Jul 8, 2026

The PC I Didn't Buy

A few days of hardware research, one month's salary saved, and the realization that the machine was never the point.

It's past 11pm. My phone is on Do Not Disturb, battery at 69%, and I'm doing the thing I always do — scrolling chợ máy tính PC Hà Nội on Facebook, evaluating a build I have no real reason to buy. If you're the kind of person who treats "researching hardware" as a hobby in itself, you already know how this goes. Or you think you do.

This is a full recap of a deliberation that started as a straightforward "is this a good deal?" and slowly turned into something I wasn't expecting: a mirror. I'm writing it all down — the numbers, the rabbit holes, the rationalizations — because the technical stuff is genuinely useful, but the real lesson was about me, and probably about a lot of us who do this.

The build that started it

A used PC popped up on Facebook Marketplace, one month old, warranty to May 2029:

CPU: Intel Core Ultra 9 285K
Board: Asus PRIME Z890-P-CSM DDR5
RAM: G.Skill Trident Z5 RGB 32GB DDR5-6000
SSD: Samsung 990 Evo Plus 1TB
Cooler: DeepCool LE360 V2 ARGB (360mm AIO)
GPU: MSI RTX 5060 8GB Shadow 2X OC
PSU: Cooler Master MWE Gold 850W V3 (ATX 3.1)
Case: KENOO Esport AK400 + 2 fans

Asking price: 44 triệu. My plan was to immediately swap the weak RTX 5060 for a 5070 Ti 16GB — roughly a +17m move once you account for reselling the 5060 — bringing the whole thing to about 61 triệu.

My first instinct was that it was overpriced. I added up what I thought each part cost new and landed around 40–45m, which would make a one-month-old used build a terrible value — you'd expect maybe 33–37m for something like this. I was ready to write it off.

Then I got one number badly wrong, and it flipped the entire analysis.

The plot twist was RAM

I had the 32GB DDR5 kit priced in my head at around 3–3.5 triệu. That was old information. In mid-2026, that exact G.Skill Trident Z5 kit is closer to 11.4 triệu.

There's a serious memory crunch happening right now. DRAM and NAND prices are climbing — reports of up to 25% increases — driven by AI-datacenter demand vacuuming up supply. Some shops in Hanoi have stopped selling high-capacity kits standalone and only offer them bundled with a full PC build, because stock is that tight. The SSD is caught in the same squeeze.

Re-running the build with correct numbers put the new-retail total at roughly 50–53 triệu, not the ~44 I'd estimated. That means the 44m asking price was actually 12–17% below new — on a sealed-condition, one-month-old machine with two-plus years of warranty left.

More subtly: the seller bought that RAM before the spike. Buying this used build meant inheriting a component that had literally appreciated in value since purchase. In a normal market, "buy used to save money" is the play. In this market, buying used was also a way to dodge paying peak prices for RAM I'd otherwise have to buy new.

Lesson one, and it's a boring one: your mental price list is always stale. I've built enough machines to feel confident about component costs, and I was off by 8 million dong on a single stick pair. The market doesn't care what things cost last year. Re-check current prices before you judge any deal — the whole verdict can flip on one line item.

Then came the rabbit holes

This is the part where the deal-hunter brain really lights up. Once a purchase feels plausible, I don't research to decide — I research to savor. I went deep on five different angles, and every one of them was fascinating, and not one of them was the actual question.

The dead-end socket

The 285K sits on Intel's LGA 1851. I'd flagged it as a "dead-end socket," and when I looked at the roadmap, that held up — with nuance. LGA 1851 launched, lost its originally-planned first chips (Meteor Lake desktop got cancelled), and Arrow Lake became the first thing to actually ship on it. There's a minor "Arrow Lake Refresh" (clock bumps on the K/KF models, no new architecture). But the next real desktop generation, Nova Lake, moves to a brand-new socket, LGA 1954, with no backward compatibility. Panther Lake is laptop-only and never comes to desktop. So the 285K is essentially the last meaningful CPU its socket will ever hold.

That sounds damning — until I remembered something about myself: I have never once swapped a CPU on an existing board, because CPUs are the expensive part. I buy a machine and run it for years. So the entire "no upgrade path" flaw costs me exactly nothing. This is worth stating as a general principle: a weakness that doesn't apply to how you actually use things is not a weakness for you. Half the internet's AMD-vs-Intel arguments evaporate once you're honest about which benefits you'll never actually use.

Intel vs Ryzen

I compared against AM5. The Ryzen 9 9950X is the true same-league rival — they trade blows, with AMD holding a slight multi-threaded edge in mixed workloads and better performance-per-dollar. The 9950X3D is the smarter buy if gaming matters, since 3D V-Cache puts it at the top of gaming charts while matching the productivity. One detail genuinely relevant to my simulation work: the 9950X has full AVX-512 support, which the 285K lacks — useful for scientific/Monte-Carlo-style workloads.

But here's the thing: AMD's two biggest structural wins are gaming-per-dollar (via X3D) and platform longevity (AM5 is supported for years, so you can drop in a faster chip later). The longevity advantage is, again, something I established I will never use. And going AMD wouldn't dodge the RAM spike — an AM5 build needs the same expensive DDR5. So the AMD case shrank to "marginally better mixed-workload performance," while the 285K actually won on idle power efficiency, which matters for an always-on box.

The chip is almost too good

Benchmarks put the 285K firmly S-tier for productivity — it trades the multicore crown with the 9950X and sits behind only the halo Threadripper / mega-cache parts. For gaming it's more like A-tier: the X3D chips beat it, and even the regular 9950X edges it in some CPU-bound titles. But at 1440p/4K with a real GPU, games become GPU-bound and that gap mostly disappears. For my kind of use — sims, dev, VMs, limited dad-time gaming at normal resolutions — I'd never feel the CPU as a limit. If anything, pairing a 15-triệu flagship CPU with a budget 8GB GPU is backwards; the GPU was always the real bottleneck in this build, not the chip.

The Windows tax

A hybrid chip like the 285K (8 performance + 16 efficiency cores) leans hard on the Windows Thread Director scheduler to place threads correctly. Get it wrong and you lose performance — Arrow Lake had a rough launch partly for scheduling/latency reasons that needed OS and firmware patches. A uniform-core chip like the 9950X gives the scheduler nothing to mess up. On top of that, a 24/7 Windows box constantly burns cycles on Defender scans, Search indexing, telemetry, and forced updates, which raises the idle floor and interrupts you. The honest conclusion: this hardware wants to be a Linux box that occasionally boots Windows for games — which, given I already live in Proxmox / WSL2 / Tailscale land, actually suits me. And it saves the ~1.5m Windows license.

The power bill

I run things 24/7, so I did the math. At the wall, this config idles around 80W — the 285K idles okay, but a Z890 board, a 360mm AIO pump, three radiator fans, case fans, and RGB RAM all stack up. That's ~58 kWh/month. At Hanoi's top EVN tariff tier (my marginal kWh, with a household and summer AC, lands there — roughly 4,200–4,300 đ/kWh with VAT), that's about 250k/month, or ~3 triệu a year just to leave it idling. If it's actually doing work a good chunk of the time (~150W average), closer to 5–6 triệu a year. Not catastrophic, but not nothing — and notably, this workstation idles far higher than my HP Elite Mini homelab box, which sips 10–15W.

Every one of these rabbit holes was interesting. I could have written a separate post on each. But collectively they were a very elaborate way of avoiding the one question that actually mattered.

Buy, or keep the money?

I finally asked it plainly. With the GPU swap, this was a 61 triệu machine. That is a month's salary. For a want — not a need. I made myself say that out loud.

Then the follow-up I didn't want to answer: does my current setup actually stop me from doing anything I want to do? No. I have a working desktop with an RTX 5060 Ti 16GB and a homelab mini PC running Proxmox. Between them, they cover everything I actually do today.

There's a trap I walked right up to the edge of, and I want to name it clearly because it's the whole game: a good deal on something you don't need is still money you didn't need to spend. The discount is the bait. "It's 15% under retail and dodges the RAM spike" is a genuinely excellent reason to buy if you were buying anyway. It is not, by itself, a reason to convert a "maybe" into a "yes." The good-deal-ness and the do-I-need-it-ness are two completely different questions, and it's dangerously easy to let a strong answer to the first one masquerade as an answer to the second.

The part that actually mattered

Somewhere in here I said what I really wanted the machine for: I want to constantly backtest trading strategies. And then, almost as a throwaway, I admitted the real situation — I can't find new strategies.

That one sentence quietly dismantled the entire purchase. Here's why:

Backtesting is I/O and single-thread bound, not core-bound. Reading historical price data and running a strategy loop over it runs fine on a laptop. A vectorized backtest rips through years of data in seconds. The only time it gets heavy is huge tick-level datasets or massive parameter sweeps — and even then, renting cloud compute for the occasional big run beats owning a 24/7 workstation that idles 95% of the time.
My bottleneck was never compute. It was ideas. And no CPU on earth generates an edge. The constraint in algo trading is almost never "my backtests are too slow" — it's "I don't have a hypothesis with real, economically-grounded edge."

A faster machine would just let me test bad strategies faster. Worse, it would make it easier to overfit — to torture the same data until something looks profitable, then watch it evaporate the moment it meets a live market. I was about to spend a month's salary solving a problem I don't have (speed) while leaving completely untouched the one problem I do have (no genuine edge). You cannot buy your way out of an ideas problem with hardware.

The upstream work — reading the factor-investing literature, understanding why an edge might exist, testing honestly with out-of-sample and walk-forward validation and realistic costs — none of it needs new silicon. I can do all of it on hardware I already own, tonight. (The freqtrade bots in the lab run on a container with less compute than my phone.)

The relapse, and the machine that isn't even out yet

Decision made: money stays in the bank. And then — of course — barely a message later, I sent one more listing. A sealed Lenovo ThinkStation P3 Ultra SFF Gen 2: Ultra 9 285, 128GB DDR5, 2TB SSD, RTX A1000, Win 11 Pro, "chỉ 62 triệu — half US retail!" My deal-hunter brain does not stop just because I made a decision at 11pm.

And on absolute value, it was a strong price — that config retails around $4,100–4,700 in the US (~105–120 triệu), and it's loaded with 128GB of DDR5 that's worth a fortune at today's spiked prices. But it was the wrong tool, and in a way that was almost educational:

Its GPU, the RTX A1000, is an entry workstation card — 8GB, low compute, weaker than even the 5060 I was trying to swap out. Useless for gaming, weak for local AI, and in that 3.9L chassis you physically can't upgrade it to a real GPU.
128GB RAM is luxury overkill for anything I actually do.
Its genuine strength — a tiny, quiet, low-idle-power SFF box — is a better 24/7 homelab fit than the tower, sure, but my mini PC already does that job for a fraction of the price.

A great price on the wrong thing is still the wrong thing. Every strength that machine had served a workload I don't have, and its one weakness killed the only interest I do have.

And here's the genuinely funny ending. The one machine that would actually serve my real interest — running serious local AI models without slamming into an 8-to-16GB VRAM wall — is NVIDIA's RTX Spark, announced at Computex 2026 and shipping this fall. It's not a graphics card; it's a full ARM SoC: a 20-core Grace CPU, a Blackwell GPU (~RTX 5070-class, 6,144 CUDA cores), and up to 128GB of unified memory on a single 3nm chip, delivering about 1 petaflop of AI performance. The headline is that the huge unified-memory pool can run 120-billion-parameter models locally — exactly the VRAM wall every box I looked at kept hitting.

It has real caveats: the GPU is mid-tier, the memory is bandwidth-limited LPDDR5X (big models fit but don't run at datacenter speed), it's Windows-on-ARM (compatibility questions for my Linux stack), it's premium-priced (128GB configs will be expensive), and it's a first-gen product on a new architecture, so early buyers are beta testers.

But it's the first machine in this entire saga that would actually serve the itch — if the itch turns out to be real. And it doesn't ship until fall. So the smartest possible move fell right into my lap: I decided not to buy, and the genuinely interesting hardware for me isn't even available yet. Waiting costs nothing and might hand me the right tool in a few months, with real third-party benchmarks instead of launch-day hype.

What I actually learned

Re-check current prices before judging any deal. Your gut is running on last year's data. One stale line item flipped my entire verdict.
A flaw that doesn't apply to how you use things isn't your flaw. Dead-end socket, no upgrade path, platform longevity — all irrelevant if you buy whole machines and run them for years.
Good-deal-ness and do-I-need-it-ness are different questions. The discount is bait. A great price is only great if you were buying anyway.
A great deal on the wrong tool is still the wrong tool — and sometimes it's the right tool for a problem you don't actually have.
Find your real bottleneck before you spend. Mine was ideas, not compute, and no amount of silicon fixes that. Most "you need this PC" content is made by people whose income depends on the machine; you're watching someone's job and absorbing the feeling that you need their gear. You usually don't.
The projects don't need the hardware to start. If a hobby is real, prove it on what you own first. If you hit a genuine wall, then buy — probably cheaper, once this RAM madness cools down.

So I closed the laptop and kept a month's salary in the bank. Every hobby I fantasized about, I can start today on hardware already sitting in my house. If one of them actually grabs me and hits a real wall, the machine will still be there — and by then, maybe the right one (RTX Spark, or whatever follows) will have real reviews and saner prices.

Đóng Facebook, đi ngủ. The best build I made this week was the case for not buying anything at all.