Physics / mechanism
A data flywheel is the claim that deployed AI systems generate proprietary interaction data (user corrections, agent trajectories, evaluation outcomes, sensor logs) which is fed back into post-training to improve the model, which in turn improves the product and generates more data. The mechanism is not a physical one but an economic one: it converts an operational activity (serving inference) into an accumulating asset that cannot be bought or downloaded. Its strength depends on whether the feedback signal is dense enough and specific enough that a competitor starting from the same open weights cannot replicate the result.
The key contested parameter is where in the stack the flywheel accrues. The same fine-tune plus reinforcement-learning plus serve loop can be operated by the serving vendor, the application owner, or the model lab, and the party that holds the loop is the party that captures the margin. Evidence that specialisation demand is real: Fireworks reported that 95% of tokens it serves come from customer-specialised models, meaning fine-tuned open weights, adapters and distillations rather than off-the-shelf third-party checkpoints ref. Evidence that the loop is not owned by the serving layer: Stripe cut inference costs by 73 per cent by serving open models on vLLM, handling 50 million daily API calls on one-third of its previous GPU fleet, suggesting value can migrate to the orchestration layer rather than to a differentiated serving vendor ref.
The observable signature of a working flywheel is gross margin. Products that wrap a frontier model with minimal added value run 50 to 60 per cent gross margin, while those with proprietary models, fine-tuning or a real data moat clear 70 per cent or more ref. Cursor reached slight gross-margin profitability in April 2026, attributed to its proprietary Composer model and cheaper model routing, with net dollar retention reported above 90 per cent and ARR reported at roughly $2bn in February 2026 rising to roughly $4bn in May 2026 ref.
In robotics the same argument is made about physical interaction data: open-source is expected to commoditise model architecture, while data and deployment layers remain proprietary and defensible, with hardware cost compression shifting value away from OEMs ref.
Competitive landscape
| Claimed flywheel owner | Supporting evidence | Counter-evidence |
|---|---|---|
| Serving/inference vendor | 95 per cent of Fireworks tokens from customer-specialised models ref | Stripe self-served open models on vLLM at 73 per cent lower cost ref |
| Application owner | Cursor margin inflection via proprietary Composer model ref | Wrapper products stuck at 50-60 per cent gross margin ref |
| Robotics data/eval layer | Data and deployment layers remain defensible as architecture commoditises ref | Ubtech humanoid gross margin 54.6 per cent in 2025 with management guiding 40-43 per cent for 2026, without an established data flywueel argument ref |
The adjacent claim is that value accrues instead to the component layer. LiDAR and sensing companies rallied in 2026 (Ouster +28.28 per cent, Aeva +22.89 per cent year to date), consistent with a photonics component moat rather than a data moat ref. A separate pressure on flywheel arguments is open-weight capability convergence: GLM-5.2 (744bn total parameters, 40bn active, June 2026) was reported to match or exceed proprietary flagships on long-horizon coding and agentic benchmarks at one-sixth the serving price, which compresses the head start any single loop can hold ref.
Evidence base
- 2026-07-18: Fireworks reported 95 per cent of served tokens come from customer-specialised models (fine-tuned open weights, adapters, distillations) ref.
- 2026-07-20: Stripe cut inference costs 73 per cent serving open models on vLLM, running 50 million daily API calls on one-third the GPU fleet ref.
- 2026-07-26: Wrapper products run 50-60 per cent gross margin; proprietary-model or data-moat products clear 70 per cent or more; 2026 Series A benchmark is ~$3.5M ARR, >120 per cent NRR, >60 per cent gross margin ref.
- 2026-07-26: Cursor reached slight gross-margin profitability in April 2026 via its proprietary Composer model and cheaper routing, with NDR above 90 per cent ref.
- 2026-07-07: GLM-5.2 (744bn total / 40bn active parameters, June 2026) matched or exceeded proprietary flagships on long-horizon coding and agentic benchmarks at one-sixth the serving price ref.
- 2026-04-20: Bessemer predicted open source commoditises model architecture while data and deployment layers stay proprietary, shifting value from robot OEMs ref.
- 2026-03-31: Ubtech reported 54.6 per cent humanoid gross margin for 2025 and guided 40-43 per cent for 2026, against the OEM margin-compression thesis ref.
Frontier (open questions)
- Does any pure application-layer company sustain >70 per cent gross margin for four consecutive quarters attributable to proprietary post-training data, rather than to model routing or price arbitrage?
- Do serving vendors’ specialisation shares (e.g. the Fireworks 95 per cent figure) translate into pricing power, or does self-hosting on vLLM-class stacks continue to erode them as it did for Stripe?
- Can a robot OEM or standalone robot-foundation-model company reach durable >40 per cent gross margin at scale without a proprietary data flywheel or component moat?
- How quickly does an open-weight release such as GLM-5.2 close the gap on a task where a competitor has been running a closed data flywheel, measured in months of accumulated interaction data equivalent?
Synthesised 2026-08-31 from 7 KB sources by the resynth pipeline; citations are KB source slugs.