AI Data Governance Infrastructure: What Bezos Got Right About the Horizontal Layer
- May 6
- 16 min read
Abstract
Jeff Bezos has repeatedly characterised artificial intelligence as "a horizontal enabling layer," drawing an explicit analogy to electricity. This framing maps precisely onto General Purpose Technology (GPT) theory, first formalised by Bresnahan and Trajtenberg (1995). However, applying the GPT framework to AI requires more than historical pattern-matching. AI differs structurally from prior GPTs in two critical respects: its software-native diffusion velocity, and the emergence of agentic multi-agent systems that could compress the complementarity lag that defined earlier technology transitions. This article examines AI through the GPT framework while accounting for these structural differences. Drawing on Brynjolfsson's recent productivity estimates, Acemoglu's task-based framework, Aghion and Bunel's growth models, and Jones and Tonetti's work on data nonrivalry, it argues that the most durable value will accrue not to the capability layer but to the governance infrastructure that makes the horizontal layer safe for enterprise deployment. That governance layer is itself more complex than any previous GPT infrastructure, encompassing data control, energy policy, export controls, and the coordination of autonomous agents operating across organisational boundaries.
1. Introduction
In December 2024, at the New York Times DealBook Summit, Jeff Bezos offered a characterisation of AI that has since become one of the most cited framings in technology investment discourse: "Modern AI is a horizontal enabling layer. It can be used to improve everything. It will be in everything. This is most like electricity."
The statement was not new. Bezos had used nearly identical language in 2017, telling the Internet Association that "machine learning and AI is a horizontal enabling layer" that would "empower and improve every business, every government organization, every philanthropy." The consistency of the framing across seven years suggests a considered structural view rather than a passing analogy.
What followed the 2024 iteration was a wave of popular commentary that took the concept and extrapolated it into claims that the underlying economic theory does not support: that horizontal layers "price everything to zero," that professional expertise would "evaporate," and that every company selling AI as a product would be rendered worthless. These claims misrepresent both the historical record and the economics of general purpose technologies.
This article attempts to bridge the gap between Bezos's framing (which is, I think, structurally correct) and the analytical literature that gives it substance. However, it also addresses a limitation of the pure historical analogy. AI is not electricity. The differences between software-native and physical-infrastructure GPTs are material, and the emergence of agentic AI introduces a dynamic that the canonical GPT literature did not anticipate. The aim is to identify what GPT theory predicts, where AI diverges from the pattern, and what new infrastructure requirements the horizontal layer thesis implies.
2. General Purpose Technology Theory: The Analytical Framework
The concept Bezos describes has a formal name in economics: a General Purpose Technology, or GPT, a term introduced by Bresnahan and Trajtenberg in their 1995 paper "General Purpose Technologies: Engines of Growth?" published in the Journal of Econometrics.
Bresnahan and Trajtenberg identified three defining characteristics of a GPT:
Pervasiveness. The technology is used as an input by many downstream sectors, not confined to a single industry or application domain.
Inherent potential for technical improvement. The technology improves substantially over time, sustaining and amplifying its economic impact across adoption cycles.
Innovational complementarities. The productivity of R&D in downstream sectors increases as a consequence of innovation in the GPT. The technology enables the creation of new processes, products, and organisational forms that could not have existed without it.
The canonical examples in the literature are the steam engine, the electric motor, and semiconductors. Each exhibited all three characteristics. Each produced economy-wide restructuring that took decades to fully manifest.
AI clearly satisfies all three criteria. Large language models, computer vision systems, and generative AI tools are finding applications across virtually every sector. Foundation models are improving rapidly, with capability doublings measured in months rather than years. And the downstream innovation they enable, from drug discovery to code generation to legal analysis, is already visible.
The critical insight of GPT theory, and the one most frequently omitted from popular accounts, is that the relationship between the GPT and its application sectors is characterised by what Bresnahan and Trajtenberg termed "increasing returns to scale." This creates a virtuous cycle, but one that a decentralised economy may struggle to exploit efficiently. Arms-length market transactions between the GPT and its users, the authors argued, could result in "too little, too late" innovation in both sectors.
This is a different picture from the popular narrative of inevitable, rapid disruption. GPT theory predicts powerful but delayed economic transformation, mediated by complementary investment and organisational adaptation.
The question is whether the delays that characterised physical GPTs will apply with the same force to a software-native one.
3. The Productivity Paradox: What Electricity Teaches Us, and Where the Analogy Breaks
3.1 The Historical Pattern
Paul David's influential 1990 paper, "The Dynamo and the Computer," documented a striking fact. Although the electric motor was commercially available in the 1880s, measurable productivity gains from electrification did not appear in aggregate statistics until the 1920s, a lag of roughly thirty to forty years.
The reason was that early adopters overlaid the new technology onto existing organisational structures. Factories designed around centralised steam power replaced the steam engine with an electric motor but kept the same layout and workflow. Transformative gains only emerged when a new generation of managers redesigned plants around "unit drive," with individual electric motors powering each piece of equipment. This enabled fundamentally different factory layouts organised around material flow rather than proximity to a central power source.
Brynjolfsson, Rock, and Syverson formalised this insight in their 2021 paper "The Productivity J-Curve" (American Economic Journal: Macroeconomics). They demonstrated that GPTs require substantial complementary investments, including process redesign, workforce retraining, and organisational restructuring, that are poorly measured in national accounts. In the early years of adoption, these intangible investments depress measured productivity. Only later does productivity accelerate, producing a J-shaped curve. Their empirical analysis found that adjusting for unmeasured intangibles yielded a total factor productivity level 15.9% higher than official measures by the end of 2017.
3.2 Where the Analogy Weakens
The electricity parallel is instructive but imperfect, and overstating the continuity between physical and software-native GPTs risks misallocating capital and attention.
Two structural differences warrant careful examination.
Diffusion velocity. Electricity required physical rewiring: new generators, new transmission infrastructure, new factory layouts. The capital expenditure was enormous and the deployment timeline was measured in decades. AI is software-first and cloud-delivered. Adoption can and does occur in weeks. ChatGPT reached 100 million users in two months, a diffusion rate roughly double that of the internet over a comparable period following its introduction. This does not eliminate the complementarity bottleneck, but it compresses the timeline over which the bottleneck operates.
The shrinking complementarity gap. David's key insight was that electricity's productivity benefits were delayed because organisations needed to co-invent new processes alongside the technology. This remains true for AI, but the co-invention burden is arguably lighter. Foundation models arrive pre-trained on vast corpora and can be fine-tuned or prompted with far less organisational restructuring than was required to redesign a factory floor. The "unit drive" equivalent for AI, embedding models into specific workflows, is a software configuration problem rather than a physical engineering one. This does not mean the organisational adaptation is trivial; McKinsey research suggests that companies treating AI as a bolt-on tool rather than a trigger for process redesign consistently fail to scale. But the adaptation cycle is faster.
3.3 Emerging Empirical Evidence
Recent data suggests the J-curve may be steeper and shorter for software-based GPTs than the historical pattern would predict.
Writing in the Financial Times in February 2026, Brynjolfsson estimated US productivity growth at approximately 2.7% for 2025, nearly double the 1.4% annual average of the prior decade. His analysis drew on revised Bureau of Labor Statistics data showing total payroll growth revised downward by approximately 403,000 jobs while real GDP remained robust at 3.7% in Q4.
Brynjolfsson characterised the combination of high output with lower labour input as "the hallmark of productivity growth" and argued that the US is "transitioning out of this investment phase into a harvest phase."
These findings should be treated with appropriate caution. Productivity data is noisy, short-term readings are susceptible to statistical revisions, and causation is notoriously difficult to establish at the macro level. As several observers have noted, a portion of the GDP growth Brynjolfsson cites could be driven by massive capital expenditure on AI infrastructure itself rather than by productivity gains from using AI. Jason Furman's parallel analysis arrived at a somewhat lower but still elevated figure of 2.2% over a six-year peak-to-peak measurement.
The broader academic landscape shows significant dispersion in forecasts. Acemoglu (2024), using his task-based framework, estimates an upper-bound impact of approximately 0.66% TFP over ten years, a figure that has been influential but is at the conservative end of the spectrum.
Aghion and Bunel (2024), applying the same task-based methodology but with different assumptions about AI exposure and cost savings, estimate productivity growth increases of 0.8 to 1.3 percentage points per year over the next decade. Their analysis draws on historical parallels with the electricity wave of the 1920s and the digital technology wave of the late 1990s. Jones and Tonetti's work on data nonrivalry (2020) and their more recent growth model (2026) suggest that once new-task creation and scientific-acceleration channels are included, the effects could be substantially larger than either Acemoglu or Aghion's frameworks predict, with automation projected to accelerate economic growth over the coming decades.
The honest assessment is that we do not yet know which end of this range will prove correct. What the emerging data does suggest is that assuming a multi-decade productivity lag analogous to electricity may overstate the continuity between physical and software-native GPTs.
4. The Complexity Question: Automation, Augmentation, and the Task-Based Framework
One of the more persistent claims in popular AI commentary is that horizontal enabling layers dissolve complexity. In this framing, expertise in law, medicine, and finance is characterised as navigational friction that AI eliminates entirely.
Acemoglu and Restrepo's task-based framework (2018, 2019) provides a more rigorous analytical structure. In their model, production is decomposed into discrete tasks, some of which can be automated and some of which remain the domain of human labour. Automation operates through two countervailing channels:
The displacement effect. Automation replaces human labour in specific tasks, reducing demand for associated skills and exerting downward pressure on wages.
The productivity effect. Cost savings from automation increase overall output, raising demand for labour in non-automated tasks. This is complemented by the creation of entirely new tasks that could not have existed before the automating technology was available.
The net effect depends on the relative magnitude of these forces, and critically, on whether the technology is deployed primarily to automate existing tasks or to augment human capability in ways that create new categories of work.
Acemoglu's 2024 paper, "The Simple Macroeconomics of AI," reached notably more conservative conclusions than popular commentary would suggest, estimating modest medium-term GDP increments. His model has been rightly influential, but it may underweight two channels: the creation of genuinely new tasks (which Acemoglu himself acknowledges he does not focus on), and the acceleration of scientific discovery that several empirical studies are beginning to document.
The task-based framework nonetheless provides an important corrective to the "expertise evaporates" narrative. Complexity in regulated industries is not primarily navigational friction amenable to automation. Much of it is irreducible: context-dependent, novel in its specific configurations, and carrying consequences where errors are not easily reversible. AI will compress the pattern-matching components. The judgment, accountability, and contextual reasoning that constitute the difficult residual may actually increase in value as the routine components become cheaper.
The more precise prediction is: routine expertise compresses, while contextual judgment reprices upward.
5. The Agentic Phase Shift: A GPT Within a GPT
The analysis thus far has treated AI primarily as a tool that augments or automates discrete tasks within existing organisational structures. This framing, while useful, may already be insufficient.
The emergence of agentic and multi-agent AI systems, visible in research since 2024 and accelerating through 2025 and 2026, represents what could be described as a GPT-within-a-GPT dynamic: a qualitative shift within the broader horizontal layer that the canonical GPT literature does not fully capture.
Current large language models are, in essence, reasoning engines that process data when prompted. Agentic systems go further. They exhibit goal-directed behaviour, adaptive learning, and the capacity to orchestrate multi-step processes. Multi-agent architectures take this further still: specialised agents coordinating with each other across functional and organisational boundaries, with minimal human oversight of individual steps.
Gartner has reported a 1,445% increase in multi-agent system inquiries between Q1 2024 and Q2 2025, and projects that 40% of enterprise applications will embed task-specific AI agents by the end of 2026. Salesforce/Deloitte research (2025) found that organisations currently deploy an average of 12 agents, with that number projected to grow 67% within two years.
The significance for GPT theory is that agentic systems could compress the complementarity lag dramatically. David's electricity research showed that the bottleneck was organisational redesign: humans needed decades to figure out how to restructure processes around cheap power. Multi-agent systems could, in principle, design and implement organisational adaptations themselves, creating emergent workflow configurations that no single human team would have conceived.
This is speculative, and caution is warranted. Gartner also predicts that over 40% of agentic AI projects will fail by 2027, largely because legacy systems lack the architecture for true agentic integration. Fewer than one in four organisations have successfully scaled agents to production. The gap between experimental deployment and reliable, governed operation remains large.
But the directional implication matters for capital allocation. If the complementarity lag that defined every prior GPT era is itself being automated, then forecasts based on historical lag patterns may substantially underestimate the speed of the transition. The J-curve could be not merely steeper but qualitatively different in shape.
The governance implications are equally significant. When autonomous agents operate across organisational boundaries, coordinating actions that affect data flows, contractual obligations, and regulatory compliance, the "control plane" challenge becomes orders of magnitude more complex than managing human users of AI tools. The question is no longer merely "what data can the model access?" but "what can autonomous agents do with data across multiple systems, organisations, and jurisdictions, without continuous human oversight?"
6. The Control Plane: Why AI Data Governance Infrastructure Is the Durable Moat
6.1 The Historical Pattern
Every horizontal enabling layer generates infrastructure requirements that did not exist before its arrival. Electrification required grid management, safety regulation, metering, and standardised voltage. The internet required DNS, SSL/TLS, content delivery networks, and an entire cybersecurity industry. Cloud computing required identity management, API governance, and data sovereignty frameworks.
The companies that occupied these infrastructure positions were rarely the most visible players in their respective technology eras but were among the most durable. Infrastructure positions are defensible precisely because they become embedded in the operational fabric of every ecosystem participant.
6.2 AI's Governance Layer Is Structurally More Complex
For AI, the governance challenge is arguably more complex than any previous horizontal shift, for reasons that extend well beyond enterprise data boundaries.
The substrate reasons with data. Previous horizontal layers transported, stored, or processed information. AI reasons with it. A system that extracts patterns, synthesises outputs, and in some architectures retains learned information indefinitely, creates a governance problem qualitatively different from one that merely moves or stores data. The question of how enterprises manage the boundary between their sensitive information and an AI substrate that learns from that information is not a compliance checkbox. It is infrastructure.
This is the structural gap I have written about elsewhere as the Prompt Gap: the space between what an enterprise wants to ask an AI system and the unresolved data exposure that the question creates.
Energy and compute infrastructure is geopolitically intensive. The physical infrastructure required for AI operates at a scale and geopolitical complexity that has no real precedent in prior GPT eras. Training a single large model can consume as much electricity annually as 200 average US households. AWS has committed up to $50 billion in AI data-centre capacity for US government agencies alone. Google has announced capital expenditure plans of $175-185 billion for 2026, far exceeding analyst expectations.
Bezos himself, speaking at the 2026 DealBook Summit, warned that firms investing in private computing facilities are repeating a historical error from the pre-grid era. He recalled visiting a 300-year-old brewery in Luxembourg that once housed its own electric generator; once centralised power grids became available, the private plant became an expensive burden. The technology sector, Bezos argued, is in a similar "generator phase," and energy availability is the ultimate bottleneck for AI growth.
Yet many enterprises are pursuing private infrastructure precisely for data sovereignty reasons, creating a tension between the efficiency logic of centralised compute grids and the control logic of keeping sensitive data off shared infrastructure. This tension does not resolve itself. It requires a governance architecture that allows enterprises to use centralised compute without surrendering control of their data, a problem that the electricity era never faced because electrons are fungible and data is not.
Export controls and supply chain concentration. The semiconductor supply chain underpinning AI compute is concentrated in ways that create fragility and geopolitical leverage. Advanced chip manufacturing is effectively limited to a single company (TSMC), located in a geopolitically sensitive region. US export controls on advanced AI chips to China, tightened repeatedly since 2022, have introduced a policy dimension to AI infrastructure that has no analogue in the electricity or internet eras. The "control plane" for AI is not merely a technical governance layer; it is entangled with industrial policy, trade restrictions, and national security considerations.
Open-source model weights diffuse governance challenges beyond enterprise boundaries. The availability of open-weight foundation models (Llama, Mistral, and others) means that capable AI systems operate well beyond the perimeter of any single enterprise's governance framework. When a model is downloaded, fine-tuned, and deployed independently, the data governance assumptions embedded in the original training pipeline no longer hold. The proliferation of open models creates a surface area for governance failures that enterprise-focused control architectures, however well-designed, cannot fully address.
Multi-agent coordination creates emergent risk. As discussed in Section 5, autonomous agents operating across organisational boundaries introduce governance challenges that are combinatorial in nature. An agent authorised to access data in System A and execute actions in System B may create exposure that neither system's individual governance framework anticipated. The interaction effects multiply as the number of agents and the number of organisational boundaries increase. This is a coordination problem, not merely an access-control problem, and it requires governance primitives that do not yet exist at scale.
6.3 The Value of the Governor Layer
If the historical GPT pattern holds, the organisations and technologies that solve these governance problems effectively will occupy positions of substantial and durable strategic value. The more horizontal AI becomes, the more critical the control plane becomes, because the attack surface, the regulatory surface, and the liability surface all scale with pervasiveness.
Jones and Tonetti's work on data nonrivalry (2020) provides an additional theoretical anchor. Their model demonstrates that broad data access generates social gains, but firms are incentivised to hoard proprietary data to mitigate creative destruction, leading to inefficient use of a nonrival resource. The resolution they propose, giving data property rights to consumers, requires precisely the kind of governance infrastructure that is currently missing: systems that enable data to flow to where it is productive while maintaining enforceable controls over its use.
The control plane for AI is not a single product or technology. It is an infrastructure layer encompassing data governance within the AI substrate, energy and compute allocation across public and private infrastructure, regulatory compliance across jurisdictions, provenance and accountability for autonomous agent actions, and the coordination of multi-agent systems operating at enterprise boundaries. Companies that solve one slice of this problem will be useful. Those that solve the integration problem across these dimensions will be, I think, genuinely consequential.
7. Implications for Capital Allocation
If the analysis above is broadly correct, several implications follow.
The J-Curve may be shorter than the electricity precedent suggests, but the organisational bottleneck is real. Brynjolfsson's early harvest-phase data is encouraging, and the software-native character of AI compresses diffusion timelines. But the gap between experimental adoption and scaled operational deployment remains large, as the agentic AI failure rates suggest. Investors should expect faster returns than the electricity precedent implied but should not assume the organisational co-invention problem has been solved.
Product-layer AI is likely to commoditise. The GPT framework suggests that companies selling AI as a standalone product are building on a commodity. As the horizontal layer matures and capability becomes abundant, differentiation at the product layer will become increasingly difficult to sustain. This does not mean these businesses will fail in the short term, but it suggests their long-term value is likely to erode absent a defensible position in infrastructure, vertical depth, or governance.
Infrastructure and governance positions are undervalued. The market appears to be allocating capital primarily to the capability layer (foundation models, AI applications) and substantially underweighting the infrastructure layer (data governance, energy, compute orchestration, agent coordination). If the historical GPT pattern repeats, and the structural arguments above suggest it should, this represents a significant misallocation.
The governance challenge is broader than enterprise data control. Capital allocation strategies focused solely on enterprise data governance will miss the energy dimension, the geopolitical dimension, and the multi-agent coordination dimension. The full control plane is more expansive than what most current AI governance frameworks address.
The agentic transition could be the most consequential variable. If multi-agent systems compress the complementarity lag, the timeline for value redistribution across the AI stack accelerates. Early positioning in the governance layer becomes more urgent, not less, as the pace of deployment increases.
8. Conclusion
Bezos's characterisation of AI as a horizontal enabling layer is not merely a useful metaphor. It maps onto a body of economic research, General Purpose Technology theory, that has been developed over three decades. When that research is applied to the current AI transition, with appropriate adjustments for AI's structural differences from physical GPTs, the picture that emerges is both more nuanced and more consequential than popular commentary suggests.
The nuance: AI satisfies the GPT criteria but differs from prior GPTs in diffusion velocity, complementarity dynamics, and the emergence of agentic systems that could automate the organisational co-invention that historically defined the productivity lag. Early empirical data, while noisy, suggests the J-curve may be steeper and shorter than historical precedent implied. The range of credible productivity forecasts remains wide, from Acemoglu's conservative estimates to the larger effects predicted by Aghion and Bunel and by Jones and Tonetti's growth models.
The consequence: the governance infrastructure required for AI is structurally more complex than for any previous horizontal layer. It encompasses data control within the AI substrate, energy and compute allocation, regulatory compliance, export controls, open-model proliferation, and the coordination of autonomous agents operating across organisational and jurisdictional boundaries. This is not a single product category. It is an infrastructure layer, and the historical record suggests that infrastructure layers generate the most durable value in any GPT era.
The companies and investors who are currently most visible in the AI landscape, those building and selling the capability itself, may not be the ones who ultimately capture the most durable value. If the GPT framework is any guide, that value will accrue to whoever builds the governance infrastructure that makes the horizontal layer safe for enterprise deployment on the world's most sensitive data and most consequential decisions.
Every grid needs a governor. The question, as it has been in every previous technology transition, is who builds it, and how quickly.
References
Acemoglu, D. (2024). "The Simple Macroeconomics of AI." NBER Working Paper No. 32487.
Acemoglu, D. and Restrepo, P. (2018). "Artificial Intelligence, Automation and Work." NBER Working Paper No. 24196.
Acemoglu, D. and Restrepo, P. (2019). "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives, 33(2), pp. 3-30.
Aghion, P. and Bunel, S. (2024). "AI and Growth: Where Do We Stand?" Federal Reserve Bank of San Francisco Working Paper.
Aghion, P., Jones, B., and Jones, C. (2019). "Artificial Intelligence and Economic Growth." In The Economics of Artificial Intelligence: An Agenda, pp. 237-290. University of Chicago Press.
Bresnahan, T.F. (2010). "General Purpose Technologies." In Hall, B.H. and Rosenberg, N. (eds.), Handbook of the Economics of Innovation, Vol. 2, pp. 761-791. Elsevier.
Bresnahan, T.F. and Trajtenberg, M. (1995). "General Purpose Technologies: Engines of Growth?" Journal of Econometrics, 65(1), pp. 83-108.
Brynjolfsson, E. (2026). "The AI Productivity Take-off Is Finally Visible." Financial Times, February.
Brynjolfsson, E., Rock, D., and Syverson, C. (2018). "Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics." In Agrawal, A., Gans, J., and Goldfarb, A. (eds.), The Economics of Artificial Intelligence: An Agenda, pp. 23-57. University of Chicago Press.
Brynjolfsson, E., Rock, D., and Syverson, C. (2021). "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies." American Economic Journal: Macroeconomics, 13(1), pp. 333-372.
David, P.A. (1990). "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox." American Economic Review, 80(2), pp. 355-361.
Jones, C.I. and Tonetti, C. (2020). "Nonrivalry and the Economics of Data." American Economic Review, 110(9), pp. 2819-2858.
Jones, C.I. and Tonetti, C. (2026). "Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion." NBER Working Paper No. 34779.
Jovanovic, B. and Rousseau, P.L. (2005). "General Purpose Technologies." In Aghion, P. and Durlauf, S. (eds.), Handbook of Economic Growth, Vol. 1, Chapter 18, pp. 1181-1224. Elsevier.
McElheran, K., Yang, M-J., Kroff, Z., and Brynjolfsson, E. (2025). "The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s)." US Census Bureau Working Paper
Comments