How to Price 2027 Inference Capacity: Verda 250MW NVL72

The trade-off is reserved power, not chip count

Before locking 2027 inference spend, the trade-off is between paying for guaranteed capacity and betting on spot markets. Verda, a Helsinki-based AI cloud company, has a capacity plan that is a useful test case because it pairs early rack-scale deployments with in-house serving software. If a GPU-hour has a cost and your stack serves a given number of tokens per GPU-hour, the floor is that cost divided by token throughput before overhead. The rest of the deal is whether the vendor can hold that number at your latency target.

The money behind the plan

Verda, formerly DataCrunch, raised $189 million in early-stage financing to expand AI cloud infrastructure. The Verda Series B was led by Emergence Capital, an investor, with additional participants. The round pushed its valuation above $1 billion and lifted cumulative equity and debt past $450 million.

Revenue is not a substitute for your workload. Verda said its annualized revenue run rate reached $165 million in July, up from more than $60 million reported at the time of its April round. It serves clients in 50 countries, including sovereign AI, media-generation, and radiology-models workloads. That breadth shows demand, not your latency curve. The question is whether the plan survives your token mix, your latency budget, and your contract term.

The arithmetic before the signature

Start with prices you can verify today. A public GPU market listed H100 SXM5 spot at $1.46 per GPU-hour in June 2026, down from $1.66 in May 2026, while on-demand pricing was about $2.54 per GPU-hour. AWS p5.48xlarge with eight H100 SXM5 GPUs cost about $55.04 per hour on-demand, or about $6.88 per GPU-hour. Spot can be cheaper, but it carries interruption risk.

Turn that into tokens. If your serving stack sustains 10 million tokens per GPU-hour, that on-demand price is $0.254 per million tokens before overhead. At the AWS on-demand GPU-hour price, the same 10 million tokens per GPU-hour is $0.688 per million tokens. The spread is the entry fee. Your model, batch size, prompt length, and cache behavior decide where you land.

Now add the supply constraint. Nvidia said Vera Rubin NVL72 would begin production ramp in Q3 2026, with first allocations for AWS, Google Cloud, and Azure, and independent cloud access not expected before Q4 2026 or Q1 2027. Custom NVL72 rack orders faced lead times of 40 or more weeks. HBM3e supply constraints had already pushed H200 and GB200 NVL72 custom-config lead times to 36-52 weeks as of May 2026. If you need 2027 capacity, you are negotiating against a supply queue, not a price list.

Verda's plan fits that queue. It plans to operate more than 250 megawatts of data-center capacity in 2027, with live capacity in Finland and future capacity in Europe, the UK, the US, and Asia. It said it will begin early deployments of Nvidia VR200 NVL72 rack-scale systems within the coming months. It also builds in-house compiler and serving software, which lets it manage performance and reliability engineering for customer workloads.

The pre-signing scorecard

The market is not uniform. A September 2026 industry assessment placed only 19 neoclouds, GPU-focused cloud providers, in a Medallion rating and put 15 providers in a Participation Ribbon tier for meeting only minimum expectations. Use the list below as a quick test for an inference cloud capacity plan. Each item is a gate for finance, users, procurement, launch, legal, hardware, and exit.

• Cost per token: ask for the unit price for your model, prompt length, output length, and batch size, then compare it to your current provider.
• p99 latency: require a recent p99 at target utilization, not median, and ask for the load profile used.
• SLA/reliability: ask for uptime, error budget, and the credit formula that applies when you miss the budget.
• Egress: ask for the per-gigabyte price and whether model weights, logs, or checkpoints count as egress.
• Sovereignty: ask for data residency, jurisdiction, key custody, and whether your workload can stay in a named region for sovereign AI inference.
• NVL72 utilization: ask for rack allocation, expected utilization, upgrade path, and whether your contract includes rack-scale systems.
• Exit terms: ask for cancellation, data return, model weight return, and transition support.