Nvidia Rubin memory cuts and HBM shortage

China reported blocking most H200 imports with approximately 600,000 chips in Hong Kong; requiring domestic chips for inference, permitting Nvidia only for training

Oracle scrapped a five-year, $7 billion contract to supply Tencent with chips via southeast Asian data centers after US officials objected, closing a route Chinese firms had used to access advanced processors unavailable domestically.

Nvidia is separately reported to be financing new CoWoS packaging capacity at SPIL, an ASE Technology subsidiary, to relieve a packaging bottleneck that persists even after processors and memory are secured, as TSMC's CoWoS lines are fully allocated through at least mid-2027.

TSMC denied it was drafting 3D DRAM plans at Nvidia's request, closing off a reported bid to add TSMC as a fourth memory supplier.

It said DRAM entry could not ease shortages for at least five years. Nvidia is reportedly asking Samsung, SK Hynix, and Micron to build its NVHBM base dies on TSMC's 3nm process instead, giving TSMC memory revenue without direct DRAM exposure. Chinese AI firms ordered about 6 million RTX Pro 5500 GPUs totaling $70 billion; Nvidia told partners it can produce at most 3 million cards in 2027.

Micron paused construction of Fab 11E, a $3-4.5 billion HBM and DRAM facility in Taoyuan, Taiwan, after a union demanded 15% of operating profit in cash, removing a planned source of 2028 capacity.

TrendForce reported top-end GPU lead times jumped to 45 weeks as Nvidia's Rubin transition begins, bottlenecked by substrates and capacitors running at more than three times balanced lead times. Hard drive lead times reached 50 weeks, ABF substrate lead times 48-56 weeks, and server CPU shortages are also emerging.

Nvidia publishes blog arguing full-stack codesign delivers up to 30x higher agentic throughput per dollar on Vera Rubin NVL72 vs GB300 NVL72

Nvidia's shipping Rubin GPU delivers HBM4 bandwidth of 19.2 TB/s and NVLink of 3.0 TB/s, below the previously announced 22 TB/s and 3.6 TB/s, Nvidia noting earlier figures were labeled preliminary.

Micron's 2026 HBM4 output is sold out at 60-70% of ordered volumes, Samsung is seeking $4 per gigabit for 2027 HBM4, triple its HBM3E price, and TSMC is reportedly raising CoWoS prices up to 20% with capacity booked through mid-2027. Apple CEO John Ternus called memory pricing 'unprecedented' and forecast holiday gross margin at roughly 45%, down from 48%.

TrendForce: top-end GPU lead times jumped to 45 weeks; substrate and capacitor bottlenecks delaying Rubin server builds

Sam Altman met roughly 50 Samsung executives in South Korea to strengthen memory chip and data center partnerships

Epoch AI data shows AI chip compute in use growing 3.4x per year; quarterly chip sales growth slowed to 2.5x since early 2024

SK Hynix reports 104.6 trillion won quarterly revenue at 80% operating margin

DRAM contract prices reported up ~80% in Q2 and ~40% in Q3, with Q4 projected at ~20%; Bernstein argues slowdown narrative is misleading

Jensen Huang said at GTC Berlin that memory oversupply by 2029 is unlikely, pushing back on the South Korean analyst forecast

Micron CEO Sanjay Mehrotra tells investors demand for memory will exceed supply through at least 2028

Micron reported in high-volume production of 36 GB 12-layer HBM4 for Vera Rubin at above 11 Gb/s per pin

Silicon Analysts reports Nvidia testing at least three downgraded HBM configurations for Rubin Ultra

Micron's entire 2026 HBM4 production sold out; customers receiving only 60-70% of ordered volumes

Nvidia halved SOCAMM capacity on Vera Rubin Superchip due to LPDDR5X constraints; SemiAnalysis reports mainstream Rubin Ultra SKU at 192 GB 8-Hi HBM4

TrendForce reports Nvidia began evaluating lower-spec HBM configs for Rubin Ultra in early Q3 2026, with DRAM shortage expected through 2027

TSMC CoWoS packaging capacity reported fully allocated through at least mid-2027