May 16, 2025

Round-up

Highlights

1. UAE gets a legal path to buy cutting-edge U.S. AI chips. Less than 24 hours after Washington scrapped the global “AI-Diffusion” rule (covered in yesterday’s DCP), presidents Trump and Sheikh Mohamed unveiled a framework that lets Emirati cloud operator G42 purchase top-tier Nvidia GB-series parts while keeping U.S. compliance teams in the loop ¹²³.
2. Tariffs slam the tool chain. Applied Materials’ Q3 outlook came in a half-billion dollars light and management blamed “customer caution linked to trade policy” ⁴. Equipment slowdowns tend to ripple through foundry cap-ex six to nine months later.
3. Nvidia doubles down on China R&D. Nvidia has started leasing a Shanghai R&D centre even as its export-restricted H20 struggles to gain traction, signalling it will court Chinese talent—and regulators—rather than retreat ⁵.

Other developments

• India approved a $435 m Foxconn-HCL display-driver fab in Uttar Pradesh ⁶⁷
• Vietnam green-lit phase II of CT Semiconductor’s OSAT plant; Amkor accelerates its $1.6 bn Bac Ninh package line ⁸
• Pliops unveiled a 3-D-NAND-based PCIe card that extends GPU memory “context windows” 8× for AI inference ⁹
• Silicon Motion’s new SM2324 USB4 controller enables 32 TB portable SSDs at 4 GB/s ¹⁰
• Nvidia postponed its Blackwell-era SOCAMM stacked-memory technology to “Rubin” GPUs ¹¹
• Malaysia’s E&E trade group warned local suppliers cannot absorb new U.S. tariff costs ¹²

Did you know? The UAE-U.S. accord guarantees G42 up to 500 000 top-spec Nvidia GPUs per year—equal to roughly one-third of Nvidia’s total 2024 H100/H200 shipment run-rate ³.

In-depth

1. Government & Corporate Policy

• UAE-U.S. AI-chip pathway ¹²³

  • Framework requires chips to reside in data-centres managed by approved U.S. cloud vendors.
  • Follows yesterday’s rollback of the global AI-Diffusion rule; Washington is betting on inspections over blanket bans.

• India’s sixth “Semicon Mission” project ⁶⁷

  • Foxconn–HCL JV will make 20 000 wafers / month of display-driver ICs; commercial output slated for 2027.
  • Reboots Foxconn’s foundry ambitions after last year’s Vedanta collapse.

• Intel fights €376 m EU fine ¹³

  • Chipmaker argued in Luxembourg that the lowered penalty still “double-counts” alleged rebates from 2002-2006.
  • A final ruling could set precedent for future EU tech cases.

• Malaysia sounds tariff alarm ¹²

  • Industry association says local OSAT houses cannot hold prices on U.S. contracts if new levies persist.

2. Economics, Finance & Business Outlook

• Applied Materials trims Q3 guide ⁴

  • Forecast midpoint of $7.2 bn misses bullish analyst models; management cites China-related demand uncertainty.
  • Shares fell 5.8 % in after-hours trading.

• Wistron adds U.S. server capacity for Nvidia ¹⁴

  • Taiwanese ODM will have two Texas plants online by 2H 2026 to build GB-series liquid-cooled clusters.
  • Signals OEMs are localizing AI-server assembly to sidestep future export red-tape.

• Computex preview: AI everywhere ¹⁵

  • Next week’s Taipei show will feature Arm-based PC chips and Blackwell boards; Jensen Huang keynotes again.
  • Taiwanese officials expect record foreign attendance despite tariff jitters.

• TrendForce: Vietnam OSAT boom ⁸

  • Country aims for $25 bn annual semiconductor revenue by 2030; government courting foreign back-end players.

3. Technology & R&D

• Nvidia Shanghai R&D hub ⁵

  • Leased space supports autonomous-driving and deep-learning projects tailored to Chinese partners.
  • Part of a broader strategy to keep a $17 bn China revenue stream alive despite curbs.

• Pliops “memory-tier” accelerator ⁹

  • PCIe Gen5 card caches embeddings on NAND, boosting LLM inference throughput up to 8× and cutting GPU RAM needs.
  • Early trials target retrieval-augmented-generation clusters.

• Silicon Motion SM2324 controller ¹⁰

  • Single-chip USB4 design allows fan-less 32 TB external SSDs at 40 Gb/s, aimed at AI workstation backups.
  • Engineering samples ship Q3 2025.

• Nvidia pushes SOCAMM to Rubin ¹¹

  • New stacked-LPDDR6+ module, originally slated for Blackwell Ultra, delayed one generation as packaging yield lags.
  • Keeps pressure on Micron and Samsung to finalize 36 Gb GDDR7X parts.

Footnotes