The Burning Question: Can the AI Economy Afford Itself?
The Burning Question
Token generation is growing exponentially. The energy bill, the infrastructure cost, and the revenue models are not keeping pace — and nobody wants to say it out loud.
Every coding agent session described in The Great Inversion — the specification that becomes the product, the test suite that constrains the agent, the supervisor watching parallel AI streams generate thousands of lines of code — consumes electricity. Not the modest electricity of a text search or a database query. The electricity of something running a GPU cluster at sustained load: a chain of probabilistic calculations, each one multiplied by every step of the agent's multi-turn reasoning, executed at a scale that would have been unrecognisable as software five years ago.
A single ChatGPT query already consumes roughly ten times the energy of a Google search[1]. A coding agent completing a feature request — reading a codebase, drafting a specification, generating code, running tests, iterating on failures — is not one query. It is dozens or hundreds, running sequentially and sometimes in parallel, each drawing from hardware that consumes between 700 watts and several kilowatts per GPU. Multiply that by the adoption curve described in every market forecast and you arrive at a number that looks less like a software industry statistic and more like a national energy policy problem.
The International Energy Agency projects that global datacenter electricity consumption will more than double — from approximately 415 TWh in 2024 to roughly 945 TWh by 2030[1]. That figure, roughly equivalent to Japan's total annual electricity use today, is not primarily driven by streaming video or cloud storage. It is driven by AI, and specifically by the agentic, multi-step AI workloads that are growing fastest. Electricity demand from AI-focused datacenters climbed well ahead of overall datacenter demand in 2025, and the IEA projects power use from AI-focused facilities will triple by 2030[2].
But energy is only part of the story. Behind the electricity bill is a capital expenditure arms race with no historical precedent in the software industry. Behind the capital expenditure is a revenue model that, for most participants, still does not cover its costs. And behind the revenue model is a set of unanswered questions about who ultimately pays for the infrastructure the AI economy is being built on — and whether the answers arrive before the money runs out.
A Note on Evidence and Claims
AI infrastructure economics is a fast-moving field where the most dramatic claims tend to come from parties with interests in either the alarm or the reassurance. This article draws primarily on IEA reports, peer-reviewed energy research, public financial filings, and named executive statements. Where financial figures are cited from secondary sources — particularly analyst estimates of private companies' costs and revenues — the provenance is noted and the reader should treat magnitudes as directional rather than precise. The Critical Evaluation appendix grades each major evidence type.
I. The Physics of a Token
To understand what is at stake, it helps to start with the physical reality of what "generating a token" actually means. A large language model inference call requires the model's billions of parameters to be loaded into GPU memory and traversed for every token generated. The Nvidia H100, the current workhorse of enterprise AI inference, draws 700 watts at full load. A standard eight-GPU H100 server runs at roughly 5.6 kilowatts. At modest utilisation, a single such server consumes something in the range of a typical American home's annual electricity use — every month[3].
For frontier models, the arithmetic compounds quickly. Larger models require more GPUs per inference call. Reasoning models — the kind that pause to think through intermediate steps before answering — generate more tokens per user request. Agentic workflows can consume an order of magnitude more tokens than a simple question-and-answer interaction. Researchers at the University of California, Riverside estimated that a 100-word AI prompt uses roughly 500 millilitres of water for cooling[4]; that figure scales with token count, model size, and the number of reasoning steps involved.
There is genuine good news on efficiency. Per-token energy costs are falling: the IEA notes that power consumption per AI task is declining "at a rate unprecedented in energy history"[2]. Better model architectures, quantisation, distillation, and hardware advances have dramatically reduced the compute required per unit of intelligence produced. Mixture-of-experts models like DeepSeek activate only a fraction of their parameters per token, reducing computational overhead without sacrificing output quality[5].
The problem is that efficiency gains are being overwhelmed by demand growth. This is not speculation — it is a 161-year-old observation about resource economics, playing out in real time in the AI industry.
The Jevons Paradox, Revisited
In 1865, the economist William Stanley Jevons observed that more efficient steam engines did not reduce Britain's coal consumption. They increased it, because lower per-unit costs made coal economical for applications that had previously been unviable, expanding demand faster than efficiency gains reduced per-unit consumption. The paradox has since been documented in electricity, transportation, agriculture, and computing — wherever efficiency improvements lower the marginal cost of a resource, aggregate consumption tends to rise rather than fall.
Microsoft CEO Satya Nadella invoked it explicitly the day DeepSeek's efficiency claims circulated in January 2025: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket."[6] His framing was self-serving but also analytically accurate. Days after DeepSeek's announcement, Meta raised its 2025 AI spending to $60–65 billion, a 50% year-over-year increase[7]. No hyperscaler reduced its infrastructure investment in response to a competitor's efficiency claims. Every major player increased it.
A ByteDance computer architect published an analysis in 2025 arguing that the Jevons rebound effect in AI datacenters may already exceed 100% — meaning efficiency improvements are being more than offset by demand growth[8]. Whether that precise figure holds is uncertain; the direction is not. Cheaper inference enables more inference. More inference enables new applications. New applications create demand for frontier capabilities. The loop has no obvious internal brake.
Every time tokens get cheaper, developers build agents that use more of them. Efficiency is a floor, not a ceiling.
II. The Arms Race: $400 Billion and Counting
The capital expenditure of five large technology companies surged to more than $400 billion in 2025, and is set to increase by a further 75% in 2026[2]. This is not software development spending. It is physical infrastructure: concrete, steel, power substations, cooling systems, land, and Nvidia GPUs. The companies building AI are not becoming better software companies. They are becoming, functionally, electricity utilities with a software problem.
OpenAI's Stargate initiative targets $500 billion in infrastructure investment over four years. Microsoft committed to spending $80 billion on datacenters in fiscal 2025 alone[9]. Google announced $75 billion in 2025 capex, substantially driven by AI infrastructure. Meta raised its guidance to $60–65 billion. Amazon's datacenter commitments across AWS regions run to hundreds of billions when aggregated. These figures have outgrown the venture capital model. They require sovereign wealth fund backing, national government incentives, and utility-scale power purchase agreements that lock in 20-year commitments.
The NVIDIA Chokepoint
Every dollar of this infrastructure spending eventually flows through one company's supply chain. Nvidia's datacenter GPU business generated revenue of roughly $115 billion in fiscal year 2025, with gross margins above 70%[10]. The H100 at $25,000–40,000 per card remains the workhorse of frontier AI. Eight-GPU DGX B200 systems cost in excess of $500,000 each[11]. The entire industry runs on CUDA, Nvidia's proprietary software stack, and on HBM memory on TSMC fabrication nodes. AMD's MI-series GPUs and custom accelerators from Google, Amazon, and Microsoft represent genuine competition — but none has yet displaced Nvidia's ecosystem at scale for frontier model training. The hardware layer is the one part of the AI stack where competition has genuinely not arrived, and every participant in the industry knows it.
The Hardware Tax
When an enterprise pays per month for a Claude or GPT subscription, a portion of that revenue pays for compute, which pays for GPU time, which ultimately accrues to Nvidia at 70%+ gross margin. The AI value chain is, at its base, a mechanism for converting software revenue into semiconductor profit. The labs and hyperscalers compete fiercely over the middle; the hardware manufacturer collects a toll from everyone who passes through.
III. The Revenue Gap Nobody Discusses Plainly
The arithmetic of AI inference economics is not difficult to state, though it is awkward to acknowledge. Building and operating the infrastructure to serve frontier models costs more, in aggregate, than the revenue those models currently generate. OpenAI spent an estimated $1.70 for every dollar it earned in 2025, with inference costs alone approaching $7 billion annually against total operating costs of roughly $8.5 billion[12]. Anthropic's CFO stated in a March 2026 legal filing that the company had spent "over" $10 billion on inference and training against $5 billion in cumulative revenue[13].
A Note on Financial Figures
The revenue and cost estimates in this section come primarily from secondary analyses of public statements, legal filings, and reported documents. OpenAI and Anthropic are private companies with no obligation to publish detailed financials. The directional conclusion — that both companies currently spend more than they earn at the company level — is consistent across all credible analyses reviewed. Specific figures should be treated as estimates, not verified accounts.
Revenue growth has been extraordinary by any software industry standard. Anthropic grew from $1 billion in annualised revenue in December 2024 to an estimated $30 billion by April 2026, representing roughly 10× annual growth sustained for three consecutive years[14]. Salesforce took 24 years to reach $30 billion. These are remarkable numbers — but they cannot be separated from a cost structure that is also growing. OpenAI is projected to spend $121 billion on AI model training by 2028[15]. That is not a software development budget. It is a particle accelerator that happens to autocomplete emails.
The Subsidy Chain
| Subsidy Source | Mechanism | Sustainability Concern |
|---|---|---|
| Venture capital / private equity | Investors provide capital at valuations embedding future profitability assumptions. Anthropic's $380B post-money valuation implies 27× revenue[14]. | Requires revenue growth and margin expansion on a precise, aggressive timeline. Dario Amodei told Fortune that a twelve-month delay in AI progress would make Anthropic bankrupt[14]. |
| Hyperscaler cloud margins | Microsoft and Google absorb inference losses on consumer AI products as customer acquisition cost for their cloud businesses. | Works as long as cloud margins remain robust and AI adoption drives incremental cloud spend. Vulnerable if AI commoditises cloud compute itself. |
| Enterprise premium pricing | Enterprise customers pay significantly higher prices than consumer API users. Fortune 500 relationships carry margins that partially offset consumer-tier losses. | Sustainable but limited in scale. Enterprise budgets are finite, and the number of large customers paying large premiums is not proportional to total token volume. |
| Government / sovereign investment | National AI programs, sovereign funds (Saudi PIF, UAE G42), and industrial policy subsidies provide capital without near-term return requirements. | Politically contingent. Subject to diplomatic risk, regulatory conditions, and shifts in government priority. |
| Consumer subscriptions | $20–200/month subscriptions provide predictable revenue, but per-user inference cost for heavy users significantly exceeds the subscription price for the most compute-intensive workloads. | The structural mismatch between flat-rate pricing and variable compute cost is unresolved. Rate limits address the symptom. |
The critical question is not whether any individual subsidy is sustainable — several are — but whether the aggregate subsidy structure converges on positive unit economics before the available capital is exhausted. The optimistic reading holds that token prices will continue to fall and the expanding market for AI-native applications will generate revenue at scale that eventually covers costs. The pessimistic reading notes that this story has been told for three years and the gap between costs and revenues has not closed — it has grown.
IV. The Grid Cannot Keep Up
AI datacenters cluster in a small number of locations — Northern Virginia, Silicon Valley, Oregon, Dublin, Singapore, Tokyo — chosen for fibre connectivity, land cost, and access to cheap power. The result is that a global infrastructure buildout is concentrating its demand stress on a handful of regional grids that were not designed to absorb it. Northern Virginia, home to more than 27 million square feet of existing data center space, has datacenters consuming an estimated 26% of the state's total electricity[16]. Ireland used approximately 21% of its national electricity for datacenters in 2024, with projections of 32% by 2026[16]. A peer-reviewed study published in 2025 identified Power Stress Index values exceeding 0.25 in Oregon, Virginia, and Ireland, indicating local grid vulnerability from concentrated AI infrastructure[17].
Grid interconnection operates on timescales measured in decades, not quarters. The IEA reports that datacenters are advancing a large number of projects with onsite natural gas-based power generation, primarily in the United States[2]. The renewable energy purchases that hyperscalers advertise in their sustainability reports are real — they accounted for roughly 40% of all corporate power purchase agreements for renewables signed in 2025[2]. But wind and solar are intermittent, and a datacenter GPU cluster running a training job cannot wait for wind.
Water: The Invisible Constraint
Energy receives most of the public attention, but water may prove the more binding physical constraint in key regions. A medium-sized facility uses roughly 110 million gallons of water per year, and the largest can consume up to 5 million gallons per day[4]. Training GPT-3 is estimated to have evaporated 700,000 litres of freshwater at Microsoft's US datacenters[18]. Google reported withdrawing 6.4 billion gallons of water in 2023, with 95% used to cool datacenters[19]. Roughly two-thirds of the datacenters built since 2022 have been located in water-stressed regions, according to Bloomberg analysis[20]. A University of Houston study projected that Texas data centers alone will use 399 billion gallons of water by 2030 — equivalent to drawing down Lake Mead by more than 16 feet in a year[20].
A ScienceDirect study estimated AI systems' water footprint at between 312.5 and 764.6 billion litres in 2025 — a range approaching the global annual consumption of bottled water[21]. The Lawrence Berkeley National Laboratory estimated that indirect water use — the water consumed by the power plants supplying datacenter electricity — is twelve times larger than direct cooling use, and almost universally unreported[4].
The Transparency Gap
Amazon does not disclose how much water its datacenters consume. Microsoft provides overall figures without site detail. The ISO/IEC only recently published its first international standard on sustainable AI, including water footprint as a metric. Regulation requiring consistent, comparable, facility-level disclosure of both energy and water use does not yet exist in most jurisdictions. The public and policymakers are making decisions about datacenter approvals without the information required to assess cumulative impact[18].
The engineering solutions exist. Microsoft launched a new datacenter design in August 2024 that consumes zero water for cooling, using closed-loop chip-level liquid cooling that reduces annual water use by more than 125 million litres per facility[22]. These designs are just beginning to be deployed at scale; the existing fleet cannot be retrofit quickly.
V. The Nuclear Bet
If the grid cannot keep pace with AI demand, and if renewable energy cannot provide the 24/7 baseload power that a datacenter running training jobs requires, then the logical answer is to own the power source. This is the conclusion that the major hyperscalers have reached, and the resulting pivot to nuclear energy represents one of the most extraordinary corporate infrastructure plays since the railroad era.
Microsoft signed a 20-year, 835-megawatt power purchase agreement with Constellation Energy for the restart of Three Mile Island Unit 1, dormant since 2019, targeting return to service in 2028[23]. The decision is entirely rational: Unit 1 never had a safety incident, and a 20-year PPA at a known cost provides exactly the price certainty and baseload reliability that intermittent renewables cannot match. Google made history in October 2024 with the world's first corporate small modular reactor purchase agreement, partnering with Kairos Power to deploy 500 megawatts of advanced nuclear capacity near its datacenters, with the first unit targeting 2030[24]. Amazon is investing more than $20 billion in a campus adjacent to the Susquehanna nuclear plant. Amazon, Google, and Microsoft have collectively committed over $10 billion to nuclear partnerships, with 22 gigawatts of projects in development globally[25].
The nuclear bet is rational but carries significant execution risk. Small modular reactors have not yet been deployed commercially at scale. The most advanced first-of-kind projects face regulatory timelines and construction complexity that make 2030 target dates optimistic. Between now and then, the gap is being filled by natural gas. There is also a deeper structural consequence: if AI labs need to secure nuclear power plants to run their models, then the barrier to entry for frontier AI includes energy infrastructure that takes decades to build. Nuclear-secured AI compute is a moat that cannot be replicated by a well-funded startup. It is the kind of advantage that large incumbents accumulate over time.
VI. The Carbon Accounting Gap
The major AI providers publish sustainability reports containing real data about energy consumption and renewable energy purchases. They do not contain the complete picture. Standard corporate carbon accounting covers Scope 1 and 2 emissions; Scope 3 — which includes GPU manufacturing, datacenter construction, supply chain for rare earth materials, and the embodied carbon in the infrastructure buildout — is rarely reported comprehensively. GPU manufacturing is itself extraordinarily energy-intensive: TSMC's fabrication plants operate at the frontier of semiconductor physics, and the carbon in an H100 GPU appears in TSMC's emissions reports, not in Microsoft's or Google's.
A ScienceDirect study estimated that AI systems' carbon footprint could be between 32.6 and 79.7 million metric tonnes of CO2 equivalent in 2025 — roughly equivalent to New York City's total annual emissions[21]. The wide range reflects the absence of standardised disclosure. The midpoint reflects a credible attempt to aggregate available data across the supply chain.
The Real Carbon Footprint
A useful estimate from environmental researchers: the actual carbon footprint of a model or datacenter is likely 3–5× what is typically reported in corporate sustainability disclosures, once supply chain manufacturing, infrastructure construction, and indirect water-for-power use are included. This is not a criticism of deliberate misrepresentation — Scope 3 is genuinely difficult to measure. It is an argument for mandatory, standardised, third-party-verified environmental reporting before the infrastructure scales further.
Carbon regulation is coming. The EU's Carbon Border Adjustment Mechanism, SEC climate disclosure rules, and emerging AI-specific environmental reporting frameworks will eventually force fuller accounting. Public opposition to datacenter construction — already visible in Ireland, the Netherlands, and parts of the American Southwest — is partly driven by energy and water concerns. An industry that cannot provide accurate, standardised, comparable environmental data is poorly positioned to manage that opposition.
VII. The Open-Source Wildcard
Every assumption in the preceding sections rests on a hypothesis that deserves explicit examination: that frontier AI capability remains gated behind infrastructure only hyperscalers can build. If that hypothesis is correct, the nuclear-PPA model makes sense — the moat is the megawatts. If it is wrong, the economics change substantially.
DeepSeek's January 2025 announcement was the most vivid challenge to date. A Chinese lab trained a model matching GPT-4-class performance at a reported training cost of $6 million — a fraction of what OpenAI, Google, and Anthropic spend on comparable capability[26]. The claimed figures attracted scepticism — full infrastructure costs are difficult to attribute — but the directional message was taken seriously: algorithmic innovation can, at least partially, substitute for raw compute. Nvidia's stock fell 17% in a single day[26]. Meta's Llama series has made frontier-adjacent capability available as open weights, runnable on hardware that organisations can own outright. A single server with four consumer-grade RTX 5090 GPUs can now run models that would have required cloud infrastructure a year ago[27].
The incumbents' capex bet only pays off if model quality at the frontier remains gated behind scale. That assumption has been correct so far: frontier models demonstrably outperform open-weight alternatives on the hardest tasks. But the gap is narrowing. The question is not whether open-source will eventually match frontier capability. It is whether it will do so before the current infrastructure investments are amortised.
VIII. Three Possible Futures
Consolidation and Recovery
Two or three infrastructure-backed providers survive at scale. Training costs continue to fall faster than capability requirements rise. Enterprise revenue reaches margins sufficient to cover costs. Standalone AI labs either achieve profitability or are absorbed by their computing partners.
What has to be true: Token price declines slow. Model efficiency improvements outpace demand growth. Enterprise adoption sustains premium pricing long enough for cost structures to normalise.
Commoditisation
Inference becomes a utility. Margins compress to near zero. Only infrastructure owners win. Standalone labs without infrastructure parents face terminal economics. Open-source captures the mid-market; frontier capability remains a small, expensive niche.
What has to be true: Open-source quality continues converging with frontier. Enterprise customers prove unwilling to pay frontier premiums as capable open alternatives emerge.
Regulatory Constraint
Energy consumption triggers mandatory carbon accounting, grid impact fees, or environmental review requirements. Water scarcity restrictions limit siting options. The buildout slows not because of economics but because of physical and regulatory constraints the industry cannot engineer around quickly enough.
What has to be true: Local opposition to datacenter siting reaches political scale. Frameworks currently exempting AI infrastructure are amended. Carbon rules include Scope 3 emissions.
These scenarios are not mutually exclusive. The most probable near-term outcome is a version of Scenario A for hyperscaler-backed AI (infrastructure ownership provides durable advantage) and a version of Scenario B for standalone API businesses (price competition and open-source erode margins). The regulatory scenario is lower probability in the near term but growing.
What Survival Looks Like
| Structural Advantage | Why It Matters | Who Has It |
|---|---|---|
| Infrastructure ownership | Converts variable compute cost to fixed amortised cost, improving unit economics at scale. | Microsoft, Google, Amazon. Partially: Oracle, Meta. |
| Revenue diversification | AI sold as part of a broader cloud or enterprise platform can absorb inference losses as a cost of the overall relationship. | Microsoft (Azure + Office), Google (Workspace + GCP), Amazon (AWS). |
| Efficiency leadership | The lab that delivers frontier capability at the lowest cost per token can price competitively without the largest infrastructure base. | Contested. Anthropic's revenue-per-gigawatt metrics suggest efficiency leadership; DeepSeek's training costs suggest external challengers. |
| Long-term energy contracts | 20-year nuclear PPAs lock in baseload power at known cost, insulating against grid price volatility. | Microsoft (Three Mile Island), Google (Kairos), Amazon (Susquehanna). |
| Customer lock-in | Enterprise customers embedded in a provider's toolchain face high switching costs, sustaining premium pricing under competitive pressure. | Microsoft (GitHub Copilot + Azure), Anthropic (Claude Code enterprise contracts), Google (Workspace + Gemini). |
Conclusion: Paying the Bill
The story of AI infrastructure is, at its core, a story about who pays for a transformation whose costs arrive before its returns. The productivity gains described in The Great Inversion are real: agents are generating more code, faster, with fewer human hours. This piece has examined whether the physical and economic infrastructure supporting those gains can sustain itself — and the honest answer is: not yet, and it is not clear exactly how it will.
The energy demand is real and growing faster than grids can absorb it. The water consumption is real and concentrated in regions that cannot easily afford it. The capital expenditure is real and requires a scale of financing that has moved AI infrastructure from the venture-backed software model into the industrial infrastructure model. The revenue is real and growing at rates no industry has previously sustained — but it is growing into a cost base that is also growing, and the convergence point remains uncertain.
What is not in doubt is the direction. Electricity consumed by AI-focused datacenters is set to triple by 2030[2]. Hyperscaler capex is set to increase 75% in 2026 alone[2]. The nuclear pipeline has grown from 25 to 45 gigawatts in conditional commitments in eighteen months[2]. These are the signals of an industry that has concluded it is building physical infrastructure for a generation, not a product cycle.
The AI economy is not a software business that happens to use a lot of electricity. It is an electricity business that happens to produce software.
The Jevons Paradox means that making AI cheaper does not reduce its aggregate energy cost. It expands the range of applications AI can economically address, which expands demand, which expands the infrastructure required to serve it. Every efficiency improvement that makes tokens cheaper is, simultaneously, an argument for building more datacenters. The efficiency optimists and the infrastructure bears are both right, and they are right about the same thing, just from different vantage points.
For developers and organisations deploying AI — the audience of The Great Inversion — the practical implication is this: the tools you are building on are subsidised. Structurally subsidised by capital that is betting on a future where inference revenue covers infrastructure cost. That bet may well pay off. But organisations that plan their technical architecture around the assumption of indefinitely cheap, indefinitely available AI compute are making a bet of their own — on the same future, with less information and less capital than the companies making it.
The bill for the AI economy is not yet due. When it arrives, the question of who pays — end users, enterprise customers, investors, governments, or some combination — will be the most important business question in the industry. The answer will determine which providers survive, at what prices, under what regulatory conditions, and in which geographies. Planning for that question is not pessimism. It is the kind of systems thinking that The Great Inversion argued is the defining skill of the AI-native engineer. In the energy economy, the specification is a power purchase agreement, and the rigour is a 20-year nuclear contract.
Appendix: Critical Evaluation and Evidence Grading
What Holds Under Scrutiny
| Proposition | Evidence Base | Confidence |
|---|---|---|
| Global datacenter electricity demand doubles by 2030 | IEA Energy and AI (2025), ABI Research, peer-reviewed modelling[1][3][17] | High — multiple independent projections converge |
| Efficiency gains are outpaced by demand growth (Jevons) | IEA 2026 update, ByteDance SIGARCH analysis, hyperscaler capex trajectories[2][8] | Medium-High — directionally robust, magnitude uncertain |
| AI labs currently spend more than they earn at company level | Legal filings, analyst estimates, executive statements[12][13][15] | Medium — directional consensus; precise figures unverified |
| Hyperscalers are securing nuclear power at unprecedented scale | Public PPAs, regulatory filings, company announcements[23][24][25] | High — documented in public filings |
| Water consumption is a material constraint in water-stressed regions | EESI analysis, Bloomberg, Lincoln Institute, ScienceDirect[4][20][21] | High for direction, Medium for magnitude |
Structural Weaknesses
Weakness 1 — Private financials are estimates, not accounts. Revenue and cost figures for OpenAI and Anthropic are drawn from analyst reconstructions of public statements, legal filings, and reported leaks. Neither company publishes audited financial statements. Readers should treat specific numbers as directional rather than precise.
Weakness 2 — Efficiency projections may be too pessimistic. This article takes the Jevons Paradox as its central analytical frame, which assumes efficiency gains are consistently overwhelmed by demand growth. That assumption holds in the historical data so far but is not a law of nature. Dramatic breakthroughs in model efficiency could inflect the energy trajectory downward faster than current projections suggest.
Weakness 3 — Geopolitical uncertainty is underweighted. AI infrastructure economics is deeply entangled with export controls, tariff regimes, national AI programs, and diplomatic relationships in active flux. A shift in semiconductor export rules or energy regulation could materially alter the cost structure for any major player.
Weakness 4 — The open-source wildcard is hard to model. If open-weight models reach frontier capability within two to three years, much of the infrastructure investment discussed here becomes less defensible as a competitive moat. This outcome is genuinely uncertain, and the analysis does not resolve it — it merely names it as the highest-impact variable in the industry's long-term economics.
What This Article Does Not Prove
| Claim | Why the evidence does not support it |
|---|---|
| The current AI business model is unsustainable and will fail | Revenue is growing at rates that have historically supported eventual profitability. The timing question is genuinely open. |
| Energy constraints will cap AI capability growth | The industry has consistently found ways around physical constraints faster than pessimists expected. |
| Specific financial figures are accurate | These are analyst estimates from secondary sources. The direction is credible; the precise figures are not verified. |
| The Jevons Paradox will dominate at all efficiency levels indefinitely | The paradox describes a tendency, not a deterministic outcome. Historical effects have sometimes been offset at absolute cost floors. |
Methodological note: sources and their limitations
Energy demand figures are drawn primarily from the IEA's April 2025 Energy and AI report and its April 2026 update — the most authoritative available sources, based on comprehensive primary data collection. These are treated as the highest-confidence sources in the article.
Financial figures for AI labs are drawn from analyst reconstructions, legal filings, and executive public statements. These are treated as directional signals, not verified accounts. Where analyses conflict, the article cites the range.
Water consumption figures are drawn from corporate sustainability reports, academic research (University of California Riverside, Lawrence Berkeley National Laboratory), and environmental journalism. Corporate sustainability reports vary widely in scope and comparability.
Nuclear and energy infrastructure figures are drawn from public power purchase agreements, regulatory filings, and verified company announcements — treated as high-confidence primary sources.
References
- International Energy Agency (2025). Energy and AI. IEA, Paris. iea.org/reports/energy-and-ai
- International Energy Agency (2026). Key Questions on Energy and AI. IEA, Paris. iea.org
- ABI Research (2025). "Data Center Energy Consumption Forecast, 2024–2030." abiresearch.com
- Environmental and Energy Study Institute (2025). "Data Centers and Water Consumption." eesi.org
- Bain & Company (2025). "DeepSeek: A Game Changer in AI Efficiency?" bain.com
- Northeastern University News (2025). "What is Jevons Paradox? And why it may — or may not — predict AI's future." news.northeastern.edu
- Shao, G. / AI Proem (2025). "The Jevons Paradox in AI Infrastructure." aiproem.substack.com
- HumaI Blog (2025). "The Secret of Cheap AI: DeepSeek Uses 4.6x More Energy Than Claude at Inference." Citing Wang, W., SIGARCH (2025). humai.blog
- Data Center Dynamics (2025). "Microsoft & OpenAI consider $100bn, 5GW 'Stargate' AI data center." datacenterdynamics.com
- NVIDIA Corporation (2025). Annual Report / Investor Relations. investor.nvidia.com
- Intuition Labs (2026). "NVIDIA AI GPU Prices: H100 and H200 Cost Guide." intuitionlabs.ai
- AI2Work (2025). "AI Inference Economics in 2025: Why OpenAI and Anthropic Are Still Losing Billions." ai2.work (secondary analysis — figures are estimates)
- Zitron, E. / Where's Your Ed (2026). "Why Are We Still Doing This?" wheresyoured.at (commentary; cites Anthropic CFO legal filing, March 2026)
- Perera, S.A. (2026). "The Growth Miracle and the Six Fractures: Anthropic at $380 Billion." Substack (investor analysis — figures are estimates)
- AI After Hours Substack (2026). "OpenAI vs Anthropic: The $121 Billion Question." aiafterhours.substack.com (analyst reconstruction — treat as directional)
- AIM Multiple (2026). "AI Energy Consumption Statistics." aimultiple.com
- Yin, J. et al. (2025). "Concentrated siting of AI data centers drives regional power-system stress under rising global compute demand." arXiv:2604.06198. arxiv.org
- Environmental Law Institute (2025). "AI's Cooling Problem: How Data Centers Are Transforming Water Use." eli.org
- Malota Studio (2025). "The Hidden Thirst of AI: Evaluating Data Center Water Usage and Paths to Sustainability." malotastudio.net
- Lincoln Institute of Land Policy (2026). "Data Drain: The Land and Water Impacts of the AI Boom." lincolninst.edu
- de Vries-Gao, A. (2025). "The carbon and water footprints of data centers and what this could mean for artificial intelligence." ScienceDirect / Cell Press. sciencedirect.com
- Microsoft Cloud Blog (2024). "Sustainable by design: Next-generation datacenters consume zero water for cooling." microsoft.com
- Data Center Dynamics (2025). "Three Mile Island nuclear power plant to return as Microsoft signs 20-year, 835MW AI data center PPA." datacenterdynamics.com
- Introl Blog (2025). "SMRs Power AI: $10B Nuclear Data Center Revolution." introl.com
- Introl Blog (2025). "Nuclear power for AI: inside the data center energy deals." introl.com
- FinancialContent (2025). "The $6 Million Revolution: How DeepSeek R1 Rewrote the Economics of Artificial Intelligence." financialcontent.com (secondary source; training cost claim is contested)
- Silicon Analysts (2026). "NVIDIA GPU Prices 2026." siliconanalysts.com
This article was researched and written collaboratively with Claude Sonnet 4.6 (Anthropic), following the supervisory model described in The Great Inversion: human specification and critical review, machine research synthesis and drafting, iterative refinement through structured dialogue. The research phase involved live web searches across IEA reports, peer-reviewed energy studies, public financial filings, and analyst commentary. All primary sources are linked in the bibliography. Financial figures for private companies are clearly flagged as estimates.
The method is, of course, recursive: the article examines the economics of the AI infrastructure that produced it. The irony is not lost on the author.
Physicist & Developer
.PNG)
.PNG)
Comentários
Enviar um comentário