May 13, 2025

Yield breakthroughs, export chess moves, and a $600 million lifeline

Round-up

Highlights

  1. Samsung’s 2 nm bid wins fresh credibility. Yield on the gate‑all‑around node has reportedly pushed past 40 %, putting the foundry within reach of first external 2 nm orders from Nvidia GPUs and Qualcomm’s next Snapdragon flagship and threatening TSMC’s lock‑in. 1
  2. Washington’s export net tightens from the tooling side. Korean equipment maker Hanmi Semiconductor is said to have told Chinese clients it will stop shipping its market‑dominant thermal‑compression bonders—critical for stacking HBM3E—just as Beijing races to build its own AI memory supply‑chain. 2
  3. Gallium whiplash: Beijing softens—slightly. One day after yesterday’s DCP newsletter flagged China’s new traceability regime on strategic minerals, rare‑earth insiders now say U.S. buyers are getting fast‑track export licences as part of the 90‑day tariff truce. 3

Other developments

  • Wolfspeed’s junior creditors table a $600 m rescue to avert bankruptcy4
  • Apple wafer supplier IQE eyes moving output to the U.S. to dodge looming chip tariffs5
  • OSAT house ChipMOS launches a $16 m buy‑back amid tariff‑clouded Q1 results6
  • Saudi wealth fund unveils AI champion Humain—and lines up Groq inference silicon7
  • Microsoft Build leak hints at “Maia 2” accelerator reveal later today8

Did you know? Hanmi controls about 90 % of the global market for TC bonders used in HBM3E stacking—one vendor, one chokepoint.2


In‑depth

1 · Government & Corporate Policy

  • Hanmi TC‑bonder halt foreshadows a new front in export controls

    • EE Times China says Chinese DRAM fabs were told shipments are suspended, likely under U.S. pressure.2
    • A stoppage would slow Chinese HBM plans by up to two years and raise wafer costs ~30 %.
  • Rare‑earth paperwork eases for U.S. buyers

    • MOFCOM insiders expect licence approvals in “under 45 days,” a sharp contrast to the indefinite holds reported last week.3
    • Tesla and Micron are among early applicants for neodymium and terbium quotas.
  • Riyadh‑Washington chip corridor

    • During President Trump’s Gulf tour, a draft MoU would grant Saudi Arabia priority access to U.S. AI accelerators in exchange for critical‑minerals supply to U.S. fabs.9
    • Parallel FT reporting shows the PIF‑backed Humain venture anchoring the demand side.7
  • IQE weighs U.S. fab footprint

    • New CEO Jutta Meier says dual‑sourcing inside America is “on the table” to blunt potential 25 % tariffs.5
    • Strategy includes offloading its Taiwan wafer unit to fund cap‑ex.

2 · Economics, Finance & Business Outlook

  • Samsung inches toward 2 nm revenue

    • Chosun Biz and Sedaily claim NVIDIA/Qualcomm test wafers have passed PVT at Hwaseong.1
    • Orders would diversify Samsung Foundry’s book for the first time since 2022.
  • Wolfspeed rescue proposal

    • Junior creditors offer $600 m to refinance a 2026 convertible and head off Apollo‑led senior lenders.4
    • The SiC wafer specialist still awaits $750 m in delayed CHIPS‑Act support.
  • ChipMOS buys time with buy‑back

    • Board authorizes NT$525 m (US$16.4 m) repurchase as Q1 free‑cash flow hits US$25 m.6
    • Management signals opportunistic support ahead of expected Q2 margin dip on tariff uncertainty.
  • Smartphone demand watch

    • Samsung’s ultraslim S25 Edge launches May 30; execs warn tariffs could dent unit volume even as AI features lift ASPs.10

3 · Technology & R&D

  • 2 nm yield milestone

    • Internal data peg Samsung’s 3 nm GAA yield at > 60 %, 2 nm at > 40 %, narrowing the gap with TSMC N2 (~80 %).1
    • Qualcomm’s ‘Snapdragon 8 Elite 2’ could debut on Samsung’s node in 2026.
  • Thermal‑compression bonding as a geopolitical lever

    • Hanmi’s tool freeze underscores how advanced packaging equipment joins EUV scanners on the strategic‑tech list.2
    • Chinese rival PrecisioNext still limited to TC‑NCF processes, stalling 12‑high stacks.
  • Microsoft’s Maia 2 rumor

    • Pre‑Build leaks point to a successor to the 5‑nm Maia custom AI chip, aiming at 250 W envelope for LLM inference.8
    • Coupled with new “agentic” Copilot features, Redmond edges closer to vertical silicon control.
  • Groq wins Saudi inference deal

    • Newly launched Humain will deploy California‑made Groq LPU clusters in a 200 MW Riyadh datacenter, bypassing Nvidia export hurdles.7
    • Phase‑one capacity targets 500 petaOPS on chat‑optimized INT‑8 workloads.

Footnotes