
Jetson Thor is not sitting in a vacuum. Partners including Boston Dynamics and Amazon Robotics are building on the platform, while LG has confirmed that it is exploring a strategic collaboration with Nvidia around physical AI and robotics.
Then there is DRIVE AGX Thor, Nvidia’s Blackwell-based automotive system-on-chip. That means robotics and automotive hardware are now competing for TSMC 3nm wafer capacity alongside Blackwell data centre GPUs.
The good news is that these physical AI products do not require TSMC’s CoWoS advanced packaging, which remains the big bottleneck for Nvidia’s data centre GPU production. The not-so-good news: they still need limited 3nm wafer capacity and Asian-sourced LPDDR5X memory, both of which are already under pressure.
Older Jetson modules are getting squeezed out
The memory crunch is also affecting Nvidia’s older hardware. At the end of April, reports said Nvidia had sped up end-of-life plans for its Jetson TX2 and Xavier modules because LPDDR4 supply had become too limited to keep production going.
Samsung has shifted away from LPDDR4 manufacturing, while AI demand is pulling memory capacity toward higher-margin products. That pushes Jetson customers toward newer Orin or Thor modules, which use LPDDR5X from the same suppliers already stretched by demand for HBM and data centre DRAM.
TSMC’s CoWoS advanced packaging capacity for data centre GPUs is reportedly growing at an 80% compound annual growth rate, according to comments from TSMC’s head of North American packaging to CNBC. Still, chips made at TSMC’s Arizona Fab 21 currently ship back to Taiwan for packaging.
Nvidia has committed to US$500 billion in U.S. server manufacturing with Foxconn and Wistron, while Amkor and SPIL are building advanced packaging facilities in Arizona. But those operations are not yet running at full production scale.
For SEA, the takeaway is simple: Asia is not just supporting the AI boom — Asia is the AI boom’s hardware backbone. And as Nvidia expands from cloud GPUs into robots and cars, the pressure on chip and memory supply could stay gila tight for a while.
Source: Tom's Hardware