Everyone wants a piece of NVIDIA.

01 / THE NVIDIA QUESTION

What got me looking at the libraries

I got interested in this while looking through NVIDIA’s physics libraries: Modulus, now PhysicsNeMo, and Warp. A lot of the underlying methods were familiar from research papers. Seeing NVIDIA put them together into software made me wonder what it expected to get out of the effort. [S46][S48]

My first reaction was: do we really need another toolkit? A collection of papers wrapped in an API doesn’t automatically excite me. I’m more interested in whether it saves someone enough work that they actually keep using it. Warp was the part I wanted to dig into more, so I wondered why NVIDIA was also putting effort into the broader collection.

Looking at what the tools actually do helped. PhysicsNeMo includes neural operators, graph networks and physics-informed neural networks, along with data pipelines, training utilities and examples. Getting those pieces to work together is useful work in its own right. Warp sits lower down: it compiles Python functions into CPU or CUDA kernels for things like simulation, and supports differentiation and connections to ML frameworks. They give a developer different starting points. [S46][S47]

That gave me a possible business explanation too. If the software makes a project easier to get running, more teams may try it. Some of those projects could grow into workloads that need more NVIDIA hardware. The library could be worth maintaining even if nobody pays for it directly. That’s the connection I wanted to investigate; the filings can help test it, but they can’t tell me every reason NVIDIA made the decision. [S01]

I’m still more drawn to something like Warp, where a developer gets building blocks and can decide what to do with them. But I can see the appeal of a more complete example if it gets a team past the setup work. My test for either would be pretty practical: does anyone keep using it once the demo is over, and does that turn into more hardware spending?

The documentation includes applications and contributors, so there are people finding uses for the tools. What I wanted to understand was how much of that usefulness comes back to NVIDIA when a customer has more hardware options. I started putting the company results alongside each other to get some context.

Each prior-year quarter is indexed to 100. Different periods and revenue definitions: NVIDIA and AMD total revenue, Broadcom AI semiconductor revenue, Cerebras non-GAAP core revenue. These bars cannot be added together or used to calculate accelerator market share.

The growth figures made the question more interesting. A competitor can grow quickly while NVIDIA keeps growing too, particularly when the overall market is expanding. That means I need to look at pricing, margins and where new spending goes, rather than treat a competitor’s growth as an equal loss for NVIDIA.

02 / THE NVIDIA QUESTION

What a competing chip has to get right

I get skeptical when a “NVIDIA killer” pitch stops at the chip specs. I want to know how the thing gets manufactured, packaged, delivered and put to work. A strong chip result is worth paying attention to, but the customer still has a whole system to run.

The engineering choices also affect each other. Adding compute can add heat and power demand. More memory bandwidth changes the package and its cost. A larger design can make manufacturing losses more expensive. Then there’s verification before fabrication, and software that has to make good use of the finished hardware. [S01][S44]

One thing I had to separate was the difficulty of making chips from the advantages NVIDIA itself owns. NVIDIA is fabless. TSMC runs the manufacturing process and fabs, while memory and packaging suppliers bring their own expertise. NVIDIA benefits from its planning, scale and work with those partners, but competitors can use some of the same suppliers. A manufacturing bottleneck can make entry difficult without belonging exclusively to NVIDIA. [S01][S44]

Try moving yield and volume separately in the worksheet. Lower yield makes each usable part more expensive; lower volume spreads the design bill across fewer sales. That helped me see why cheaper engineering alone might not be enough. It also leaves room for a specialist that can make the numbers work for a narrower set of customers.

For me, the comparison gets interesting when someone can show the cost of running the customer’s actual workload at the required performance and reliability. I’d include power, utilization and the engineers needed to keep it working. If a supposedly cheaper chip leaves a team doing months of extra software work, I want that in the comparison too.

03 / THE NVIDIA QUESTION

How much of this can AI make easier?

This was the next thing I wanted to check. If AI is getting better at coding and engineering work, does a smaller chip team have a much better shot now? There are useful examples: Google’s AlphaChip works on floorplanning and has been used in TPU designs. NVIDIA’s ChipNeMo looks at engineering assistance, EDA scripts and bug analysis. Those cover particular tasks, so I’d be careful about extending the results to a whole chip program. [S42][S43]

I’m excited about AI taking some of the tedious engineering work out of this. I just want to be honest about which work it removes. In the example below, speeding up 12 months of work by 2× while leaving eight months unchanged takes the total from 20 months to 14. I’d happily take those six months back. I wouldn’t describe the whole program as twice as fast.

Fabrication, packaging and validation can still take time after the design work speeds up. AI may help with those stages too; TSMC and NVIDIA describe applications in manufacturing, simulation and inspection. I just wouldn’t use a simulation speedup as a stand-in for how quickly a supplier can deliver qualified hardware. [S44][S45]

The part I’m wary of is assuming better AI mainly helps the challenger. NVIDIA gets to use it too, and ChipNeMo shows it adapting models to chip-design work. I’d expect a pile of design history and engineers who know what went wrong last time to be useful here. I don’t have a number for that advantage, which is why the two speedups stay separate in the worksheet. [S43]

04 / THE NVIDIA QUESTION

Who is trying to build an alternative?

The companies in this group aren’t all trying to do the same thing. AMD offers a GPU alternative. A cloud provider can build a chip mainly to replace some of its own spending. A specialist might only need to be good at one workload to find a business.

I’m interested in who gets funded, but I get a lot more interested when someone pays to run production work and then comes back for more. A funding round buys a team the chance to prove something. I don’t want to give it credit for already having proved it. The explorer keeps those stages separate.

I’d use Crunchbase, PitchBook or Dealroom to find companies worth following, then look for filings and evidence of deployments, utilization and repeat business. Funding alone doesn’t tell me which chip is taking work from NVIDIA. I’d also check whether a reported deal is a new round, a revision to an earlier announcement or debt, so the same money doesn’t get counted twice.

05 / THE NVIDIA QUESTION

What would it take to switch?

This brought me back to the software question. If a team already has a working system, what would make it worth moving? The comparison includes its code, debugging tools and all the things the operations team has learned about keeping it running.

This is where I think NVIDIA’s software spending starts to look pretty sensible. A team picks up a useful library, adds its own optimizations and builds more work around it. Eventually, changing the chip means touching a bunch of things that already work. I can see why a team would need a fairly good reason to volunteer for that.

Reading the physics software also made me pay closer attention to backends. The license tells me what I’m allowed to do with the code; the supported backends tell me where I can run it. NVIDIA has open projects and products with specific GPU requirements, so I need to check the particular tool rather than infer portability from its license. [S28][S30][S34][S35]

That’s the part of the moat I find most convincing so far. It doesn’t require an elaborate plan to close the code later. Being the system people already know how to build and operate can be a pretty good position. Whether it holds up depends on how much better or cheaper the alternative really is.

06 / THE NVIDIA QUESTION

The funding relationships were interesting too

The financing is where I started opening more tabs. Suppliers investing in customers, capacity agreements helping fund data centers, warrants encouraging a buyer to commit: there’s a lot going on around the hardware order. I put a few of the documented relationships into the map because I wanted to keep track of who was funding whom.

There are sensible reasons for those arrangements. They can help get infrastructure built and give customers access to compute. What interests me financially is how much expansion depends on continued financing, and whether several companies end up exposed to the same slowdown.

I haven’t added the relationships into one dollar total. Buying equity, lending to a customer, issuing warrants and committing to capacity create different obligations. I’d rather keep the terms visible than put unlike figures into a single headline number.

I’m not going to call a customer investment fake demand just because the diagram loops back on itself. I do want to know whether the customer can pay without another financing round. That sends me back to cash collection, receivables, payment terms and customer concentration, along with evidence that the compute is actually being used. [S01][S02]

07 / THE NVIDIA QUESTION

Working through the numbers

By this point I wanted to put some numbers around the argument. The worksheet separates market growth, NVIDIA’s share and the margins on that revenue. They can move in different directions, which is easy to lose sight of when reading competition headlines.

Try taking share away from NVIDIA while leaving margins alone, then start moving the margin and cost inputs. This is why I’m reluctant to turn “more competitors” straight into “bad investment.” The revenue can keep growing. The part I’m less willing to take for granted is how much profit NVIDIA keeps. These presets are examples, not cases I’ve assigned odds to.

It also makes me cautious about treating the semiconductor industry as one investment. Chip designers, foundries, memory suppliers and cloud operators can all benefit from more spending, but they have different costs and bargaining positions. I’d want to work through those separately before deciding who keeps the return.

08 / THE NVIDIA QUESTION

Where I’ve landed so far

I think the software effort makes a lot of commercial sense. If the tools help people get useful work running, NVIDIA can benefit well before it ever charges for the software. And once a team has built around that stack, a competing chip has more work to do than win a benchmark. That’s where I’ve landed on the moat argument. The stronger claim that NVIDIA plans to make all physical AI software NVIDIA-only still needs evidence I haven’t found.

I’m going to pay more attention to production migrations, repeat purchases and cash collection than to the next claim that someone has killed the GPU. I think the alternatives are worth taking seriously. I also think making NVIDIA easier to leave is a much bigger job than designing another chip. Whether that keeps its current economics intact is the part I still want to watch.