OYASS × GPU Pathway

Rent → Own · Budget → Butter

Your pathway from pay-as-you-go to all-inclusive power. Every tier shows the hardware, the monthly cost, what models it runs, and where it fits in the OYASS model.

R
Rent
Pay as you go
H
Hybrid
Mix & match
O
Own
Buy the box
B
Butter
All-inclusive

☁️Rent

Pay-as-you-go · No commitment
$0.09–$2.27
Per hour
$65–$1,635
Per month (24/7)
0
Upfront cost
∞
Model size

What you get

  • Access to any GPU — RTX 3060 to B200
  • No upfront, no commitment, cancel anytime
  • Best for variable workloads & experimentation
  • Try before you buy — test models before committing to hardware

Tradeoffs

  • Highest long-term cost — renting 24/7 for a year costs more than buying
  • No equity — you own nothing at the end
  • Provider dependency — prices change, GPUs get delisted
  • Data leaves your premises
☁️

OYASS Fit: Onboarding & Trial

OYASS customers start here. We build the worker, you run it on rented GPUs to validate the workflow. Zero hardware commitment — just the $3k–$10k build fee + pass-through compute. Cancel anytime.

⚖️Hybrid

Mix rental + owned hardware
$700
Upfront (3090)
$30–$100
Monthly rental
$0.09
Per hr (owned)
27B Q4
Local models

What you get

  • Own a used RTX 3090 ($700) for daily 27B-class inference
  • Rent H100s on Vast.ai for 70B+ models when needed
  • Best of both worlds — low base cost + burst capacity
  • Your 3090 pays for itself in 2–4 months vs. renting

Tradeoffs

  • Two cost centers to manage (owned + rental)
  • 3090 is 350W — not silent, not cool
  • 24GB VRAM wall — anything >32B at Q4 needs rental
  • You manage the local box (drivers, updates, cooling)
⚖️

OYASS Fit: Active Production

Your OYASS worker runs daily on your owned 3090 — $0.09/hr effective cost. When you need to run a larger model or batch process, burst to rented H100s. The OYASS upkeep ($500–$1.5k/mo) covers tuning both paths.

🏗️Own

Buy the hardware · Lowest long-term cost
$700–$4,200
Upfront
$0.02–$0.09
Per hr (amortized)
$12–$50
Electricity / mo
70B Q3
Max local model

What you get

  • Own the hardware — it's yours, it doesn't disappear
  • Lowest per-hour cost after break-even (2–4 months)
  • Full control — no provider dependency, no API changes
  • Data stays on your hardware, full stop
  • Choose your build: 3090 ($700), V100 32GB ($700), or Mac Studio ($4,200+)

Tradeoffs

  • Upfront cost — $700–$4,200+ depending on build
  • You manage it — drivers, updates, cooling, power
  • VRAM is fixed — can't burst to a bigger GPU
  • Hardware depreciates — resale value drops over time
🏗️

OYASS Fit: Full Ownership

Your OYASS worker runs on hardware you own. No monthly compute bill — just the one-time build ($3k–$10k) + optional upkeep. The 🫏 runs where you run it. If OYASS vanished tomorrow, your worker keeps working on your hardware.

🧈Butter

All-inclusive power mode
$7,500+
Upfront
$0.01–$0.04
Per hr (amortized)
192GB
Unified memory
Anything
Model size

What you get

  • Mac Studio M3 Ultra 192GB — runs any model that fits in 192GB unified memory
  • 30W TDP — near-silent, runs 24/7 in your office
  • No GPU lottery — unified memory means no VRAM fragmentation
  • 70B Q4, 120B Q3, even 200B+ at low quant — all fit
  • Or: full rack with multi-GPU V100/5090 for maximum throughput

Tradeoffs

  • Highest upfront cost — $7,500+ for the max config
  • Mac Studio can't use CUDA — limited to MLX / Metal / llama.cpp
  • No upgrade path — what you buy is what you have for 5+ years
  • Overkill if you only run 8B–27B models
🧈

OYASS Fit: The Ultimate 🫏

Your OYASS worker runs on the most capable consumer hardware available. No compromises — run any model, any time, silently, in your own space. The $7,500 upfront replaces $200+/mo in API subscriptions. After 3 years, you've saved $7,200 and own the hardware.

Budget → Butter · Pick your tier

Budget ~$300
Sweet Spot ~$700
Serious ~$1,500
Premium ~$4,200
Butter $7,500+
Budget
~$300
Tesla P40 24GB + DIY cooling
Runs 8B–13B models · ~15 tok/sec
Entry
Sweet Spot
~$700
Used RTX 3090 24GB
Runs 27B Q4, 70B Q3 · 30–50 tok/sec
Default
Serious
~$1,500
New build + used 3090
Same as sweet spot + reliability
Build
Premium
~$4,200
Mac Studio M3 Ultra 96GB
Runs 70B Q4 · near-silent · 30W
Silent
🧈 Butter
$7,500+
Mac Studio 192GB / multi-GPU
Runs anything that fits in VRAM
All-in

12-Month Cost Comparison

Rent
$2,628
Hybrid
$1,260
Own
$844
Butter
$1,524
Rent (24/7 on Salad 3090)
Hybrid (own 3090 + occasional rental)
Own (3090 + electricity)
Butter (Mac Studio amortized)

Break-Even Analysis · When does owning beat renting?

RTX 3090 ($700) vs. renting
2–4 months
At $0.20–$0.30/hr rental rate
Mac Studio ($4,200) vs. renting
8–14 months
At $0.50–$0.90/hr H100 rental
Butter ($7,500) vs. API subs
~3 years
Replaces $200+/mo in API costs
OYASS build ($5k avg)
Day one
You own the worker forever