CoreWeave Trains DeepSeek-V3 in 2.02 Minutes on 8,192 GB300 GPUs
Blackwell Ultra's GB300 NVL72 delivers 1.6x training speedup over GB200, enabling CoreWeave's 2.02-minute DeepSeek-V3 run at 8,192 GPUs.
Read more
Blackwell Ultra's GB300 NVL72 delivers 1.6x training speedup over GB200, enabling CoreWeave's 2.02-minute DeepSeek-V3 run at 8,192 GPUs.
Read moreTYLsemi exits stealth with $43M and a chiplet platform targeting custom XPUs, aiming to cut cost and time by up to 50%.
Read more
A $60B debt package, $125.2B TPU lease, and 73% gross margin guide show how custom AI silicon financing works.
Read more
Engineering breakdown of Arm, Qualcomm, Apple and Tesla AI chips across CES 2026 silicon disclosures, data centre inference cards and physical AI robotics.
Read more
ASUS ramps Vera Rubin NVL72 with 72 Rubin GPUs, 36 Vera CPUs, 227 kW peak power, full liquid cooling and nine NVLink 6 switches.
Read more
How BIS export controls govern AI chips, HBM3E/HBM4 memory stacks, 300mm foundry tape-outs and China data centre accelerator architectures.
Read more
Comparing cloud AI chips across AWS, Google and Microsoft: HBM memory bandwidth, optical pod interconnects, liquid cooling and tokens-per-dollar economics.
Read moreOur chronological newsroom feed tracks AI chips news today across 2025, December 2025 and 2026, indexing verified hardware dispatches by recency and engineering beat. Whether you are auditing recent 3-day semiconductor supply-chain alerts or comparing multi-year processor transitions, the timeline below summarizes the key architectural handoffs across our archive:
| Timeline Window | Key GPU & Custom ASIC Milestones | Memory, Packaging & Rack Infrastructure | Primary Archive Coverage |
|---|---|---|---|
| 2025 & December 2025 Production Ramp | NVIDIA Blackwell B200/GB200 & GB300 Ultra NVFP4; AMD Instinct MI325X (256 GB HBM3E); AWS Trainium2 UltraServers | 8-high & 12-high HBM3E stacks; TSMC CoWoS-L 3.3x–5.5x reticle interposers; 120–142 kW direct-to-chip liquid cooling | GB300 to Vera Rubin Rack Planning & Blackwell Ultra NVFP4 MLPerf |
| Early-to-Mid 2026 Volume Transition | AMD Instinct MI355X (CDNA 4, 3nm, FP4/FP6); Google Cloud Ironwood (7th-Gen TPU); Meta MTIA v2/v3 custom silicon | SK hynix 12-layer HBM4 (2,048-bit interface, 2+ TB/s per stack); 800G & 1.6T OSFP optical transceivers | SK hynix HBM4 Architecture & Meta MTIA & AMD MI355X Roadmap |
| Late 2026 & Next-Gen Roadmaps | NVIDIA Vera Rubin NVL72 & AMD Instinct MI400 rack-scale systems; Intel 18A/14A High-NA EUV foundry silicon | NVLink 6 (3.6 TB/s per GPU), UALink & Ultra Ethernet fabrics; 227 kW liquid-cooled racks; 9.5x–14x CoWoS roadmaps | Vera Rubin NVL72 227kW Racks & TSMC CoWoS & HBM Stack Roadmap |