Index Company
AMD
10 CÈ pieces since 10 June 2026: 2 deals, 8 intel items. Newest first. All companies →
-
The AI chip bottleneck moves to Singapore: Powertech takes 30% of a $5.66B packaging venture
Powertech puts in $400 million for 30 percent of XCIP Technologies, a Broadcom joint venture in Singapore. Its fan-out panel packaging fits about 45 large AI dies per pass against eight or nine today. AMD ships first, in late 2027.
-
A fully open MoE trained without a single Nvidia GPU: AMD ships Instella-MoE-16B-A3B
AMD released Instella-MoE-16B-A3B, a 16-billion-parameter mixture-of-experts model with 2.8 billion active per token, pretrained on 7.1 trillion tokens across Instinct MI300X and MI325X GPUs, with the training code under MIT.
-
The ceiling is megawatts: AMD and Cerebras split inference for 5x tokens per watt
AMD and Cerebras split inference across two machines: Helios racks handle prompts and long context, the Wafer-Scale Engine generates tokens. Modelling shows up to five times more tokens per second per kilowatt. Cerebras Cloud ships it this half.
-
AI's real chip bottleneck is packaging, not silicon: Powertech puts $400M into a Broadcom JV in Singapore
Taiwan's Powertech will invest $400M in a Singapore joint venture with Broadcom to build fan-out panel-level packaging — the fine-geometry routing behind Broadcom's and AMD's AI chips. The deal is pending Taiwan regulatory sign-off.
-
China drafts $295 billion AI data-center plan mandating 80% domestic chip sourcing
National Development and Reform Commission blueprint calls for networked computing hubs by 2028, with state telecoms China Mobile and China Telecom operating facilities powered predominantly by Huawei, Alibaba, and Biren chips instead of Nvidia or AMD.
-
Zyphra ships ZAYA1-8B sparse reasoning model trained end-to-end on AMD Instinct hardware
Released under Apache 2.0 license, the 8-billion-parameter model routes only 760 million active parameters per token and was built entirely on AMD GPUs, proving a viable non-Nvidia path for high-efficiency training and breaking the assumption that frontier open models require Nvidia infrastructure.
-
Zyphra ZAYA1-8B sparse model trained on AMD hardware, ships under Apache 2.0
Zyphra released ZAYA1-8B in June 2026 under Apache 2.0, an 8-billion-parameter sparse-routing model with only 760 million active parameters per token, trained from scratch on AMD Instinct MI300 GPUs to demonstrate inference efficiency outside Nvidia-dependent pipelines.
-
Zyphra ships ZAYA1-8B reasoning model trained entirely on AMD Instinct MI300X GPUs
Released May 6 under Apache 2.0 license, the 8-billion-parameter mixture-of-experts model was pretrained, midtrained, and fine-tuned on a 1,024-GPU AMD cluster with zero Nvidia hardware, matching far larger proprietary models on math and coding benchmarks.
-
China announces $295 billion five-year plan to build nationwide AI data centers
Beijing is preparing to spend 2 trillion yuan over five years on interconnected computing hubs operated by China Mobile and China Telecom, with 80 percent of AI chips sourced from domestic suppliers including Huawei. The network is designed to fuel domestic AI sector growth and reduce dependence on Nvidia and AMD.
-
Zyphra releases ZAYA1-8B sparse model trained from scratch on AMD Instinct hardware under Apache 2.0
Zyphra's ZAYA1-8B model uses a sparse routing architecture with 8 billion total parameters but only 760 million active per token. Trained entirely on AMD Instinct hardware without NVIDIA dependency, it ships under Apache 2.0 license, demonstrating decentralised training infrastructure.