Cloud AI

On July 1, 2026, AWS raised EC2 Capacity Blocks for ML prices by roughly 20% per accelerator-hour, its second hike in six months after a ~15% increase in January (Converge Digest). A reserved P6-B300 (Blackwell Ultra) now costs $14.04/hr. Cloud GPU capacity — the scarce fabric every foundation-model training and production inference run depends on — just got measurably more expensive for the platform teams that book it up to eight weeks ahead.

What AWS Actually Changed

Capacity Blocks are AWS’s marketplace for guaranteed AI accelerator capacity, hosted in EC2 UltraClusters and reservable up to eight weeks in advance for windows of up to six months, in clusters of one to 64 instances (AWS). Unlike Reserved Instances, they demand no one- or three-year commitment — you pay the prevailing reservation fee upfront and get guaranteed, low-latency, high-throughput fabric for the window you booked. After July 1, 2026, the headline rates per accelerator moved decisively upward.

The critical wrinkle: AWS bills the reservation fee upfront when you book, and OS charges separately during runtime, with prices “updated periodically based on supply and demand” (AWS). You pay the rate in force at purchase time even if the reserved window starts after a future increase. That clause turns pricing announcements into a procurement clock — every team with a training run landing in Q3 just had its budget recalculated. The P6e-GB200 (NVIDIA GB200 NVL72) and Trainium Trn2 instances are also supported, signaling that AWS is pricing the entire accelerated fleet on a dynamic, demand-revealing curve rather than a static rate card.

Capacity Blocks vs Commitments

The price hike forces a clearer split between two reservation models. Capacity Blocks sell you a guaranteed, dated window of fabric — ideal for time-boxed training and checkpointed experimentation — but at a dynamic premium that has now compounded ~38% across two increases in 2026. Reserved Instances and Committed Use Discounts (on GCP) sell you a discount on usage over a one- or three-year horizon, without guaranteeing capacity on a specific date. They are cheaper per hour but useless the day your job can’t get a host.

The engineering rule that falls out: reserve the baseline inference and steady-state serving with RI/CUD commitments where capacity is rarely the binding constraint, and reserve time-critical, capacity-constrained training runs with Capacity Blocks where the alternative is no GPUs at all. Spot instances remain the third lever, delivering 70–91% reductions for fault-tolerant, checkpointed training workloads (nOps) — though, as we’ve unpacked before, spot GPUs save 70% on paper and often far less in production. The trap is treating these as interchangeable — a team that books a Capacity Block for a one-off fine-tune is paying a scarcity premium for a job that spot could have absorbed, while a team that runs production inference on on-demand is paying 2–3x what a commitment would have locked in.

The Real Inference Bill

Reservation rates are only the first line of a GPU cloud invoice, and the second line is the one most teams under-model: egress. A reproducible 2026 cost model across AWS, Azure, GCP, and a sovereign regional provider, running one GPU 24/7 for a month (730 GPU-hours) with 30 TB of outbound traffic, found that compute is “at most, half of an inference bill” — and on a token-serving API, network egress alone can be a third of the total (HackerNoon).

That model put on-demand AWS P6-B200 (Blackwell) at $12.36/GPU-hr and P5 (H100) at $12.29/GPU-hr, normalized per GPU from eight-GPU flagship instances — meaning the cheapest GPU-hour rarely wins, because the providers that discount compute often recover it on network line items that never appear on the headline rate card (HackerNoon) — a dynamic we examined in detail when cloud egress fees surpassed GPU compute costs for AI workloads. For platform teams, the implication is structural: any GPU procurement decision that quotes only dollars-per-GPU-hour is incomplete by design. A 20% capacity-block hike on compute, layered on a workload where egress is already a third of spend, compresses margins in a way no single line item reveals.

Multi-Cloud Procurement Gets Harder

The hike is not an AWS-specific story; it is a market signal. At FinOps X 2026, the opening keynote reported that many organizations had already burned through 3x their entire annual AI budget by June, and that token consumption is forecast to grow 20x by 2030 while the cost-per-token floor hit a wall in late 2025 and is not coming down (nOps). The four largest cloud providers collectively hold over $2 trillion in infrastructure backlog, neoclouds have extended minimum commitments from one year to 3–5 years to secure their own capacity, and industry forecasts see no relief until 2028 (nOps).

For engineers, this reframes “multi-cloud” from a resilience checkbox into a capacity-arbitrage necessity. When a single hyperscaler can dynamically reprice reserved fabric up 20% mid-quarter, the ability to shift a training run to a sovereign provider, a neocloud, or an on-premises cluster becomes a real hedge, not a vendor-neutrality slogan. The trade is real: sovereign and regional providers may quote a competitive per-GPU-hour rate but cap cluster size, raise egress, or demand longer minimum commitments. The teams that win are the ones that model compute, egress, and commitment horizon as a single equation — and keep at least one secondary capacity path warm before they need it.

An Engineering Playbook

The combined pressure — dynamic reservation pricing, egress-blind budgets, and a 2028 capacity horizon — calls for a concrete set of moves rather than a slogan.

• Split the reservation stack. Steady inference → RI/CUD commitments. Time-critical training → Capacity Blocks. Fault-tolerant batch → spot with checkpointing at 70–91% off (nOps).
• Model egress as a first-class cost. Quote every GPU workload in compute-plus-network terms before booking. A workload serving 30 TB/month can see egress at a third of spend (HackerNoon).
• Keep a secondary capacity path warm. A neocloud or on-premises cluster that has been validated end-to-end (model artifacts, container images, networking) is the only real hedge against a mid-quarter hyperscaler repricing.
• Buy before the next window. Because Capacity Block rates are fixed at purchase time, booking a dated window before an announced increase freezes the lower rate for that reservation — turning the procurement calendar into a cost-control lever.
• Attribute spend to the workload. Tagging GPU and token spend to the product or model that consumed it is the prerequisite for the kind of chargeback that catches a 38% compounded hike before it reaches finance.

The deeper lesson is that cloud GPU capacity is now priced like a commodity in structural shortage, not a metered utility. AWS’s July 1 hike is the market telling platform teams that guaranteed, high-fabric compute is a premium resource whose price will keep revealing demand until supply catches up — and that the engineering response is procurement discipline, not a bigger budget. For teams weighing the self-host exit ramp, NVIDIA NIM economics show exactly where self-hosting beats every API.

Related reading on cloudai.pt:

• Spot GPUs Save 70% on Paper. Production Says Otherwise.
• Cloud Egress Fees Now Surpass GPU Compute Costs for AI
• NVIDIA NIM Economics: Where Self-Host Beats Every API

References

• Converge Digest — AWS Raises EC2 Prices by About 20% for AI GPU Reservations
• AWS — Amazon EC2 Capacity Blocks for ML
• HackerNoon — I Priced the Same Inference Workload on 4 GPU Clouds. Egress Was the Catch
• nOps — GPU Cost Optimization: 5 Proven Strategies to Cut Your AI Infrastructure Cost