NVIDIA H200 exports to China were blocked by 25%, and HBM skyrocketed by 150%！
Nvidia H200 Faces 25% Export Hurdles to China; HBM Prices Surge 150%! Samsung, SK Hynix, and Micron Ramp Up HBM Production for Fierce Competition; DRAM Shortage Hits 40%, Domestic Substitution Reaches 15%
I. Nvidia H200 Chip Deliveries to China
In February 2026, Nvidia plans to deliver 5,000 to 10,000 sets of H200 AI chip modules (approximately 40,000 to 80,000 chips) to China. This marks the first substantive move following the Trump administration's policy of "allowing H200 exports to China but imposing a 25% levy". Although the H200 has been replaced by the Blackwell series, its performance remains six times that of the "customized" H20, enabling Chinese AI laboratories to build supercomputers approaching the top-level standards in the United States. However, Chinese authorities have not yet approved any purchase orders, leaving the final delivery schedule uncertain.
Supplementary Perspectives:
- Accelerated Domestic Substitution: Domestic AI chips such as Huawei Ascend and Cambricon have been included in official procurement lists and have achieved commercial application in government and financial sectors. In 2025, the market size of domestic AI chips grew by over 120% year-on-year, with the technical route shifting from "emergency breakthroughs" to "systematic formation".
- Backdoor Concerns Triggered by Locator Requirements: The U.S. requirement to install locators in H200 chips has raised concerns about "backdoors". In July 2025, China's Cyberspace Administration of China (CAC) interviewed Nvidia regarding vulnerabilities in the H20 chips.
II. Storage Cost Transmission Behind Cloud Gaming Price Hikes
Starting from January 2026, Nvidia's GeForce Now cloud gaming service will implement a "100-hour monthly usage cap" for all paid members, with additional hours charged at a rate of 15 hours per block. For users gaming 4 hours per day on average, the monthly fee for the premium plan will increase from $19.99 to $31.97, representing a 60% price hike.
The surging demand for AI servers has squeezed consumer-grade storage production capacity. The skyrocketing price of DDR5 memory has led to a sharp increase in hardware procurement costs for cloud service providers. In Q4 2025, the contract price of DDR5 rose by 35% quarter-on-quarter, increasing the cost per cloud gaming server by over $2,000.
Nvidia is tapping into the paying potential of high-net-worth users through a "basic service + overage charge" model, and its cloud gaming revenue is expected to exceed $5 billion in 2026.
III. Memory "Super Cycle": Global Supply Chain Disruption Amid AI Demand Siphoning
The global memory market has experienced an "earthquake-like" price surge: the spot price of DDR5 soared from $6.84 per 16Gb in September 2025 to $27.2 in December (peaking at $37 intraday). DRAM prices rose by 50% for the full year, and are projected to increase by another 20% in early 2026. The shortage cycle may extend to 2027-2028, with TeamGroup warning of a supply shortage peak in the first and second quarters of 2026.
- Explosive Demand for AI Servers: A single AI server requires 32-64 DDR5 memory modules (4-8 times that of ordinary servers). Companies like OpenAI account for 40% of global DRAM production capacity with their monthly orders, directly squeezing consumer-grade production capacity.
- HBM Technology Squeeze: The price of a single HBM4 chip has exceeded $500, representing a 150% increase compared to HBM3E, leading to a further contraction in general-purpose DRAM production capacity.
IV. Capacity Expansion by Samsung and SK Hynix: Short-Term "Remote Water" and Long-Term Game
Samsung plans to increase its monthly production capacity of 1c DRAM to 200,000 wafers by the end of 2026, while SK Hynix has advanced the mass production schedule of its new Cheongju M15X plant to 2026 (with an initial monthly production of 35,000 wafers). However, both manufacturers are diverting 70% of their production capacity to HBM, resulting in a widening supply gap in consumer-grade DRAM.
SK Hynix's 2026 HBM production capacity has already been locked in by Nvidia and Microsoft, meaning that the expansion of general-purpose DRAM production will be like "remote water cannot quench immediate thirst" for the consumer market.
- Technological Generation Gap Competition: Samsung's 1c DRAM yield rate has reached 80%, but SK Hynix is seizing the AI inference market through GDDR7 and SOCAMM2 modules, with fierce competition between the two for Nvidia's orders.
V. Micron's "Desperate Survival": All-In Bet on HBM
In the first quarter of its fiscal year 2026, Micron reported revenue of $13.64 billion (up 57% year-on-year), net profit of $5.482 billion (up 135% year-on-year), and a gross margin of 56.8%. Its core strategy is to exit the consumer-grade market (discontinuing the Crucial brand) and fully commit to the HBM business, which accounts for over 55% of its revenue.
Following Micron's exit, consumer-grade DDR5 prices have further spiraled out of control, with Samsung's products of the same specification exceeding 1,300 yuan. China's ChangXin Memory Technologies is accelerating efforts to fill the gap.
Micron's HBM4 development lags behind Samsung by two months. The company is catching up through a strategy of "overwhelming samples + low pricing", but with a yield rate of only 60%, it faces significant cost pressures.
VI. Long-Term Transformation from "Cyclical Fluctuations" to "AI-Driven"
- Consumer Electronics Under Pressure: The cost proportion of memory in mid-to-low-end mobile phones has risen to 30%, forcing brands like Redmi to potentially increase prices by 10%-15%.
- Soaring Data Center Costs: The proportion of storage costs in the BOM (Bill of Materials) of AI servers has increased from 15% to 35%, with OpenAI's "Stargate" project incurring monthly storage expenses exceeding $1 billion.
- Accelerated Technological Iteration: HBM is evolving towards HBM5 (with bandwidth exceeding 10TB/s), and compute-in-memory chips may disrupt traditional architectures.
- Domestic Substitution Breakthrough: ChangXin Memory Technologies has achieved mass production of 10nm DDR5, and Yangtze Memory Technologies' 3D NAND yield rate has exceeded 90%. The domestic storage market share is expected to surpass 25% in 2026.
Survival Rules in the Restructured Global Industrial Chain
AI is pushing the semiconductor industry towards a "dual-track system":
- High-End Track: AI-specific chips such as HBM and GDDR7 have become a battleground for industry giants, with continuously rising technological barriers and capital investment thresholds.
- Previous: DRAM has soared by more than 500% since 25 years, and Samsung/Micron/Dell/Apple have followed suit
