▸ AI in practice · 6 min read
Is it feasible to own your own AI models?
Everyone is telling you to own your AI.
The security-advisory crowd is saying it. Big corporates are saying it. It shows up in the client projects I am on, usually with the word “privacy” attached: run your own model, keep your data in house, stop renting from the big labs. I run an edtech company, so I hear it constantly.
I get the appeal, and I agree with parts of it. But most people saying “let’s just host our own” have no idea what they are actually signing up for. So it is worth sitting with the real question, the way I do with the founders I work with: is it feasible to own your own AI models?
Let me walk through both halves of the answer: what it costs to buy, and the much bigger cost nobody quotes.
The sticker shock
There is a memory supercycle happening, and it is brutal. AI datacenters are eating the world’s memory. Samsung, SK Hynix and Micron have shifted the vast majority of their output to HBM for AI chips, and standard RAM is the collateral damage.
The numbers are not subtle. A 32GB DDR5 kit that cost around 80 US dollars in mid-2025 crossed 375 to 430 dollars by early 2026, a 300 to 500 percent jump in under a year. TrendForce logged DRAM contract prices rising roughly 90 percent in a single quarter, and one supplier told customers to plan for 10 to 20 percent increases every month through the end of the year. Gartner expects it to push average PC prices up 17 percent and drive the worst year for PC shipments in over a decade.
And it is spreading. GPU prices are up roughly 20 percent; the RTX 5090 that launched at 1,999 dollars now sells on the street for 3,700 to over 5,000. On some cards the memory now costs more than the whole chip did a year ago. The rumors say CPUs and motherboards are next, and board makers are quietly reviving old DDR4 platforms just to keep a cheaper entry point alive. Nobody serious expects real relief before 2028.
So before we even talk about running a model: the door to owning the hardware is getting more expensive by the month, and the average person will struggle to run anything locally unless it is heavily quantized.
What it actually takes to run one
Here is the ladder, using the 4-bit quantization most people run. Rough rule: about half a gigabyte of memory per billion parameters, plus overhead.
| You want to run | Realistic setup | Memory (Q4) | Rough build cost (US) |
|---|---|---|---|
| 4B–8B | one 16GB GPU (RTX 4060 Ti 16GB) + 32GB RAM | ~3–8GB | ~US$1,300–1,700 |
| 12B–14B | 16GB GPU, or 24GB for headroom | ~9–10GB | ~US$1,400–1,900 |
| 32B | one 24GB card (used RTX 3090 or RTX 4090) + 64GB RAM | ~20–24GB | ~US$2,800–5,500 |
| 70B | two 24GB cards, a 48GB pro card, or a 96GB Mac Studio | ~42–48GB | ~US$4,500–10,000 |
| 200B+ (e.g. Qwen3-235B) | ~130–140GB of memory: a used big-RAM Mac, or a multi-GPU server | ~130–140GB | ~US$10,000–40,000+ |
A 4B or 8B model on a single card is genuinely accessible and useful. Everything above 32B is where the crisis bites, and the top of the ladder just got worse: a 235B model needs around 140GB of memory, which used to mean a big unified-memory Mac Studio. Then in March 2026 Apple quietly pulled the 256GB and 512GB Mac Studio options as the DRAM squeeze reached even them, so a new M3 Ultra now tops out at 96GB. Running a 200B-class model at home actually got harder this year, not easier.
And because I am in Hong Kong, the geography matters:
| Component (mid-2026, indicative) | USA | EU (incl. VAT) | Hong Kong (free port) |
|---|---|---|---|
| DDR5 32GB kit | ~US$400 | ~€460 | ~HK$3,100 |
| DDR5 64GB kit | ~US$750 | ~€860 | ~HK$5,800 |
| RTX 5090 (32GB) | ~US$3,900 | ~€4,600 | ~HK$31,000 |
| RTX 4060 Ti 16GB | ~US$425 | ~€490 | ~HK$3,300 |
| Mac Studio M3 Ultra (96GB, now the max) | ~US$5,299 | ~€6,200 | ~HK$41,000 |
Hong Kong has a real structural edge: it is a free port, so no VAT, no GST, no import duty on this gear. A buyer on price.com.hk generally pays less than a European eating about 20 percent VAT, and lands close to or under a US buyer once you add American sales tax and tariffs. It is one of the few places the shortage stings a little less.
Sources and caveats: price figures are drawn from TrendForce, Tom’s Hardware, Gartner, Newegg, TechPowerUp and Apple, cross-checked against public GPU and RAM price trackers, and are indicative for mid-2026 (prices in this market move weekly). The component table lists the key parts only, the RAM, the GPU, or the Mac itself, not a full working machine: a complete build still needs a CPU, motherboard, power supply, storage, case and cooling on top. The Hong Kong column is computed on Hong Kong’s free-port basis, no VAT or import duty, rather than scraped live, so confirm the day’s numbers on price.com.hk before you buy.
The hardware is the cheap part
Now the part almost nobody preaching “own your AI” mentions.
Buying the machine is not owning the capability. Running a local model well is not prompt engineering. You are building AI workflows on top of it, managing serving and context and updates, and that is real, ongoing engineering. I watch people stand up subpar setups and then wonder why the product feels worse than the API they replaced. It feels worse because it is worse. The hard part was never the download.
The honest total-cost studies say it plainly. The real bill is human time: deploying, monitoring, patching, updating models, and being the one who gets up when a GPU dies at 2am on a Saturday. That is 10 to 20 engineer-hours a month at the low end, and 20 to 30 percent of a senior engineer, 3,000 to 6,000 dollars a month, for anything production grade. Add electricity, and a quieter kind of depreciation: your rig does not lose its value because a new model dropped, it loses value because the frontier keeps moving to larger models it cannot run, while the big providers keep their API prices subsidized below what you can match at home. The economics drift away from you while the machine sits in the corner. Analysts land on a true cost of three to five times the number on the spreadsheet that started the conversation.
I have watched large companies spend hundreds of thousands of dollars on equipment that then sits severely underutilized, because the setup was never something you just switch on. It needs expertise to build and expertise to keep alive, and they budgeted for neither.
When local actually makes sense
I want to be clear, because I am not anti-local. I run local models myself for menial tasks, within what my hardware allows. I am genuinely excited about where this goes, especially on device. The day your phone runs a capable local model for the small, private, constant things has real value, and the beauty of that case is that it is specific and you are not maintaining a server farm to get it.
So local wins in specific situations: truly sensitive or regulated data; volume high enough that the economics flip, which is roughly millions of tokens a day before a self-hosted setup breaks even against the API, often 18 to 24 months out; a workload narrow enough that you set it up once and it just runs.
Where it goes wrong is the blanket “for privacy, we must run everything local,” usually said early, usually by people discounting that privacy-safe cloud setups exist and serve regulated businesses every day. And I would almost never do it in early prototyping. Unless the local setup is the entire point of your project, prove the AI part works on a managed service first, then migrate if the case is really there. Do it the other way around and you will burn your first three months on GPU hours and infrastructure instead of finding out whether the thing is even worth building.
So, is it feasible?
Yes, and no.
Yes, if you have a specific, real reason: genuinely sensitive data, volume high enough that the economics flip, or a narrow workload you set up once and leave running. And yes, increasingly, for the small on-device things, which is the part of this future I am most excited about.
No, if you are doing it because a slide told you to own your AI. No, in early prototyping, before you even know the AI part works. And no, if “we need privacy” is the whole argument, because privacy-safe cloud exists and serves regulated businesses every day.
The memory supercycle is quietly pricing regular people out of the hardware while Big AI hoards it, and that is worth watching, because the whole promise of this era was access. Open weights and on-device models are what keep that door open, and I am betting they will.
But owning your own AI is like owning anything else. The purchase price is the part everyone talks about. The upkeep is the part that decides whether it was a good idea. The hardware was never the hard part. Owning it is.