China's export rule hands the open-weight enterprise market to US labs

Both product ideas already have competitors, so neither space is empty. Managed RL on owned weights is the more contested of the two. Adaptive ML already sells RL post-training delivered in-VPC or air-gapped on-prem with forward-deployed engineers. Applied Compute is pursuing bespoke enterprise RL. River AI has a reported $1.1B to sell "train and own your model". Prime Intellect, Together, Fireworks and free open-source RL stacks all compete for the same budget. There is also a credible argument that many enterprises get more from RAG and workflow changes than from RL, which narrows the paying market. A new entrant could still win in the very top tier: frontier-scale models (hundreds of billions to trillion-parameter MoE) trained on the customer's own clusters, including classified and defence-accredited environments such as FedRAMP High, IL5/IL6 and ITAR. Most current players work at smaller scale or on their own cloud, and none clearly sells "full RL plus serving plus governance, delivered on hardware you own, weights handed back" as one package with outcome-based pricing. The provenance ledger faces a different risk: big security and supply-chain vendors are absorbing it as a feature. Cisco already fingerprints lineage and publishes country-of-origin and licence data for free, while Mend, Black Duck and PointGuard put model licences into AI bills of materials. The opening is a procurement-grade product that meets federal acquisition rules. It would provide signed, auditable attestations of lineage, including derivatives and distillations of Chinese base models, mapped to specific rules such as a DeepSeek contractor ban if enacted (only the bill was found, not whether it passed), the relevant NDAA provisions and agency rules. It would also track every in-house fine-tune as its own lineage node. Today that would most likely sell as a wedge or add-on to the RL and serving platform rather than as a standalone company, because the scanning layer is turning into a commodity.

• Adaptive ML (Adaptive Engine)

  This is the closest match to the managed RL idea. Its AWS Marketplace listing says it runs RL post-training (GRPO, PPO, DPO and others) on open models inside the customer's VPC or fully air-gapped on-premises, with forward-deployed engineers who build training recipes and data, tune and evaluate models, and run production deployments. The customer keeps the weights. These claims come from the vendor and secondary listings.

• River AI

  Igor Babuschkin's startup. Coverage reports a $1.1B first round, led by General Catalyst with NVIDIA and AMD Ventures taking part. It sells managed LoRA fine-tuning and RL on open-weight models, pitched as letting companies train and own their models. Billing is metered on tokens. Hardware is planned but has not shipped, and the service runs on River's infrastructure rather than on customer hardware.

• Applied Compute

  Its thesis is bespoke RL for enterprises, building proprietary models. One product directory lists VPC deployment, SOC 2 and ISO 42001, and full data ownership. I found no source confirming on-prem deployment.

• Prime Intellect

  Full-stack open training platform with the prime-rl framework, the Verifiers library and an Environments Hub listing more than 2,500 RL environments. It is pitched as reducing dependence on closed APIs and was used by Ramp to train a 35B agent. It is mostly hosted compute plus open-source tooling, not on-prem managed service.

• Deep Cogito

  Raised $43M for post-training research on open models, which puts it next to the managed RL idea.

• CoreWeave reinforcement learning solution

  Managed GPU infrastructure for RL training, serverless or provisioned. It covers the infrastructure layer only and does not run RL for the customer or hand back weights.

• Together AI / Fireworks AI fine-tuning

  Together offers managed fine-tuning with downloadable LoRA adapters or full weights, and Fireworks offers reinforcement fine-tuning. Both are cloud-hosted and focused on inference rather than the customer's own hardware.

• Open-source RL stacks (Miles, SkyRL, prime-rl, OpenRLHF)

  Free frameworks an enterprise ML team can run on its own hardware for RL post-training. They are a do-it-yourself substitute for a managed service.

• Cisco AI Defense model provenance kit and AI Supply Chain Provenance Explorer

  This overlaps directly with the provenance ledger idea. It fingerprints weights to find hidden lineage (for example, Chinese base models under American fine-tunes) and offers a free explorer that reports licence restrictions, country of origin and security notes for many models. Cisco argues for a model bill of materials in procurement.

• Mend.io AI governance / AI-BOM

  Automatically finds the models in use, including shadow AI, and keeps them in an AI bill of materials. It flags licence terms for Hugging Face models and can block or escalate based on licence risk. It uses the Cursor and Kimi lineage episode as a marketing case.

• Black Duck SCA AI Model Risk Scanning

  Its software composition analysis tool has identified and analysed open-source AI models (use, licensing, data origins) since release 2025.10.

• PointGuard AI

  Offers model discovery plus provenance and licensing review for open-source models as part of an AI security platform.

Checked 2026-10-10