Cloud Capex Will Double to $690B in 2026 While Google Cuts Gemini Costs 78%

Hyperscalers Double Down on AI Infrastructure Despite Unit Cost Gains

Microsoft, Alphabet, Amazon, Meta, and Oracle will spend between $660 billion and $690 billion on AI infrastructure in 2026, nearly double the $380 billion projected for 2025, according to Futurum Group analysis. At the same time, Google reported cutting Gemini serving costs by 78% through model optimization over the past year. The combination creates a new procurement reality: infrastructure capacity will remain scarce and competitive even as vendors lower the cost per token or per workload.

For enterprise buyers, this shifts the decision framework. The question is no longer "how cheap is compute?" but "how predictable is access, and what is the actual cost per token after optimization?" When hyperscalers prioritize AI capacity over broad price cuts, procurement teams must negotiate on utilization, reserved access, and total cost per workload rather than list price alone.

Token Economics Now a Line Item in Infrastructure Decisions

Google's 78% reduction in Gemini serving costs demonstrates that model choice and routing strategy can materially change operating budgets. Inference is now a recurring expense large enough to justify dedicated optimization efforts. The cost improvement came from model architecture changes, not hardware alone, which means buyers who benchmark token economics and negotiate usage-based contracts will see lower run rates than those who buy on model capability without testing efficiency.

This creates pressure on OpenAI, Anthropic, Microsoft, AWS Bedrock partners, and other model providers to show comparable cost reductions. Enterprise buyers should demand proof of token-cost trends before committing to multi-year model contracts. A vendor that cannot demonstrate sustained cost-per-inference improvement is a budget risk in a market where Google just cut serving costs by more than three-quarters.

AI ROI Cases Shift From Capability to Cost Reduction

NVIDIA's 2026 State of AI survey found that 87% of respondents reported AI reduced annual costs, with 25% seeing decreases greater than 10%. More telling, 42% made optimizing AI workflows and production cycles their top spending priority for 2026. This changes the competitive dynamic among AI infrastructure vendors, MLOps platforms, and cloud providers. Buyers now justify infrastructure purchases through operational savings and productivity gains, not model accuracy or feature velocity.

The implication for procurement is that FinOps controls, model governance, and workload optimization tools are no longer optional. If nearly half of enterprises prioritize workflow efficiency, vendors that cannot show measurable cost reduction through orchestration, automation, or utilization gains will lose deals to platforms that can. Infrastructure decisions should include evidence of optimization capability—such as workload routing, dynamic scaling, or batch processing improvements—before expanding consumption.

Resilience Overtakes Pure Cost Optimization

IDC reports that enterprise IT leaders have shifted from cost optimization to operational resilience. Concerns about hardware supply constraints rose more than 15%, and interest in multi-region cloud architectures and backup strategies increased. Cheaper infrastructure is no longer the sole decision factor. Procurement now prices in outage risk, sovereignty requirements, and migration flexibility, which can increase total spend while lowering operational risk.

This favors cloud, colocation, backup, and disaster-recovery vendors that demonstrate geographic redundancy, portability, and faster recovery over single-region cost leaders. For buyers, the calculation changes: a 15% higher infrastructure cost may be justified if it eliminates a multi-hour outage or provides regulatory compliance. The trade-off is between optimizing for uptime and optimizing for unit price, and the market is moving toward the former.

What to Watch

The $660 billion to $690 billion capex wave and the 78% cost reduction are not contradictory signals—they define a scarce-capacity, high-utilization cycle where vendors compete on efficiency rather than list price. Buyers who benchmark workloads, use committed-use discounts intelligently, and demand evidence of optimization will extract more value than those who chase headline discounts. Procurement teams should track three metrics before expanding cloud or AI infrastructure: cost per token or workload after optimization, utilization rates under actual production load, and recovery-time commitments with penalties. The infrastructure market in 2026 will reward buyers who treat access, efficiency, and resilience as negotiable variables, not fixed costs.