US tightens export controls on Nvidia AI chips to China

🔑 Enhanced Key Takeaways

•The loophole allowing advanced AI chips to reach Chinese entities via overseas subsidiaries was created in May 2025 when the Trump administration declined to enforce the Biden-era AI Diffusion rule.
•Industry sources estimate that hundreds of thousands of advanced AI chips, including Nvidia's Blackwell, may have been exported through this loophole before its closure.
•The new guidance extends license requirements to entities headquartered in China, irrespective of their physical location, to prevent circumvention of export controls.
•Nvidia has reportedly ceased production of its H200 chips specifically for the Chinese market, reallocating that capacity to its next-generation Vera Rubin platform due to these controls.
•The US export control strategy has evolved to primarily target computing performance, moving beyond earlier considerations of interconnection bandwidth as a key parameter.

📊 Competitor Analysis▸ Show

Feature	Nvidia Blackwell (GB200)	Nvidia Rubin (R100)	AMD Instinct MI350X
Architecture	Blackwell	Rubin	CDNA4
Process Node	TSMC 4NP	TSMC 3nm	TSMC 3nm / 6nm FinFET
Transistors	208 billion	336 billion	185 billion
FP4 Performance	20 PFLOPS (single chip)	50 PFLOPS (single GPU)	9.2 PFLOPs
HBM Memory	186GB HBM3e	288GB HBM4	288GB HBM3E
HBM Bandwidth	8 TB/s	22 TB/s	8 TB/s
Interconnect	NVLink 5 (1.8 TB/s per GPU)	NVLink 6 (3.6 TB/s per GPU)	Infinity Fabric Link (153.6 GB/s bidirectional per GPU)
TDP	Up to 1000W (air-cooled PCIe)	Max Q (~1.8 kW), Max P (~2.3 kW)	1000W

🛠️ Technical Deep Dive

Nvidia Blackwell Architecture (GB200):
- Manufactured using a custom TSMC 4NP process, featuring 208 billion transistors.
- Employs a dual-die design connected by a 10 terabytes per second (TB/s) NV-HBI interface, allowing it to function as a unified single GPU.
- Includes a second-generation Transformer Engine with custom Tensor Core technology, supporting FP4 AI for accelerated inference and training of large language models (LLMs) and Mixture-of-Experts (MoE) models.
- Features fifth-generation NVLink, providing 1.8 TB/s of total bandwidth per GPU, and a NVLink Switch supporting up to 130 TB/s GPU bandwidth for scaling across multiple servers.
- Incorporates a dedicated Decompression Engine capable of decompressing data at up to 800GB/s, and a Reliability, Availability, and Serviceability (RAS) Engine for fault identification and minimized downtime.
Nvidia Rubin Architecture (R100):
- Built on TSMC's 3nm process with a dual-die design, containing 336 billion transistors.
- Features 288GB of HBM4 memory, delivering up to 22 TB/s of bandwidth per GPU.
- Utilizes NVLink 6, NVIDIA's next-generation interconnect, which doubles bandwidth to 3.6 TB/s per GPU compared to Blackwell's NVLink 5.
- The Rubin platform integrates the Vera CPU (227 billion transistors, 88 custom Arm Olympus cores) and can include the Groq 3 LPU as a dedicated low-latency inference accelerator.
- Designed for agentic AI workloads, offering 5x rack-level inference performance and 10x lower inference token cost compared to Blackwell.
AMD Instinct MI350X:
- Based on the 4th Gen AMD CDNA™ architecture (CDNA4), built on TSMC's 3nm and 6nm FinFET processes, with 185 billion transistors.
- Features 16384 shading units, 1024 texture mapping units, and 1024 tensor cores.
- Equipped with 288 GB of HBM3e memory, connected via an 8192-bit memory interface, providing 8 TB/s of peak memory bandwidth.
- Supports expanded MXFP6 and MXFP4 datatype for maximizing computational throughput and energy efficiency in AI inference.
- Interconnected using AMD Infinity Fabric™ Link, enabling high-bandwidth GPU-to-GPU communication, and compatible with Universal Base Board (UBB 2.0).
- Integrates seamlessly with the AMD ROCm™ software stack, supporting major AI and HPC frameworks.

🔮 Future ImplicationsAI analysis grounded in cited sources

China will accelerate its domestic chip development and production capabilities.

US export controls incentivize China to invest heavily in indigenous semiconductor R&D and manufacturing to achieve self-sufficiency.

US chipmakers like Nvidia and AMD will face continued market volatility and strategic challenges.

Shifting US export policies create uncertainty, impact revenue projections, and necessitate product modifications for specific markets, affecting global sales strategies.

The US will intensify efforts to secure multilateral cooperation from allies on export controls.

The effectiveness of US export restrictions on advanced AI chips is highly dependent on the coordinated efforts and enforcement by key allies like Japan and the Netherlands.

⏳ Timeline

2022-10-07

US implements initial sweeping export controls on advanced computing chips and semiconductor manufacturing equipment to China.

2023-10

US further restricts types of semiconductors sold to China, including Nvidia's A800 and H800 chips, and adds 13 Chinese firms to the Entity List.

2024-03-18

Nvidia officially announces the Blackwell architecture at GTC 2024.

2024-12-02

US BIS issues a Final Rule adding 140 entities to the Entity List and an Interim Final Rule introducing new Foreign Direct Product rules, including controls on high-bandwidth memory.

2025-05

The Trump administration declines to enforce the Biden-era AI Diffusion rule, creating a loophole for advanced chip exports to Chinese overseas subsidiaries.

2025-06-12

AMD launches the Radeon Instinct MI350X.

US tightens export controls on Nvidia AI chips to China

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (24)

👉Related Updates

Samsung Secures $200B Broadcom AI Infrastructure Deal

CATL Q2 earnings show growth despite margin pressure

Dongshan Precision Faces Security Challenges