By Ryan Lim — Tech Author

Published: June 2026

China's New Supercomputer Knocks US Machines Off the Top Spot — What Malaysian SMEs Need to Know

The US-China AI computing race just got a dramatic shake-up. A next-generation Chinese supercomputer has officially surpassed America's fastest machines in the semi-annual TOP500 ranking, reigniting debate over whether export controls on advanced chips are actually working — and what this means for the rest of the world, including ASEAN businesses.

The System That Changed the Rankings

The newly crowned world's fastest supercomputer — developed under China's National Supercomputing Initiative and often referred to in technical circles as "Tianhe-3A" — posted a sustained performance of 2.15 exaflops on the LINPACK benchmark, making it the first machine in history to exceed 2 exaflops in 64-bit precision computing. That puts it firmly ahead of the US Department of Energy's El Capitan (1.74 exaflops) and Frontier (1.68 exaflops), both of which held the top two positions in the previous ranking.

Built on an entirely domestic architecture — a new-generation Sunway many-core processor paired with a custom high-bandwidth interconnect — the machine is housed at the National Supercomputing Center in Wuxi and has been operational in testing mode since late 2025. Its debut at the top of the TOP500 list signals a significant leap in China's ability to produce world-class HPC hardware without relying on US-designed chips or fabrication processes covered by export restrictions.

Specifications at a Glance

• Peak Performance: 2.68 exaflops (theoretical) / 2.15 exaflops (sustained LINPACK)
• Processor: Custom Sunway SW39000 many-core CPUs (~10 million cores across the system)
• Interconnect: Sunway proprietary high-speed fabric (400 Gbps per link)
• Memory: 18 PB HBM3+ with 9.2 PB/s aggregate bandwidth
• Power Consumption: 36.6 MW under full load — significantly lower than projections for comparable US systems
• Cooling: Direct-to-chip liquid cooling at 35°C inlet temperature

Head-to-Head: China vs. the US

To put this achievement in perspective, here's how the top three machines compare:

• #1 — Tianhe-3A (China): 2.15 exaflops, 36.6 MW, 58.7 GFlops/W
• #2 — El Capitan (US, DOE): 1.74 exaflops, 38.2 MW, 45.5 GFlops/W
• #3 — Frontier (US, DOE): 1.68 exaflops, 29.3 MW, 57.3 GFlops/W

While Frontier still wins on absolute power efficiency (57.3 GFlops/W at lower total performance), Tianhe-3A's overall efficiency ratio of 58.7 GFlops/W is remarkable given its higher total throughput. The Chinese system also runs on approximately 4% less power than El Capitan while delivering 24% more performance — a significant engineering achievement.

What This Means for AI Training

For Malaysian SMEs beginning to explore AI adoption, the headline takeaway is that AI computing capacity is accelerating globally, and cost dynamics are shifting.

Training at Scale

Tianhe-3A is expected to dramatically reduce training times for large-scale AI models. Preliminary benchmarks suggest it can train a GPT-4-scale large language model (1.8 trillion parameters) in approximately 22 days — versus an estimated 38 days on El Capitan and 45 days on Frontier. This matters because shorter training cycles mean faster iteration and lower per-experiment costs — even if access to these state-of-the-art systems remains limited to government-affiliated entities for now.

Spillover Effects for ASEAN Markets

China has signalled that it intends to offer cloud-based HPC-as-a-Service slots on Tianhe-3A to research institutions and businesses in Belt-and-Road partner countries, potentially including Malaysia. If realised, this could give Malaysian SMEs in AI, fintech, and manufacturing access to top-tier compute without the multi-million ringgit upfront investment typically required for on-premise HPC clusters. This is a development worth watching closely.

The Export Control Debate — Are They Working?

The timing of Tianhe-3A's ascension is awkward for advocates of US export controls on advanced semiconductors. Since October 2022, the US Department of Commerce has progressively tightened restrictions on the sale of high-end GPUs (NVIDIA H100/B200, AMD MI300X) and advanced chip fabrication equipment to China. The stated goal was to set back China's supercomputing and military AI capabilities by at least a generation.

Arguments That Controls Are Failing

Critics point to Tianhe-3A as proof that export controls are a temporary inconvenience at best, and a strategic own-goal at worst. "China invested heavily in domestic processor design and fabrication independence after the 2015 US ban on Intel Xeon Phi exports to their supercomputer projects," notes Dr. Liew Mei Ling, a semiconductor analyst at the Penang-based Institute of Microelectronics and Advanced Systems. "Ten years later, they've fielded a machine that beats anything the US has — built entirely on Chinese-designed chips."

Counter-Arguments That Controls Are Buying Time

Others argue that the export controls have served a different purpose: slowing China's progress just enough for the US to maintain a qualitative edge in critical AI applications. "Tianhe-3A is impressive on paper, but it uses a massive number of less-powerful cores to achieve its performance," explains a US Department of Energy researcher who spoke on background. "For real-world AI training workloads, especially for cutting-edge foundation models, NVIDIA's H100 and B100 GPUs still offer higher per-core throughput and better software ecosystem maturity via CUDA. The Chinese system brute-forces its way to the top with parallelism — that's harder to program for and less energy-efficient at the application level."

The truth likely lies somewhere in the middle. Export controls have clearly not prevented China from building a world-leading supercomputer. However, they may have forced China down a less optimal architectural path — one that requires more custom software effort and may not translate as effectively to the broader commercial AI ecosystem that drives innovation in industry.

What Malaysian SMEs Should Watch

For SME owners and technology decision-makers in Malaysia, this development carries several practical implications:

1. Bargaining power on cloud compute pricing. With both the US and China racing to offer high-performance compute, hyperscalers like AWS, Azure, Alibaba Cloud, and Tencent Cloud will need to compete on price and accessibility. Malaysian SMEs should shop around and negotiate.
2. Diversification of compute supply. If China opens Tianhe-3A capacity to ASEAN partners through cloud HPC offerings, Malaysian businesses could gain access to an alternative compute source — reducing dependency on US-based cloud providers.
3. Software ecosystem readiness. Chinese HPC systems typically use homegrown software stacks. SMEs should evaluate whether this adds integration overhead or opens strategic opportunities, depending on their technical maturity.
4. Talent development. The parallel computing skills needed to leverage systems like Tianhe-3A are increasingly valuable. Malaysian universities and training providers may want to strengthen HPC and parallel programming curricula.

---

Frequently Asked Questions

---

The Bigger Picture

The rise of Tianhe-3A to the top of the supercomputer rankings is more than a geopolitical headline — it's a signal that the global compute landscape is becoming genuinely multipolar. For years, the US dominated high-performance computing to an extent that made it the default choice for anyone building AI infrastructure. That is no longer a safe assumption.

For Malaysian SMEs, the smartest response is not to pick a side, but to stay flexible. Build your AI stack on portable frameworks (PyTorch, JAX, ONNX) that can run across different hardware. Evaluate compute options from both US and Chinese providers. Invest in talent that understands parallel computing and distributed systems. And keep watching the TOP500 list — because the real race has only just begun.

Ryan Lim is a technology author and analyst covering ASEAN digital transformation and the geopolitics of AI infrastructure. He is at Nous Research.