AWS raises GPU instance prices 20%; spot-vs-reserved split widens as memory bottleneck persists

AWS raised hourly prices on its most powerful GPU instances by 20% effective July 1, 2026, with P6-B300 now at $14.04/hour, P6-B200 at $12.36/hour, and P5 at $5.19/hour. This is the second AWS GPU price increase in 2026, following a 15% hike earlier in the year, bringing cumulative increases to roughly 38% in six months. For a 32-GPU training job on P6-B200 instances for two weeks, the bill jumped from ~$99.5K to $119.4K before discounts.

Memory costs are driving the increases: DRAM and GDDR6 spot prices have roughly tripled since late 2025 as AI data centers consume the majority of global memory supply. AMD raised its GDDR memory kit prices 10% in July 2026 (second increase in six months), and long-term fixed-price contracts that insulated GPU makers through 2025 have now expired with 2026 renewals landing at significantly higher rates. GamersComputing consumers are paying collateral damage: RTX 5090s are selling at $2.9-5K (50% over MSRP), and memory now accounts for 80%+ of a GPU's bill of materials.

For architects: AWS's 20% hike applies only to the on-demand and spot markets; reserved-instance and multi-year contract customers do NOT see these increases, widening the bifurcation between hyperscaler-backed labs (Apple, Meta, OpenAI at negotiated rates) and smaller labs buying by the hour. CME and ICE are launching GPU futures contracts in 2026 to enable hedging on this volatile spot tier. The structural constraint is memory supply: new capacity from Micron and SK Hynix won't reach volume until 2027-2028, so price relief is unlikely before then.