A $3 trillion rally in NVIDIA’s shares since ChatGPT launched – triggering colossal capital expenditure globally on AI hardware – means its earnings are now market-moving macroeconomic events in their own right.
Yet as questions increase about how long capex outlays can continue at red hot levels – and as pressure mounts on CIOs and their peers to demonstrate ROI from AI investments – analysts had tough questions for NVIDIA’s executives on a Q2 call that saw it report $30 billion in revenues.
The Stack listened in. From AI ROI to Blackwall racks, generative AI efforts in industry and its new NIM agent blueprints, here’s our top X takeaways.
NVIDIA earnings key takeaways
1: AI ROI
Are giddy $30 billion/quarter revenue levels sustainable as scrutiny deepens on the ROI of AI infrastructure and application investment?
To CEO Jensen Huang the answer – asked a few ways by a few analysts – was simple: “The people who are investing in NVIDIA infrastructure are getting returns on it right away. It's the best ROI computing infrastructure investment you can make today… Our capacity gets rented right away.”
There’s a race to deliver improved model performance, myriad AI startups emerging and new architectures are set to deliver new breakthroughs, he said; from sovereign AI efforts, via AI startups and established enterprises.
“Recommender systems, custom ad generation targeting ads at very large scale… search, and user-generated content; these are all very large-scale applications that have now evolved to generative AI” Huang added.
AI or no, investment in data centre modernisation has a return, he said.
“Accelerated computing speeds up applications. It also enables you to do computing at a much larger scale, for example, scientific simulations or database processing; that translates directly to lower cost and lower energy consumed…That's really the core of the first platform transition, going from general-purpose computing to accelerated computing.
“It's not unusual to see someone save 90% of their computing cost. And the reason for that is, of course, you just sped up an application 50x…”
It is not, say critics, quite that simple: As Gartner's George Brocklehurst noted to The Stack: "The challenge we see – at least focussing on Enterprise segment – is with generative AI projects consistently getting to business value.
"Our data currently suggests a downward trend for AI projects successfully making it into production. In the short term we expect over 50% of generative AI solution deployments in enterprise will fail through 2026..."
2: What the hell is a “NIM”?
NVIDIA Inference Microservices, or NIMs, are basically a container runtime for AI; think images containing a model and all of its dependencies that can be used to spin up containerised AI models across either NVIDIA’s own DGX Cloud or elsewhere. Talking of ROI… “AT&T realised 70% cost savings and eight times latency reduction after moving into NIMs for generative AI, call transcription, and classification” boasted Huang.
This week the company added new NIM “agent blueprints” for enterprise developers looking to build AI apps. Customisable early blueprints include a “digital human workflow for customer service, a generative virtual screening workflow for computer-aided drug discovery and a multimodal PDF data extraction workflow” for enterprise RAG applications.
Ffrom a customer perspective that’s a “jump start” to get building; from an NVIDIA side it’s arguably a way to further lock customers into not just its hardware but a more ARR-centric software ecosystem.
3: Blackwell progress
Amid rumours of issues with its next-generation Blackwell platform, NVIDIA admitted on the call that it had made some design changes but Huang claimed that “there were no functional changes necessary.” (Tom’s Hardware has a more detailed look at some of the issues/changes here.)
“We're sampling functional samples of Blackwell, Grace Blackwell, and a variety of system configurations as we speak. Blackwell production ramp is scheduled to begin in the fourth quarter and continue into fiscal year '26. In Q4, we expect to get several billion dollars in Blackwell revenue…
“There are something like 100 different types of Blackwell-based systems; we're enabling our ecosystem to start sampling those,” Huang added.
4: “The supply chain hates us doing integration”
NVIDIA shows pictures of Blackwell as a rack. It’s also selling more and more interconnect equipment. It’s selling a GB200 NVL72 “exascale computer in a single rack”. Media images tend to show integrated racks.
Is NVIDIA going to start eating into OEM partner space here?
Huang was quick to dismiss this suggestion.
“The Blackwell rack system is designed and architected as a rack but it's sold in disaggregated system components. We don't sell the whole rack.
Our customers hate that we do integration. The supply chain hates us doing integration. They want to do the integration. That's their value-add – Jensen Huang
“The reason for that is because everybody's rack's a little different. Some of them are OCP standards, some of them are not. Some of them are enterprise. And the power limits for everybody could be a little different.
“The way we designed it, we architected the whole rack. The software is going to work perfectly across the whole rack. Then we provide the system components; for example, the CPU and GPU compute board is then integrated into an MGX. It's a modular system architecture…
“That 3,000-pound rack… has to be integrated and assembled close to the data center because it's fairly heavy. And so… from the moment that we ship the GPU, CPUs, the switches, the NICs, from that point forward, the integration is done quite close to the location of the CSPs and the locations of the data centers. You can imagine how many data centers in the world there are and how many logistics hubs we've scaled out!
“I think because we show it as one rack and because it's always rendered that way and shown that way, we might have left the impression that we're doing the integration. Our customers hate that we do integration.
“The supply chain hates us doing integration. They want to do the integration. That's their value-add. There's a final design-in, if you will.
“It's not quite as simple as shimmy into a data centre; the design fit-in is really complicated… the installation, the bring-up, the repair and replace, that entire cycle is done all over the world. We have a sprawling network of ODM and OEM partners that does this incredibly well. We don't want to be an integrator, we want to be a technology provider,” he emphasised.
5: Software – and networking
The vast majority ($26.3 billion) of NVIDIA’s Q2 sales were to data centres.
They’re buying not just GPUs, but networking kit and increasingly, software. Networking revenue increased 16% sequentially and NVIDIA’s Spectrum-X Ethernet networking platform is “well on track to begin a multibillion-dollar product line within a year” said CFO Colette Kress.
(NVIDIA sells NVLink for GPU scale-up, Quantum InfiniBand for supercomputing and “AI factories”) and Spectrum-X for AI on Ethernet.)
“We expect our software, SaaS, and support revenue to approach a $2 billion annual run rate exiting this year” she added, significantly for NVIDIA, which offers a host of software libraries and increasingly things like low-rank adaptation (LoRA) model fine-tuning software support.
6: Industry AI users: What’s cooking?
Great, CSPs are buying up NVIDIA hardware and renting it out. Who’s actually using it in a big way and what for in the enterprise?
Firstly note said CFO Kress, “over the trailing four quarters, we estimate that inference drove more than 40% of our Data Center revenue.” (Inference being where powers learned in LLM training are put to work.)
As well as the advertising and search industries, automotive, healthcare and manufacturing are all investing heavily, NVIDIA executives said.
“[Like automotive], healthcare is on its way to being a multi-billion-dollar business as AI revolutionizes medical imaging, surgical robots, patient care, electronic health record processing, and drug discovery” said Kress.
On the manufacturing front, she name-checked electronics manufacturer, Foxconn, and Mercedes-Benz as using its NVIDIA Omniverse Cloud (3D software) to build industrial digital twins, with advertising agency WPP using its NIM software proposition for “generative AI-enabled content creation pipeline for customers such as The Coca-Cola Company…”
7: The robots are coming
Boston Dynamics, BYD Electronics, Figure, Intrinsyc, Siemens, and Teradyne Robotics are now among the companies using the NVIDIA Isaac robotics platform “for autonomous robot arms, humanoids, and mobile robots” added the company (Huang earlier this year said “There's just a giant suite of robotics companies that are emerging”). Watch this space.
NVIDIA’s Q3 revenue is expected to be $32.5 billion. Full-year operating expenses will be $4.3 billion and grow “in the mid- to upper 40% range as we work on developing our next generation of products” the firm added.
