1. What happened
Bolt de-identifies opted-in Forge sessions inside its infrastructure and licenses resulting datasets to AI developers under a data license agreement. Arcee AI is the first named recipient; Bolt says the datasets may also be licensed to other AI developers.
2. Buyer / seller
Arcee AI is the AI buyer or model-development side. Bolt / StackBlitz owns or contributes the proprietary corpus.
3. Economics
Individual Pro users receive up to 50× additional Forge usage during the preview; Bolt describes the allocation as payment for sharing. Bolt may also receive cash for licensed datasets, but no amount is disclosed. Economics scope: In-kind usage allocation to contributing users; any Arcee-to-Bolt license payment is undisclosed.. Strategic Signal has not estimated undisclosed consideration.
4. Proprietary data
Opted-in Bolt Forge sessions containing developer prompts, code, project files and configuration, tool calls, edit histories and fix traces from individual Pro users during the September 14–October 14, 2026 preview window. Historical depth: Newly collected preview corpus beginning September 14, 2026; no long operating history is disclosed.. Refreshability: Renewable while users opt into Forge; the first disclosed Arcee training window is September 14–October 14, 2026..
5. Rights and restrictions
One-tap opt-in is required before collection; switching away from Forge stops new collection. Ownership: No ownership transfer is disclosed; the announced structure is a data license agreement.. Training: Explicit: the licensed sessions feed Arcee AI's first training run for a trillion-parameter-class model.. Post-training: not disclosed.
6. Why the data matters for AI
Training open-weight coding and software-building models to complete real application-development workflows in fewer attempts. The corpus records software-building sequences including attempts, edits, tool calls, errors and fixes, creating observable short-cycle action and outcome traces.
7. Asset Score — 72/100
Strategic Signal analysis. The workflow traces are unusually useful for software-agent learning, but the corpus is new, excludes enterprise workspaces and has no independent uplift evidence. Score coverage: 100%.
| Factor | Weight | Rating / 5 | Points |
|---|---|---|---|
| Uniqueness / scarcity | 12 | 4 | 9.6 |
| Historical depth | 8 | 0.5 | 0.8 |
| Decision → outcome richness | 13 | 4 | 10.4 |
| Decision Value Density | 13 | 2.5 | 6.5 |
| Real-world grounding | 10 | 4 | 8.0 |
| Domain economic value | 9 | 4 | 7.2 |
| Refreshability | 8 | 4.5 | 7.2 |
| Proprietary advantage | 9 | 4 | 7.2 |
| Model-learning usefulness | 10 | 4.5 | 9.0 |
| Non-replicability | 4 | 3.5 | 2.8 |
| Rights usability | 4 | 3.5 | 2.8 |
8. Transaction Signal Score — 66/100
Strategic Signal analysis. The arrangement cleanly proves training use and non-cash contributor consideration, but provides weak cash price discovery and no measured uplift. Score coverage: 85%.
| Factor | Weight | Rating / 5 | Points |
|---|---|---|---|
| Disclosed economics | 16 | 3 | 9.6 |
| Clean price discovery | 16 | 1 | 3.2 |
| Separability of data consideration | 15 | 3.5 | 10.5 |
| Explicit training / model-improvement use | 14 | 5 | 14.0 |
| Clarity of rights purchased | 12 | 4 | 9.6 |
| Competing bids | 10 | Unknown | — |
| Strategic buyer quality | 6 | 3.5 | 4.2 |
| Repeatability / precedent value | 6 | 4.5 | 5.4 |
| Independent model-uplift evidence | 5 | Unknown | — |
9. M&A / valuation implications
Shows that a software platform can monetize consented workflow telemetry separately from subscriptions and can compensate contributors with compute or usage rather than cash.
10. Evidence boundary
The data-license structure, corpus fields, consent controls, initial Arcee recipient and model-training purpose are disclosed by Bolt. Cash economics, raw-data retention, derived-weight ownership, Arcee sublicensing and measured model uplift are not disclosed.