Power Without End: Where AI Gets Its Electricity —
Strategic Analysis — Energy Infrastructure Part III: Power Sources & Bubble Risk | May 2026
Who Pays If the Bubble Bursts
AI data centers demand 24/7 firm power that solar and wind cannot reliably deliver alone. The honest answer — natural gas — collides head-on with California's climate commitments. Meanwhile, credible voices from hedge funds to Harvard Law are warning that the infrastructure being built to serve the AI boom may become the next great stranded-asset disaster, with utility ratepayers left holding the bill.
The most important engineering fact about AI data centers is one their proponents prefer to soft-pedal: they cannot run on solar panels. Not primarily. Not reliably. A single large-scale AI training facility requires between 100 megawatts and 1 gigawatt of continuous, uninterrupted power — 24 hours a day, 365 days a year, regardless of cloud cover or wind conditions. The sun shines on average six hours a day in California. A wind turbine reaches its rated output less than 40% of the time. Battery storage, at current commercial scale, can bridge gaps of hours, not days of adverse weather.
This is the 24/7 baseload problem — and it is the central, largely unresolved energy challenge behind every gigawatt of AI data center demand now flooding California utility interconnection queues. How it gets resolved — or whether it gets resolved before the investment cycle exhausts itself — has profound consequences for San Diego ratepayers being asked to fund the grid infrastructure to serve it.
- 40% Natural gas share of global data center electricity in 2024
- 20% Nuclear share — the only carbon-free 24/7 option at scale today
- 24% Renewables share — growing but inherently intermittent
- 1,000+ Terawatt-hours global data center consumption projected by end of 2026 — equal to Japan's entire annual electricity use (IEA)
Where the Power Actually Comes From Now
The International Energy Agency's comprehensive 2025 Energy and AI report provides the clearest global picture. In 2024, natural gas accounted for 40% of data center electricity, followed by renewables at 24%, nuclear at 20%, and coal at 15%. The coal figure reflects primarily Asian markets — in the U.S., coal's share is smaller, but natural gas dominance is even more pronounced. The Brookings Institution confirmed the 40% gas figure in a 2026 update, adding that 64% of incremental generation to serve growing data center demand through 2035 is projected to come from fossil fuels, even as the share from new renewable installations grows.
| Source | 2024 Global Share | 24/7 Firm Capability | California Status | Outlook for Data Centers |
|---|---|---|---|---|
| Natural Gas | ✅ Fully dispatchable | Abundant; contradicts state climate goals | Bridge fuel for next 5–10 years; fastest to deploy (12–18 months) | |
| Nuclear (existing) | ✅ Always-on baseload | One plant (Diablo Canyon); no expansion planned | Ideal solution; in short supply; tech giants racing to sign deals | |
| Renewables (solar/wind) | ⚠️ Intermittent; needs storage | Abundant in Imperial Valley; can't provide 24/7 alone | Growing for daytime load; must be paired with firm power | |
| Coal | ✅ Firm; dispatchable | Effectively zero (state policy) | China-dominant; minimal U.S. role going forward | |
| SMR / Advanced Nuclear | ✅ Ideal if built | None operating; regulatory framework nascent | Post-2030 at best; 10+ GW contracted globally by tech companies |
The Nuclear Bet: Big Tech's $100 Billion Wager
The technology industry has made its position clear: nuclear power is the only credible long-term solution to the 24/7 baseload requirement that doesn't generate carbon. In a remarkable reversal of decades of corporate anti-nuclear sentiment, the hyperscalers are now the industry's most aggressive investors in nuclear revival.
Microsoft signed a 20-year, $16 billion agreement to restart the 835-megawatt Three Mile Island Unit 1 — the same facility site as the 1979 partial meltdown — targeting 2028 operation to power its Azure data centers. Google struck a deal with NextEra to restart the 615 MW Duane Arnold nuclear facility and signed the first U.S. corporate Small Modular Reactor (SMR) fleet deal with Kairos Power for 500 MW of capacity by 2030. Amazon has committed over $20 billion toward repurposing the Susquehanna nuclear plant as an AI campus. Meta issued a request for proposals for 1 to 4 gigawatts of new nuclear capacity.
Together, Big Tech signed more than 10 gigawatts of new U.S. nuclear capacity contracts in the past year alone — a figure that dwarfs the commercial nuclear order books of the prior two decades combined. The U.S. Department of Energy approved a $1 billion loan guarantee for the Three Mile Island restart, signaling federal alignment with this strategy.
The catch is timing. The IEA projects that SMRs will begin entering the mix "after 2030" — with the most optimistic deployment estimates placing early commercial operation in the early-to-mid 2030s. The NuScale US 460, the first SMR to receive a Standard Design Approval (issued May 2025), is a promising development, but no SMR is operating at commercial scale in the United States today. The gap between what tech companies need now and what nuclear can provide by the time new plants are built is measured in years.
Natural gas and coal together are expected to meet over 40% of the additional electricity demand from data centres until 2030. After 2030 SMRs enter the mix, providing a source of baseload low-emissions electricity to data centre operators.
The California Contradiction
California's position is particularly acute. The state has mandated 100% clean electricity by 2045, banned new natural gas plant approvals in most circumstances, and closed San Onofre Nuclear Generating Station in 2013. Its one remaining nuclear plant, Diablo Canyon (2.2 GW), was narrowly saved from early closure in 2022 — largely because grid planners recognized how badly California needs firm baseload capacity. Diablo Canyon's current operating license runs to 2030.
California has no operating SMRs, no new large nuclear plants in development, and no realistic pathway to significant new nuclear capacity before 2035 at the earliest. Yet its data center operators — including those now proposed for Imperial Valley — require 24/7 firm power that the state's renewable portfolio cannot reliably provide, particularly during evening peak hours, high-pressure weather systems that suppress wind generation, and the occasional multi-day fog events that reduce solar output across Southern California.
The proposed Imperial Valley data centers address this contradiction with varying degrees of credibility. CalEthos claims its complex will be "powered by geothermal and solar energy." Imperial Valley does have substantial geothermal potential — it sits atop the Salton Sea geothermal field, one of the largest in the world — but California's total installed geothermal capacity is approximately 2.7 gigawatts statewide, and expanding it significantly requires complex permitting and resource development timelines that rival nuclear. Solar, as noted, cannot provide firm 24/7 power without storage that doesn't exist at the scale of a 330-megawatt data center campus operating around the clock.
The IVCM project — the $10 billion, 330-megawatt facility — has not made similarly specific clean energy claims. That 330 MW demand, running continuously, is equivalent to roughly 2.9 billion kilowatt-hours per year. Serving it with battery storage alone would require approximately 14,000 megawatt-hours of storage — more than four times California's current total grid-scale battery capacity — just for overnight backup.
For a 330 MW data center in Imperial Valley to genuinely run on clean power 24/7 today, it would need: approximately 1,000 MW of dedicated solar capacity (to generate enough surplus to charge batteries during peak hours); approximately 3,000 to 5,000 MWh of grid-scale battery storage (to provide overnight power); and dedicated geothermal backup for multi-day solar shortfalls. The capital cost of this clean package alone — excluding the data center itself — would be several billion dollars. No proposed Imperial Valley data center has announced a credible plan at this scale. The realistic near-term power source, absent a dedicated nuclear or geothermal deal, is the California gas fleet.
The Efficiency Wildcard: DeepSeek and Jevons's Paradox
In January 2025, Chinese AI startup DeepSeek released its R1 model — trained on approximately 2,000 Nvidia H800 chips for roughly $6 million, achieving performance comparable to OpenAI's GPT-4, which required an estimated 20,000 higher-powered chips and cost over $100 million to train. The efficiency implications were staggering: DeepSeek demonstrated that comparable AI capability could potentially require 40 to 75% less power than assumed in hyperscaler demand forecasts.
The immediate reaction was a stock market shock: Nvidia's market cap dropped $600 billion in a single day, reflecting fear that the assumptions behind trillions of dollars of AI infrastructure investment might be wrong. If AI models can be trained and run at a fraction of the power cost, why build 1-gigawatt data centers?
The answer lies in a principle economists call Jevons's Paradox: when a technology becomes more efficient, its lower cost per unit typically drives dramatically higher consumption, resulting in greater total resource use. The history of computing confirms this pattern relentlessly. Personal computers became thousands of times more energy-efficient per calculation between 1980 and 2020 — and global computing energy consumption grew exponentially throughout that same period. AI token costs have already dropped approximately 280-fold over the past two years as models improved. That dramatic cost reduction has been matched by an explosion in usage volume, not a reduction in total computational demand.
There is also a subtlety in DeepSeek's efficiency claims that matters for power calculations: its chain-of-thought reasoning outputs consume approximately 41% more energy per query than standard model responses. And DeepSeek's open-source availability means its techniques are being adopted across the industry — simultaneously reducing per-query energy cost and enabling a vastly larger user base to run AI queries. The net effect on total power demand is deeply uncertain.
A peer-reviewed analysis published in 2025 in the National Institutes of Health's publication archive concluded that highly efficient AI models create a "counterintuitive phenomenon": while individual queries use less power, the resulting price drop and accessibility expansion drives extensive utilization that paradoxically increases total energy consumption. DeepSeek's rise may accelerate the shift toward distributed edge computing — smaller facilities closer to users — rather than reducing total data center demand.
A significant trend largely absent from California grid planning discussions: approximately 30% of anticipated new data center energy capacity is now designed to operate from on-site generation sources, up from effectively zero just a year ago, according to a February 2026 Cleanview report. Hyperscalers are bypassing grid interconnection queues by building their own gas turbines, fuel cells, or microreactor installations directly at data center sites. This could partially decouple data center growth from grid-scale transmission expansion — but it also means natural gas combustion is increasingly located in communities, not at remote generation sites, raising local air quality concerns.
The Bubble Question: Dot-Com Redux or Permanent Infrastructure?
The comparison to the late-1990s internet bubble is now being made by voices serious enough to demand engagement: not just contrarian bloggers, but Man Group (one of the world's largest hedge funds), Brookings Institution researchers, Harvard Law School's Environmental and Energy Law Program, and Goldman Sachs economists.
| Dimension | Late-1990s Telecom / Dot-Com | 2024–2026 AI Data Center | Verdict |
|---|---|---|---|
| Revenue vs. Investment | Companies had near-zero revenue; pure speculation | Hyperscalers (Amazon, Google, Microsoft) are enormously profitable; AI revenue is real but growing | Less fragile than dot-com, but smaller players (CoreWeave) deeply leveraged |
| Infrastructure Overbuilding | ~$500B in fiber optic cable; much unused for a decade | Wood Mackenzie tracking 134 GW of proposed U.S. data centers (up from 50 GW in 2024); Morgan Stanley projects $800B hyperscaler capex in 2026 alone | Classic speculative overbuild pattern — though fiber eventually proved useful |
| Debt Structure | Telecom companies overleveraged; debt markets closed | Data center developers using asset-backed securities, private credit; CoreWeave IPO's $23B valuation vs. 62% revenue from one customer (Microsoft) | Leverage mounting; risk migrating to utilities, pension funds, and private credit |
| Technology Obsolescence | Fiber technology was durable; protocol shifts (not hardware) disrupted business models | Air-cooled data centers already becoming obsolete as liquid cooling required for AI GPUs; retrofitting costs 7–10% premium over new builds | Rapid hardware obsolescence creates stranded asset risk within 3–5 years of construction |
| Demand Verification | No systematic vetting of demand projections | CAISO: "Nobody really knows. Nobody has been looking carefully enough at the forecast." (Joe Bowring, mid-Atlantic grid watchdog) | Speculative demand inflating utility forecasts; no standard vetting practice |
| Long-term utility of infrastructure | Fiber networks eventually used; internet still growing | AI computing demand is real and structural; question is timing and scale | Technology not vaporware, but current investment scale likely exceeds near-term monetizable demand |
The Bear Case: Man Group and Harvard Law
Man Group, the London-headquartered hedge fund managing over $170 billion in assets, published a detailed analysis in early 2026 characterizing the AI investment cycle as an "Inflection Bubble" — a pattern in which a genuine technological breakthrough becomes overlaid with speculative financial excess. Their assessment: "Power infrastructure built for 2024–2025 demand levels risks becoming stranded assets by 2027–2028." They identified a critical structural risk: "Risk is increasingly migrating away from tech company balance sheets and into institutions — utilities, insurers, data centre operators, private credit funds, pensions, and retail vehicles — that do not see themselves as betting on GPU cycles."
Harvard Law School's Environmental and Energy Law Program published a paper in March 2025 titled "Extracting Profits From the Public: How Utility Ratepayers Are Paying for Big Tech's Power," documenting how utility interconnection and upgrade costs for large data center loads become part of the utility rate base — paid by customers. The Brookings Institution corroborated the timing mismatch: data center developers are demanding power in 2 to 3 years, while planning, permitting, and constructing new generating and transmission facilities takes 8 to 10 years. "Where we're at today is in a world of constraint," as one former Microsoft energy executive put it.
The OpenAI financial picture illustrates the underlying monetization gap. OpenAI faces an estimated $1.4 trillion in projected computing costs over the next eight years while currently generating approximately $13 billion in annual revenue. The structural loss is enormous and persistent. OpenAI is not alone: most AI startups are operating at negative unit economics, relying on venture capital and hyperscaler cloud subsidies rather than profitable AI revenue.
The Bull Case: KKR and Infrastructure Permanence
KKR, one of the world's largest private equity firms and a major data center investor, published a detailed counter-argument in late 2025. Their key points deserve engagement. First, available data center space is forecast to be scarce through at least 2027, suggesting no current overbuilding in active markets. Second, AI hardware accelerators age quickly — each new GPU generation offers step-function performance gains — so excess capacity gets absorbed rapidly by new workloads rather than sitting idle. Third, the barriers to entry in data centers are substantial: power, land, grid connections, and permits are structural chokepoints that limit speculative overbuilding in ways the fiber optic boom was not constrained.
KKR's qualification is important: "Those who control the moats should reap the compounding returns" — but the stranded asset risk falls on those who don't. A hyperscaler-backed campus with a 20-year power purchase agreement and a committed enterprise customer base is a fundamentally different risk profile from a speculative developer like IVCM in Imperial Valley, whose developer's previous business history includes unresolved legal disputes in another state and whose financing, water commitments, and power sourcing remain unclear.
The forecasts are eye-popping: utilities saying they'll need two or three times more electricity within a few years to power massive new data centers. But the challenges — some say the impossibility — of building new power plants to meet that demand so quickly has set off alarm bells. One burning question: whether the forecasts are based on data center projects that may never get built — eliciting concern that regular ratepayers could be stuck with the bill.
Who Holds the Stranded Asset Risk?
This is the question that connects the AI bubble debate directly to San Diego ratepayers. When the dust from speculative overbuild settles — as it invariably does — who is left holding infrastructure that was built for load that never materialized?
The CPUC's Public Advocates Office identified the mechanism with precision: data center developers submit multiple speculative interconnection requests without financial commitment. The CEC uses these requests to develop official load forecasts. Those forecasts drive CAISO's transmission planning. CAISO approves transmission projects. Utilities build the projects. Costs are socialized across all CAISO ratepayers under FERC rules. If the data centers never build — or build smaller than projected — the transmission infrastructure remains, financed by residential customers who received no benefit from it.
Virginia provides a current-tense preview. Dominion Energy petitioned state regulators to raise base rates by 15% over two years, citing data center load growth. The mid-Atlantic grid watchdog Joe Bowring stated flatly: "There's speculation in there. Nobody really knows. Nobody has been looking carefully enough at the forecast to know what's speculative, what's double-counting, what's real, what's not." Residents in Philadelphia's data-center-adjacent districts have already absorbed electricity rate increases attributed to wholesale power cost increases driven by data center demand — before any of the speculative projects that drove the forecast ever connected to the grid.
Demand Lands at 60% of Projections
Some Imperial Valley data centers built; others stall or scale back. Powerlink serves real but oversized needs. Ratepayers pay for modest excess capacity. Cost borne over decades; bill impact manageable but real. This is roughly the dot-com fiber outcome — painful but eventually absorbed.
AI Demand Grows Into Capacity
Multiple large data centers build out in Imperial Valley. Powerlink operates at high utilization. Revenue from large-load customers offsets costs; downward pressure on per-unit rates. Renewable energy integration benefit realized. Ratepayers break even or marginally benefit over 20-year asset life.
AI Capex Cycle Contracts Sharply
AI monetization disappoints; hyperscaler capex cut; speculative data centers in Imperial Valley never built or abandoned mid-construction. Powerlink built but underutilized. Billions in stranded transmission costs socialized across CAISO ratepayers. San Diego customers — already highest-rate in nation — absorb accelerated cost increases. No regulatory mechanism in California currently prevents this outcome.
What California Has Failed to Do
The legislative record is damning. California entered 2025 with broad political consensus that data centers should pay their own grid costs. It exited 2025 with a law (SB 57) that merely authorizes the CPUC to study the question — after industry lobbyists stripped the bill's substantive cost-allocation provisions. A 46-page Little Hoover Commission report issued in February 2026 recommended more than a dozen protective measures for ratepayers and remains unimplemented.
Meanwhile, the CAISO's 2025–26 transmission plan — pending board approval as of May 2026 — recommends 38 additional transmission upgrades at an estimated cost of $7 billion, driven in part by data center load projections that the grid's own planning paper acknowledges may be inflated by speculative applications. The cost allocation methodology that governs how this $7 billion is split among ratepayers remains unchanged: transmission network upgrades are socialized across all CAISO customers under FERC rules.
The Brookings Institution identified the timing trap with precision: data center developers need power in 2–3 years. Utilities need 8–10 years to build the infrastructure to serve them. If the transmission is built on the 8–10 year timeline and the data center demand lands 3 years after request, there is a 5–7 year window in which ratepayers are servicing debt on infrastructure that isn't generating revenue. If the data center demand never fully materializes, that window is permanent.
Where will the power come from? For the next 5–10 years, the honest answer is: primarily natural gas, supplemented by Imperial Valley solar and geothermal where genuinely contracted and built. Nuclear — the only credible long-term clean baseload solution — is a post-2030 technology play at best, and California has no new nuclear in its pipeline. AI data centers that claim clean 24/7 power without a verified nuclear or geothermal contract are making a promise their physics cannot keep without a gas backstop.
Is the AI boom a bubble? The technology is real. The underlying demand for AI capability is real. But the financial architecture surrounding the infrastructure build-out bears the classic hallmarks of speculative excess: debt-financed overbuilding based on unverified demand forecasts, rapid hardware obsolescence creating stranded asset risk within years, monetization gaps between revenue and investment at key players, and risk migrating from well-capitalized tech companies to utilities, pension funds, and ratepayers who do not understand they are exposed.
The worst case for San Diego ratepayers is specific and concrete: the Golden Pacific Powerlink is built on a demand forecast that includes substantial speculative data center load; one or both major Imperial Valley data center projects fail or scale back dramatically; the Powerlink operates below design utilization; and the $1.3–2.3 billion capital cost is recovered over decades from SDG&E customers already paying the highest electricity rates in the continental United States — with no mechanism in current California law to recover those costs from the data center operators whose speculative applications inflated the forecasts that justified the project.
The comment period for the Golden Pacific Powerlink is open through early November 2026. The formal CPUC proceeding — where these questions can be placed on the official record and formally challenged — begins when SDG&E files its Certificate of Public Convenience and Necessity application before year-end. That filing triggers a 30-day protest window available to any member of the public. Formal intervenors in CPUC proceedings can demand that CAISO produce the demand modeling behind the need finding — and can challenge both the data center load assumptions and the cost socialization structure that puts residential ratepayers at risk.
Sources
-
Official — IEA
International Energy Agency. (2025). Energy and AI — Energy Supply for AI. [Global data center generation mix; natural gas 40%; IEA Base Case: 460 TWh 2024 → 1,000+ TWh 2030 → 1,300 TWh 2035.]
https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai -
Policy — Brookings Institution
Brookings Institution. (April 2, 2026). "Global energy demands within the AI regulatory landscape." [2024 data center power mix; 64% of incremental generation from fossil fuels through 2035.] Brookings.edu.
https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/ -
Policy — Brookings Institution
Brookings Institution. (November 3, 2025). "Boom or bust: How to protect ratepayers from the AI bubble." [8–10 year infrastructure timeline vs. 2–3 year data center power request; Lawrence Berkeley Lab / Brattle Group study context; Virginia Dominion 15% rate increase.] Brookings.edu.
https://www.brookings.edu/articles/boom-or-bust-how-to-protect-ratepayers-from-the-ai-bubble/ -
Financial — Man Group
Man Group. (2026). "The AI Bubble: Hidden Risks and Opportunities." [Inflection Bubble analysis; stranded asset risk 2027–2028; risk migration to utilities and pension funds.] Man.com.
https://www.man.com/insights/the-ai-bubble -
Financial — KKR
KKR Global Macro, Asset Allocation & Real Assets. (November 2025). "Beyond the Bubble: Why AI Infrastructure Will Compound Long after the Hype." [Bull case; hardware refresh cycles; JLL vacancy data through 2027; moat analysis.] KKR.com.
https://www.kkr.com/insights/ai-infrastructure -
Financial — Investing.com / Morgan Stanley
Morgan Stanley data cited in: "Is the AI Boom Starting to Look Like a Credit Bubble?" Investing.com. May 2026. [Big 5 hyperscaler capex: ~$800B in 2026, ~$1.1T in 2027.]
https://www.investing.com/analysis/ -
News — Fortune
Fortune. (November 15, 2025). "The AI bubble may be showing up in an unexpected place." [Utility demand forecast vetting; Joe Bowring quote; PPL Corp CEO quote; speculative interconnection requests.] Fortune.com.
https://fortune.com/2025/11/15/ai-bubble-electrical-grid-forecasts-demand-speculative/ -
Analysis — Inflection Bubble Report
Pascal's Substack. (December 2025). "The analysis confirms that the AI sector in late 2025 exhibits the classic hallmarks of an 'Inflection Bubble.'" [Stranded asset analysis; CoreWeave IPO data; Goldman Sachs 19% share; air-cooled to liquid-cooled retrofit cost premium.] p4sc4l.substack.com.
https://p4sc4l.substack.com/p/the-analysis-confirms-that-the-ai -
Academic — PMC / NIH
National Institutes of Health, PubMed Central. (2025). "Does DeepSeek curb the surge of energy consumption in data centers?" [Jevons Paradox in AI; centralized vs. edge computing shift; peer-reviewed analysis.] pmc.ncbi.nlm.nih.gov.
https://pmc.ncbi.nlm.nih.gov/articles/PMC12447567/ -
Analysis — S&P Global
S&P Global Market Intelligence. (March 2025). "Potential impacts of DeepSeek on datacenters and energy demand." [GPU utilization analysis; 15–18 GW/year global data center additions 2025–2029; 30–40% GPU-dedicated.] spglobal.com.
https://www.spglobal.com/market-intelligence/en/news-insights/research/potential-impacts-of-deepseek-on-datacentars-and-energy-demand -
Nuclear Deals — Introl / Multiple Primary Sources
Introl.com. (January 2026). "Nuclear power for AI: inside the data center energy deals." [Microsoft/Three Mile Island 835 MW; Google/Kairos SMR fleet; Amazon/Susquehanna; Meta 1–4 GW RFP; 10 GW+ contracted in 2025.] Introl.com.
https://introl.com/blog/nuclear-power-ai-data-centers-microsoft-google-amazon-2025 -
Official — CAISO
California Independent System Operator. (April 7, 2026). "Draft 2025–26 Transmission Plan: 38 infrastructure upgrades, $7 billion." [Load growth 15 GW by 2035, 20 GW by 2040; 74 GW and 107 GW installed resource capacity needed.] CAISO.com.
https://www.caiso.com/about/news/news-releases/ -
Legal / Harvard Law
Harvard Law School — Environmental & Energy Law Program. (March 2025). "Extracting Profits From the Public: How Utility Ratepayers Are Paying for Big Tech's Power." Cited in: Food and Water Watch. (March 2026). "The Urgent Case Against Data Centers." foodandwaterwatch.org.
https://www.foodandwaterwatch.org/wp-content/uploads/2026/03/RPT2_2602_DataCenterMoratorium.pdf -
News — Interesting Engineering
Interesting Engineering. (May 2026). "Why AI data centers are overwhelming the US power grid." [Google/Duane Arnold nuclear deal; DOE transmission gap study; nuclear as leading zero-carbon option.] interestingengineering.com.
https://interestingengineering.com/energy/ai-power-grid-data-centers-us -
Analysis — Tech Insider
Tech Insider. (April 2026). "AI data center power crisis." [Global data center consumption exceeds 1,000 TWh by end 2026; rack density 8 kW to 50+ kW trend; 30% on-site generation share.] tech-insider.org.
https://tech-insider.org/ai-data-center-power-crisis-2026/ -
Analysis — Sightline Climate / CTVC
Sightline Climate / CTVC. (February 2025). "DeepSeek hype vs. the hyperscalers." [Chain-of-thought 41% energy premium; AI token cost drop 280× in 2 years; Jevons Paradox analysis.] sightlineclimate.com.
https://www.sightlineclimate.com/research/deepseek-hype-vs-the-hyperscalers
Comments
Post a Comment