Chinese think tank advises data centers to prioritize NVIDIA GPUs over domestically-produced alternatives due to superior cost-performance and lower engineering risk.
On Sunday, the Chinese-government-backed research institute and think tank CAICT issued a detailed report regarding China's computational power, warning data centers to keep using Nvidia-based solutions. The institute cited "high costs" and "complex engineering" as key hurdles that data centers would face if they attempted to transfer their models to Chinese chips. This recommendation comes despite significant pressure from US export restrictions: the US government has barred Nvidia from exporting its high-performance A100 and H100 GPUs to China, and another round of restrictions came last year when the Biden Administration restricted exports of Nvidia's A800 and H800 alternatives to China. Most recently, Nvidia appears to have stopped taking orders for its currently permitted HGX H20 GPU.
While many domestic Chinese chip manufacturers have responded to these constraints by developing their own alternatives, CAICT advises data centers to continue relying on Nvidia. The institute stated that "it involves complex engineering to transfer models trained on Nvidia GPUs to domestic solutions due to differences in hardware and software." This technical barrier represents a significant obstacle to transitioning away from Nvidia infrastructure, even as supply becomes increasingly constrained.
New data centers in China face numerous compounding challenges, including the lack of GPU solutions, supply mismanagement, and rising computing power fragmentation—essentially the inefficient use of computer resources like GPUs. Data centers must engage in a "mix-and-match" game, merging solutions from different vendors in hopes of getting their systems operational. CAICT warned that "there are big discrepancies on hardware in IDCs, such as in GPUs, AI accelerators and network structures, which made it harder to manage and dispatch hardware resources to accommodate for differential computing needs of AI tasks, further impeding the use rate." Much of this inefficiency stems from incompatible software and varying hardware across vendors.
Despite ambitious goals of semiconductor self-sufficiency, China remains heavily dependent on imported chips. The country, despite investments worth tens of billions of dollars, still imported $350 billion of chips in 2023. Players like Intel and AMD continue struggling to keep pace with Nvidia due to Team Green's strong integration between hardware and software in the AI market, suggesting that homegrown Chinese GPUs might, at best, achieve parity with Nvidia in a few years.