Earnings reviews from this season:

Nvidia designs semiconductors for data center, gaming and other use cases. It’s considered the technology leader in chips meant for accelerated compute and generative AI (GenAI) use cases. While that’s where it specializes, it does a lot more. Its toolkit includes chips, servers, switches, networking, AI models and cutting-edge software to optimize the hardware it provides. Owning more pieces of GenAI infrastructure means opportunity for more software-based product optimization.

The following items are important acronyms and definitions to know for this company:

Chips:

Connectivity:

NVLink Fusion allows companies to build “semi-custom” AI infrastructure with Nvidia and its integration ecosystem. GPUs are general-purpose in nature. They’re not granularly designed for every single niche use case like an Application-Specific Integrated Circuit (ASIC). This can help Nvidia capture more of that demand by pairing with Marvell and a few other partners to more easily emulate purpose-built hardware.

The Nvidia GB300 NVLink72 is its rack-scale computing system. Rack scale means the entire server rack powers computation rather than a single server. Because this includes Blackwell chips and NVLink switches, it’s partially in the compute bucket and partially in networking. This aggregated product is the core revenue driver right now.

Software, Models & More:

DGX: Nvidia’s full-stack platform combining its chipsets and software services.

Gross margin fell Y/Y as expected due to the Blackwell-ramp. That revenue stream is replacing a lot of demand for its more mature Hopper products, which is weighing on margins. The headwind eased Q/Q, which enabled the sequential margin expansion. OpEx rose by 38% Y/Y to support higher growth-related infrastructure costs, as well as more engineering and research headcount. 

Let’s take a moment to appreciate what we see below, using AMD for comparison. Nvidia’s 66% operating margin is far higher than AMD’s 54% gross margin. That’s not because AMD is a bad company. Lisa Su is a great CEO, they dominate the CPU landscape and they’re making great progress with GPUs as well. This simply highlights two very powerful things. First, Nvidia’s GPUs are excellent and yield considerable pricing power. Second, their ecosystem is wonderfully sticky and margin-accretive as they layer software and other products on top of their GPU-based offerings. We can bicker all we want to about how quickly AMD is closing the technological and ecosystem gaps. But? If those gaps were truly closed, Nvidia pricing power would wane and the giant margin lead would start to shrink a lot more quickly than it is.  Again… AMD is a great company and they are winning their fair share of GPU deals. Nvidia is just a different animal.

They reiterated $500B in Blackwell and Rubin combined revenue for calendar 2025 and 2026. This implies about $350B in Blackwell and Rubin revenue over the next 14 months (it's a calendar year guide & their fiscal year doesn’t match the calendar year). Analysts currently expect to generate roughly $330B in total revenue over those same 14 months. That should mean 6% upside to these forecasts from Blackwell and Rubin alone, with some modest incremental upside from non-Blackwell-and-Rubin-related segments.

Nvidia trades for 32x forward EPS. EPS is expected to grow by 56% this year and by 47% next year.

Data Center – Compute:

Demand for the Nvidia GB300 NVLink72 was the star of the quarter. Blackwell Ultra, the newest member of the Blackwell GPU family, is thriving. Why? Because it’s an incredible product. Jensen cited Semi-Analysis InferenceMAX benchmarks that rank Blackwell #1. It also topped MLPerf’s inference and training ranking leaderboards across every category.

Blackwell’s technological strength is powering “off the charts” sales momentum and 56% Y/Y compute revenue growth. It’s also why “cloud GPUs remain sold out,” which is so important. A massive piece of Blackwell demand comes from giant public clouds like Oracle Cloud Infrastructure (OCI) and Azure purchasing large sums of GPUs and then renting that capacity to cloud customers. To keep that demand source humming, these public cloud players need to be able to turn around and rent access for high enough rates to make the CapEx worth their while. All of the giant cloud vendors say these purchases are tied to near-term revenue opportunities and clearly they think those opportunities still look great.  They’re purchasing as much as Nvidia can sell them.

How about another positive indirect demand signal? Multi-year cloud service agreements rose by roughly 100% sequentially. These are contracts Nvidia has with cloud vendors to provide services that support its own scaling. They clearly feel the need to vastly bolster their commitments, which is good news for the public cloud and the entire AI cycle.

Looking ahead, Rubin remains on track for scaled production during the second half of 2026. This will include the Rubin Context Processing Unit Extended (Rubin CPX) that will specialize in (as the name indicates) long-context processing (thinking harder/longer to give a better answer). It’s expected to deliver “x-factor performance leap increases in performance per dollar and costs” vs. Blackwell. They weren’t ready to precisely quantify that.

Data Center – Networking:

Networking revenue soared by 162% Y/Y as demand for the eXtreme Data Rate (XDR) InfiniBand products remained stellar. XDR doubles the bandwidth of the previous Next Data Rate (NDR) Infiniband offering. Just like with GPUs, massive leaps forward for networking performance will inherently drive stronger motivation and sense of urgency to upgrade. Pace of progress is so important for cycle longevity. Without it, the incentive to replace expensive assets and accept some operational disruption becomes less powerful. It’s great to see them advancing so quickly. 

It’s not just InfiniBand that’s rocking and rolling. Relatedly, NVLink and Spectrum-X Ethernet AI connectivity both remained very strong too. Whether it’s Meta, Oracle, Microsoft or others using Spectrum-X to enhance data center scale-out (connecting more server racks in a data center), Nvidia has quickly built a commanding market position in this highly important data center subcategory.

More on Data Centers:

Sovereign AI momentum remains notable. They’re working with Microsoft and a few others on AI infrastructure in the UK, while Nvidia also announced a £2 billion investment in that nation. In the USA, Nvidia and Oracle are building a new supercomputer with 100,000 Blackwell GPUs for the U.S. Department of Energy. And finally, Nvidia and South Korea are working with other industry titans in that nation, like SK Group and NAVER Cloud, to boost that nation’s AI infrastructure. This will include a 250,000-GPU deployment.

It introduced the BlueField-4. This is a data processing unit (DPU) that’s well suited for assuming infrastructure-related tasks from CPUs and completing that type of work more efficiently.

Data Center – Depreciation Debate:

Nvidia leadership was clearly prepared to counter Michael Burry’s argument about unfair lengthening of depreciation schedules for GPU useful lives. According to CFO Colette Kress, Nvidia’s Ampere GPU fleet still has a 100% utilization rate. This product came out 5.5 years ago. Hyperscaler depreciation schedules for GPUs range from 5-6 years. Seems fair to me! One of two things is true. Either longer useful lives are entirely fine and Burry is wrong. Or? Every single Ampere customer is incompetent, doesn’t understand their own businesses and is unaware that they’re irrationally using assets that carry negative value. I find that hard to believe. I’m quite comforted by this new disclosure and the fact that Hopper is still generating $2B in quarterly revenue as of Q3.

Why are useful lives for GPUs lengthening? I think (and so does Nvidia) it’s because the software systems in place to extract performance gains from this hardware are so much better. In the words of Jensen Huang, CUDA “extends the life of Nvidia systems well beyond original useful life.” That’s why they’re so frequently called CUDA GPUs. All of their chips use this software layer to extract significant incremental performance from these GPUs, augmenting efficiency and delaying antiquation.  It’s also worth noting that the pace of GPU progress from generation to generation will naturally slow as Nvidia rolls out more advanced technology. It’s easier to deliver performance gains from a lower base. Meaning? Older models will become out-dated more and more slowly over time. That should support even longer useful lives over the coming years if anything.

Nvidia was also adamant that its GPUs + CUDA are the only way to extend the useful life of chips to 5-6 years. Without this combination, performance bottlenecks quickly appear in just a few years. I’m sure Alphabet, Broadcom and other custom accelerator vendors would take issue with this statement, but this is Huang’s point of view. They were asked about custom accelerator market positioning vs. Nvidia’s more general-purpose GPUs. The combination of fantastic performance and the deeply integrated software makes the team confident they’ll keep winning a lot of these large deals.

Data Center – Cycle Runway Progress Report:

Automotive & Robotics:

Nvidia unveiled the DRIVE AGX Hyperion platform for autonomous vehicles (AV). This provides the compute, sensors and other tooling needed to accelerate AV progress and deployment. As covered already, it also partnered with Uber to build one of the largest AV fleets on the planet starting in 2027.

In robotics, Amazon, Caterpillar, Figure, Foxconn, Toyota and others are all using Nvidia’s physical AI tools like Cosmos and Omniverse. Speaking of Omniverse, Nvidia shipped its “Omniverse DSX” as a data center digital twin offering to optimize AI factory performance.

Robotics will become an increasingly powerful growth driver for Nvidia in the coming years as humanoid models proliferate and physical AI takes hold. That’s why it’s so important for the mega-cap to have a highly capable synthetic data generation offering. Collecting physical data costs a ton. This makes that data collection a lot cheaper, as you don’t need nearly as much of it to generate the same amount of insight. In turn, that ensures Nvidia is the best company to run simulations and physical AI with and also expedites global adoption.

Other Notes:

Nvidia is looking to use its balance sheet to fund some equity investments that expand CUDA’s scale, support AI adoption and leverage the potential footprint for its hardware. It also likes the idea of working more closely and directly with what it views as the most exciting disruptors in AI. That was the motivation behind the (10 gigawatt) OpenAI deal & other investments like Mistral and Anthropic. This should deepen reliance on Nvidia’s suite of products and strengthen its position in the market. It’s also worth noting that while Nvidia may invest up to $100B in OpenAI, they have not invested that full amount at least yet. 

Great quarter. The company continues to comfortably surpass expectations despite training the entire analyst community to expect large outperformance for the last few years. The demand signals for the all-important data center segment look good and their technological lead in the category, despite AMD’s progress, remains strong based on 3rd party evaluations. This is an iconic American growth story with a legendary founder and a few years of delivering the most impressive scaled revenue ramp I’ve ever seen. 

I continue to place this in my too hard pile solely because I do not want to make timing the end of a cycle a vital part of my process. Semiconductors will always be cyclical and while Nvidia has insulated itself a bit with software-related growth, they are not immune to cyclical ebbs and flows. I’d rather get exposure to these tailwinds elsewhere and continue praising Nvidia’s amazing results from the sidelines. Congratulations to bulls on another remarkable performance for this company. Take a bow.

From a market perspective, I found this report encouraging. The AI GDP tailwinds are fully intact. I’m incrementally more comfortable with GPU depreciation schedule fairness, which is so important for Meta, Amazon, Google and Microsoft. Customers are enjoying real value from these AI investments. All good news.