The Computer Still Exists, Anon
“You spent $10,000 on local inference? Do you know how many tokens that buys?”
I just came across this site that claims you're under water financially for 40+ years if you buy inference hardware to run local models. It completely ignores the fact that you still own the hardware, and that hardware is not worth $0 the moment you buy it.
Buy a machine for $10,000, run it for a year, sell it for $8,000 after selling costs. The depreciation was $2,000. A comparison that charges the full $10,000 against one year of inference has misplaced $8,000.
People make the same mistake with a $40,000 setup. The bigger number gets a louder reaction. It doesn’t improve the arithmetic.
What the year actually cost
Owning has other costs. Assume 5% in the market: parking $10,000 in a GPU for a year is $500 in opportunity cost. Say $300 in electricity and $600 for setup and upkeep. Against that $8,000 resale, the year costs $3,400. Call it $283 a month.
That number is the one worth comparing to a provider’s bill.
Annual ownership cost = purchase price − net resale proceeds + opportunity cost + operating costs + time.
An API bill buys inference that gets consumed. The output can be worth millions. The tokens are gone either way. Hardware does the work and keeps most of what you paid for it.
Sometimes the box goes up
NVIDIA’s 96GB RTX PRO 6000 Blackwell showed up in US retail at $8,565 in March 2025. Earlier this year you could buy them for $9800. NVIDIA lists it at $16,000 now. Up about 87%. Those are new-card prices, so they don’t tell you what you’d get selling yours. They do show which way this market can move. (VideoCardz)
Now let that $10,000 setup sell for $15,000 net. After the same electricity, time and opportunity costs, the owner is $3,600 ahead before taxes, and a full year of 24/7 tokens.
Appreciation isn’t guaranteed. It’s also not fictional. An argument that can only recognize losses is going to say strange things about owning things.
The bigger hole is privacy
A lot of the data that would make an assistant genuinely useful is data I don’t want to upload: medical records, tax returns, financial statements, private correspondence, proprietary research. An assistant that has the whole file does work a chatbot fed a few sanitized excerpts can’t.
Local inference runs offline. Records, prompts, results: none of it leaves your machine. LM Studio supports exactly that workflow today. (LM Studio, Offline Operation)
That means years of medical history, a pile of tax documents, a private financial picture, worked on without handing person data to labs that can be data breached. It also means not scrubbing documents, which usually means deleting the details that explained the problem.
For data that can’t leave your systems, a public API isn’t expensive. It’s unavailable. At any token price. That capability is part of what the machine is buying, and it never shows up in the token math.
The other stale input is capability
As of September 15, 2026, Artificial Analysis lists these on its Intelligence Index v4.3:
- Qwen3.8-Flash-Next: 40
- Claude Opus 4.6, adaptive reasoning, max effort: 32 (estimated)
Qwen scores above Artificial Analysis’s estimated Opus 4.6 baseline. Try any of these models yourself and it's clear they're not toy chat bots, they do real work. And the gap between the latest SOTA model release and open-weights is slowly closing all the time. Practically speaking you can expect locally hosted models to be a few months behind SOTA.
A model at that score covers what most people actually use AI for: drafting, summarizing, extracting, working through documents. Comparing against a requirement to run the newest frontier model for every one of those tasks inflates the case for renting.
And the hardware fits inside the budget being mocked. Unsloth lists 96–114GB of total memory for its 4-bit versions of Qwen3.8-Flash-Next. NVIDIA’s 128GB DGX Spark is $4,699. I run this model on a single 96GB RTX 6000 at ~130 tok/s per session, up to 4 concurrently (Unsloth, NVIDIA Marketplace)
You don’t have to cancel the cloud
Owning an assistant doesn’t mean canceling everything. Keep the $60 per month Cursor Pro+ sub for the heavy coding, plus whatever usage runs over, for development work you’re comfortable hosting elsewhere. The workstation has no obligation to replace every cloud service you use. (Cursor pricing)
Control over the model is worth money too. Abliterated models are modified to reduce learned refusals. Research into model behavior, moderation and security runs on experiments mainstream hosted assistants won’t perform. Weights let you inspect, modify, and pin a fixed version so experiments repeat. GPT-6 won't even read logs for cyberbench runs without tripping "safety" refusals (What is abliteration?)
Specialist hosts like abliteration.ai and rented servers exist, but you'll spend about $60 an hour doing anything meaningful. Owning the box lets those experiments run next to the sensitive dataset, on hardware you control.
Cheap hosting still belongs in the comparison
Cheap hosted inference belongs in this comparison, fully. Artificial Analysis lists hosted Qwen3.8-Flash-Next at $0.15 per million input and $0.47 per million output. For modest usage on data you’re willing to upload, hosting can win comfortably. (Artificial Analysis)
That proves hosting wins for that workload. It says nothing about a machine used for private records, repeated experiments, and modified models.
Ownership also changes the next decision. Once the hardware is paid for, another overnight eval costs electricity and occupies the machine. No new invoice. The upside is whatever that extra run turns up.
Setup and maintenance take time and the power bill is real. Learning to run these models builds skills you keep. All of that belongs in the account, resale value included.
If you’re going to call the buyer a moron, at least account for the computer. All of these comparisons completely ignore the real economics, which are much more favorable than they're being framed.