Nvidia’s AI Factory Strategy: More Tokens per Megawatt, Not Just More GPUs

AI infrastructure is increasingly constrained by available power, not simply by the number of accelerators that can fit in a rack. Nvidia used the opening day of its 2026 AI Infra Summit to argue that the key output metric for an AI factory is agentic tokens per megawatt: how much useful inference work a facility produces within its power envelope. 23

Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, presented the strategy at the Santa Clara event. The throughline was broader than GPU performance: AI factories need coordinated compute, networking, inference software, cooling, power controls and grid operations if they are to remain productive as electricity becomes a binding constraint. 123

The practical evidence: Lambda’s DSX MaxLPS validation

The most concrete result came from AI cloud provider Lambda’s first deployment-environment validation of Nvidia DSX MaxLPS on HGX B200 GPU servers. Nvidia says DSX MaxLPS monitors GPU and rack-level power consumption, then reallocates available headroom across nodes. 23

In the reported five-rack, 19-node cluster, Lambda operated 19 nodes within a power budget that would normally accommodate 16 full-power nodes. The result was a reported 24% increase in cluster token throughput, from roughly 4 million to 5 million tokens per second, alongside a 23% improvement in performance per watt. 23

The significance is operational rather than theoretical. Data-center capacity is often limited by delivered site power. If workloads have variable power profiles, software that uses unused headroom can increase useful output without a new electrical service or a larger facility power budget.

Vera Rubin claims are projections, not production results

Nvidia also outlined potential gains for future Vera Rubin NVL72 deployments. It said DSX MaxLPS could enable up to 40% more GPU capacity within a given site-power envelope and up to 35% higher token throughput, depending on deployment conditions. 23

Those figures should be read differently from Lambda’s Blackwell-based validation. The Lambda measurements are reported deployment results; the Vera Rubin figures are Nvidia projections for a future platform. 23

Making AI factories responsive to the grid

Nvidia’s second major theme was flexibility: an AI factory should be able to respond to grid conditions without treating every workload as equally urgent.

At Nvidia’s Eos AI factory, Emerald AI’s Conductor platform worked with Silicon Valley Power on a commercial flexible-load program. Nvidia reported that the system responded to more than 200 utility demand signals and, in a demonstrated event, reduced facility consumption from 4 MW to 3 MW in less than a minute. The system did this by deprioritizing flexible computing work while allowing higher-priority inference to continue. 23

Nvidia positioned this program as an early example of the capability planned for DSX Flex. The company describes DSX Flex as a way to receive load-shedding, demand-response and power-price signals, then apply workload priorities so that deferrable work can pause and resume while critical services remain active. The Eos demonstration itself was an earlier commercial-scale proof point rather than a DSX Flex installation. 23

This approach matters for operators facing utility constraints: not every training, batch processing or offline task needs the same response time as production inference. Treating compute as a flexible load could help a facility participate in grid programs while protecting higher-value workloads.

A full-stack pitch, from silicon to operations

The summit announcements framed Nvidia less as a component supplier and more as an AI-factory platform provider. Its proposed stack spans Vera Rubin compute, NVLink scale-up networking, Dynamo inference software, Ethernet and SuperNIC networking, BlueField infrastructure components, rack-level power optimization and grid-aware workload operations. 2

Groq 3 LPX was part of that broader positioning. Nvidia describes it as an interactive inference accelerator for Vera Rubin intended for the low-latency and large-context requirements of agentic systems. 112 The supplied evidence supports that role, but does not independently substantiate more specific claims about release timing, a 35-times-throughput-per-megawatt comparison or performance on models above two trillion parameters.

Why the strategy has business weight

The infrastructure message comes as Nvidia’s data-center business continues to expand. For the second quarter of fiscal 2027, ended July 26, 2026, Nvidia reported $96.2 billion in revenue, up 106% year over year, including $89.0 billion in Data Center revenue. 17 The company guided for third-quarter revenue of $108 billion, plus or minus 2%. 2830

That financial context helps explain the company’s emphasis on factory-scale operations. As customers deploy more AI capacity, the practical questions increasingly become how to secure electricity, raise output within fixed power limits and keep inference services running through power and operational disruptions.

The takeaway

Nvidia’s opening-day case was not that raw accelerator performance has stopped mattering. It was that peak performance alone is no longer sufficient for AI-factory economics.

The company’s proposed measure—validated tokens per megawatt—links hardware performance to the constraints operators face in production: site power, grid events, workload priority and system availability. Lambda’s reported 24% throughput gain offers an early deployment-based data point for that thesis; Nvidia’s wider Vera Rubin and DSX claims remain a roadmap that will need to be validated in customer operations. 23