| Methods | same valve | Model | Capability growth is the model layer's product; inference price at fixed capability ⇄ price-per-token pass-through; open-weights lag ⇄ open-vs-closed gap. | Frontier-lab revenue run-ratesOpen-weights gap to the closed frontier (ECI points)METR 50% time horizonMETR 80% time horizon50%/80% horizon ratioFrontier training compute growth (doubling time)Inference price at fixed capability (halving time)Open-weights lag behind the closed frontier (months)ARC-AGI-2 best scoreOrd half-life misfit (observed 50/80 ratio vs constant hazard)Progress against the AI 2027 quantitative predictions |
| Methods | input to | Compute & physical | Compute growth and hardware perf-per-dollar feed the methods stock. | Training power draw growth (doubling time)Hardware price-performance growth (doubling time) |
| Products | same valve | Serving & orchestration | Productisation runs through serving, routing and agent infrastructure. | |
| Products | same valve | Deployment & application | Uplift RCTs, pilot→production and app-layer margin net of inference measure the same valve. | Application layer's share of enterprise generative-AI spend (Menlo)Developer productivity uplift (METR RCTs)Enterprise pilots with measurable P&L impactFrontier models completing expert legal tasks end to end (Harvey LAB) |
| Early adoption | same valve | Adopters | BTOS, BBD, Ramp and the Anthropic Economic Index are adoption from both ends. | Share of work hours assisted by generative AIUS firms using AI (Census BTOS)US businesses paying for AI (Ramp AI Index)Augmentation vs automation share (Anthropic Economic Index)Work-related share of ChatGPT consumer messagesUsage as a share of theoretical exposureUS adults who have used generative AIOccupations with a quarter of their tasks using ClaudeUS workers using generative AI for work weeklyUS workers using generative AI for work every day |
| Adaptation | same valve | Labour & consumers | Productivity, wages, employment and consumer surplus. | Consumer surplus from generative AI (willingness to accept)US labour productivity, year on yearUS total factor productivity, private nonfarm businessEmployment in AI-exposed occupations, year on yearEntry-level employment shortfall in AI-exposed occupationsExposed vs unexposed occupation employment gap (Revelio)California unemployment claims from AI-exposed occupationsRecent-graduate unemployment rate (NY Fed)Cross-tracker concordance (labour)US labour share of income, year on yearWorkers highly exposed to AI with low adaptive capacityOccupational-mix dissimilarity since ChatGPT (Yale Budget Lab)Chatbot effect on earnings and hours in exposed occupations (Denmark)Freelance writing earnings after ChatGPT (Upwork)Wage premium for AI skills (PwC AI Jobs Barometer) |
| Adaptation | leak from | Deployment & application / Deployment / FDE / implementation services | Deployment/FDE services capture part of the surplus that adaptation creates. | |
| Adaptation | same valve | Deployment & application | Adaptation measured inside deployed products: regulated devices built on language models and paid robotaxi rides. | FDA-authorised devices built on a large language modelWaymo paid rides per week |
| Return arrow | return arrow | Training input | Expert data, RL environments and continual-learning infra are the deployment→invention channel; lab recoupment ⇄ capex-to-revenue. | Continual-learning ladder (highest production rung)Active RL-environment vendors (directory count)Expert-data market run-rate (Mercor)Vendors claiming shipped continual learningPer-customer adapters per productRetention after continual updatesLearning-curve slope in deploymentFastest production refresh of a continually updated model (hours)Enterprise sessions allowed to train vendor modelsA benchmark for learning on the job |
| Leak | leak from | Compute & physical | Margin stacking, HHI, circular financing: surplus exiting the chain at semis and cloud. Interpretive spine: Teece, Nordhaus and Perez in docs/interpretation; Ding through the bottleneck list. | |
| Return arrow | return arrow | Model | Inside the labs: agent work on the labs' own research and lab recoupment are deployment feeding back into invention. | Lab recoupment ratioAgent-workdays per human workday in frontier research (self-reported)Human interventions on 4–8 hour agent tasks (self-reported)Deployment data share of training |
| Adaptation | input to | Model | Backlash, safety brakes and paused runs are adaptation acting back on the model layer. | Safety brakes inside the labs (events, trailing year) |
| Methods | input to | Training input | The public text stock and other training inputs bound how fast the methods stock can grow. | Public text data exhaustion year (Epoch projection) |
| Adaptation | same valve | Adopters | Adaptation inside adopting organisations and sectors: regulated devices, state bills, call-centre uplift, reported executive impact. | FDA-authorised AI-enabled devices, growthUS state AI bills introduced (legislative year)Customer-support productivity uplift from a generative-AI assistantExecutives reporting no AI impact on their own firm |
| Products | same valve | Model | Reliability, calibration and the gap between internal and external deployment decide what the model layer can productise. | Reliability-science framework adoptionAgent calibrationAgent operational safety incidentsAgent reliability across repeated runs (tau-bench pass^k)Internal-to-public deployment gap at the frontier |
| Early adoption | same valve | Model | Buyers spreading work across several labs is adoption breadth read from the model layer. | Enterprise multi-model usage share (Menlo) |
| Early adoption | same valve | Labour & consumers | People adopting directly: consumer and worker use counted from the household side. | |
| Leak | leak from | Labour & consumers | Consumer surplus is where most of the value goes if the capture thesis fails: the leak to labour and consumers. | |