Budget Scenarios: $50, $150, $500 per month¶
Three honest playbooks for the 13-month timeline. Pick one and stick to it.
Scenario A: $50/month (₹4,200/mo) — the minimum viable path¶
One-time capex: $700 for a used RTX 3090.
Monthly opex: $50 (₹4,200)
$30 RunPod H100 burst (~15 hrs at $500-1200/hr) — for FA3/WGMMA/FP8 experiments.
$10 Modal free-tier overrun (rarely) — for demo hosting of your Rung 5 mini-engine.
$10 HuggingFace Pro (optional) — for private models + Spaces GPU minutes.
What you can do:
Everything in Phases 0-4 fully locally on the 3090.
Rent H100 for Phase 3 FA3 experiments (once, ~2 hrs = $4).
Rent H100 for Phase 4 vLLM/SGLang benchmarks vs your mini-engine (~4x during Phase 4 = $20).
Rent H100 for Phase 5 FP8 quant experiments (~2 hrs = $4).
Phase 6 distributed: rent 2x H100 or 2x A100 once for TP experiments (~2 hrs = $6).
What you cannot do:
Sustained multi-GPU training runs.
Playing with 405B or larger models.
MI300X exploration.
8x H100 cluster experiments (a single hour = $16).
This is the correct path for 90% of readers. You will not learn faster by spending more. You will learn faster by writing more code on your 3090.
Scenario B: $150/month (₹12,500/mo) — the comfortable path¶
One-time capex: ~$1600 (used 3090 + upgrade to build parts) OR ~$1400 dual-3090 rig.
Monthly opex: $150 (₹12,500)
$80 RunPod (~40 hrs H100) — more frequent benchmark runs.
$30 Modal / Vast (~15 hrs, on-demand + serverless demo hosting).
$20 HuggingFace Pro (+ optional Enterprise tier for Spaces GPU minutes).
$10 Cursor / GitHub Copilot / whatever paid AI tools.
$10 domain + Cloudflare + Substack Pro / Ghost hosting for your blog.
What you can do (in addition to A):
Dual-3090 with NVLink; run 70B-4bit locally without offload.
Weekly 2-hour H100 sessions for continuous validation.
8x H100 pod for a single multi-day experiment during Phase 6 or Phase 7 (~$200 for a 12-hr session).
Publish 1-2 substantial blog posts per phase (hosting + graphics not free).
Occasional MI300X hour to write “same kernel on AMD” writeup ($4).
Sustainability: at Zoho L3/L4 salary in India this is roughly 3-5% of take-home. Realistic. If you’re above L5, it’s noise.
Best-case outcome: you finish the 13 months with a portfolio of blog posts + 2-3 merged PRs + a mini-engine + a reference architecture doc. Job offer material.
Scenario C: $500/month (₹42,000/mo) — the accelerated path¶
One-time capex: ~$3000 (dual 3090 with matching NVLink build + peripherals + monitor + ergonomics).
Monthly opex: $500 (₹42,000)
$300 sustained multi-GPU rental (Lambda 1CC or Prime Intellect reserved) — 8x H100 access on demand.
$80 RunPod ad-hoc.
$50 Modal for hosting demos + always-on inference endpoints.
$30 HuggingFace Pro + Enterprise.
$40 tools (Cursor Ultra, Claude Max, GitHub Enterprise if you split billing).
Only justify this if:
You are single OR have partner buy-in and disposable savings.
You are in month 9+ of the roadmap and have proof-of-work that justifies acceleration.
You are treating this as a targeted 6-month career sprint, not a lifestyle.
What you can do (in addition to B):
Sustained 405B-scale experiments across a rented 8x H100 pod.
Full multi-node distributed inference test-beds (~$40/hr for 2 nodes of 8x H100).
Contribute meaningful benchmarks to vLLM/SGLang leaderboards.
Own a compute footprint big enough to support 1-2 side projects on top of the roadmap.
Warning: this is where most people burn out. You will feel pressure to “use” the compute you’re paying for. If the compute is not producing measurable output within 2 weeks, cut back to Scenario B.
The wrong choices (don’t do these)¶
The trap: “$0/mo, only free tiers”¶
You can technically do Phase 0-1 on Colab and Kaggle. You cannot build muscle memory jumping between forced-idle sessions. You will spend 40% of your time restarting notebooks. Spend $50/mo. It is not optional.
The trap: “buy an H100 personally”¶
H100 SXM = $30,000+. Even H100 PCIe = $22,000+. You cannot cool it, you cannot power it (750W), you cannot use it in a US/EU home without a dedicated 240V circuit. Nobody does this. Rent by the hour.
The trap: “buy a 2x RTX 6000 Ada workstation for $10K”¶
Don’t. If you have that kind of money to spend, put $2K into a dual-3090 build and put the other $8K into an emergency fund. You’ll get 80% of the compute for 20% of the cost.
The trap: “AWS/GCP/Azure GPU instances”¶
p4d.24xlarge (8x A100) on AWS on-demand = ~$32/hr. Same instance on RunPod = ~$12/hr. AWS pricing is designed for enterprise customers with committed spend and audit compliance. For personal work, it’s 2-3x overpay. Only use AWS if your employer is paying and you need their networking/security profile.
Cost tracking discipline¶
Set up a spreadsheet on day 1. Columns:
Date
Vendor (RunPod / Modal / etc.)
GPU type + count
Hours
Total $
What I was doing (1 sentence)
Was it worth it? (Y/N)
At the end of every month, review. If “Was it worth it?” is N > 3 times, you’re using cloud as an IDE.
Effective cost of one job offer¶
Do the math: 13 months × $150 = $1950. Plus $1400 rig. Total: ~$3350.
One inference-engineer offer at any US/EU AI company (or a good Indian AI startup like Sarvam / Krutrim / Yellow) is +30-50% comp over Zoho L4-L5 baseline. In year-1 salary delta, this pays for itself ~10-30x. This is the most cost-effective career move you can make in 2026.
But only if you actually do the work. The GPUs don’t do it for you.
Cross-references¶
Own GPU:
01_own_gpu_choice.mdRental strategy:
02_rent_gpu_strategy.mdMotivation to keep going:
13_discipline/07_motivation_sustainment.mdstudy conversion (the ROI justification):
13_discipline/09_interview_conversion.md
One line: aim for Scenario B ($150/mo) as your default. Downshift to A if life gets in the way. Upshift to C only after Month 9 with real portfolio to show.