September 18, 2026 — curated links and takeaways.
1. AMD Warns Partners Of A 10% Price Hike On AI Chips, Sparing Ryzen
AMD announced a 10% price increase on AI accelerators and Radeon GPUs effective Q4 2026, citing TSMC wafer cost pressures. Builders pricing inference and fine-tuning workloads need to lock in GPU contracts now before costs rise across the supplier base.
2. The HBM Bottleneck: 3 Tech Companies Have Locked Up 85% of the Supply of AI's Scarcest Asset
HBM memory supply is concentrated among three vendors and remains the structural constraint on inference scaling; AMD is deploying LPDDR5X memory-on-package as a workaround. Teams building inference systems must understand memory bandwidth limits now dominate chip selection over raw FLOPS.
3. AI Inference: The New Driver of Semiconductor Spend
Demand has shifted from raw training compute to power-efficient inference chips; major cloud providers are accelerating ASIC development rather than relying on off-the-shelf GPUs. Builders must assume inference economics—energy per token, not peak throughput—will dictate margin structures for the next 18 months.
4. House overwhelmingly passes bill to shield ratepayers from data centers
House passed legislation requiring AI data centers with 100+ MW demand to fund their own power infrastructure rather than passing costs to ratepayers. Infrastructure teams must budget for direct power procurement and grid build-out; regional power scarcity will become a hard constraint on capacity expansion.
5. AMD Advancing AI 2026: Helios Rack Hits $5.25M
AMD's MI455X GPU delivers 50% more HBM4 capacity (432GB vs. 288GB) than Nvidia Vera Rubin at 40 PFLOPS FP4, with Helios infrastructure deploying in Q4 2026. On-premises training and inference alternatives are now available with meaningful memory advantages; teams can evaluate non-cloud options for captive workloads.