"depreciating" is not being used in the right sense.
There is depreciation, which is taking the purchase price and dividing it across N number of years (typically 5). That's the D in EBITDA and is mostly used as a profitability calculation.
The depreciation of a GPU also gets mucked up in the current GPU financed market as well. DDTL loans. The people running the GPUs often don't even own the GPU, they lease it, so there is nothing for them to depreciate (D).
The analogy that a GPU is like a used car makes zero sense. There is no oil or tires to change on a GPU. They don't wear out in the same way that a rental car would. They are housed in climate controlled locations with clean power. They just don't fail the way that is portrayed in the press.
Useful life of a GPU is based on profitability. When does opex cost more than profitability?
Some companies, like mine, also have support contracts. Anything goes wrong with the GPU (or any part of the system), Dell comes and fixes it at no extra charge. We just migrate customers and workloads to hot spares while the parts are replaced.
As for compute going down in value... the 122TB of enterprise nvme and 2GB of ram in each server that I bought 2 years ago is now worth vastly more than I paid for it. I'm also renting my GPUs out for more money now due to supply being so tight and demand being so high.
If you're on a Mac, I've been hacking on my own fork of Restic Scheduler [0] to make it work almost exactly like Backblaze Desktop, except better (and open source). I wanted something I could just install as a menubar app, point it at my home directory, and forget. Turns out doing it this way is about half the yearly cost of BD, and way better performance. I've finally gotten it to the point where it actually works pretty darn well, especially if you're using B2 on the backend.
I modified it to auto generate a list of excludes based on a whole bunch of criteria. It only backs up what you need, as well as your system Brewfile. Turn on "smart backup" in the preferences, point it at your home folder and let it rip.
It'll show you how much is backed up and even estimated costs. For ~750GB, it is about $52/year, not bad at all.
I'm a 30+ year software engineer, but I don't know swift at all, so I've been using Codex. I've done my best to not make it full on slop, but there is probably some in there.
I need to improve the usability for building it yourself, but right now you just clone the project and tell Codex to build it with whatever developer account you're logged into xcode with (you don't need the apple subscription) and install it into /Applications.
Cerebras is a good example here. Largest IPO of 2026 and as of Friday, down 33% from their top and about $15 away from their initial price.
CFO was at Bird (a SPAC flop) and CEO was previously charged by the SEC with a felony... for cooking the books.
Everyone wants you to believe that a giant wafer is the future (and soon enough layers of wafers), but a P/E of $500, just doesn't make sense for a company selling AI fast tokens.
Especially with a whole bunch of other solutions just waiting for tapout and competing with everyone else for more and more memory allocations to be able to hold the models.
So a PE of 500 means it would take 500 years for the earnings of the company to equal the current market cap (price per share X number of shares). This implies absurd (almost certainly impossible) growth over the next 500 years. Of course anyone expecting to pull their investment out and spend it on retirement can’t be looking at a 500-year investment horizon. I suppose the 1% can, though. What the hell else are they going to spend their cash on?
I know it isn't FOSS, but I just plug a $500 Mac mini into my LG TV and use it with a wireless backlit keyboard/trackpad combo I got for $35 off the zon. IINA is a fantastic player. I rarely use the tv os.
More accurately... I'm long on a viable alternative to the current monopoly. We have two OS's for phones (android and ios), there is no reason why we shouldn't have the same for all AI hardware and software. The only one even close, is AMD.
I agree with you though, serving up inference is secret sauce for a lot of teams and not everyone publishes how to do it because of the costs involved in doing so. They need an ROI.
I'm reading the website and nothing about this addresses the compute running the models. If that's going to a third party (just like openrouter is), then there are no guarantees, other than words on paper.
Proving my point. Your prompt gets sent through TR, to another provider on the other end.
There are zero guarantees beyond "trust me bro" that the inference provider isn't taking your prompts and selling them to promptbase or one of a dozen other similar services.
Venice claims no logs, which may or may not be true, but what happens to your prompt after they proxy it to the service running the GPUs?
From their website:
"The GPUs that process your inference requests come from multiple decentralized providers, and while each specific provider can see the text of one specific conversation, it never sees your entire history, nor knows your identity."
Which is an absurd claim if your prompt has your company name in it.
It doesn't matter if it is encrypted in transport, once it hits the company running the GPUs, it is open season for them.
TR is end to end encrypted to the provider, and we offer providers like Tinfoil that are also end to end that we also attested using Secure Enclaves. OpenRouter doesn't provide that guarantee.
This is entirely transparent. TR is also 100% open source.
It'd be nice if you acknowledged the value of this.
There is depreciation, which is taking the purchase price and dividing it across N number of years (typically 5). That's the D in EBITDA and is mostly used as a profitability calculation.
The depreciation of a GPU also gets mucked up in the current GPU financed market as well. DDTL loans. The people running the GPUs often don't even own the GPU, they lease it, so there is nothing for them to depreciate (D).
The analogy that a GPU is like a used car makes zero sense. There is no oil or tires to change on a GPU. They don't wear out in the same way that a rental car would. They are housed in climate controlled locations with clean power. They just don't fail the way that is portrayed in the press.
Useful life of a GPU is based on profitability. When does opex cost more than profitability?
Some companies, like mine, also have support contracts. Anything goes wrong with the GPU (or any part of the system), Dell comes and fixes it at no extra charge. We just migrate customers and workloads to hot spares while the parts are replaced.
As for compute going down in value... the 122TB of enterprise nvme and 2GB of ram in each server that I bought 2 years ago is now worth vastly more than I paid for it. I'm also renting my GPUs out for more money now due to supply being so tight and demand being so high.
reply