Your assessment is missing the simple fact that FPGA can do things a GPU cannot faster

Yes, there are corner cases (many of which no longer exist because of software/compiler enhancements but…). But there is always the argument of “Okay. So we run at 40% efficiency but our GPU is 500% faster so…”

Nvidia is the Ford F-150 of the data center world, sure. It’s stupidly huge, ridiculously expensive, and generally not needed unless it’s being used at full utilization all the time. That’s like the only time it makes sense.

You are thinking of this like a consumer where those thoughts are completely valid (just look at how often I pack my hatchback dangerously full on the way to and from Lowes…). But also… everyone should have that one friend with a pickup truck for when they need to move or take a load of stuff down to the dump or whatever. Owning a truck yourself is stupid but knowing someone who does…

Which gets to the idea of having a fleet of work vehicles versus a personal vehicle. There is a reason so many companies have pickup trucks (maybe not an f150 but something actually practical). Because, yeah, the gas consumption when you are just driving to the office is expensive. But when you don’t have to drive back to headquarters to swap out vehicles when you realize you need to go buy some pipe and get all the fun tools? It pays off pretty fast and the question stops becoming “Are we wasting gas money?” and more “Why do we have a car that we just use for giving quotes on jobs once a month?”

Which gets back to the data center issue. The vast majority DO have a good range of cards either due to outright buying AMD/Intel or just having older generations of cards that are still in use. And, as a consumer, you can save a lot of money by using a cheaper node. But… they are going to still need the big chonky boys which means they are still going to be paying for Jensen’s new jacket. At which point… how many of the older cards do they REALLY need to keep in service?

Which gets back down to “is it actually cost effective?” when you likely need

Not small but… smaller than you would expect.

Most companies aren’t, and shouldn’t be, training their own models. Especially with stuff like RAG where you can use the highly trained model with your proprietary offline data with only a minimal performance hit.

What matters is inference and accuracy/validity. Inference being ridiculously cheap (the reason why AI/ML got so popular) and the latter being a whole different can of worms that industry and researchers don’t want you to think about (in part because “correct” might still be blatant lies because it is based on human data which is often blatant lies but…).

And for the companies that ARE going to train their own models? They make enough bank that ordering the latest Box from Jensen is a drop in the bucket.

That said, this DOES open the door back up for tiered training and the like where someone might use a cheaper commodity GPU to enhance an off the shelf model with local data or preferences. But it is unclear how much industry cares about that.

Prices rarely, if ever, go down and there is a push across the board to offload things “to the cloud” for a range of reasons.

That said: If your focus is on gaming, AMD is REAL good these days and, if you can get past their completely nonsensical naming scheme, you can often get a really good GPU using “last year’s” technology for 500-800 USD (discounted to 400-600 or so).

From a “compute” perspective (so not consumer graphics), power… doesn’t really matter. There have been decades of research on the topic and it almost always boils down to “Run it at full bore for a shorter period of time” being better (outside of the kinds of corner cases that make for “top tier” thesis work).

AMD (and Intel) are very popular for their cost to performance ratios. Jensen is the big dog and he prices accordingly. But… while there is a lot of money in adapting models and middleware to AMD, the problem is still that not ALL models and middleware are ported. So it becomes a question of whether it is worth buying AMD when you’ll still want/need nVidia for the latest and greatest. Which tends to be why those orgs tend to be closer to an Azure or AWS where they are selling tiered hardware.

Which… is the same issue for FPGAs. There is a reason that EVERYBODY did their best to vilify and kill opencl and it is not just because most code was thousands of lines of boilerplate and tens of lines of kernels. Which gets back to “Well. I can run this older model cheap but I still want nvidia for the new stuff…”

Which is why I think nvidia’s stock dropping is likely more about traders gaming the system than anything else. Because the work to use older models more efficiently and cheaply has already been a thing. And for the new stuff? You still want all the chooch.

Captchas don’t do shit and have actually been training for computer vision for probably over a decade at this point.

Also: Any “simple test” is fixed in the next version. It is similar to how people still insist “AI can’t do feet” (much like rob liefeld). That was fixed pretty quick it is just that much of the freeware out there is using very outdated models.