In the race to secure power for AI data centers, Amazon leads on scale while Google differentiates through vertical integration by acquiring an energy developer, according to a report. Power supply has emerged as the biggest constraint on AI infrastructure expansion.
The Power Race · Amazon vs Google
Electricity, Not Chips, Becomes AI's Biggest Bottleneck
As compute demand explodes, two incumbents pull ahead — Amazon on raw scale and cost-cutting custom silicon, Google on in-house chip innovation built for the "agentic era."
~9 GW
Power drawn at Amazon's U.S. data centers — comparable to a state's generating capacity
500k+
Trainium2 chips in "Project Rainier," scaling past 1 million by end of 2025
70-80%
Lower inference cost on Amazon's Inferentia versus GPUs
Generational Leap in Training Compute
Each company's new silicon vs its own prior generation — months of training now collapse into weeks
prior compute
1× baseline
5×
Amazon Project Rainier
Amazon — Scale & Cost
Dedicated chips: Trainium + Inferentia
20+ years of data-center buildout
Cuts dependence on NVIDIA GPUs
Snap: 70% cost cut on AR filters
Claude 3.5 Haiku: 60% faster on Trainium2
Anchored by Anthropic, Uber, Databricks
Google — In-House Innovation
8th-gen TPU split into 8t + 8i
8i triples SRAM to 384MB
HBM up 50% to 288GB, MoE-optimized
Built for autonomous "agentic" AI
Ironwood TPU: 3.7× better carbon intensity
Vertical integration + research depth
The Takeaway
Enterprises lean to Amazon's chips for cost; Google and Microsoft lead in consumer-facing models — but custom silicon still trails NVIDIA on versatility and ecosystem maturity.
Power constraints and in-house chip strategy — not any single product launch — will decide where the AI race goes next.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…