Skip to content

RESEARCH.md

Latest commit

 

History

History
155 lines (113 loc) · 7.04 KB

RESEARCH.md

File metadata and controls

155 lines (113 loc) · 7.04 KB
Raw

DockVision — Competitive Research

Synthesized 2026-05-12. Sources at bottom.

Closest competitors

Neurosnap (neurosnap.ai)

Pricing (annual, EUR):

Plan Price/yr Credits Parallel jobs
Free €0 1 credit 2
Academic Budget €93.50 84 7
Academic Standard €185.90 168 14
Academic €280.50 300 30
Professional €880 720 60
Enterprise A/B/C €5,500-10,956 1,200-4,000 5-10 seats

Implied credit cost ≈ €1.10-1.22 per credit at academic tier. Credits do not expire. Each job shows "Estimated Credits" before launch.

Tool catalog: Boltz-1, Chai-1, Protenix, AlphaFold2, RoseTTAFold2, DiffDock-L, GNINA, DynamicBind, FlowDock, ProteinMPNN, RFdiffusion2, GROMACS.

Strengths: broad tool catalog, no-card free tier for academics, built-in 3D viewer, persistent credits.

Weaknesses: opaque per-run cost until login; Euro pricing confuses US users; "unlimited storage" is soft-quota'd in practice.

Tamarind Bio (tamarind.bio)

Pricing hidden behind "Contact us". YC-backed, $13.6M Series A. Tools: AlphaFold, Chai-1, RFdiffusion, MPNN, GROMACS, Smina, EquiDock, GNINA, RFantibody.

Sentiment: free tier perceived as demo-only; reviewers cite "tens of thousands/year" for real use. Strong on antibody/binder tooling. Generally reads as enterprise-focused.

Fovus

Runs Boltz-1 at $0.10/structure (21 min, includes complex cases). Public price floor for Boltz-class workloads. Not a full bio-SaaS — general compute that happens to host Boltz.

Others reviewed

  • NVIDIA BioNeMo — NIM microservices, opaque enterprise pricing.
  • Lightning AI — general-purpose; ~$49/mo + $5-15/hr credits.
  • Chai Discovery (hosted) — own free playground with rate limits. No marketplace.
  • Helixon / OpenADMET / Inductive Bio — adjacent, not direct competitors.

Pricing benchmarks for our tools

Inferred from Neurosnap credit cost and observed runtimes:

Tool Typical runtime (H100) Suggested user price Notes
Boltz-2 2-8 min $0.20 - $0.80 Fovus floor = $0.10
Protenix 3-10 min $0.40 - $1.20
Chai-1 2-6 min $0.20 - $0.70 Chai's own playground is free, rate-limited
DynamicBind 5-20 min $0.80 - $2.50 License blocker for commercial
GNINA 1-5 min (4090 OK) $0.05 - $0.30

At our 1.5× cost-plus markup on RunPod Flex:

  • GNINA 3-min on RTX 4090: $1.10/hr × 0.05 hr × 1.5 = $0.083
  • Boltz-2 5-min on H100: $4.18/hr × 0.083 hr × 1.5 = $0.52
  • Protenix 6-min on H100: $4.18/hr × 0.1 hr × 1.5 = $0.63

Competitive with Fovus on Boltz; undercuts Neurosnap on transparency.

RunPod product taxonomy (May 2026)

Three distinct products:

  1. Serverless Endpoints (mature) — Docker + handler, billed per-second. What we use.
  2. Hub — open-source model marketplace; beta. Not for custom docking.
  3. Flash (GA Apr 30 2026) — Pure-Python @Endpoint decorator, no Docker. Limit: can't install conda-heavy stacks (OpenMM, RDKit). Wrong fit for our tools.

Flex pricing (May 2026):

GPU Flex $/sec Active $/sec $/hr Flex
B200 (180 GB) $0.00240 $0.00190 $8.64
H200 (141 GB) $0.00155 $0.00124 $5.58
H100 (80 GB) $0.00116 $0.00093 $4.18
A100 (80 GB) $0.00076 $0.00060 $2.72
L40S (48 GB) $0.00053 $0.00037 $1.90
RTX 4090 (24 GB) $0.00031 $0.00021 $1.10

Network volume: $0.07/GB-mo (<1TB), $0.05 (>1TB), $0.14 (high-perf). No egress.

Cold starts: marketing claims sub-200ms for 48% of starts. Reality with 30+GB images + 5-50GB weights: 30-60s typical, 1-2min worst case.

SDK / cost capture: runpod.Endpoint(id).run(...) returns {id, status, delayTime, executionTime, output}. Billed = executionTime on Flex (queue time free). Multiply by GPU rate for exact cost.

Limits: /runsync max payload 20 MB. Async output retained 30 min. Default max job time 600s, configurable to 7 days. Default 3 max workers per endpoint (request increase).

Output > 20 MB: configure S3 upload — response becomes {"type":"s3_url","data":"..."}. We use our own S3 bucket (dockvision-prod); RunPod doesn't provide artifact storage.

Alternative compute (briefly)

  • Modal — best DX, Python-native, ~$3.95/hr H100, 2-4s cold starts. Pricier, can't BYO datacenter. Real contender for v2.
  • Replicate — $5.49/hr H100, great for public demos, weak for custom pipelines.
  • Beam — ~$3.50/hr H100, Modal-like, smaller.
  • Lambda Cloud — VM-style only, no scale-to-zero. Skip.

RunPod wins on per-second price and exact billed-seconds in response. Modal is a plausible v2 alternative if cold-start UX becomes a complaint.

CLI-style UI prior art

Nobody in bio-SaaS does this. All competitors use forms/wizards. Closest precedent: Warp, Modal CLI, fly.io, replit shell. CASP audience is heavily Linux/bash-fluent — a CLI-style web UI is a real differentiator.

Risk: alienates wet-lab scientists. Mitigation: wizard <tool> drops into form mode.

Auth/payment stack consensus (2026)

  • Stripe Billing with metered events. 2026 AI-usage SKUs are the right primitive.
  • Auth.js (formerly NextAuth) for SvelteKit — self-hosted, OSS, fits AGPL ethos. Clerk is the SaaS alternative if Auth.js becomes painful.
  • Postgres for users/jobs/billing metadata. Trigger-maintained storage_bytes denorm.
  • AWS S3 for artifacts (co-located with www0 in us-east-1; free intra-region transfer; $0.09/GB egress to internet).

Sources

  • Neurosnap pricing: graphstats.net/neurosnap-pricing; neurosnap.ai/services
  • Tamarind pricing: tamarind.bio/pricing
  • Fovus Boltz price: fovus.co/2025/03/03/achieving-cost-efficient-boltz-1-simulations
  • RunPod docs: docs.runpod.io/serverless/pricing; runpod.io/pricing
  • RunPod Flash launch: runpod.io/blog/flash-is-ga
  • FlashBoot: runpod.io/blog/introducing-flashboot-serverless-cold-start
  • Serverless GPU comparison: introl.com/blog/serverless-gpu-platforms-runpod-modal-beam-comparison-guide-2025
  • BioStars Neurosnap vs Tamarind: biostars.org/p/9597524/
  • Protein folding SaaS pricing: getmonetizely.com/articles/what-is-protein-folding-saas-and-how-to-navigate-biological-structure-prediction-pricing
  • Boolean Biotech (AF3, Boltz, Chai-1 overview): blog.booleanbiotech.com/alphafold3-boltz-chai1