NVIDIA's AI Infra Summit Power Efficiency Talk: What's New With Vera Rubin and DSX

On September 15, 2026, according to a roundup post NVIDIA's official blog published that same day, NVIDIA vice president Ian Buck spoke on AI factory efficiency at the AI Infra Summit, held at the Santa Clara Convention Center in California, making it the centerpiece of his infrastructure keynote.

This article was fact-checked on September 17, 2026, against the full text of the two official NVIDIA blog posts published that day, plus two official press releases from January 5 and August 24 used as a baseline for comparison. This site has not run any tests of its own, and offers no purchasing or upgrade advice; every number below is labeled as either something NVIDIA measured itself, something NVIDIA projects, or a result NVIDIA attributes to a third-party testing organization.

What Happened on September 15: From Peak Performance to Tokens per Megawatt

NVIDIA states in this roundup post that the metric for AI infrastructure is quickly shifting from peak performance to validated agentic tokens per megawatt, and that AI factories must now be codesigned from silicon to grid. That statement is what separates September 15 from the January CES announcement: January was about chip specifications and the cost per token, while September is about how much output an entire data center can squeeze from each megawatt of power.

Neither of the two September 15 announcements introduced a new chip; instead, NVIDIA listed ecosystem partnerships and partner test results. NVIDIA says Amazon's Annapurna Labs is working with it on the NVHBM custom high-bandwidth memory technology; d-Matrix is integrating NVLink Fusion to combine NVIDIA Vera CPUs with its own Raptor XPUs for ultra-low-latency inference at scale; and Pinterest is using the Blackwell platform and Dynamo inference software to bring conversational AI to visual discovery. NVIDIA gives each of these three items only a single sentence, with no specifications or timeline; the same post also includes Vera CPU test results from several startups and a section on NVLink 6 reliability.

Compared With January: The Chips Haven't Changed, the Metric Has

On January 5, NVIDIA unveiled the Rubin platform at CES, describing it as six new chips, including the Vera CPU and Rubin GPU. The headline numbers were about performance: up to a 10x reduction in per-token inference cost and a 4x reduction in the number of GPUs needed to train mixture-of-experts models, both compared with the Blackwell platform; the press release used future tense, saying partner products were expected in the second half of 2026. At the Hot Chips conference on August 24, NVIDIA announced that Groq 3 LPX had reached full production, describing the platform as spanning seven chips and five purpose-built racks; the only trademark note states that the Groq and LPU names are used under license from Groq, Inc., and the official documents checked for this article make no mention of NVIDIA acquiring or merging with Groq.

By September 15, neither official blog post introduced any new chip name; the January press release never used megawatts as a unit of measurement anywhere in its text, while the two September 15 posts repeatedly describe results in terms of tokens produced per megawatt. In other words, what's new on September 15 isn't hardware — it's that the DSX power-management software has reached the stage of its first round of partner testing and commercial demonstrations.

Why Power Is the Bottleneck: What DSX MaxLPS and DSX Flex Do

NVIDIA quotes founder and CEO Jensen Huang to explain why power is the bottleneck: “A one-gigawatt factory will never become a two-gigawatt factory,” Huang has said. NVIDIA writes in the same post that power is the constraint in the AI factory economy. DSX software was not launched on September 15 — NVIDIA states that DSX was introduced at GTC Taipei, an NVIDIA event held in Taiwan, in May, without giving a year; neither of the two September 15 announcements says anything further about deployments, partners, or supply chains in Taiwan, and this article does not extend the record on NVIDIA's behalf.

DSX MaxLPS handles power dispatch within a data center. NVIDIA says that at the factory level, MaxLPS dynamically shifts power across racks as demand rises and falls, delivering up to 40% more GPUs within the same site-power envelope, and up to 35% higher token throughput without requiring new power lines; at the rack level, separate power-smoothing software and expanded energy buffering absorb short spikes, though NVIDIA gives no percentage figure at that level. The one partner-measured result in this set of power figures is Lambda's: running 19 nodes under an 85% power policy within the power budget of 16 nodes at full power, NVIDIA says the cluster measured 24% more throughput and 23% better performance per watt — but that test ran on Blackwell-generation HGX B200 servers, not Vera Rubin, and Lambda's president of cloud services calls it a proof of concept. When that same 40% figure is attached to the Vera Rubin NVL72, the companion post states that it is based on NVIDIA's own projections and limited to suitable deployment environments; that post's figures table further notes it requires pairing MaxLPS with data center power planning, and is not a measured result.

DSX Flex handles communication between a data center and the power grid: it receives load-shedding requests, demand-response events, and pricing signals from the grid, and reacts automatically within a predefined workload hierarchy, keeping the most critical jobs running while everything else pauses and later resumes. NVIDIA's example is its own Eos data center, which took part in a flexible load interconnect program with Silicon Valley Power, paired with Emerald AI's Conductor: the text says power automatically dropped from 4 megawatts to 3 megawatts one August evening, and a photo caption on the same page states the same two figures; but a separate figures table in the same post describes that same demonstration as a 40% reduction in power demand within one minute, and NVIDIA does not explain the relationship between the two. NVIDIA says Silicon Valley Power has since sent more than 200 demand signals, every one of them successful, but this demonstration is not yet a formal DSX Flex installation — only proof that the concept works.

How to Read NVIDIA's Numbers: What's Measured, What's Projected

With Groq 3 LPX attached, NVIDIA's headline number is that for models with more than two trillion parameters at long context, the combined platform delivers up to 35 times the token throughput per megawatt of the GB200 NVL72; a more specific figure is that on a 100,000-token-context Qwen 3.8 27B workload, Groq 3 LPX measured 2,529 output tokens per second per user — both figures carry their own comparison baseline, and quoting either without its conditions would distort it. NVIDIA also cites a result from third-party firm SemiAnalysis, measured with its AgentX methodology: on the DeepSeek V4 Pro model, the Vera Rubin NVL72 delivers up to 30 times the throughput per megawatt of the GB300 NVL72 — this is NVIDIA relaying another organization's test, and this site has not obtained the underlying data.

On availability, the two September 15 blog posts give only one sentence: Vera Rubin is in full production and scaling across the ecosystem, and performance will keep improving with ongoing software optimization, with no shipment volume, customer list, or list of cloud regions. This particular post is a living document that keeps being updated: on September 16, the day after the event, it added a section of MLPerf Inference v6.1 results, which NVIDIA describes as the Vera Rubin NVL72's first MLPerf preview submission, delivering up to 3.7 times the throughput of the GB300 NVL72. That content belongs to September 16, not to what happened on September 15 itself.

What This Means for Cloud Service Users, and What NVIDIA Didn't Say

Every number above is a data-center-level metric — tokens produced per megawatt, performance per watt, how many GPUs fit within a given power allotment — and none of them is a promise about end-user cloud service speed, pricing, or delivery timing. Whether hardware efficiency ever shows up in a lower monthly subscription bill is a gap between inference cost and end-user pricing this site's earlier CES coverage already broke down, and it will not be repeated here; neither of the two September 15 announcements mentions a price or licensing fee for Vera Rubin, Groq 3 LPX, or DSX at any point, and the word “price” only appears when describing grid pricing signals.

As checked on September 17, 2026, neither post contains the words “Taiwan” or “TSMC”; there is no description of any Taiwan deployment, Taiwan partner, or supply chain. The only Taiwan connection is the one noted in the previous section — that DSX was introduced at GTC Taipei. Readers who want to confirm the current state for themselves can go back to the two blog posts directly; the summit post's last-updated timestamp moved forward again during the course of fact-checking. If readers later see an updated version, they should treat whatever the official page says at that time as authoritative.

2026 Tech News Roundup: Key Points on Hardware, Platforms, Telecom, and Regulation

NVIDIA Adds CUDA-Q Logical to Its Open Source CUDA-Q Platform: A Resource Estimator, Not a Quantum Computer

• Tech & industry
• Tech news