Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How NVIDIA Transformed Into an Operations Powerhouse
Transcript
- Lucas: When people talk about NVIDIA, they usually focus on the stock price or the AI boom. But from an operations standpoint, what they've done over the past five years is genuinely underappreciated. Luna: Because everyone just sees 'chip company' — but it's so much more than that now. What's the real story? Lucas: The real story is that NVIDIA has transformed from a fabless semiconductor designer — think of a company that just designs chips and hands them off to a foundry — into a full-stack AI infrastructure operator. They now design the chips, but they also design the systems, the networking, the software stack, and they operate large-scale data centers to test and deploy their own hardware. Luna: So it's vertical integration, but not in the old-school way of owning factories. It's more like they control the entire user experience from silicon to service. Lucas: Exactly. And this shift didn't happen overnight. It really accelerated around 2020 with the acquisition of Mellanox, the networking company, for about $7 billion. That gave NVIDIA high-speed interconnect technology that's critical for tying together thousands of GPUs into a single compute cluster. Then in 2022, they closed the deal for Arm — which ultimately fell through — but the ambition was clear. They wanted to control the compute architecture from end to end. Luna: And that Mellanox acquisition is a great operations example. Because before that, NVIDIA was selling chips to data center operators who then had to figure out the networking themselves. After, they could sell an integrated system. Lucas: Right. And the DGX line — those are the big GPU servers — they're essentially supercomputers in a box. But what a lot of people miss is the logistics and supply chain complexity behind them. A single DGX system contains over 30,000 components. And NVIDIA doesn't just assemble them; they orchestrate a global network of suppliers and contract manufacturers. Luna: So how do they manage that? I mean, most companies would outsource the whole thing to Foxconn or something. Lucas: They do use partners like Foxconn and Wistron for final assembly. But NVIDIA keeps tight control over the critical path — the GPU itself, the memory, the networking chips. They also run their own internal supply chain planning software, which they call 'NVIDIA's digital twin.' It's a simulation of the entire supply chain that lets them run what-if scenarios. For example, if a factory in Taiwan shuts down for a week, they can instantly see the impact on delivery schedules for every customer. Luna: That's the kind of operational sophistication you'd expect from a company like Toyota, not a chip designer. How did they build that capability? Lucas: A lot of it came from hiring people with backgrounds in aerospace and automotive supply chains. Jensen Huang, the CEO, has always said that 'a computer is not a chip.' He means that the value is in the system — the cooling, the power delivery, the software — all of it has to work together. So NVIDIA has invested heavily in what they call 'full-stack computing.' They have teams that design the printed circuit boards, the power supplies, even the liquid cooling systems for their data center racks. Luna: And that's where the operations challenge gets really interesting. Because now you're not just managing chip fabrication; you're managing the production of complex electro-mechanical systems. Lucas: Exactly. And the newest generation, Blackwell, is even more complex. The Blackwell GPU is actually two dies connected by a high-speed bridge. That means NVIDIA has to source the dies from TSMC, then do the chiplet packaging — which is a specialised process — before the GPU even goes into a server. They've had to work closely with TSMC to ramp up CoWoS packaging capacity, which is a bottleneck for the entire industry right now. Luna: And TSMC itself is a fascinating operations story. But for NVIDIA, the relationship is strategic. They're TSMC's second-largest customer after Apple, and they get priority access. Lucas: Yes, but priority access isn't just about paying more. It's about operational alignment. NVIDIA shares its demand forecasts with TSMC years in advance. They co-invest in capacity. They put their own engineers on-site at TSMC's fabs in Taiwan. The level of integration between the two companies is remarkable. Luna: So it's not just a vendor relationship — it's almost a joint venture in operations. Lucas: That's exactly right. And this is where NVIDIA's operational philosophy differs from a traditional fabless company. They treat their foundry partner almost as an extension of their own factory. They've even developed their own process recipes with TSMC, which gives them a competitive edge in performance per watt. Luna: But let's talk about the demand side. AI companies are buying these systems as fast as NVIDIA can make them. How do you manage allocation in a supply-constrained environment without alienating customers? Lucas: That's a huge operational challenge. NVIDIA uses a system called 'fair share allocation' that's based on historical orders and strategic importance. They also have a program called 'NVIDIA Partner Network' for cloud providers and system integrators. But they've been criticised for not being transparent about how they prioritise. Some startups say they can't get GPUs while large cloud providers get thousands. Luna: Yeah, there was that whole controversy with CoreWeave and other GPU cloud providers. How did NVIDIA respond? Lucas: They've tried to increase transparency. They now publish a blog post each quarter with a supply chain update. But at the end of the day, they have to allocate to the customers who generate the most revenue — that's just business. What's interesting is how they've started to use their own data centers as a kind of buffer. They have this network of internal labs — called 'NVIDIA AI Labs' — where they test new hardware and software. But they also offer a cloud service called 'DGX Cloud' that lets customers access NVIDIA hardware without buying it. Luna: So that's a way to serve smaller customers without having to allocate physical hardware. And it gives NVIDIA real-time data on how their systems perform in production. Lucas: Exactly. It's a feedback loop. The operational data they get from DGX Cloud helps them improve the next generation of hardware. And it also helps them forecast demand more accurately. They can see which workloads are growing fastest — large language models, scientific computing, whatever — and adjust their production priorities. Luna: So in a way, NVIDIA has built an operations machine that learns from itself. That's pretty advanced. Lucas: And it's not just about the hardware. There's a whole software component. They have a team of over 20,000 software engineers who maintain CUDA, the AI software platform. That's an operational challenge in itself — managing a codebase that spans decades, while also enabling new architectures like Blackwell. Luna: Right, because every new hardware generation requires software optimization to get the full performance. It's a co-development process. Lucas: Exactly. And that brings us back to the core operations principle: everything is connected. You can't just design a great chip; you have to design the system, the software, the supply chain, the data center infrastructure. And NVIDIA has managed to do all of that simultaneously. Luna: You know, it's episodes like this that really show why operations is such a fascinating field. It's not just about moving boxes — it's about orchestrating complex systems at global scale. Lucas: If today's deep dive into NVIDIA's operations was useful to you, here's something I want to mention. The reason we can do episodes like this without any ads or sponsored content is that some listeners choose to support the show directly. It's a small thing — you can go to buy me a coffee dot com slash fexingo — and it helps us keep doing what we do. Luna: Yeah, it really makes a difference. And it means we can spend the time to research stories like this one — the real operational details that you don't hear in the mainstream press. Lucas: All right, back to NVIDIA. I want to talk about something specific: how they handled the transition from Hopper to Blackwell. That was a massive operational undertaking. Luna: Because you have to keep shipping Hopper while ramping up Blackwell, and you can't cannibalise your own sales too quickly. Lucas: Exactly. NVIDIA managed that by pricing the two lines differently — Blackwell is positioned as a premium product for the largest AI clusters, while Hopper continues to serve the mid-range. They also offered trade-in programs for some customers to upgrade. But the real operational trick was in the supply chain: they designed Blackwell to be pin-compatible with Hopper in some systems, so data center operators could swap out GPUs without redesigning their entire setup. Luna: That's a smart operational move. It reduces friction for the customer and smooths the transition for NVIDIA's own production lines. Lucas: And it shows how operations thinking can drive product design. The engineering team had to coordinate with supply chain, manufacturing, and customer support to make that backward compatibility happen. That's not easy in a company that's growing as fast as NVIDIA. Luna: So what's the takeaway for other companies? What can they learn from NVIDIA's operations playbook? Lucas: I think the biggest lesson is that operations is not a support function — it's a strategic differentiator. NVIDIA didn't become the most valuable company in the world just by designing great chips. They did it by building an operations machine that could deliver those chips at scale, with the right software, in the right systems, to the right customers. Luna: And they did it by redefining what a chip company even is. It's a reminder that your business model — and your operations — have to evolve as the market changes. Lucas: Absolutely. That's a good note to end on. Next time, we're going to look at a completely different kind of operations challenge — how a major airline turned around its on-time performance. Luna: Looking forward to it. Until then.