[P0] DePIN cost benchmark — Akash vs io.net vs Nosana on 3 workloads

[abhicris]

The "50% cheaper than Replicate/fal.ai" claim is unbacked. Decides whether Layer 1 compute marketplace is a viable revenue stream at all.

Scope

Run these three workloads on Akash, io.net, and Nosana. Compare wall-clock time and $-cost:

Workload	Size
SDXL image generation	10 images at 1024x1024, 30 steps
Llama 3.1 70B inference	1000-token completion, batch of 10
LoRA fine-tune	200-step LoRA on a 500-image dataset, SDXL base

Comparison baselines

• Replicate on-demand pricing for each workload.
• fal.ai on-demand pricing for each workload.
• Bare-metal reference (Lambda / RunPod) if relevant.

Output

cr8-depin-benchmark.md:

• Full pricing table: provider × workload × total cost × wall time × reliability note.
• Verdict per workload: "DePIN wins by X%" / "DePIN loses by X%" / "DePIN on par".
• Overall verdict: go/no-go on the compute-marketplace narrative.

Why

• gtm-review-2026-04-13.md §Use cases: "Without this, the '50% cheaper' claim is hand-waved."
• Feeds the compute-marketplace scaffold (abhicris/CR8#16) with realistic price copy.
• Feeds grant applications (Arbitrum, EF) with defensible numbers.

Label

documentation — this is a research output, not code.

[abhicris]

DePIN cost benchmark draft landed on branch research/depin-benchmark — file: research/depin-cost-benchmark-2026-04.md.

Summary:

• Methodology, workloads (SDXL / Llama-3.1-70B / 13B LoRA), pricing snapshot 2026-04-18.
• GPU $/hr table across Akash / io.net / Nosana vs Replicate / Modal / RunPod / Lambda / Together AI.
• Cost-per-unit arithmetic ($/image batch 1 & 8, $/1K tokens, $/fine-tune run).
• Reliability + cold-start column — the honest tax that naked $/hr tables hide.
• Go/no-go at 2.5% take: conditional GO contingent on CR8 shipping an orchestration layer (auto-restart, checkpoint replication, provider-health routing) and positioning as batch & training, not real-time inference.
• §9: every cell marked (unverified) or (estimated) and paired with a verification plan (≈2-3 operator days, ~$150-300 real compute).

Headline verdicts:

• W1 — SDXL batch: DePIN wins ~15-70× on naked cost. Best wedge for CR8.
• W2 — Llama-70B inference: DePIN beats Replicate/Modal but loses ~2-3× to Together AI / Groq. Do not pitch against shared-tenant managed inference.
• W3 — LoRA fine-tune: DePIN wins 2-10× on cost; blocked on pipeline tooling that CR8 must ship.

Status: DRAFT — not citable publicly until a human runs the verification jobs in §9.

PR blocker: repo is still single-branch (research/depin-benchmark is both the head and the default); opening a PR requires a base branch to exist first. Options: (a) promote this branch to main directly (since nothing has merged yet), (b) seed an empty main and retroactively PR the research branch in. Flagging for a human call — either is fine.