June 13, 2025

Round-up

Highlights

1. Beijing delays the $35 billion Synopsys-Ansys merger. China’s antitrust regulator paused the deal just as Washington tightened EDA export rules, raising the stakes for U.S.–China tech détente.¹
2. AMD doubles down on AI: new MI350 GPUs ship on TSMC N3P and ROCm 7 jumps 3.5×. Silicon, software and an OpenAI-influenced roadmap aim squarely at Nvidia’s datacenter crown.²³
3. U.S. Commerce says Huawei can build “no more than 200 k” AI chips next year. Officials warn the gap could close quickly, signaling no near-term easing of export controls.⁴

Other developments

• Deutsche Telekom teams with Nvidia on Europe’s first dedicated industrial-AI cloud.⁵
• Micron issues DDR4 end-of-life notices and hints at price hikes to fund next-gen memory.⁶
• AMD courts Cohere and other AI startups; lessons already shaping the MI450 architecture.⁷

Did you know? Imec’s new enhancement-mode GaN-on-Si transistor sets a record-low contact resistance of 0.024 Ω·mm, a milestone for efficient 6 G radios.⁸

In-depth

1. Government & Corporate Policy

• SAMR halts Synopsys-Ansys approval

  • China’s regulator cited “complexity,” but timing coincides with fresh U.S. curbs on chip-design software.¹
  • Insiders say concessions on Chinese access to IP tools could revive the deal; deadline remains 1 H 2026.

• Huawei AI-chip cap acknowledged in Congress

  • Under-Secretary Jeffrey Kessler testified Huawei can make ≤200 k 7 nm-class Ascend 910C parts in 2025.⁴
  • Commerce is “not planning immediate new rules,” but will keep tightening loopholes.

• EU industrial-AI cloud endorsed

  • Berlin backs Deutsche Telekom-Nvidia plan supplying 10 000 GPUs for factories by 2026—part of a sovereignty push.⁵
  • Project sits outside U.S. export-control perimeter by sourcing chips through Germany.

• AMD–OpenAI design feedback institutionalised

  • OpenAI’s requests on memory topology and FP formats now flow into AMD’s silicon sign-off process, signalling a new corporate-to-government lobbying vector for compute policy.⁷

2. Economics, Finance & Business Outlook

• Micron winds down DDR4

  • EOL notices sent; spot quotes already up 50 %, with tight supply expected through 2026.⁶
  • Long-life sectors (auto/industrial) get carve-outs but at “premium” pricing.

• AMD taps AI startups for ecosystem lift

  • Cohere cuts porting time from weeks to “days” on ROCm; acquisitions of ZT Systems and niche software outfits expand talent bench.⁷
  • MI450 server “Helios” seeded for 2026, targeting rack-scale deals now dominated by Nvidia.

• Nvidia lands German cloud flagship

  • 10 k-GPU order underpins a sovereign EU AI platform; Deutsche Telekom provides infra and sales.⁵
  • Move echoes Microsoft’s and Amazon’s U.S. vertical stacks, but with European data-residency guarantees.

• Synopsys faces merger cost overhang

  • Delay triggers ticking termination fee and higher integration risk, pressuring FY25 guidance even as limited tool sales to China resume.¹

3. Technology & R&D

• AMD Instinct MI350 / MI355 debut

  • Built on TSMC N3P with Samsung and Micron 12-high HBM3E; AMD claims 4× generational throughput and parity with Nvidia B200 in FP4 inference.²
  • Supply-chain win for Samsung after recent HBM3E validation hiccups.

• ROCm 7 expands to Windows + Radeon

  • Adds FP4/FP6, distributed inference via vLLM, and promises local LLMs on Ryzen PCs in 2H25.³
  • Internal tests show 3.5× inference uplift versus ROCm 6 on MI300X.

• Imec’s record GaN-on-Si RF transistor

  • 7–24 GHz E-mode MOS-HEMT hits >70 % power-added efficiency at <20 V, crucial for 6 G FR3 handsets.⁸
  • Contact resistance record paves the way for mainstream CMOS back-end integration.

• Ferroelectric gate-stack memory prototype

  • Researchers demonstrated high-density FeFET storage with antiferroelectric erase needing only “tiny” net charge—promising cooler, denser embedded NVM.⁹

Footnotes