Innovation rarely waits for tidy data. When generative AI burst onto the scene, analysts struggled to say whether the opportunity was a few billion dollars of API calls or a multi‑trillion‑dollar reshaping of software. Similar fog envelops carbon‑removal credits, quantum‑computing cloud services, synthetic biology toolkits, Web3 identity wallets, and space‑based solar power. In such frontier arenas, classic top‑down or bottom‑up approaches falter because market boundaries shift monthly, value chains coagulate and dissolve, and price signals flip from scarcity premia to commoditized discounts overnight.
9.1 Framework for Markets with Undefined Value Chains
When pipelines, suppliers, and customer segments are still fluid, traditional “market = producer → distributor → end‑user” diagrams mislead more than they clarify. Instead, sizing must revolve around four anchoring questions:
- What core job‑to‑be‑done is the innovation vying to satisfy?
Is a carbon‑removal credit a compliance tool, a voluntary CSR signal, or a speculative asset? Each job implies a different demand curve, willingness‑to‑pay, and buyer set. - What are the conserved scarcities?
Even in chaos, constraints anchor analysis—GPU hours, regulated bandwidth, asteroid‑launch slots, skilled quantum engineers. Quantify these scarcities to bound maximum supply. - Where does economic gravity pull value?
In many nascent ecosystems the profit pool shifts: from hardware (mobile handsets) to operating systems to app stores; from search engines to ad exchanges to influencer platforms. Map plausible value‑capture nodes even if the actors are not yet clear. - Which trigger events could bend the trajectory?
Regulatory approvals, cost‑curve breakthroughs, or standard‑setting consortiums often flip markets from niche to mainstream. Enumerate triggers and assign probability‑timing windows.
Five‑Step Methodology
Step 1 – Boundary Hypothesis Building
Start with jobs‑to‑be‑done workshops: brainstorm all customer problems the innovation might solve. Cluster them into primary and adjacent use cases. Each cluster becomes a boundary hypothesis—a stake in the ground that can be tested and redrawn as evidence accrues.
Step 2 – Constraint‑First Sizing
Instead of total addressable demand, begin with conserved scarcities. For quantum compute, tally qubits available worldwide under realistic error rates. For DAC (direct air capture), estimate maximum sorbent production and CO₂ storage injection capacity. This yields a physics‑bounded TAM—a firm ceiling no hype can breach.
Step 3 – Ecosystem Mapping via Option Trees
Draw an option tree rather than a value chain. Nodes represent possible value‑capture roles (infrastructure provider, orchestration layer, application integrator). Branch probabilities reflect industry precedent and venture‑capital flows. The tree visualizes multiple future industry structures without locking into one.
Step 4 – Dynamic Adoption Modeling
Classic S‑curves assume homogenous markets; frontier categories fracture by sub‑persona. Build adaptive S‑curves that allow different segments to jump curves when enabling layers to mature. For instance, enterprise adoption of generative‑AI coding assistants may lag consumer chatbots until security layers harden.
Step 5 – Trigger‑Based Scenario Grid
Overlay the option tree with trigger events: “$50/MWh renewable PPA,” “FAA clears autonomous drone corridors,” “EU taxonomy mandates carbon disclosure.” Create a 2×2 or 3×3 grid of trigger combinations. Size each cell using the constraint‑first TAM scaled by adaptive adoption. Weight by trigger probabilities to produce a composite forecast with explicit upside tails and downside floors.
Practical Tips
- Use proxy data aggressively but transparently. GPU lead‑time orders, GitHub repo stars, and VC term‑sheet volumes often predict demand earlier than customer invoices.
- Reforecast frequently. In undefined markets, a quarterly refresh cycle is prudent; monthly if funding rounds or regulatory dockets move fast.
- Maintain narrative logs. Document every boundary change and trigger update—future reviewers need to see the reasoning chain, not just numbers that jumped.
Checklist for Undefined‑Value‑Chain Sizing
- Jobs‑to‑be‑done clusters agreed and documented.
- Conserved‑scarcity metrics identified and quantified.
- Option tree of value‑capture roles drawn with branch probabilities.
- Adaptive adoption curves parameterized for each segment.
- Trigger events enumerated with probability‑time windows.
- Scenario grid sized, weighted, and rolled into composite forecast.
- Refresh cadence and narrative log established.
This framework accepts uncertainty as a feature, not a flaw. By bounding the physics, visualizing multiple structural futures, and tying growth to explicit triggers, it delivers decision‑grade insight while staying humble about what cannot yet be known.
9.2 Adoption‑Curve and Diffusion‑of‑Innovation Models
When a technology is still negotiating its relevance—think quantum‑computing clouds or carbon‑removal credits—traditional linear growth assumptions mislead. Uptake tends instead to follow non‑linear patterns: slow ignition among innovators, rapid acceleration as social proof mounts, and eventual tapering once saturation nears. Diffusion‑of‑innovation theory provides the mathematical vocabulary to capture those patterns and translate them into market‑sizing forecasts.
The Classic Bass Model—Quick, Transparent, Imperfect
Developed in 1969, the Bass diffusion equation models cumulative adopters N(t)N(t)N(t) as driven by innovation (ppp) and imitation (qqq) coefficients:
- M= ultimate market potential (units or firms)
- p = coefficient of external influence (marketing, subsidies)
- q = coefficient of internal influence (word‑of‑mouth, network effects)
Strengths: needs only three parameters, fits many historical series (color TVs, solar PV).
Limitations: assumes homogeneous adopters, ignores supply constraints, treats MMM as static.
Practical calibration: use early sales data or analog markets (EVs vs. hybrids) with non‑linear least squares. If data is scarce, set p=0.03p = 0.03p=0.03, q=0.38q = 0.38q=0.38 (consumer durables median) and stress‑test.
Logistic & Gompertz Curves—Flexible Shape Control
When adoption starts slower or saturates earlier than Bass predicts, logistic or Gompertz functions offer alternatives:
(Gompertz, asymmetric—fast early, slower late)
Choose form by inspecting semi‑log plots: symmetric S‑shape → logistic; right‑skewed peak → Gompertz.
Multi‑Segment & Layered S‑Curves
Frontier markets rarely adopt en masse; segments flip at different trigger points. Build layered S‑curves: one per segment (e.g., Tier‑1 automakers, mid‑market OEMs, aftermarket retrofits). Calibrate each to segment drivers—capital‑expenditure cycles, regulatory deadlines. Aggregate for total adoption. This preserves realism without drowning in micro‑detail.
Supply‑Constrained Adoption
If capacity lags demand (vaccines, GPUs), adoption is the minimum of demand curve and cumulative supply. Overlay a capacity ramp curve (see Chapter 6) on the diffusion model; actual uptake follows the lower bound until supply catches up.
Agent‑Based Simulation—When Interactions Drive Outcomes
Peer effects, platform two‑sidedness, or localized externalities (charging networks) break analytical curves. Agent‑based models (ABM) simulate thousands of heterogeneous agents—each with utility thresholds, network connections, budget constraints—and observe emergent adoption.
Toolkits: NetLogo for prototypes; mesa (Python) or AnyLogic for scalable runs.
Input needs: distribution of willingness‑to‑pay, network topology (scale‑free for social media, lattice for geographic contagion), policy levers (subsidy per adopter).
Outputs: adoption over time, clustering heatmaps, sensitivity to policy shocks.
Downside: data‑hungry, computational, harder to explain to non‑technical execs.
Linking Diffusion Models to Revenue
Adoption alone is not revenue. Multiply cumulative units by:
- Usage intensity if pay‑per‑use (API tokens, carbon credit redemptions).
- ASPs declining on experience curves for hardware.
- Subscription churn curves for SaaS layers atop the innovation.
Then layer price/volume scenarios (Chapter 7) and stochastic drivers (Chapter 8) for risk bands.
Calibration Shortcuts When Data Is Scarce
Situation | Proxy Calibration |
Zero sales data | Use analog tech with similar complexity & customer type; down‑weight qqq if network effects weaker |
Few early adopters | Calibrate ppp to observed trial rate; estimate qqq from social‑media buzz (growth of mentions) |
Moving regulatory target | Shift t0t_0t0 per policy timeline; simulate alternate start years |
Validation and Iteration
- Back‑cast using the first 10 % of adoption to predict the next 10 %; adjust parameters if error >20 %.
- Cross‑method triangulation: compare modeled units in year X to capacity‑based supply (Chapter 6) and NLP‑extracted forward‑guidance (Chapter 8.4).
- Real‑time signals: feed monthly KPI deltas—PO approvals, prototype pilots—into Bayesian updates of ppp and qqq.
Common Pitfalls—and Remedies
Pitfall | Symptom | Fix |
Over‑fitting Bass to sparse data | Unrealistically high qqq | Use hierarchical Bayesian pooling across regions/segments |
Static market potential MMM | Market suddenly outgrows ceiling | Tie MMM to external variables (e.g., GDP, regulation) |
Ignoring price feedback | Adoption stalling at sticker shock | Embed price elasticity loops—lower ASP boosts ppp, qqq |
Quick‑Action Checklist
- Select diffusion model: Bass, logistic, Gompertz, or ABM.
- Calibrate p,q,Mp, q, Mp,q,M or equivalent using early data or analogs; document sources.
- Segment of adoption drivers differ materially; stack curves.
- Overlay supply‑capacity curve when relevant.
- Convert adoption to revenue with ASP, usage, churn inputs.
- Validate with out‑of‑sample back‑cast; iterate quarterly.
- Integrate into scenario grid (Section 9.1) and Monte‑Carlo risk bands.
Applied with rigor and humility, diffusion models transform the fog of “disruptive potential” into structured, testable adoption forecasts—helping leaders stage investments, time capacity, and set go‑to‑market cadence in markets where yesterday’s data offers little guidance on tomorrow’s prize.
9.3 Regulatory and Technology‑Trigger Considerations
In nascent markets, growth often pivots not on incremental consumer sentiment or marginal price drops but on a handful of “big bang” events: a government mandate, a breakthrough cost threshold, or an industry‑wide interoperability standard. These triggers determine when adoption curves bend, when funding floods in, and when supply chains scale. Ignoring them yields forecasts that glide smoothly past an inflection point that reality treats like a cliff.
Two Classes of Triggers—Policy and Technology
- Regulatory triggers
- Mandates and bans – Internal‑combustion sales phase‑outs, single‑use‑plastic bans, minimum cyber‑resilience standards.
- Subsidies and tax incentives – ITC for solar, 45Q carbon‑capture credits, R&D super‑deductions.
- Market‑based instruments – Cap‑and‑trade, clean fuel standards, tradable permit schemes.
- Disclosure or compliance rules – SEC Scope‑3 emissions, EU CSRD sustainability reporting.
- Technology triggers
- Cost parity milestones – Solid‑state batteries <$80 kWh, green hydrogen <$1.50 kg, LEO launch <$500 kg.
- Performance breakthroughs – Quantum error rates below 1e‑3, AI model context windows >100 k tokens.
- Standards ratification – 5G Release 18, USB‑C uniform charging, OPC‑UA TSN for industrial Ethernet.
- Ecosystem enablers – Open‑source frameworks, SaaS APIs, manufacturing process IP release.
Triggers often interplay: a subsidy accelerates cumulative production, which pushes cost down the learning curve, which flips a cost‑parity trigger, which in turn prompts tighter regulation. Modeling must capture such feedback.
Framework for Incorporating Triggers into Market Sizing
1 Trigger identification matrix
List potential triggers in rows; categorize by certainty (high, medium, speculative) and impact (low, medium, high). High‑certainty/high‑impact items become baseline milestones; speculative/high‑impact ones define upside or downside tails.
2 Time‑window probability curves
For each trigger, assign a probability distribution over time—triangular or PERT. Example: 80 % chance the EU approves e‑fuel quotas by 2027, 15 % by 2026, 5 % slip to 2029. Document sources—legislative calendars, patent pipeline, prototype roadmaps.
3 Parameter linkage
Map each trigger to model drivers:
- Mandate → adoption S‑curve inflection t0t_0t0 shifts earlier by ∆ years.
- Cost parity → ASP declines accelerate (learning‑rate slope change).
- Standard ratification → reduces perceived risk, raising imitation coefficient qqq in the Bass model.
- Subsidy sunset → price effective net‑of‑incentive rises, dampening demand elasticity.
Link via explicit formulas or scenario switches—no hidden overrides.
4 Scenario synthesis
Combine triggers using logical relationships:
- Conjunctive – both A and B must occur for market opening (e‑VTOL requires FAA airworthiness and urban vertiport zoning).
- Disjunctive – either A or B suffices (carbon price or corporate net‑zero pledges could tip CCS adoption).
- Sequence dependent – cost‑parity cannot arrive before learning‑rate acceleration, which itself needs volume, which depends on subsidies.
Encode sequences in Monte‑Carlo with correlated random draws or event‑driven simulation (e.g., Bayesian belief networks).
5 Refresh and revision governance
Set trigger review checkpoints aligned with legislative sessions, industry conferences, or earnings seasons. Use NLP alerts (Chapter 8.4) to flag new bills, patents, or prototype demos so the model updates automatically or queues an analyst review.
Practical Tips for Trigger Modeling
- Use Bayesian updating. As bills clear committees or prototypes hit benchmarks, update prior probabilities to posteriors—keeps forecasts responsive without wholesale rebuilds.
- Model threshold effects, not linear drifts. A subsidy dropping from 30 % to 26 % may be negligible; its expiry to 0 % is a cliff.
- Stress for unintended consequences. Regulations can backfire—internal‑combustion bans may spur older‑vehicle life extension, slowing fleet turnover. Include rebound loops where plausible.
- Differentiate geographic timelines. Chinese and EU mandates often advance faster than US federal rules; segment triggers regionally.
Common Pitfalls—and Mitigations
Pitfall | Symptom | Mitigation |
Trigger optimism bias | Aggressive timelines for tech breakthroughs | Benchmark against historical analogs—lithium‑ion, mRNA—then haircut |
Binary thinking | Treating probability 0 % or 100 % | Use distribution curves; even “certain” triggers face implementation delays |
Hidden dependencies | Subsidy assumed effective without supply‑chain capacity | Add supply‑constraint overlay; trigger only activates when capacity meets demand |
Policy whiplash | Ignoring possibility of repeal or legal challenge | Include reversal scenarios with tail probability (e.g., 10 %) |
Quick‑Action Checklist
- Trigger matrix populated; certainty vs. impact charted.
- Probability‑time distributions assigned and sourced.
- Triggers explicitly linked to model parameters via formulas.
- Logical relationships (AND, OR, sequence) encoded in scenario engine.
- Bayesian update mechanism or review cadence established.
- Downside and upside tails quantify trigger failures or accelerations.
- Narrative summarizing trigger path communicated to decision‑makers.
By embedding regulatory and technology triggers into your sizing logic, you convert uncertainty from a vague hand‑wave into a structured risk‑reward calculus—arming leaders with a clear view of what must happen, by when, for their bets to pay off, and what contingencies to prepare if reality charts a different course.
9.4 Case Study: Generative‑AI Foundation Models
Few markets have sprinted from laboratory curiosity to board‑level priority faster than generative‑AI foundation models (FMs). In just three years, usage exploded from novelty chatbots to core productivity tools, synthetic‑media engines, code copilots, and verticalized copilots for legal, design, and biotech. Yet published estimates of the sector’s future value still swing from <$50 billion to >$1 trillion. Applying the framework developed in Sections 9.1–9.3 brings order to the chaos and yields a decision‑grade forecast executives can bank on.
1 Boundary Hypothesis—Jobs to Be Done
Foundation models address three primary jobs, each spawning adjacent use cases:
- Cognitive augmentation—drafting, summarizing, coding, translating.
- Creative generation—images, video, 3‑D, music, synthetic data.
- Semantic orchestration—agentic workflows that chain tools or APIs.
Adjacencies include model‑based fine‑tuning services, domain‑specialized copilots (medical, legal), and embedded inference in edge devices. For sizing, we define the core FM services market as API‑ or licensing‑based access to large‑scale text, image, and multimodal models, excluding downstream SaaS wrappers that merely call the API.
2 Conserved Scarcities—Bounding Supply
- Compute: H100‑class GPUs or equivalent ASICs. Published cloud provider roadmaps plus hyperscaler CapEx filings point to ~25 million high‑end accelerators installed by 2027, doubling to ~55 million by 2030.
- Energy: Inference clusters draw ≈0.5 kWh per 1 000 tokens; global data‑center energy growth forecasts cap feasible power at ~2 % of total generation.
- Data: High‑quality, non‑copyright text and code plateau around 10 trillion tokens; synthetic self‑play may alleviate but not fully replace scarcity.
These scarcities bound the physics‑limited TAM at ≈500 quadrillion tokens served annually by 2030, translating to roughly $290 billion revenue at a blended $0.60 per million tokens.
3 Option Tree—Future Value‑Capture Roles
- Model builders (OpenAI, Anthropic, Google DeepMind)
- Accelerated‑compute cloud vendors (AWS, Azure, Google Cloud, OCI)
- Fine‑tuners/platforms (Databricks, Snowflake, open‑source community)
- Vertical solution integrators (law tech, med tech, design suites)
Branch probabilities—derived from venture funding flows and early gross‑margin data—suggest 55 % of long‑run profit pools accrue to compute and infrastructure layers, 30 % to model builders, and 15 % to fine‑tuners. Vertical SaaS captures revenue but low gross profit once royalty and compute pass‑throughs are netted.
4 Adaptive Adoption Curves
We deploy layered S‑curves for four customer archetypes:
Segment | Baseline Adoption (2024) | Trigger to Rapid Uptake | Asymptote |
Developers (code copilots) | 18 % | <$20/month seat cost | 75 % |
Knowledge workers | 6 % | Native integration in MS 365 / Google Workspace | 65 % |
Creative professionals | 12 % | IP indemnity standards | 70 % |
Regulated industries | 2 % | SOC 2+/ISO 42001 compliance | 40 % |
Weighted across the global workforce and factoring seat expansion within organizations, cumulative demand reaches ~2.8 trillion FM calls/day by 2030—comfortably within the compute cap above.
5 Trigger Matrix and Probabilities
Trigger | Median Year | P‑window (±2 y) | Impact |
IP‑safe harbor legislation (US/EU) | 2026 | 70 % | +0.08 q (imitation) |
$15/GPU‑hour cost milestone | 2027 | 65 % | −35 % inference ASP |
ISO 42001 AI management standard | 2025 | 85 % | +5 pp adoption in regulated industries |
50 % energy‑carbon penalty | 2029 | 30 % | −10 % adoption growth |
Monte‑Carlo draws over these triggers feed into the adaptive curves, generating a probability distribution of adoption timing and effective pricing.
6 Composite Market‑Sizing Result (Reference Case)
- Total tokens served 2030: 465 peta‑tokens (P50).
- Blended net price: $0.62 per million tokens after anticipated price decline.
- API and licensing revenue: $288 billion.
- Incremental platform services (fine‑tuning, vector DBs): +$35 billion.
- Infrastructure margin capture: 55 % of gross profit, implying meaningful upside for compute vendors.
P10/P90 band: $195 – $410 billion, driven primarily by GPU supply elasticity and IP‑safe‑harbor timing.
7 Cross‑Validation Checks
- Supply–demand reconciliation: 55 million H100‑equivalent GPUs × 200 infer‑TFlops × 70 % utilization yields ~480 peta‑tokens—tight but feasible.
- Energy envelope: At 0.5 kWh/1 000 tokens, 465 peta‑tokens require ~232 TWh—<0.7 % of projected 2030 global electricity, within IEA data‑center scenarios.
- NLP pipeline sentiment: Patent filings for GPU interconnect and sparse‑attention methods rose 3× YoY, supporting cost‑decline trigger plausibility.
8 Strategic Implications
- Compute scarcity remains gating through 2027; firms securing multi‑year GPU contracts gain price leverage.
- Vertical SaaS winners must differentiate on domain data and workflow integration, not core model IP.
- Regulatory engagement is a value lever: active lobbying for IP safe harbor or AI management standards can shift adoption curves materially.
- Energy procurement strategy becomes a competitive advantage in cost structure and ESG perception.
9 Key Uncertainties to Monitor (Refresh Triggers)
- Breakthroughs in sparse expert mixtures could cut token compute cost by 60 %.
- Litigation outcomes on copyright training data may raise royalty overhead.
- Quantum‑resistant encryption demands could spike computational load per token.
Quick‑Check Recap
- Physics‑bounded TAM built from GPU and energy caps = ~$290 B.
- Adoption layering reaches 2.8 T FM calls/day by 2030.
- Trigger modeling aligns median forecast at $288 B, with tails at $195 B/$410 B.
- Cross‑checks on energy and supply validate feasibility; room for upside via technology breakthroughs.
Applying the full frontier‑market framework thus transforms the generative‑AI gold rush from headline fireworks into a quantified, trigger‑linked sizing, arming strategists with a clear compass for capital allocation, partnership negotiation, and regulatory advocacy in one of the decade’s most explosive arenas.
