The oldest argument in the AI hobbyist handbook — "just run it yourself, it's cheaper" — quietly stopped being obviously true this year. Here's the honest math, with real prices, as of October 2026.
The hardware climate
Start with what the AI boom did to the parts. According to Tom's Hardware (Aug 17), memory prices climbed roughly 500% in twelve months — up to ten times the lowest prices ever tracked — and 128GB of DDR5 now runs about $3,399. J.P. Morgan's research desk estimates DRAM will have risen more than 400% from the start of 2024 through the end of 2026.
GPUs aren't better. Per TechPowerUp (Sep 2, citing TechSpot data across ten countries), graphics card prices jumped 15% in a single month; the market sits 28% above launch MSRP on average — and that's with the flagship excluded as an outlier. The RTX 5090, which NVIDIA still lists at $1,999, costs nearly $5,000 on the US street — 145% over MSRP. The midrange is infected too: the RTX 5060 Ti 16GB trades about 70% above its $429 list price. The driver, fittingly, is us: GDDR7 supply is reportedly being redirected toward AI hardware.
So the box you'd build to run serious models locally has doubled or tripled in price before you open the terminal.
The cloud price floor
Now the other side — and it's genuinely competitive:
- Runway (verified on its pricing page this week): $12/month Standard gets you 625 credits; Gen-4.5 costs about 12 credits per second of video — an effective ~$0.23/s. The $28 Pro tier works out to ~$0.15/s.
- RunComfy's pay-as-you-go marketplace lists Wan 2.5 at $0.045/s, Kling 2.5 Turbo at $0.07/s, Sora 2 at $0.10/s — per second, no hardware.
- Cloud GPU time rents for roughly $3.59–4.49/hour on the same platform.
For text, API pricing is pennies-per-million-tokens territory. For video, the per-second numbers above. No capex, no DRAM lottery, no waiting for a price tracker to flash green.
The math (my analysis, so check it)
Take that inflated midrange card — call it $730 street for a 5060 Ti 16GB — and amortize it over two years of daily use (about four hours a day, 2,900 hours): hardware cost is roughly $0.25/hour, plus power at roughly half a kilowatt under load — maybe $0.13/hour at typical Australian rates. Call it $0.38/hour to run your own machine versus **$4/hour** to rent a comparable cloud GPU. Unit economics: local wins by an order of magnitude.
But three honest caveats:
- The entry ticket. Sixteen gigabytes of VRAM runs small models. The moment you need 128GB of system RAM for real local inference, you're paying $3,399 for memory alone — before the CPU, board, or storage. The payback period stops being a rounding error.
- The utilisation trap. The math only works at high utilisation. Run inference four hours a month and that $730 card costs you more per generated minute than any API on the market.
- The trend. GPU prices rose 15% last month. DRAM contract prices are still climbing, albeit more slowly. Nobody credibly predicts a drop before the supply situation changes — some market watchers are pushing "relief" out to 2028.
Verdict
- Bursty, experimental, or business workloads: cloud, comfortably. The 2026 price floor for APIs is low and you pay for what you use.
- Sustained, daily, privacy-critical, or offline workloads: local still wins — the unit economics haven't flipped, the entry price has. If you already own the hardware, keep running it; the local tier keeps getting more capable (Qwen 3.8 Flash Next — a 125B model — reportedly runs on a single RTX 4090 at 100 tokens per second).
- Thinking of buying hardware right now specifically for AI: that's the one position the data argues against. You'd be buying at the top of a spike caused by the very demand you're joining.
The usual advice was "local is cheaper if you can stomach the setup." In October 2026 the honest version is: local is still cheaper to run, but the buy-in got repriced by the industry it serves. Watch DRAM contract prices — the moment the affordability ceiling bites and supply catches up, this piece gets a sequel with happier numbers.
Prices verified 2026-10-05; sourced from Tom's Hardware, TechPowerUp/TechSpot, J.P. Morgan Research, and the lscppdev AI-video pricing watch.