On August 26, 2026, Nvidia told investors it will raise prices on systems that use its next chips, including Vera Rubin and Grace Blackwell, early next year. Memory costs are the reason. If your plan for scalable AI infrastructure still assumes last year’s GPU bill, the plan is already wrong.
This post is for one search: ai workloads. The only product link is Scalable AI Infrastructure. If you run training jobs, real-time inference, or a bursty scoring queue and you do not want to pay hyperscaler rack prices twice, that page is the next step.
What Nvidia actually said last week
Two stories got mixed together. Keep them apart.
On August 22, Bloomberg reported that Nvidia had already told some of its largest customers that servers with its AI chips would rise by more than 15% in many cases, starting on systems shipping in early 2027. Reuters carried that report the same day and said it could not independently verify it. Nvidia did not comment that Saturday. Treat the “more than 15%” figure as a report, not a filed number.
On August 26, after the close, Nvidia reported fiscal 2027 second-quarter results: $96.2 billion in revenue, up 106% from a year ago, and $89 billion from data centre, up 117%. Gross margin was 75%. Guidance for the current quarter was $108 billion. Then the cost story landed. CFO Colette Kress said memory scarcity is being driven in large part by the AI buildout itself. She guided gross margin to about 74% this quarter, 71% to 72% in the January quarter, then 72% to 73% next year once a price increase is in place. CEO Jensen Huang confirmed that increase takes effect early next year.
Nvidia did not, on that call, confirm Bloomberg’s “more than 15%.” What it did confirm is the direction: HBM is scarce, Nvidia will not keep eating the whole bill, and Vera Rubin / Grace Blackwell systems get repriced as they ship in 2027.
Two other numbers from the same quarter matter if you buy compute instead of just watching the stock. Purchase commitments jumped from $119 billion last quarter to $279 billion, mostly to lock memory. Huang’s per-gigawatt sketch of Nvidia’s share of a customer’s data-centre spend: about $18 billion in Hopper, $25 billion in Blackwell, $40 billion in Vera Rubin. That is the cost surface your cluster sits on, whether you rent it or own it.
Sources: Reuters, August 22, 2026 (Bloomberg report); Nvidia’s August 26 earnings call, as recapped by Vested Finance, August 27, 2026 and TechTarget.
This is not a “buy more GPUs” post
A bigger purchase order is the default reaction, and it is the wrong first move for most teams that are not Microsoft, Google, or a neocloud.
Hyperscalers can negotiate, and some of them already run their own silicon. Google has TPUs. Amazon has Trainium. Everyone else is a price taker. If you are an ordinary enterprise, a SaaS product with a scoring API, or a health or fintech team with one training loop and a latency budget, you will feel the 2027 bill in cloud invoices even if you never buy a Rubin rack.
Buying a second vendor’s GPU and calling it a strategy is not enough either. The scarce part this quarter is memory, not a CUDA logo. HBM goes into the accelerator you already use. It also goes into the next server refresh you were going to do anyway. Analysts talking after the call said enterprises pay twice: AI capacity costs more, and a normal 2027 hardware cycle costs more too.
Waiting for the price to “come back” is also not a plan. Nvidia’s own margin recovery assumes the increase sticks. If demand stays ahead of HBM supply, the floor moves up, not down.
What scalable AI infrastructure actually means after a price reset
On our Scalable AI Infrastructure page, this is not a slogan. It is a workload map with a bill attached.
Name the job, not the chip. Training a custom model for a week is a different object from serving 50 million inferences a day. One needs a burst of GPUs and then almost none. The other needs a small, always-on, latency-capped pool. If those two share a cluster because “we already have H100s,” you are paying peak price for idle cards. After a 2027 reset, idle is the expensive mistake.
Autoscaling that actually turns off. Fixed provisioning is how teams waste 60% of a GPU month. Predictive scale-up is useful. Scale-down is the feature that pays for itself when the unit cost jumps. We design for both, with a floor you can explain to finance (minimum replicas, maximum dollars per hour, a kill switch for a runaway job).
Inference first, then training. Most product AI is inference. Quantization, TensorRT or equivalent, batching, and a cache in front of the model cut GPU hours without waiting for a cheaper SKU. Training gets a scheduled window, not a 24/7 reservation, unless you are actually pretraining.
More than one cloud, on purpose. Vendor flexibility is not a religion. It is a hedge when one provider passes HBM through first. We will still pick a home region for latency and data residency. We will not pretend a single-cloud “committed use” discount is a strategy if the commitment is priced off last year’s silicon.
A cost panel a human can read. GPU hours, queue wait, p99 latency, and dollars per 1,000 predictions. If you cannot see those four on one screen, you cannot react when Nvidia’s increase shows up in the invoice 90 days later.
The first slice we would actually scope
Not a 512-GPU fairy tale. One workload, one environment, a before-and-after bill.
Example: a fraud or ranking model that has to answer in under 10 ms, currently sitting on a fixed GPU pool at 40% utilization. We measure real traffic, split training off that pool, put inference on an autoscale policy with a hard dollar cap, and keep a second region cold until the first one is actually hot. If the honest work for 30 days is “stop leaving eight A100s on overnight,” that is the deliverable. The 2027 price hike will still arrive. You will meet it with a smaller denominator.
We will tell you if the model is too large for the latency budget, if the data cannot leave a region, or if you are about to rent a cluster you will use twelve hours a month. That answer is cheaper than a Terraform pile that never idle-sleeps.
This is not a remake of a generic AI services pitch, and it is not a recap of custom-software process. The date, the chips, and the cost surface are specific. The page you use next is specific too.
How to start on Scalable AI Infrastructure
Open the Scalable AI Infrastructure form. In the project details, name the workload (training, inference, or both), the current GPU or instance type, the monthly bill you can see, and whether the pain is latency, idle cost, or a 2027 budget that no longer fits. Pick the infrastructure type that matches: GPU training clusters, real-time inference, multi-cloud orchestration, or a full stack. Budget can be a range.
Bring one week of utilization numbers if you have them, or a written description of the jobs if you do not. We reply with a yes, a not-yet, or a smaller first slice. If you only wanted a shopping list of H100s, you do not need us. If you want the cluster to shrink when the job is done, start on that page today.