NVIDIA Gets the Headlines. These Companies Get the Checks.
NVIDIA just reported another staggering quarter: $96.2 billion in revenue, up 106% from a year earlier, with Data Center contributing $89.0 billion and gross margins holding at 75%. The company sits at the center of the AI conversation for good reason. Its GPUs power the largest training runs and the inference clusters that keep turning model demos into actual products. Yet there is a quiet fact that still surprises many investors: NVIDIA does not manufacture the semiconductor wafers inside those GPUs.
NVIDIA is fabless. It designs the chips, owns the architecture, and has spent nearly two decades building the software and interconnect layers that make the hardware useful. The actual silicon is fabricated, packaged, and tested by other companies. Think of NVIDIA as the architect and general contractor. It owns the plans and gets its name on the mailbox, but an awful lot of electricians, plumbers, and specialized contractors have to show up before anyone can move in—and none of them work for free. Every time a hyperscaler or AI lab writes a multi-billion-dollar check for accelerated computing, a long chain of suppliers gets paid along the way. Following one of those GPUs from design to data-center floor reveals an ecosystem far larger than any single stock.
The wafers themselves are produced primarily at TSMC, the world’s dominant pure-play foundry. TSMC holds roughly 70-plus percent of the global foundry market and an even higher share of the most advanced nodes that AI accelerators require. NVIDIA’s high-end GPUs run on TSMC processes; packaging capacity, especially TSMC’s CoWoS technology, has repeatedly been a tighter bottleneck than raw wafer starts. CoWoS places compute dies and stacks of high-bandwidth memory onto a silicon interposer so the parts can talk to one another at extreme speeds. As AI processors grow more complex—multiple large dies, eight or more HBM stacks—packaging has become both more valuable and more constrained. TSMC has been expanding CoWoS capacity aggressively, yet demand continues to press against supply.
Those advanced nodes cannot be manufactured without extraordinarily sophisticated equipment. ASML’s extreme-ultraviolet lithography machines are the only tools that can print the finest features at commercial scale; each system costs hundreds of millions of dollars. Applied Materials, Lam Research, and KLA supply the deposition, etch, and inspection systems that turn a patterned wafer into a working chip and keep yields high enough to be economic. When TSMC or a memory maker builds a new fab or adds capacity, these equipment companies receive large, recurring orders. The tools are so specialized that there are few realistic substitutes.
Memory is another critical and expensive piece. Modern AI accelerators rely on high-bandwidth memory—HBM—stacked vertically and placed close to the compute dies so data can move at enormous rates. Without it, the GPU sits idle waiting for information, rather like buying Ferraris for the whole family only to get stuck behind Grandpa doing 48 in the fast lane. SK Hynix has led the HBM market, with Micron and Samsung also supplying major volumes. NVIDIA works with all three. The recent tightness in HBM supply has been one of the clearer reminders that the AI boom is still capacity-constrained in places.
Once the packages leave the factory, they must be connected. Thousands of GPUs in a cluster need extremely fast, low-latency links. NVIDIA’s own NVLink handles high-bandwidth communication within a rack or domain. For larger fabrics it offers InfiniBand and its Spectrum-X Ethernet platform, purpose-built for AI traffic patterns. Networking is no longer an afterthought; it is a material portion of system cost and performance. Bottlenecks here waste expensive compute the same way a clogged kitchen sink wastes a perfectly good dinner.
The story does not stop at the chip. AI is increasingly an infrastructure story. A single large training cluster draws enormous power, requires sophisticated liquid cooling, transformers, switchgear, backup generation, fiber, and buildings measured in acres rather than square feet. In an earlier note, “Why Nvidia Is Willing to Backstop $250 Billion for One Data Center,” we examined the scale of a proposed roughly 10-gigawatt campus—land, generation, transmission, construction, cooling, and the GPUs themselves. The larger lesson was straightforward: the infrastructure surrounding the chips can cost extraordinary amounts of money. If investors follow the capital outward from NVIDIA, they discover an entire ecosystem of companies getting paid to manufacture, package, connect, cool, and power AI.
We appear to be in the early-to-middle innings of this buildout. Capital spending by hyperscalers and specialized AI cloud providers remains elevated, and new applications continue to surface. That does not mean every company touched by AI is a good investment, nor does it mean NVIDIA itself should be avoided. NVIDIA’s position is unusual. It increasingly sells a full computing platform—processors, networking, interconnects, and the CUDA software layer that has become the de-facto standard for accelerated computing. The installed base of CUDA-optimized code and the talent familiar with it create a genuine moat. Switching costs are real.
Risks remain. Valuations across the group can become demanding. Competition is intensifying as AMD improves its offerings and hyperscalers design custom silicon. Semiconductor manufacturing remains heavily concentrated in Taiwan. Capital spending cycles eventually moderate; today’s infrastructure boom will not compound at the same rate forever. And the oldest investment principle still applies: a wonderful company and a wonderful stock are not necessarily the same thing if you pay too much for it.
The useful question is therefore not whether to own NVIDIA or ignore it. It is simpler and more practical: if NVIDIA succeeds, who else gets paid? That question points toward foundries, equipment makers, memory producers, packaging specialists, networking businesses, cooling and electrical-equipment suppliers, and the power producers that keep the lights on. The AI investment opportunity is an ecosystem, not a single ticker.
As always, please call or email us if you’d like to discuss how the AI infrastructure buildout, the companies supplying it, and the enormous amounts of capital flowing into it may intersect with your financial plan or how we are positioning portfolios around the firms enabling this transformation.
Just remember—when the kids finally move out, the electric bill doesn’t drop nearly as much as you hoped. The same is true of data centers.
Your capital, our expertise, a bespoke creation.