External AI Dependence and Startup Financial Survivability

01What this research is about and why it matters

Executive Summary

For independent founders and early-stage ventures, the practical implication is not that AI tools are unaffordable -- it is that the financial model underpinning AI adoption was designed around large-organisation economics in high-income economies. Every independent business making a build-versus-buy decision, every founder evaluating AI infrastructure spend against runway, and every investor modelling unit economics in an AI-intensive sector is operating inside an asymmetric cost structure that existing frameworks do not adequately surface. The Specialised Digital Assets framework proposed in this paper -- the principle that independent businesses should own rather than rent their critical infrastructure -- is the practical complement to the theoretical foundation that Judgment Layer Theory (Agarwal, 2026) establishes at the organisational governance level: organisations that do not own what they depend on face compounding structural vulnerability.

How does dependence on externally owned AI infrastructure affect the financial survivability of independent businesses operating in AI-intensive sectors?

Primary Finding

Independent businesses across ten economies pay the same nominal USD price for AI infrastructure but absorb materially different income-adjusted costs. A representative AI tool stack costing approximately USD 1,131 per year consumes roughly 1.5% of average annual income in the United States but rises to 15-20% in Brazil -- a tenfold asymmetry driven entirely by currency denomination and income-level differences. This structural cost gap, compounded by price trajectory risk, competitive asymmetry, platform lock-in, and AI project failure amplification, creates measurable and widening survivability pressure for independent founders in lower-EADI economies.

~1.5%

US Income Share

USD 1,131/yr AI stack as share of average US annual income

~15-20%

Brazil Income Share

Same USD 1,131/yr stack as share of average Brazilian annual income

US EADI Score

Highest EADI among ten studied economies; AI infrastructure is domestic (Source classification)

Brazil EADI Score

Lowest EADI among ten studied economies; near-total dependence on externally owned infrastructure

80%+

AI Project Failure Rate

Share of AI projects that fail to reach production deployment (RAND 2024)

95%

GenAI Pilots with No ROI

Share of GenAI pilots generating no measurable return on investment (MIT NANDA 2025)

~63%

Hyperscaler Cloud Concentration

AWS (~28%), Azure (~21%), and GCP (~14%) combined share of global cloud compute market

Economies Studied

US, China, South Korea, France, Germany, Japan, Canada, UK, India, Brazil

Who this research serves

Independent Founders

To understand the true income-adjusted cost of AI infrastructure dependency and make informed build-versus-buy decisions with full survivability implications modelled into financial planning.

Investors and VCs

To incorporate EADI-adjusted cost structures into unit economics models, due diligence templates, and portfolio risk assessment for AI-intensive ventures across different economies.

Policymakers

To design export control frameworks, digital sovereignty programmes, and SME support schemes that account for the asymmetric cost burden placed on independent businesses in lower-EADI economies.

02What is being lost and why

The Problem

The contemporary AI landscape presents a surface-level uniformity that obscures a deep structural asymmetry. Every independent founder -- whether operating from San Francisco, Seoul, or São Paulo -- faces the same advertised API pricing from OpenAI, Anthropic, Google, and Microsoft. The nominal dollar figure appears equal. The economic reality is not. When income-adjusted costs are calculated, a founder in Brazil spending USD 1,131 per year on a representative AI tool stack is committing 15-20% of average annual income to infrastructure she does not own, cannot modify, and has no contractual guarantee of continued access to. Her counterpart in the United States commits roughly 1.5% of comparable income to the same stack. This tenfold asymmetry does not appear in any standard financial model, pitch deck, or due diligence template. It is the central empirical finding of this paper and the motivation for the External AI Dependence Index (EADI) framework.

The AI industry has converged on a pricing model that is geographically uniform in nominal terms but structurally regressive in real terms. Foundation model providers set prices in USD and adjust them infrequently relative to exchange rate movements. Cloud compute providers -- dominated by AWS (~28%), Azure (~21%), and GCP (~14%), representing approximately 63% of global market share -- operate similarly. Hardware supply chains are concentrated in Nvidia for GPU design and TSMC for fabrication, with both subject to US export control regimes that govern which economies can access the most advanced chip architectures. The combined effect is a three-layer infrastructure stack where pricing power, supply decisions, and geopolitical risk all rest outside the operating jurisdiction of most AI-dependent businesses. Large enterprises partially offset this through volume discounts and negotiated enterprise contracts; independent businesses do not. BCG and McKinsey data indicate that large firms are adopting AI at 1.5-2x the rate of small and medium enterprises, partly because the income-adjusted cost burden is proportionally lower at scale.

Existing frameworks for assessing AI risk -- including OECD AI principles, national AI strategies, and standard venture due diligence -- focus on ethical risk, regulatory compliance, and technical feasibility. None of the eighteen primary data sources reviewed for this paper contained a cross-economy measure of income-adjusted AI infrastructure cost burden. The closest approximations are the OECD's digital trade policy work and the World Bank's digital development indices, but neither operationalises infrastructure dependency as a financial survivability variable. This gap is the motivating observation for the EADI framework and the five-mechanism model developed in this paper.

A representative independent business operating in AI-intensive sectors would reasonably require: a foundation model API subscription (ChatGPT Plus or Claude Pro at approximately USD 240/yr), a cloud compute instance for inference and storage (approximately USD 600/yr), a developer toolchain including CI/CD and monitoring (approximately USD 180/yr), and supplementary data services (approximately USD 111/yr). This USD 1,131 annual total is used as the denominator for income-adjusted cost calculations throughout the paper. All components are priced in USD, sourced from US or UK-headquartered providers, and subject to USD-denominated price trajectories. The figure is a deliberate lower-bound estimate; established independent businesses in production commonly spend materially more.

The paper introduces the Builder-Dependent Ratio as a complementary metric to EADI. India has approximately 8,178 AI companies -- one of the largest ecosystems outside the US and China -- but operates at a Builder-Dependent Ratio exceeding 1:100, meaning more than 100 AI-dependent businesses exist for every domestic builder of foundational AI infrastructure. The UK has approximately 6,270+ AI companies at a ratio of approximately 1:50. This ratio determines the proportion of businesses in each economy that are structurally exposed to the five financial mechanisms identified in the paper, regardless of the absolute size or sophistication of the national AI ecosystem.

Stanford HAI (2025) data shows that the performance gap between proprietary and open-weight foundation models narrowed from approximately 8% to approximately 1.7% between 2023 and 2025. This convergence is relevant because it suggests that switching costs at the model layer -- one of the five financial mechanisms -- are decreasing. However, switching costs at the cloud compute and data pipeline layers remain structurally high. The fine-tuning investment required to adapt an open-weight model to a business-specific use case ranges from USD 100 to USD 5,000, representing a non-trivial barrier for capital-constrained independent businesses. Open-weight convergence at the model layer does not affect currency asymmetry, cloud concentration, or chip-layer dependency.

03The precise question this research addresses

Research Question

How does dependence on externally owned AI infrastructure affect the financial survivability of independent businesses operating in AI-intensive sectors, across economies with materially different income levels and AI infrastructure sovereignty positions?

Context

The question emerged from observing that standard frameworks for evaluating AI investment risk do not account for cross-economy income asymmetry in AI infrastructure costs. The research was motivated by the founder's operational experience building triNetra -- an AI-intensive independent business -- and the observation that infrastructure cost as a share of income varied dramatically depending on the economy in which a founder operated, even when the nominal cost was identical. The gap between the nominal USD figure and the real income-adjusted burden is the primary analytical object of this paper.

Scope

Ten economies selected to represent a range of income levels and AI infrastructure sovereignty positions: United States, China, South Korea, France, Germany, Japan, Canada, United Kingdom, India, and Brazil. The analysis covers the period 2023-2026, with primary data from 2024-2025. The unit of analysis is the independent business -- defined as a company with fewer than 50 employees that is not a subsidiary of a larger entity and that depends on externally sourced AI infrastructure for core product or service delivery.

Assumptions

AI infrastructure pricing remains USD-denominated for the foreseeable projection period.
Income-adjusted cost is a valid proxy for financial survivability pressure when controlling for revenue model differences.
The representative USD 1,131 annual tool stack is a reasonable lower-bound estimate for an AI-intensive independent business.
Open-weight model convergence does not eliminate switching costs at the cloud compute and data pipeline layers.
The five financial mechanisms identified operate with additive survivability impact across all EADI levels.
Builder-Dependent Ratio is a useful structural indicator of ecosystem dependency, independent of per-firm financial metrics.

Out of scope

Enterprise AI adoption economics and Fortune 500 infrastructure strategy.
Consumer AI product pricing and individual user affordability.
Ethical dimensions of AI development and deployment.
Technical performance benchmarking of specific AI models or platforms.
Regulatory compliance costs associated with sector-specific AI governance frameworks.
Intra-economy variation in income levels, regional cost of living, or sub-national policy differences.

04How the research was conducted- four phases

PaaF Methodology Journey

The first phase required constructing a clear taxonomy of AI infrastructure. The paper identifies three distinct dependency layers that together constitute the complete AI infrastructure stack for a typical independent business: Foundation Models, Cloud Compute, and AI Chips.

Foundation Models represent the most visible layer. The paper classifies providers into two groups: Source economies -- the US (OpenAI, Anthropic, Google, Meta) and China (Baidu, Alibaba, DeepSeek) -- which possess domestically developed foundation models at the frontier. All other eight economies studied are categorised as dependent at this layer, accessing models exclusively through API calls priced in the provider's domestic currency.

Cloud Compute represents the second layer. Synergy Research Group data (Q4 2024) indicates that AWS holds approximately 28%, Azure approximately 21%, and GCP approximately 14% of global market share -- giving the three US-headquartered hyperscalers approximately 63% of the market collectively. This concentration means that even businesses using non-US foundation models typically run inference and storage on US-owned infrastructure.

AI Chips represent the third and most structurally constrained layer. Nvidia dominates GPU design for AI training and inference. TSMC, headquartered in Taiwan, fabricates the majority of advanced chips. US export control regulations -- particularly the Entity List restrictions updated in 2024 and 2025 -- directly govern which economies can access the most advanced chip architectures, creating a sovereign dependency that no amount of commercial negotiation can fully resolve at the firm level.

Questions asked

Which economies have domestically developed foundation models at a commercially competitive level?
What share of global cloud compute is controlled by entities headquartered in each study economy?
How do US export control regimes affect chip access across the ten study economies?
Is it possible for an independent business in a non-Source economy to avoid dependency on all three layers simultaneously?

Patterns identified

No economy outside the US achieves full independence across all three infrastructure layers simultaneously.
China achieves partial independence at the Foundation Model layer but remains dependent on advanced chip fabrication via TSMC.
The Cloud Compute layer shows the highest market concentration of any layer, with three US entities controlling approximately 63% of global capacity.
Export control regimes create a hard ceiling on chip-layer independence for most non-US economies regardless of domestic investment.

The infrastructure mapping phase established that the three-layer dependency structure is essentially universal outside the United States for independent businesses. No independent business in any of the eight non-Source economies can source its complete AI infrastructure stack domestically. Even China's partial independence at the model layer is constrained by chip fabrication dependency.

Infrastructure dependency alone does not create survivability risk. The second phase focused on identifying the specific financial mechanisms through which dependency becomes a material cost or risk factor. Five mechanisms were identified through review of eighteen primary data sources and cross-validated against founder operational experience.

Mechanism 1 -- Currency-Denominated Cost Asymmetry -- is the most direct mechanism. AI infrastructure is priced in USD. Businesses in economies with weaker or more volatile currencies relative to USD absorb a structurally higher income-adjusted cost for identical nominal expenditure. A USD 1,131 annual stack consumes approximately 1.5% of average US income but approximately 15-20% of average Brazilian income. This asymmetry is compounded during periods of USD strengthening relative to local currencies.

Mechanism 2 -- Price Trajectory Risk -- reflects the forward-looking dimension of current pricing. SemiAnalysis and McKinsey data indicate that current AI API pricing is below the fully-loaded cost of delivery, subsidised by venture capital and hyperscaler balance sheets. When pricing normalises toward cost-covering levels -- a trajectory that major providers have signalled through tiered pricing introductions and volume discount restructuring -- independent businesses without volume leverage will absorb the full increase.

Mechanism 3 -- Competitive Asymmetry -- captures the AI adoption rate differential between large and small organisations. BCG and McKinsey data indicate large firms are adopting AI at 1.5-2x the rate of SMEs. Large firms also access volume discounts and negotiated enterprise agreements that are structurally unavailable to independent businesses. Over time, this creates a compounding productivity and cost-per-output gap that is not reducible to operational efficiency alone.

Mechanism 4 -- Platform Switching Costs -- addresses the lock-in dynamics that emerge once a business integrates a specific foundation model provider into its product stack. Switching costs include data pipeline migration, fine-tuning reinvestment (USD 100-5,000 per model), prompt engineering re-development, and integration testing. While open-weight model convergence has narrowed the performance gap to approximately 1.7% (Stanford HAI 2025), switching costs at the cloud compute and data pipeline layers remain structurally high.

Mechanism 5 -- AI Project Failure Cost Amplification -- addresses the compounding effect of high AI project failure rates on businesses with limited capital buffers. RAND (2024) reports 80%+ failure rates; MIT NANDA (2025) reports 95% of GenAI pilots generating no measurable ROI; S&P Global (2025) reports 42% of companies abandoned AI initiatives in 2025. For large enterprises, failed AI projects represent manageable write-offs. For independent businesses, the same failure may represent an existential capital event, particularly when infrastructure costs were non-refundable cloud compute commitments.

Questions asked

Through what specific financial channels does infrastructure dependency become a survivability variable?
Is current AI infrastructure pricing sustainable, and what is the direction of future price trajectories?
How does AI adoption rate asymmetry between large and small firms compound over a 3-5 year horizon?
What are the realistic switching costs for an independent business that has integrated a proprietary foundation model?
How do AI project failure rates interact differently with large-firm versus independent-firm capital structures?

Patterns identified

All five mechanisms are structurally asymmetric -- they impose proportionally greater burden on smaller businesses in lower-income economies.
Mechanisms 1 and 5 interact directly: currency asymmetry amplifies the survivability impact of project failures in lower-EADI economies.
Mechanism 2 (price trajectory risk) is the only mechanism not yet fully materialised -- it represents a compounding forward risk.
Open-weight convergence partially mitigates Mechanism 4 at the model layer only; cloud compute and data pipeline switching costs are unaffected.

The five-mechanism framework provides a structured account of how dependency translates into financial risk. The mechanisms are additive but not equally immediate: Currency Asymmetry and Project Failure Amplification are active today; Price Trajectory Risk and Competitive Asymmetry are compounding forward risks; Platform Switching Costs are partially mitigated by open-weight convergence at the model layer only.

The third phase synthesised the infrastructure mapping and mechanism identification into a cross-economy survivability assessment. The primary analytical tool is the External AI Dependence Index (EADI), a composite score (0-100) reflecting each economy's structural position across the three infrastructure layers, Builder-Dependent Ratio, and policy environment for AI infrastructure sovereignty.

EADI scores for the ten study economies are as follows: United States 92 (Source), China 64 (Low dependency), South Korea 45 (Moderate), France 32 (High), Germany 28 (High), Japan 24 (High), Canada 19 (Critical), United Kingdom 17 (Critical), India 12 (Critical), Brazil 7 (Critical). The classification tiers -- Source, Low, Moderate, High, Critical -- reflect qualitative breakpoints in the combined impact of the five mechanisms.

The survivability assessment maps EADI scores against income-adjusted cost burden for the representative USD 1,131 stack. The relationship is nonlinear: economies at Critical EADI levels face not only higher income-adjusted costs but also greater simultaneous exposure to all five mechanisms, with fewer institutional buffers -- government AI programmes, domestic alternatives, policy protections -- to offset them.

The paper also introduces the Specialised Digital Asset hierarchy as a strategic response framework. Independent businesses can reduce their effective dependency by investing in assets that are difficult or impossible to replicate through API access alone: Level 1 Reusable Workflows, Level 2 Structured Proprietary Knowledge, Level 3 Industry-Specific Datasets, Level 4 Business-Specific AI Systems, and Level 5 Specialised Domain Products. Assets at higher levels of the hierarchy provide stronger insulation against the five mechanisms, particularly Mechanisms 4 and 5.

Questions asked

What composite measure best captures an economy's structural AI infrastructure dependency in a comparable way across ten diverse economies?
How does EADI score correlate with income-adjusted cost burden across the study economies?
What strategic options are available to independent businesses to reduce effective dependency without requiring domestic infrastructure investment?
Does the Builder-Dependent Ratio add explanatory power beyond EADI for predicting survivability pressure?

Patterns identified

EADI scores cluster into two structurally distinct groups: Source/Low (US, China) and the remaining eight economies, which share the characteristic of no domestic alternative at the foundation model layer.
Critical-EADI economies (Canada, UK, India, Brazil) face simultaneous exposure to all five mechanisms with the fewest institutional offsets.
The Specialised Digital Asset hierarchy provides a tractable firm-level response to survivability pressure that does not require macro-level infrastructure investment.
Builder-Dependent Ratio and EADI are correlated but distinct: the UK has a higher absolute AI company count than India but a lower EADI score, reflecting different policy and institutional environments.

The EADI framework successfully differentiates the structural survivability positions of the ten study economies. Critical-EADI economies face compounding survivability pressure from all five mechanisms simultaneously. The Specialised Digital Asset hierarchy offers the most immediately tractable firm-level response strategy.

The fourth phase translates the analytical findings into structured response frameworks for three primary audiences: independent founders, investors and VCs, and policymakers. The underlying logic is that the EAD structure cannot be individually dissolved -- no independent founder can build a foundation model or hyperscaler -- but its survivability impact can be materially reduced through deliberate strategic choices at the firm and ecosystem level.

For independent founders, the primary strategic lever is the Specialised Digital Asset hierarchy. By investing in proprietary data assets, fine-tuned models, and domain-specific AI systems, founders can shift from being pure consumers of AI infrastructure to being holders of AI capital -- assets whose value is not directly reducible by provider pricing changes or export control regimes. The progression from Level 1 (Reusable Workflows) to Level 5 (Specialised Domain Products) represents increasing insulation from the five mechanisms.

For investors and VCs, the key implication is that EADI-adjusted unit economics should become a standard component of due diligence for AI-intensive investments. A business with strong nominal financials in a Critical-EADI economy may be carrying hidden survivability risk from currency asymmetry and price trajectory exposure that does not appear in standard financial models. Portfolio construction should account for EADI distribution to avoid concentration risk in Critical-EADI economies.

For policymakers, the paper identifies three tractable intervention points: sovereign compute initiatives that reduce cloud-layer dependency for domestic businesses; AI tool cost subsidy programmes calibrated to income-adjusted burden rather than nominal price; and open-weight model adoption incentives that reduce foundation-model-layer dependency. The paper notes that OECD digital trade policy frameworks do not address income-adjusted AI cost burden, and recommends this gap be addressed in future iterations of digital development indices.

Questions asked

What can an individual founder do today to reduce survivability exposure without requiring access to domestic AI infrastructure?
How should investors adjust due diligence frameworks to account for EADI-related financial risk?
Which policy interventions are most tractable for Critical-EADI economies within existing institutional constraints?
Does the Specialised Digital Asset hierarchy provide sufficient insulation at all five EADI levels, or are there ceiling effects at the infrastructure layers?

Patterns identified

Firm-level responses (Specialised Digital Assets) are tractable for all founders regardless of EADI level but deliver diminishing marginal insulation at the cloud and chip infrastructure layers.
Investor-level responses require framework adoption before survivability risk materialises -- post-hoc adjustment is too late in high-failure-rate environments.
Policy interventions are most tractable at the cloud compute layer, where sovereign compute initiatives have precedent (EU GAIA-X, India cloud programmes), and least tractable at the chip layer, which is governed by geopolitical export control regimes beyond individual economy influence.

The strategic response framework establishes that EAD is a structural condition -- not a temporary inefficiency -- and that meaningful responses require deliberate action at all three levels: founder, investor, and policymaker. The Specialised Digital Asset hierarchy provides the most immediately actionable pathway for independent founders operating today in any EADI tier.

05External AI Dependence Causal Architecture

Interactive Framework

A causal chain from the three AI infrastructure dependency layers through the EADI scoring mechanism to the five financial mechanisms and final survivability outcomes. Nodes represent structural components; edges represent causal or contributory relationships. Dashed edges indicate mitigating or constraining relationships.

Click any node to see details.

Input / Process

Output

Constraint

Ecosystem

Module

Tier

06Browse and filter the evidence base

Evidence Explorer

Evidence underlying the research findings, organised by PaaF phase and evidence strength. Use the filters to focus on the evidence most relevant to your evaluation.

Claim

Source

Strength

Phase

More than 80% of AI projects fail to reach production deployment.

RAND Corporation (2024). AI Project Failure Analysis.

Strong

95% of GenAI pilots generate no measurable return on investment.

MIT NANDA Lab (2025). Generative AI Pilot Outcomes Study.

Strong

AWS, Azure, and GCP collectively hold approximately 63% of global cloud compute market share (28%, 21%, and 14% respectively).

Synergy Research Group (Q4 2024). Cloud Market Share Report.

Strong

The performance gap between proprietary and open-weight foundation models narrowed from approximately 8% to approximately 1.7% between 2023 and 2025.

Stanford HAI (2025). AI Index Report.

Strong

42% of companies that initiated AI projects in 2024 abandoned them in 2025.

S&P Global (2025). Enterprise AI Adoption Survey.

Strong

Large firms are adopting AI at 1.5-2x the rate of small and medium enterprises, compounding the competitive asymmetry over time.

BCG and McKinsey (2024-2025). AI Adoption Rate Studies.

Moderate

Current AI API pricing is below the fully-loaded cost of delivery, subsidised by venture capital and hyperscaler balance sheets.

SemiAnalysis (2024-2025). AI Infrastructure Cost Analysis.

Moderate

Fine-tuning an open-weight model for a business-specific use case costs between USD 100 and USD 5,000, depending on dataset size and compute requirements.

Stanford HAI (2025). AI Index Report; SemiAnalysis (2024). Fine-Tuning Cost Benchmarks.

Moderate

India has approximately 8,178 AI companies operating at a Builder-Dependent Ratio exceeding 1:100.

NASSCOM and OECD (2024-2025). AI Ecosystem Assessment; author's calculation.

Moderate

The United Kingdom has approximately 6,270+ AI companies operating at a Builder-Dependent Ratio of approximately 1:50.

UK Department for Science, Innovation and Technology (2024). AI Sector Report; UK ONS (2024).

Moderate

A representative AI tool stack for an independent business costs approximately USD 1,131 per year, denominated entirely in USD regardless of the business's operating currency.

Author's calculation based on published pricing from OpenAI, Anthropic, AWS, Azure, and supplementary toolchain providers (2025).

Strong

The USD 1,131 annual AI tool stack consumes approximately 1.5% of average US annual income but approximately 15-20% of average Brazilian annual income.

Author's calculation using IMF (2025) and World Bank (2024) GNI per capita and income distribution data.

Strong

US export control regulations directly restrict access to advanced AI chip architectures for firms in multiple study economies, creating indirect geopolitical access risk for all non-US economies.

CNAS (2024). Export Control Analysis; US Bureau of Industry and Security (2024-2025). Entity List Updates.

Strong

Independent businesses in Critical-EADI economies face simultaneous exposure to all five financial mechanisms with fewer institutional buffers than Source or High-EADI economies.

OECD (2024). AI Policy Observatory; World Bank (2024). Digital Development Report; author's synthesis.

Indicative

Eighteen primary data sources reviewed for this paper contained no cross-economy measure of income-adjusted AI infrastructure cost burden, confirming the gap that motivated the EADI framework.

Author's systematic review of Stanford HAI, IMF, Federal Reserve, OECD, World Bank, BCG, McKinsey, RAND, MIT NANDA, CNAS, US BLS, UK ONS, EY, PIIE, Synergy Research Group, SemiAnalysis, S&P Global, and OECD AAW reports (2024-2025).

Indicative

07Research maturity, validation status, and roadmap

Research Dashboard

Research Maturity

Evidence Strength

Analytical Confidence

Commercial Applicability

Scores out of 100. Based on internal research assessment criteria. Not independently validated.

Validation stage: Working Paper

Implementation status: Framework Published -- Quantitative Validation Ongoing

Known limitations

▪EADI scores are composite estimates based on publicly available data and have not been independently validated by a third-party methodology audit.

▪The representative USD 1,131 cost stack is a lower-bound estimate; actual AI infrastructure spend for an established independent business may be materially higher.

▪Income-adjusted cost calculations use national average income figures, which obscure significant intra-economy income distribution differences.

▪The five financial mechanisms are treated as additive by assumption; interaction effects have not been formally modelled.

▪The analysis covers ten economies; generalisation to the full global distribution of AI-dependent businesses requires further cross-economy validation.

▪Price trajectory risk (Mechanism 2) is forward-looking and depends on assumptions about the pace of AI infrastructure pricing normalisation that cannot be validated with current data.

Open questions

?What is the empirical relationship between EADI score and actual observed business failure rates in AI-intensive sectors?

?How quickly are open-weight models narrowing switching costs at the cloud and data pipeline layers, not just the model layer?

?Does the Specialised Digital Asset hierarchy deliver measurable survivability benefit at all five levels, or do higher levels require capital thresholds inaccessible to most independents?

?How do policy interventions (sovereign compute, subsidy programmes) interact with firm-level strategic responses in practice?

?What is the distribution of Builder-Dependent Ratios across economies not included in the current ten-economy study?

Research roadmap

Initial framework development and literature review

EADI methodology design and scoring across ten economies

Five-mechanism framework articulation

Specialised Digital Asset hierarchy development

Working paper drafting and internal review

EAD-2026-01 working paper publication

Founder survey to validate income-adjusted cost estimates (500 founders, 10 economies)

Independent methodology audit of EADI scoring weights

Longitudinal tracking of open-weight convergence and switching cost dynamics

Expanded 30-economy EADI scoring study

08What this research means for different stakeholders

Commercial Implications

As an independent founder, your AI infrastructure spend is not a neutral operating cost -- it is a structural survivability variable. If you are operating in a Critical-EADI economy, the same USD you spend on AI infrastructure costs you proportionally more in income terms than your counterpart in a Source economy, and every AI project failure costs you proportionally more in capital terms. Understanding EAD turns an invisible structural risk into a manageable strategic variable that can be incorporated into financial planning, build-versus-buy decisions, and runway calculations.

Opportunities

Use the Specialised Digital Asset hierarchy as a strategic roadmap: prioritise building Level 3-5 assets (proprietary datasets, fine-tuned models, domain products) that are not directly substitutable by API access alone.
Model income-adjusted AI costs into your unit economics from day one, using EADI-adjusted cost projections rather than nominal USD figures.
Evaluate open-weight model alternatives for use cases where the ~1.7% performance gap is acceptable -- fine-tuning costs (USD 100-5,000) may be lower than the long-term cost of proprietary platform lock-in.
Treat AI infrastructure spend as a make-versus-buy decision with five-year cost trajectory projections, not a fixed-cost line item in a static model.
Build data pipelines that are infrastructure-agnostic from the outset, reducing switching costs at the cloud compute layer before lock-in compounds.

Risks

Underestimating the income-adjusted burden of AI infrastructure costs relative to your operating currency and revenue base.
Locking into proprietary foundation model APIs before evaluating open-weight alternatives, creating switching costs that compound over time.
Treating current AI API pricing as a stable input into financial models -- current pricing is below cost of delivery and is likely to increase.
Building in a high-EADI market and assuming the same unit economics will hold when expanding operations to Critical-EADI economies.
Investing in AI project development without accounting for the 80%+ failure rate in financial projections and runway planning.

Questions to ask

What is the income-adjusted cost of my current AI infrastructure stack as a percentage of my business's annual revenue?

At what level of the Specialised Digital Asset hierarchy do my current AI investments sit, and what is the path to Level 4-5 assets?

If my primary AI provider raised prices by 3x in 2027, what would be the runway impact and what are my switching options today?

Am I building AI capital (proprietary data and models) or only AI consumption (API access), and what is the ratio between the two?

09Limitations, alternative perspectives, and validation gaps

Critical Review

Research Limitations

EADI Scoring Methodology

The EADI is a composite index constructed from publicly available data using the author's weighting methodology. It has not been independently validated by a third-party methodology audit. Different weighting schemes for the three infrastructure layers and the policy environment component could produce materially different scores, particularly for economies in the Moderate and High dependency tiers where the boundaries are less clearly defined by structural characteristics.

Representative Stack Assumptions

The USD 1,131 annual tool stack is a deliberate lower-bound estimate based on publicly advertised pricing for individual subscriptions. Enterprise pricing, bundled services, and regional pricing adjustments (where they exist) are not reflected. Actual AI infrastructure spend for an established independent business in production is likely higher and may vary significantly by sector, use case, and inference volume.

Income-Adjustment Methodology

Income-adjusted cost calculations use national average income figures from IMF and World Bank sources. These averages obscure significant intra-economy income distribution differences. The relevant comparison for an independent founder in São Paulo is not Brazil's national average income but the income distribution of the urban professional class in which AI-intensive founders are concentrated, which is higher than the national average.

Mechanism Interaction Effects

The five financial mechanisms are treated as additive by assumption. In practice, interaction effects -- particularly between Currency Asymmetry (Mechanism 1) and Project Failure Amplification (Mechanism 5) -- may produce non-linear survivability impacts that the current framework does not formally model. A business in Brazil that fails an AI project absorbs the currency-amplified infrastructure cost of that failure, which is a compound effect not captured in additive treatment.

Ten-Economy Sample Scope

The analysis covers ten economies selected to represent the range of income levels and AI infrastructure sovereignty positions. However, the selection is not fully representative of the global distribution of AI-dependent businesses. Africa, Southeast Asia, and Latin America beyond Brazil are not included and may show materially different EAD dynamics, including higher income-adjusted cost burdens and lower institutional buffer capacity than any of the current study economies.

Alternative Perspectives

Claim: Standard AI risk frameworks already cover infrastructure dependency through vendor lock-in and single-source-of-failure analyses.

Response: Standard vendor lock-in frameworks address the operational risk of provider failure or service discontinuation. They do not address the structural income-adjusted cost asymmetry that is the core finding of this paper. The EAD framework is not a substitute for operational risk management -- it is an additional dimension that existing frameworks do not capture and that is particularly relevant for businesses operating across economies with different income levels.

Claim: Open-weight model convergence will eliminate EAD-related risk within 2-3 years as performance gaps close.

Response: Open-weight convergence narrows the performance gap at the foundation model layer, which mitigates Mechanism 4 (Platform Switching Costs) at that specific layer only. The cloud compute layer (~63% US-concentrated) and chip layer (Nvidia/TSMC/export controls) are structurally unaffected by open-weight model convergence. Currency asymmetry (Mechanism 1) and price trajectory risk (Mechanism 2) are also entirely independent of model-layer performance gaps.

Claim: AI pricing will decrease over time as compute costs fall, making the income-adjusted cost burden a temporary condition.

Response: While the cost of compute per FLOP has historically declined, current AI API pricing is below cost of delivery (SemiAnalysis 2024-2025), making further nominal price decreases unlikely in the near term. Future pricing is more likely to normalise upward toward cost-covering levels than to continue declining from an already-subsidised baseline. Even if nominal pricing decreases, USD denomination means exchange rate movements can offset compute cost declines for businesses in non-USD economies.

Claim: The analysis conflates infrastructure dependency with infrastructure cost, which are separate variables.

Response: The paper deliberately examines the financial manifestation of dependency, not dependency in the abstract. Dependency without a cost mechanism would not generate survivability risk. The five mechanisms are precisely the channels through which structural dependency translates into financial outcomes. The EADI score is a composite measure of structural dependency that determines exposure to the mechanisms; the mechanisms themselves are the cost variables.

Validation Gaps

No empirical validation of the relationship between EADI score and actual observed business failure rates in AI-intensive sectors.
No founder survey data to validate income-adjusted cost estimates against actual AI infrastructure expenditure patterns.
No longitudinal data to validate price trajectory risk projections (Mechanism 2) against observed pricing changes over time.
No independent methodology audit of EADI scoring weights and infrastructure layer component selection.
No controlled comparison of Specialised Digital Asset hierarchy levels against survivability outcomes in matched business samples.

Future Work

Commission a 500-founder survey across the ten study economies to validate income-adjusted cost estimates and collect actual AI infrastructure expenditure data.
Develop a 30-economy EADI scoring extension to improve global representativeness, with particular focus on African, Southeast Asian, and Latin American economies.
Build a formal interaction-effect model for the five financial mechanisms to replace the current additive assumption with empirically grounded estimates.
Engage with OECD and World Bank to develop a standardised cross-economy AI infrastructure cost burden indicator for incorporation into existing digital development frameworks.
Conduct longitudinal price tracking of the representative USD 1,131 stack to test price trajectory risk projections against observed market changes.
Develop an open-source EADI calculator that allows founders and investors to compute EADI-adjusted unit economics for their specific use cases and operating economies.

10Key terms and definitions used in this research

Glossary

EAD

External AI Dependence. The structural condition of an independent business whose core AI infrastructure -- foundation models, cloud compute, and AI chips -- is owned and controlled by entities outside its operating jurisdiction.

EADI

External AI Dependence Index. A composite score (0-100) measuring an economy's structural position across the three AI infrastructure layers, Builder-Dependent Ratio, and policy environment for AI infrastructure sovereignty. Higher scores indicate lower dependency. Source economies score near 100; Critical economies score below 20.

Builder-Dependent Ratio

The ratio of AI-dependent businesses to AI-builder businesses within an economy. A high ratio (e.g., 1:100+ for India) indicates that the vast majority of AI-active businesses in the economy are structurally dependent on externally owned infrastructure.

Currency-Denominated Cost Asymmetry

The tenfold difference in income-adjusted AI infrastructure costs between high-income Source economies (e.g., US at ~1.5% of avg income) and lower-income Critical-EADI economies (e.g., Brazil at ~15-20% of avg income) for identical nominal USD expenditure.

Price Trajectory Risk

The forward-looking risk that current AI API pricing -- which is below the fully-loaded cost of delivery -- will normalise upward, with independent businesses absorbing the full increase due to lack of volume discount leverage.

Competitive Asymmetry

The structural AI adoption rate and cost-per-output gap between large enterprises -- which adopt AI at 1.5-2x the rate of SMEs and access volume discounts -- and independent businesses operating without negotiated enterprise agreements.

Platform Switching Costs

The costs incurred when migrating from one AI infrastructure provider to another, including data pipeline migration, fine-tuning reinvestment (USD 100-5,000), prompt engineering re-development, and integration testing.

AI Project Failure Amplification

The compounding survivability impact of high AI project failure rates (80%+) on businesses with limited capital buffers, for whom a failed AI project may represent an existential capital event rather than a manageable write-off.

Specialised Digital Asset Hierarchy

A five-level hierarchy of AI-related assets ranked by strategic value and insulation from EAD mechanisms: Level 1 (Reusable Workflows), Level 2 (Structured Proprietary Knowledge), Level 3 (Industry-Specific Datasets), Level 4 (Business-Specific AI Systems), Level 5 (Specialised Domain Products).

Foundation Model Layer

The first of three AI infrastructure dependency layers. Encompasses frontier large language models and multimodal models provided by OpenAI, Anthropic, Google, Meta (US) and Baidu, Alibaba, DeepSeek (China). Access for all non-Source economies is exclusively through USD-priced APIs.

Cloud Compute Layer

The second of three AI infrastructure dependency layers. Encompasses inference, storage, and deployment infrastructure. AWS (~28%), Azure (~21%), and GCP (~14%) control approximately 63% of global market share.

AI Chip Layer

The third of three AI infrastructure dependency layers. Encompasses GPU design (Nvidia-dominant) and advanced chip fabrication (TSMC-dominant). Subject to US export control regimes that create a hard ceiling on chip-layer independence for non-US economies.

Source Economy

An economy that possesses domestically developed foundation models at the frontier level. In the ten-economy study, only the United States (EADI 92) and China (EADI 64) qualify as Source economies.

Critical-EADI Economy

An economy with an EADI score below 20, indicating near-total dependence on externally owned AI infrastructure across all three layers. In the ten-economy study: Canada (19), United Kingdom (17), India (12), Brazil (7).

Income-Adjusted Cost

AI infrastructure expenditure expressed as a percentage of the average annual income in the relevant economy. Used throughout the paper as the primary measure of financial burden, in contrast to nominal USD expenditure which obscures cross-economy asymmetry.

11How this research has evolved

Research Timeline

2026-05-01

Research initiated -- infrastructure mapping and literature review begun across 18 primary sources

2026-05-12

Eighteen primary data sources identified, catalogued, and systematically reviewed

2026-05-20

Three-layer AI infrastructure dependency taxonomy established (Foundation Models, Cloud Compute, AI Chips)

2026-06-01

EADI methodology designed and scoring applied across all ten study economies

2026-06-10

Five financial mechanisms identified and characterised through cross-source validation

2026-06-18

Specialised Digital Asset hierarchy developed as firm-level strategic response framework

2026-06-25

Working paper drafted; internal review and methodology checks completed

2026-07-05

EAD-2026-01 published as triNetra working paper

2026-09-01

Founder survey launch -- 500 founders across ten study economies to validate income-adjusted cost estimates

2026-12-01

Expanded 30-economy EADI scoring study and longitudinal price tracking report

12Download, cite, and track version history

Original Paper

Download

↓ Download Working Paper (DOCX)

Cite this research

Agarwal, S. (2026). External AI dependence and startup financial survivability (Working Paper EAD-2026-01). triNetra.

Version history

1.02026-07-05Initial publication. Ten-economy EADI analysis; five-mechanism framework; Specialised Digital Asset hierarchy; 18 primary data sources reviewed; Builder-Dependent Ratio introduced.

EAD-2026-01 establishes the individual-level structural cost asymmetry and proposes the Specialised Digital Asset hierarchy as the firm-level response. EAD-2026-02 examines the institutional architecture -- the Judgment Layer -- that is missing from organisations navigating this structural vulnerability.

Executive Summary

Who this research serves

The Problem

Research Question

PaaF Methodology Journey

Interactive Framework

Evidence Explorer

Research Dashboard

Commercial Implications

Critical Review

Glossary

Research Timeline

Original Paper

Download

Cite this research

Patterns as a Framework (PaaF)

How We Work

Who We Serve

Eagle Insight Platform – Structural Intelligence Engine

Eagle Insight Platform Documentation: Eagle Framework and RiskOpsBench 1.0 Reference

About triNetra

Research Overview

Frequently Asked Questions

Current Validation Stage

Contact triNetra Research