Thoughts on AI Buildout 1
Originally written 11.13.2025
Can the AI build-out continue?
Though the path forward will inevitably include fits and starts, interconnected forces will sustain the political, commercial, and cultural will for the massive capital expenditure necessary. This essay outlines the strategic and commercial dynamics that ensure this transformation will persist and accelerate, and touches on societal implications for what may happen after the buildout.
The Rational Imperative
From a strategic American standpoint, the massive AI buildout is not optional. It is necessary for compounding US power. Its continuity has three imperatives. The American administration, corporate, and technology sectors must clearly see that, to maintain hard and soft power and economic strength, they must: maintain computational hegemony for commercial and military use, unlock multi-sector productivity gains, and secure global dominance of frontier AI models and promote open, democratic models.
The Computational Arms Race: Quantifying Non-Linear Demand
Non-linear computation demand drives the CapEx cycle necessary for advanced reasoning from perception. A recent statement by Anthropic claims the leap from classic perception-based AI (like image classification) to Generative AI required a 1000x increase in computation for model training, a surge so explosive that the energy needed for some transformer models grew by over 18,000x even while parameters grew just 4x. The subsequent transition to Agentic AI—autonomous systems that plan and execute multi-step reasoning tasks—imposes another dramatic resource hurdle. For inference alone, Agentic AI demands over 100x more compute than single-shot generative responses due to the need for continuous processing and the synthesis of hundreds or thousands of tokens per task. Capability gaps drive obsolescence and ensure compute capacity is constantly consumed, not oversupply.
The Economic Constraint
The colossal CapEx buildout ($500b and counting in 2025) will continue. The chief limiting factor currently is electrons, followed by premium GPU, then raw materials. Roughly 5% (200+ TWh/year) of the US power supply is used by data centers. That’s enough to run 19 million homes, or a mid-sized country like Poland. Some people anticipate this growth continuing at 15-20%/year, roughly ten times the historical rate of electricity demand growth. At a 15%/year growth rate, that usage will roughly double by 2030.
A traditional compute rack consumes 5-10 kW, equivalent to the electricity usage of 3-5 central air conditioners in a single large metal cabinet. An AI rack is 100-130kW, powering 60-75 standard refrigerators or powering 50 US homes simultaneously, and holds ~16-32 GPUs. These GPU racks generate extreme heat, and traditional airflow cooling is insufficient. The newer, more advanced chips draw more power and generate more heat. Often, they require liquid cooling to prevent hardware meltdown.
A new server rack consumes as much power as 50 average homes (10kW for a traditional rack). The current US power mix is roughly 43% natural gas, 17% renewable, 23% nuclear, and 17% coal. Gas and nuclear are reliable 24/7, so they are most likely to benefit from the energy buildout. The lead time for new power is 4-7 years to bring a transmission or substation online, so most new energy additions are going to come from brownfields with existing infrastructure and power connections.
New Nuclear takes 10-15 years, or 5-10 years for a new SMR (which has not been proven out), natural gas turbine plants take 3-5 years. Securing the gas pipeline capacity is often one of the biggest challenges here. Utility-scale solar and wind are quickest, but unreliable. To build out the electric grid immediately, the only option is to increase battery storage capacity to maintain a steady power flow. Underutilization could be addressed by increasing electricity storage capacity at scale. Sufficient battery storage capacity could expand the current grid, as most power plants operate at maximum capacity only during peak hours. The buildout is as much an industrial challenge as it is technological. An interesting REIT play would be one that can nimbly house people who need to move to these brownfield sites that are ramping up.
Hyperscalers undertaking massive capital expenditure for AI and compute are fundamentally different from traditional infrastructure firms like railroads or telecom companies. While railroads and telecom companies often carry debt ratios of 2x to 6x, hyperscalers typically maintain ratios of 0.0x to 1x. A key distinction is that these technology giants are highly diversified businesses with alternative revenue streams that help mitigate risk, and the government is often heavily incentivized to support this buildout from a national defense perspective.
The development of improved AI tools creates a beneficial cascade effect, improving their internal tools, margins, and other business segments. Furthermore, the useful life of a core asset like a GPU can extend far beyond its typical depreciation schedule by transitioning from high-intensity tasks (training) to lower-intensity ones (inference) down the value chain.
Ultimately, Jevons’ paradox is expected to drive even greater demand, as computing use will increase as the technology becomes more accessible. William Stanley Jevons, observed in 1865 that technological improvements in the efficiency of the steam engine did not reduce coal consumption. Instead, because the more efficient engine made steam power cheaper, more industries adopted steam engines, and the total demand for coal skyrocketed. Similarly, new and more complex uses of compute will emerge as prices fall.
The Measurement Misalignment
Let’s put aside the concern that the AI buildout is just OpenAI’s putative $13b/year in revenue. Another core perceptual issue is that economists like Daron Acemoglu warn against using the simple math of (Service Sector % growth) * (Productivity % growth) = GDP growth, saying that too much of the AI benefit leaks out before it ever reaches the national accounts. Acemoglu acknowledges that GDP is not a sufficient measure, it is true, but his projections make for more toothsome soundbites without that disclaimer.
Essentially, there is the potential for output substitution at the cost of output expansion. For example, let’s consider a lawyer. If a lawyer can do a task x% more quickly, does he really attract x% more client work, or does his increased performance actually create underutilization? Additionally, further slippage may occur when he spends time checking that the output is actually sound or building implementation protocols and infrastructure. That x% is not distributed evenly across all tasks and roles, so simpler, lower-value add jobs accrue less real benefit from AI. A paralegal gets more out of using an AI tool than does a top-level lawyer, and the top-level lawyer’s real economic value dwarfs the value created by the increased efficiency of the paralegal, so the argument goes.
As an aside, there is the moralistic and societally impactful argument that the value created will accrue to capital and infrastructure owners. There is no immediately apparent resolution to this, aside from the Pollyannaish claim for a universal basic income that blithely disregards human psychological and cultural incentive structures. At the moment, enormous value will accrue to those who can deploy AI tools most efficiently.
The significant measurement problem here is that if a generative AI saves me 10 hours on a task, those 10 hours are effectively lost by GDP reckoning. Increased productivity and efficiency (if my salary is pro rata hourly) do not factor in if a bank integrates an AI model that identifies fraud 90% faster and saves $500mm in losses. That saving does not count positively towards GDP.
By failing to account for this increase in productivity, GDP is not an ideal measure. Total Factor Productivity growth seems more relevant, but it is a lagging indicator that takes in all the negative costs before spitting out growth on the other side.
The Flawed Analogy to Telecom 2000
The primary historical comparison cited by skeptics is the early 2000s telecom bubble, in which hundreds of billions were poured into fiber-optic and data-center equipment, resulting in massive oversupply, capacity utilization dropping to 5% by 2002, and telecom pricing plummeting by 70%. This analogy fails due to three key structural differences in the current landscape.
First, the major cloud hyperscalers like Alphabet, Amazon, and Microsoft possess diverse and highly defensible core businesses, including search, e-commerce, and enterprise software, backed by fortress balance sheets. Unlike the monocultural telecom firms of the early 2000s, low returns on a single strategic investment area, such as AI hardware, will not lead to systemic failure. These diversified revenue streams provide insulation against short-term economic downturns or periods of low utilization in the AI sector.
Second, the nature of the asset being procured is fundamentally different. Fiber-optic cable was a sunk, fixed-function asset with limited alternative use. Conversely, cutting-edge AI GPU clusters are dynamic, general-purpose computational assets. These clusters are essential not only for external generative AI applications but also for maintaining dominance in adjacent, high-margin software sectors and for running the hyperscalers’ proprietary, highly optimized cloud services. The asset’s flexibility grants it inherent strategic value beyond simple rental income.
The fundamental distinction lies in financial leverage. Hyperscalers are financing this expansion almost entirely from retained earnings rather than external debt. Their Debt-to-Equity (D/E) ratios—ranging from 0.01 (Alphabet) to 0.18 (Microsoft)—are drastically lower than the 1.1x average seen in traditional infrastructure like telecom services and railroads, and far below the debt-fueled 6x ratios of the 2000s telecom bubble. Massive Free Cash Flow (FCF) enables this internal funding. This reliance on cash flow rather than debt signals a dramatically reduced financial risk profile, confirming that the buildout is stable and self-funded. In all fairness, there is likely debt-fueled CAPEX on the way, and more intensive ROI investigations are needed.
This strategic build-out is structured to mitigate the risk of oversupply and duration actively. Microsoft CFO Amy Hood confirms that approximately half of CapEx is spent on short-lived assets (GPUs/CPUs) deliberately deployed to match the expected duration of customer contracts, ensuring the depreciation schedule aligns with revenue. Hood also clarified that the previous bottleneck wasn’t a lack of chips, but rather a lack of “space or the power” infrastructure to house them, underscoring the energy challenge again. CEO Satya Nadella highlighted the operational strategy for these assets, emphasizing the development of a “fungible fleet” that is continuously modernized and can pivot instantly across all stages of the AI lifecycle, from pre-training and R&D to inference and post-training. This strategic flexibility ensures the infrastructure is not dependent on a single application or contract duration, maximizing efficiency and minimize TCO.
This operational flexibility means the useful lifespan of a GPU is structurally longer than the typical 3-5-year depreciation schedule or the next-generation product hype cycle might suggest. Older GPUs (such as the 2016-era NVIDIA P100s) remain profitable and widely used for lower-intensity tasks, such as rendering, video processing, and inference, especially in capacity-constrained environments where their lower power consumption is a net benefit. The asset is foundational infrastructure, more akin to a power plant or a hydroelectric dam than a simple consumable, allowing for new uses to be found to “squeeze every bit of value out of it.” This extends the return on investment (ROI) over a decade or more.
The GPU lifecycle is best understood as a multi-stage value chain. The newest, most powerful chips (like the H100) start at Stage 1: Peak Training/High-Intensity, commanding the highest rates for complex LLM training. As newer generations arrive, these chips move to Stage 2: Inference and Specialized Workloads, where they continue to generate high utilization and revenue running deployed models and commercial applications. Finally, they transition to Stage 3: Legacy/Commodity Processing, where 9-year-old chips like the M4000 can still be profitable for commodity tasks such as video processing and rendering, often prized for their lower power consumption in energy-constrained data centers.
This heterogeneous compute strategy, where different generations of chips handle various components of a single advanced task, changes the economics. For the next wave of advanced “agentic AI” models, for instance, a large-scale cluster will deploy the newest chips for the most complex processing while simultaneously utilizing older GPUs to handle the less intensive, parallel components of the reasoning chain. This approach maximizes the efficiency and effective lifespan of the entire fleet, creating an ecosystem where even highly depreciated hardware remains a strategic, revenue-generating asset.
Third, the utilization metric itself is misleading. The 2000s telecom bubble saw assets built purely for external, high-volume capacity. Today, a significant portion of the acquired compute capacity is immediately internalized by the hyperscalers themselves. This capacity is deployed for proprietary large-scale model training and for re-architecting proprietary services, shielding the infrastructure from external market volatility and establishing a foundational, immediate internal return on investment.
As a millennial who grew up in the 2010s with slop-bucket BuzzFeed lists, here’s a top 10.
Shorter Asset Lifespan / Constant Obsolescence
AI Chips vs. Fiber: Fiber optic cable laid in the ‘90s remains in service decades later. Conversely, the useful life of cutting-edge AI chips (such as NVIDIA’s GPUs) is significantly shorter, driving a continuous, replacement-driven demand cycle rather than a one-time oversupply. Telecoms spent $500 billion in nominal dollars over the roughly five-year period following the Telecommunications Act of 1996 (i.e., 1996–2001) on fiber, leading to a glut because it was a one-time, durable infrastructure play. Demand didn’t materialize fast enough, resulting in “dark fiber” (unused capacity) and bankruptcies (e.g., Global Crossing). AI capex, however, is recurring and upgrade-heavy, tied to short chip lifecycles and exploding computational needs. This sustains demand rather than creating a one-and-done oversupply. The CapEx strategy is to purchase short-lived assets (CPUs/GPUs) to manage the financial risk of the rapid refresh cycle while deploying them within long-lived, fungible infrastructure (data centers) to maximize utilization and pivot workloads.
Stronger Financial Footing of Spenders
Hyperscaler Cash Flow: Companies driving the AI buildout (Alphabet, Amazon, Microsoft) possess fortress balance sheets and massive Free Cash Flow (FCF). Unlike the telecom firms of the early 2000s, which often relied on debt/vendor financing for their fiber build, the current capex is primarily internally funded. Alphabet, Amazon, and Microsoft are nearly as good, or in some cases better, credit risks than the US government.
Foundational Business Models
Monopoly/Near-Monopoly Platforms: The hyperscalers are not solely reliant on AI revenue. Their AI infrastructure serves to defend and enhance their core, highly profitable businesses (Search, Cloud, Enterprise Software), providing a built-in customer base and more immediate, diversified returns.
Greater Compute Demand Intensity
Compute Power Multiplier: Generative AI models require 1,000x the compute power of previous perception-based AI models. The shift toward more advanced “agentic AI” (AI capable of context and reasoning) is projected to require a further 30x to 100x increase in power, creating a non-linear demand curve for compute. See Jevon’s paradox.
Rapid & Widespread Adoption Rate
Faster Diffusion: Generative AI adoption is currently outpacing previous revolutionary technologies. The adoption rate for Generative AI (an estimated 54.6% of the US population as of 2025, according to a recent St. Louis Fed study) is significantly higher than for Personal Computers (19.7% in 1984) or the Internet (30.1% in 1998) at similar points in their commercial lifecycles.
Focus on Software/Model Layer
Layered Value Creation: The infrastructure is a direct enabler for the application layer (LLMs or application services) and the enterprise layer (internal use for productivity). This is a more complex, multi-sided market than the fiber optic boom, where the value proposition was almost entirely confined to a commodity data transport layer.
Productivity-First Value Driver
Trillions in Value Creation: Generative AI’s primary economic contribution is projected to be in labor productivity and automation, not just direct consumer revenue. McKinsey estimates that Generative AI could add between $2.6 trillion and $4.4 trillion annually to the global economy across 63 use cases.
Value Beyond Revenue
Non-Revenue Financial Impact: In the private equity sector, AI is emerging as a “third pillar of value enhancement” (after financial engineering and operational excellence). A majority of firms surveyed by McKinsey reported improved innovation, and nearly half reported improved customer satisfaction and competitive differentiation, even when the direct EBIT impact remained small.
Accelerated Business Efficiency
Specific Task Gains: Quantifiable gains in productivity are already being demonstrated: a St. Louis Fed study estimates a 33% increase in productivity with generative AI, and another study found a 40% increase in speed and 18% increase in output quality for basic professional writing tasks using ChatGPT-3.5, and a 56% increase in speed for JavaScript programming with GitHub Copilot.
Defense and Internal Innovation
In-House Chip Development: The capex includes spending on building proprietary AI chips (such as Alphabet’s TPUs or Amazon’s Trainium/Inferentia), which aims to reduce reliance on a single-vendor supplier (NVIDIA) and dramatically lower the long-term cost of running their own AI services. This reduces future operating expenses, which is an immediate value-add even if external revenue is low today.
The Latent Demand Thesis: Internal ROI and Strategic Option Value
The massive CAPEX is best understood as an investment in internal transformation and strategic option value, drivers that are not captured in the revenue figures of third-party application providers like OpenAI.
Large software companies have already identified and implemented internal cost-cutting uses for AI models that significantly optimize their global operations. This return on investment, which accrues through operational efficiencies and performance augmentation across global empires, often far outweighs the current, visible revenue generated from external LLM customers. The capacity is strategically utilized as an internal cost optimizer before it functions as an external revenue generator.
Furthermore, the high CAPEX secures access to the most advanced compute, ensuring the hyperscalers control the innovation velocity of the next generation of AI breakthroughs. This investment secures a strategic option to participate in future, currently undefined, high-value applications. The aggressive spending on hardware, including significant, multi-billion-dollar efforts by Alphabet, Amazon, and Microsoft to develop in-house AI chips, is not simply a capital outlay; it is a strategic maneuver to achieve autonomy. This vertical integration is designed to control innovation speed, manage the supply chain, and mitigate the formidable pricing leverage currently held by a single dominant GPU supplier, Nvidia. This CapEx is therefore fundamentally purchasing strategic technological autonomy and future adaptability.
Where I Think Things Are Heading
The Physical World AI and the Embodiment of Intelligence
The next strategic front is the shift from Software Intelligence to Embodied Intelligence (Physical World AI). Algorithms meet and manipulate the physical world. This is the stage that provides the military and materials gain required for mass production of drones and other autonomous systems. Dark factories, autonomous vehicles, and automated delivery bots will become commonplace.
How close are we to physical world AI? Embodied AI is no longer a futuristic concept. Manufacturing and logistics are ripe for early adoption. This is the integration of three elements: (1) perception and decision-making (real-time inference); (2) actuation and control (electric motors, hydraulics, and power electronics that convert decisions into motion); and (3) connectivity and coordination (private networks). The race is now defined by whoever can master the learning loop between algorithmic design and physical production. Actuators, motors, and controllers are the physical interfaces that enable the commercial proliferation of autonomous systems such as medical delivery drones, last-mile transport bots, and industrial robotics.
There’s more to talk about in these, and I look forward to diving deeper into them.
Drones: Computational Gain Enabling Material Gain. The military domain is the most visible example. The drone war, such as in Ukraine, is less about a platform breakthrough and more about an algorithmic war of attrition. The most significant material gain comes from the mass deployment of cheap, disposable, and networked systems.
The core of this development is the Civil-Military Fusion of Funding. Military necessity drives rapid innovation and large-scale funding into the fundamental technologies of autonomous systems, namely high-performance on-device inference and actuation. Once these core systems are proven reliable and scalable on the battlefield, the private sector leverages the investment to develop commercial applications such as autonomous trucking, automated last-mile delivery, and medical transport drones. This creates a powerful R&D flywheel in which defense spending accelerates technological maturity, and commercial adoption drives down costs and improves performance for the original military applications.
Tactical Revolution: The US has the advantage of access to battlefield feedback loops in Ukraine. Potential military actions in Central and South America may enable further development.
Computational Bottleneck: The primary constraint is no longer airframe production, but the ability to rapidly develop, test, and deploy autonomous flight software and AI-assisted guidance that can overcome constant electromagnetic jamming and countermeasures. The United States still lags in actual material production.
The Materials Gain: The ability to mass-produce advanced drones is directly tied to winning the computational race. The AI build-out is an investment in the capability to scale and field the most effective autonomous systems faster than any rival.
The AI Content Saturation and Cognitive Threat
The final, highly disruptive frontier is the potential for AI video slop-content—the saturation of information channels with high-volume, low-cost, algorithmically generated media.
Mass Produce: Drastically cut production costs (up to 30% for film/TV, according to some estimates) while creating high volumes of content.
Hyper-Personalize: Generate individually tailored video snippets, recommendations, and advertisements, and stream them directly to the consumer’s feed in real time. This drive for personalization is a significant revenue engine for streaming platforms, increasing engagement and retention.
(A)Musing on Culture
What will happen when there is a massive load of digital AI video slop-content?
The long-term threat is cognitive and systemic, not merely economic:
Content Devaluation: The meaningfulness of content-as-signal degrades when the marginal cost of production approaches zero, leading to an overwhelming “slop” that forces consumers to retreat from public channels or seek out certified human-generated content. If content servers and aggregators can authenticate genuine or partial human veracity, they can establish premium content creator moats. Ultimately, online culture has long tended towards homogeneity (see dead internet theory), but it is a superficial appeal. There will be real value in universal appeal through uniqueness. People far out on the normal distribution of content creation will be rewarded and develop their niches. These niches will spawn subniches, but these will largely trend towards micro-conformity. Content aggregators will reinforce the most popular content creators because these creators will demonstrate the highest appeal rating to the largest number of people. It will not be impossible, though, because the algorithms need just enough similar-but-different content to lead people to the next trendline.
Erosion of Trust and Critical Thinking: The widespread adoption of generative content, especially video, makes it increasingly challenging to distinguish human-created content from AI-generated content. This “in-distinguishability”, combined with cognitive offloading (relying on AI to analyze information), risks an overall decline in critical thinking abilities and fuels public skepticism and distrust in media and brands. People will increasingly live in augmented or fully virtual realities. Consider how powerful a stimulus the smartphone and social media are. Imagine how entrancing it will be when lifestyle content creators share three-dimensional snapshots of their time in Fiji or climbing a pristine mountain. With haptic body suits, gaming and entertainment will become permanent escapes. There will be great works of artifice, and there will be huge wastes of human potential to electronic stimuli.
The Arms Race in Authenticity: As AI-generated content floods the market, a counter-demand for authenticity is already emerging, forcing platforms to consider mandatory AIGC labels and invest in technologies (potentially blockchain-based) that can certify human origin. The race shifts from who can generate the most content to who can develop the most trusted content. Trust is primarily emotionally derived, though sometimes rational. There are four archetypes of trust: Personal Connoisseurship (expertise and intrinsic value), Second-Hand Trust (delegated editorial choice to a trusted neighbor, celebrity, or expert), Credible Institution (trust extended based on credentials from an established body like a university), and Value Marker (currency or external marker of worth, such as dollar value). The social media era developed its own value heuristic, which is a form of the value marker, and it will be interesting to see how these fundamental trust heuristics adapt and apply to AI-generated reality.