AWS Raises GPU Prices 15% as Hyperscalers Report $2.48 Trillion Cloud Backlog

AWS reprices GPU capacity for the third time in 2026

AWS raised prices for EC2 Capacity Blocks for ML by approximately 15% on October 7 for most U.S. GPU configurations. The increase follows a 15% rise on January 4, a 10% increase in April for selected configurations, and a 20% jump on July 1. Reserved GPU capacity—once positioned as a hedge against spot-market volatility—now exhibits its own price instability.

The repricing reflects demand conditions AWS cannot suppress through incremental capacity additions. Microsoft, Oracle, Amazon and Alphabet disclosed combined cloud backlogs of approximately $2.48 trillion in their latest results: Microsoft $677 billion, Oracle $664 billion, Amazon $578 billion, and Google $561 billion. These commitments represent contracted future revenue, not available infrastructure. For enterprise buyers, the gap between committed spend and deployable capacity creates procurement risk.

AWS competes with Azure, Google Cloud, Oracle Cloud Infrastructure and specialized GPU providers including CoreWeave and Nebius. Repeated GPU repricing gives rivals an opening to compete on committed-use discounts, reserved capacity and alternative accelerators. Enterprises budgeting for model training or inference should treat GPU capacity as a volatile operating cost rather than a fixed-price reservation. Procurement teams need multi-cloud capacity agreements, longer-term reservations with explicit price-protection clauses, workload portability and documented exposure limits for GPU price changes.

The risk is structural: capacity commitments reduce scarcity exposure but create stranded-cost liability if model demand, utilization or hardware economics shift faster than reservation terms allow.

Google Cloud releases cross-cloud lakehouse with zero variable egress fees

Google Cloud made cross-cloud interconnect and caching for AWS and Azure generally available in its Agentic Data Cloud. The offering queries data across clouds through private connectivity and advertises zero variable egress fees for the supported access model.

The service challenges AWS and Azure's tendency to anchor data gravity inside their own storage and analytics ecosystems. It also competes with independent lakehouse and data-movement technologies from Databricks, Snowflake and Confluent. Enterprises with data distributed across AWS, Azure and Google Cloud may reduce inter-cloud transfer costs and avoid copying datasets into a single provider.

The main purchasing question shifts from "which cloud owns the data?" to whether Google's connectivity, caching and query performance meet production requirements. "Zero variable egress fees" does not eliminate total cost. Buyers must model interconnect charges, caching costs, compute expenses, storage fees, query consumption and data-governance overhead. Measured latency and throughput results should precede any consolidation around the service.

Nebius secures $3.5 billion to expand AI-cloud infrastructure

Nebius secured a $3.5 billion senior secured facility from Citi to finance expansion of its global AI-cloud infrastructure. The financing strengthens Nebius's ability to compete with CoreWeave, Lambda, Crusoe and hyperscaler GPU clouds while positioning it as an alternative supplier for customers unable to obtain preferred capacity from AWS, Azure or Google Cloud.

More funded GPU-cloud suppliers could improve enterprise access to accelerators and provide leverage in negotiating capacity, pricing and service-level terms. A large debt-backed buildout increases execution and utilization pressure. Enterprises should avoid treating financing capacity as equivalent to delivered capacity. Buyers should demand evidence of installed GPU inventory, power availability, operational uptime, hardware fleet composition, geographic redundancy, software compatibility and exit terms.

Google invests $15 billion in Malaysian infrastructure

Alphabet reported $127.6 billion in revenue, up 25%, while Google Cloud revenue more than doubled to $30.7 billion. Google announced a $15 billion investment in Malaysian data centers and AI infrastructure during 2026. The investment competes with Microsoft and Amazon's global data-center buildouts and narrows the scale gap with AWS and Azure, particularly in AI computing and data services.

New regional capacity can improve data residency options, latency and availability for Asia-Pacific workloads. It may also increase negotiating leverage for large buyers comparing Google Cloud against AWS and Azure in the region. Data-center investment does not by itself establish usable enterprise capacity. Buyers should verify service availability by region, GPU types, quota policies, power-related constraints, local compliance certifications and production support coverage.

What to watch

AI infrastructure is now a supply-constrained, financially intensive procurement category. AWS GPU price increases, hyperscaler backlogs and billions in specialized-cloud financing point to rising value in multi-cloud architecture, workload portability and negotiated capacity commitments. Google's cross-cloud lakehouse release could reduce lock-in and data-transfer friction, but buyers should validate its advertised economics with workload-specific tests rather than assume that eliminating variable egress eliminates overall data-platform cost.

Enterprise technology buyers should model GPU capacity as a variable cost, establish multi-cloud procurement relationships and require explicit price-protection terms in reservation agreements. The supply-demand imbalance favors providers. Buyers gain leverage through portability, competitive procurement and transparent cost modeling.