/Tech22h ago

Zephyr and Teortaxes debate whether government intervention is required to scale China's domestic GPU and CPU supply by 2027

Zephyr identifies CPU and memory as the primary bottlenecks.

17271105988.6K
Zephyr
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#440
Original post
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex#440inTech

@zephyr_z9 Everything that can be made in China will eventually have supply expanded, no? Or do you mean the govt needs to backstop the demand to see that through?

Zephyr@zephyr_z9

In 2025, cost of building 1GW of compute used to be 2x-2.3x cheaper in China using Chinese hardware (GPU/ASICs) On a FLOPS basis, cost was equal cuz Nvidia hardware has a 2.5x-3x Flop/W advantage over Chinese GPUs (fp16) But if u look at recent 2026 data, component cost inflation (CPUs, storage, memory, optics/networking) and GPU scarcity has led to severe price increases Atlas 800T server used to cost 800k-1M RMB when it was launched last year Now u have to pay 1.8M-2M RMB for it

2:56 PM · Jun 9, 2026 · 2.9K Views
Sentiment
Sentiment building, check back later.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS31.1KLIKES117REPLIES8
Zephyr@zephyr_z9

The gap is much larger on a FLOP basis Ascend is 2PFLOPS on fp4 while Rubin is 50PFLOPS FLOPS/W is 21 TFLOPS/W for Rubin, while it is less than 3 TFLOPS/W for Ascend 950DT

13hViews 31.1KLikes 117Bookmarks 27
BOOKMARKS29RETWEETS7
Zephyr@zephyr_z9

"I think all lower-end component supply will expand wildly over 2026-27"

But the price rise in lower end components isn't the problem CPUs, memory/storage, GPU scarcity are the biggest problem

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Pretty crazy that it was even equal on $/fp16 FLOPS as late as in 2025 The story of Nvidia supply chain's technical lead is much more impressive I think all lower-end component supply will expand wildly over 2026-27

14hViews 25.8KLikes 113Bookmarks 29
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

Pretty crazy that it was even equal on $/fp16 FLOPS as late as in 2025 The story of Nvidia supply chain's technical lead is much more impressive I think all lower-end component supply will expand wildly over 2026-27

Zephyr@zephyr_z9

In 2025, cost of building 1GW of compute used to be 2x-2.3x cheaper in China using Chinese hardware (GPU/ASICs) On a FLOPS basis, cost was equal cuz Nvidia hardware has a 2.5x-3x Flop/W advantage over Chinese GPUs (fp16) But if u look at recent 2026 data, component cost inflation (CPUs, storage, memory, optics/networking) and GPU scarcity has led to severe price increases Atlas 800T server used to cost 800k-1M RMB when it was launched last year Now u have to pay 1.8M-2M RMB for it

22hViews 28.9KLikes 28Bookmarks 2
Frank@Lordvultures

@zephyr_z9 Do you consider CCL and MLCC lower end components?

13hViews 19