AI Cloud Procurement Splits: Inference Prices Fall 65% While GPU Costs Rise 15%

The pricing split that changes cloud budgets

Cloud infrastructure spending reached $143 billion in Q2 2026, up $43 billion year over year, but the market is now pricing AI workloads in two incompatible directions. AWS increased EC2 Capacity Block GPU reservation prices by 15% on October 7, the third increase in nine months. Google cut Gemini inference-token prices by up to 65% for certain workloads during the same period. Enterprise buyers face falling costs for model consumption and rising costs for physical infrastructure — a split that breaks traditional cloud-budgeting assumptions.

AWS generated $42.2 billion in quarterly revenue and grew 37%. Microsoft's Intelligent Cloud segment produced $39.3 billion, with Azure growth of 43%. Google Cloud generated $24.8 billion and grew 82%. Hyperscalers are competing on inference volume while charging more for constrained GPU access, power, and data-center capacity. This benefits organizations with flexible inference workloads but penalizes those that need reserved training capacity or multi-year commitments.

What the economics mean for procurement

The immediate implication is that "cloud cost" is no longer a single line item. Buyers must model inference, training, reservations, egress, security tooling, and provider-credit risk separately. Fixed-price or reserved GPU contracts now carry more risk than pay-as-you-go inference because hyperscalers can reprice capacity faster than enterprises can renegotiate contracts.

Organizations training large models should seek price caps, portability clauses, and termination rights before committing to multi-year GPU capacity. The alternative is vendor lock-in at the hardware layer, where switching costs include not just data migration but physical relocation of training workloads. Inference workloads, by contrast, become more portable as token prices fall and API compatibility improves across providers.

The pricing dynamic also explains the surge in neocloud financing. CoreWeave reported roughly $104 billion in backlog and 1.5 gigawatts of active power as of June 30. Lambda is reportedly raising up to $4 billion at a $14.5 billion pre-money valuation, with backlog rising from $15 billion to $50 billion, including a reported $35 billion commitment from Anthropic. Reported reference pricing includes approximately $49.24 for an eight-GPU CoreWeave instance, about $4.50 per GPU for Nebius, and $3.99 for an eight-H100 Lambda instance.

Neoclouds compete by offering dedicated clusters, bare-metal access, and faster deployment than hyperscalers. Their financing increasingly resembles infrastructure finance: debt and convertible instruments secured against GPUs, power, and long-term customer contracts. Buyers gain diversification but inherit counterparty, concentration, and financing risk. Contracts should address hardware substitutions, service continuity, data egress, portability, SLA remedies, and what happens if a provider cannot finance or deliver promised capacity.

Cloud-cost automation attracts capital

North.cloud raised $60 million in a Series B led by PeakSpan Capital on October 15. The company offers AI-cost-control and cloud-finops automation. Public pricing lists a Startup plan at $199 per month and a Premier plan at $1,399 per month. Its Autobot and Flexbot products charge fees tied to managed commitments or realized savings.

The financing signals investor confidence in automated commitment management rather than dashboard-only finops. North.cloud competes with Cloudability, CloudHealth, Harness Cloud Cost Management, Kubecost, and native tools from AWS, Azure, and Google Cloud. The relevant purchasing question is whether savings-based pricing produces auditable net savings after platform fees. Buyers should require controls around automated reserved-instance or savings-plan purchases, approval thresholds, rollback procedures, and treatment of multi-cloud commitments.

Security maturity lags cloud expansion

A SANS Institute cloud-security survey found that 93% of organizations run production workloads in the cloud, but only 40% describe their cloud-security programs as mature. Over the next 18 months, 62% prioritize alert correlation across the attack surface, 55% simplified vendor management, 51% cloud infrastructure entitlement management, and 44% AI-driven threat hunting.

Cloud expansion without corresponding identity and detection maturity increases operational and compliance risk. Security budgets are moving toward cross-cloud correlation, identity and entitlement controls, and consolidated platforms. This favors vendors that integrate CNAPP, CIEM, SIEM, and cloud-native detection rather than isolated point products. Procurement teams should assess whether existing tools cover AWS, Azure, and Google Cloud consistently, and should demand measurable outcomes such as reduced excessive permissions, alert volume, mean time to detect, and mean time to remediate.

What to watch

Track whether hyperscalers maintain inference price cuts beyond promotional periods or whether token pricing stabilizes once market share solidifies. Monitor neocloud financing velocity — if debt markets tighten or GPU backlog growth slows, some providers will face liquidity constraints that affect service delivery. Evaluate whether your organization's cloud security posture keeps pace with workload expansion, particularly for identity, entitlement, and cross-cloud detection. The pricing split between inference and capacity is structural, not cyclical, and will force enterprises to treat AI procurement as two separate negotiations with different risk profiles.