Hardware Strategy for the 13-Month Timeline

You need GPUs. You cannot buy H100s. This section is the honest hardware playbook for a Zoho salary in 2026.

The three-tier stack you’re building

  1. Local iteration box — a GPU on your desk that you own. Zero latency, always available, no billing anxiety. Where you write kernels, debug at 2am, and run small models. This is your primary rig.

  2. Rented burst compute — cloud GPUs you rent by the hour. Where you run 70B evals, multi-GPU experiments, and reproduce benchmarks. Do not rent to develop. Rent to finish.

  3. Apple silicon workstation (already owned) — your MacBook. MLX + Metal is now a legitimate serving stack, worth ~15% of your total kernel practice budget.

Anti-pattern: trying to do everything on rented GPUs. You will burn budget and never build the muscle memory of a local rig. Own the small stuff, rent the big stuff.

Priorities in order

  1. VRAM > FLOPs. For inference and kernel work, 24GB unlocks ~70B-4bit and gives you a real KV cache to profile. FLOPs difference between a 3090 and a 4090 doesn’t change what you can learn.

  2. Compute capability > new-ness. A 3090 is Ampere (sm_86). A 4090 is Ada (sm_89). A 5090 is Blackwell (sm_120). You want at least Ampere so you have modern warp-level intrinsics, cp.async, and BF16 tensor cores.

  3. Ecosystem maturity > peak spec. CUDA on NVIDIA is the mature ecosystem. ROCm on AMD is improving but still bites. Metal on Apple is niche. Stay on NVIDIA for at least the first 12 months.

The four files in this section

  • 01_own_gpu_choice.md — Should you buy a 3090 vs 4090 vs 5090? Dual 3090 vs single 4090?

  • 02_rent_gpu_strategy.md — RunPod / Lambda / Vast / Modal / Prime Intellect — when to rent what, cost math.

  • 03_apple_silicon_lab.md — MacBook M-series as a legitimate MLX lab.

  • 04_budget_scenarios.md — $50/mo, $150/mo, $500/mo plans.

  • 05_2026_gpu_landscape.md — H100/H200/B200/GB200/MI300X positioning (for when you’re evaluating rentals or specifying deployments).

The one-line decision

If you have $2000 and are starting today: Buy a used RTX 3090 (~$700), spend $50-100/mo on RunPod for burst H100 work, cap yourself at $150/mo total cloud burn until you have a portfolio piece that justifies more.

Everything else in these files is refinement. This is the answer.

What NOT to buy in 2026

  • RTX 3080 / 3080 Ti — 10-12GB VRAM. Not enough for 70B-4bit or meaningful KV cache experiments.

  • RTX 4070 / 4070 Ti — same VRAM problem.

  • Any Turing card (RTX 2080, T4) — pre-Ampere. Missing BF16 tensor cores, cp.async, modern async pipelines.

  • AMD Radeon consumer cards — ROCm on consumer Radeon is a QA disaster. Instinct MI-series is production ROCm; Radeon is not.

  • Mac Mini M-series as a training rig — the unified memory is a serving toy, not a kernel-development platform.

  • Data center cards on eBay (A100, V100) — cooling, power, and driver issues on desktop. Not worth the pain unless you’re building a homelab rack.

Cross-references

  • Budget breakdown: 04_budget_scenarios.md

  • Rental pricing table (which the Chinese room’s actually-2026 refresh): 02_rent_gpu_strategy.md

  • Portfolio requirements per rung: 12_portfolio/ — every rung says what hardware you need.

Verified. Actionable. Realistic for a Zoho salary in Chennai/Bangalore + INR-to-USD reality.