Where AI Value Lands

RL APIs — the commercial fine-tuning surface

3 min read

RL APIs — the commercial fine-tuning surface

One-line

The hosted services that let a non-lab actor run RL fine-tuning without building the infrastructure.

The idea

Most boutique vendors and startups do not run their own RL infrastructure. They rent it from one of a small number of API providers. The choice of API determines, in order of strategic weight:

  • Which base models are available — frontier closed-weight vs open-weight families.
  • Which training modes are supported — SFT, DPO, full RLHF, verifier-based RL.
  • How data exclusivity is contracted — does the provider train on the customer’s data? does the customer get an exclusive endpoint?
  • How much of the weights the customer ever owns — closed-weight tuning produces a tenant-scoped artifact the customer cannot extract; open-weight tuning can produce weights the customer keeps.

The landscape (sketch, mid-2026 — indicative; specific products evolve fast)

Frontier-lab tuning surfaces (closed weights):

  • OpenAI — fine-tuning across the GPT family, plus Reinforcement Fine-Tuning (RFT) with customer-supplied verifiers. Weights stay inside OpenAI.
  • Anthropic — Claude custom-model and fine-tuning programs, historically in limited preview. Closed weights.
  • Google Vertex AI — SFT and RLHF-style tuning on Gemini-family models. Closed weights.

Open-weight tuning surfaces (customer-controllable weights):

  • Together AI — fine-tuning + dedicated endpoints on open-weight families.
  • Fireworks AI — fine-tuning + low-latency serving for open-weight families.
  • Predibase — LoRA / adapter-based fine-tuning at scale.
  • Modal / Replicate / RunPod — raw compute for self-managed pipelines built on TRL, OpenRLHF, verl, etc.

Verticalised / niche — Lamini, Anyscale (Ray Serve + Ray RLlib heritage), and similar players targeting specific verticals or workloads.

Where this fits in Where AI Value Lands

In H4_rl-specialization-value-pocket, the API choice is the most concrete unit-economics decision a vendor makes:

  • RFT on a closed frontier base — best base capability, weakest IP capture. No continual-learning-paradigm compounding inside a single customer because the weights never leave the provider.
  • Open-weight + Together / Fireworks — full IP control, worse base capability per dollar at t=0. Compounding is possible because the vendor owns and re-trains the weights over the engagement.
  • Self-managed on Modal / RunPod — maximum control, maximum operational burden. Only justified at scale.

The vendor’s strategic posture in taker-shaper-maker (Taker rents, Shaper customizes, Maker trains) is largely encoded in this single choice. See ai-factory-huang for the underlying compute substrate all of these surfaces rent from.

Referenced by

Auto-generated reverse-link index — pages in the wiki that link here. Maintained by /lint and reverse-link sweeps.

Graph View

Backlinks