AWS AI chief Matt Garman stated that enterprise compute costs are declining rapidly due to efficiency gains in hardware, software optimizations, and large-scale infrastructure improvements. Speaking at a recent industry event on June 18, 2026, Garman highlighted that customers are seeing meaningful reductions in the cost of training and running AI models, with some workloads becoming up to 30-40% cheaper year-over-year.
The comments reflect AWS’s aggressive push to lower barriers for enterprise AI adoption. Advances in custom silicon like Trainium and Inferentia chips, combined with better model optimization techniques and higher utilization rates across its global data centers, are driving these savings. Garman noted that falling compute costs are enabling broader experimentation with generative AI and agentic systems across industries.
For Amazon, the trend carries mixed implications for margins. While cheaper compute could accelerate AI service uptake and increase overall cloud revenue, it may also put pressure on short-term profitability if pricing adjustments flow through to customers faster than cost savings materialize. Analysts suggest that AWS’s high incremental margins on AI workloads could help offset any near-term compression, especially as higher-value managed services and enterprise commitments grow.
Amazon Web Services continues to dominate the cloud infrastructure market, with AI contributing significantly to recent growth. The company has invested heavily in expanding capacity while focusing on efficiency to maintain leadership against competitors like Microsoft Azure and Google Cloud.
Lower compute costs are expected to fuel stronger demand in the second half of 2026, particularly among mid-sized enterprises and startups previously priced out of large-scale AI projects. AWS executives expressed confidence that volume growth and new high-margin offerings will ultimately support healthy profitability even as unit costs decline. The development underscores the intense competition and rapid innovation cycle defining the AI infrastructure race.