TJE Ventures

How Big Is the Problem, Actually?

Global data center electricity consumption in 2023: approximately 200–250 TWh, or roughly 1–1.5% of total global electricity generation. U.S. data centers alone consume an estimated 70–100 TWh per year — about 2% of U.S. electricity, comparable to the combined electricity use of all households in California and Texas.

That number sounds large and, in absolute terms, it is. Context matters, however. The entire global data center sector consumes less electricity than the U.S. residential air conditioning load. Aluminum smelting globally uses roughly twice as much electricity as data centers. The steel industry uses five times as much. Data centers are a meaningful contributor to electricity demand, but they are not the dominant industrial consumer that some coverage implies.

The more important question is the trajectory. Goldman Sachs estimated in 2024 that data center power demand will grow approximately 160% by 2030, driven almost entirely by AI infrastructure. That growth rate, if it materializes, would bring data centers to 3–4% of global electricity — a doubling of share in under a decade. The absolute numbers matter less than the direction and speed of the curve.

A useful benchmark for comparison:

• Global data centers (2023): ~240 TWh/year
• Global aviation: ~330 TWh/year (fuel equivalent)
• Global steel production: ~1,200 TWh/year
• Global residential air conditioning: ~2,000 TWh/year
• Global cryptocurrency mining (2023): ~120–150 TWh/year — roughly half of all data centers, for essentially no productive output

The cryptocurrency figure is worth sitting with. The Bitcoin network alone consumed approximately 120 TWh in 2023 — comparable to the electricity consumption of Argentina — to perform a function that consists entirely of solving computationally expensive puzzles that have no purpose except to prove that electricity was spent. This is a separate problem from productive data center use, but it occupies the same power grid and the same public perception.

The Good: What the Industry Actually Got Right

The Efficiency Revolution

The single most important environmental story in data center history is one that rarely makes headlines: between 2010 and 2020, global data center workloads increased approximately eightfold while total data center electricity consumption grew by only about 6%. The computing output per unit of electricity improved by a factor of roughly six over that decade. Nothing else in the industrial economy comes close to that efficiency trajectory.

The primary driver was the shift from distributed enterprise computing — thousands of companies running their own inefficient server rooms — to centralized hyperscale cloud data centers operated by Amazon, Microsoft, Google, and a handful of other major providers. A corporate server room from 2008 typically ran at 3–5% server utilization (processors sitting idle 95–97% of the time), consumed power regardless of workload, and operated cooling systems with a Power Usage Effectiveness (PUE) of 2.0 or worse. Every watt powering the servers was matched by another watt cooling them.

Modern hyperscale data centers operate at PUE of 1.1–1.2 — meaning 10–20 cents of cooling per dollar of computing, versus 100 cents in older facilities. Google’s flagship data centers consistently achieve PUE below 1.1. Server utilization in hyperscale facilities runs 60–80%, enabled by virtualization technology that runs dozens of independent workloads on a single physical machine. The same computation that required 10 servers in a corporate data center in 2010 runs on one server in a hyperscale facility today, using a fraction of the power.

This consolidation story deserves more credit than it receives. The migration of enterprise computing to the cloud was, among other things, one of the most significant energy efficiency improvements in the history of commercial computing. It happened quietly, driven by economics rather than regulation, and it is the primary reason data center electricity consumption did not grow in proportion to the explosion in global computing demand over the past 15 years.

Renewable Energy Procurement

The three largest data center operators — Amazon Web Services, Microsoft Azure, and Google Cloud — are also three of the largest corporate purchasers of renewable energy in the world. Their scale gives them the market power to sign long-term PPAs that fund new renewable generation capacity, not just purchase existing certificates.

• Google has matched 100% of its annual electricity consumption with renewable energy purchases since 2017 and is targeting true 24/7 carbon-free energy by 2030 across all its operations — a far more demanding standard than annual matching.
• Microsoft has committed to being carbon negative by 2030 and to removing all the carbon it has emitted historically by 2050. Its renewable energy portfolio includes long-term PPAs that have directly funded the construction of new wind and solar projects.
• Amazon reached 90% renewable energy matching in 2023 and is tracking toward 100% by 2025, backed by the largest corporate renewable energy portfolio in the world — over 500 wind and solar projects across 27 countries.

These commitments are not merely symbolic. The capital these companies commit to renewable PPAs genuinely finances new generating capacity that would not otherwise be built. When Google signs a 20-year PPA for a solar farm in Nevada, it provides the revenue certainty that makes the project bankable. The environmental benefit of that new capacity is real, even if the accounting of how it maps to the data center’s actual consumption at any given hour is less clean than the “100% renewable” label suggests.

Waste Heat Recovery

Data centers are, fundamentally, giant heat generators. Every watt of electricity consumed by a server exits as heat. For most of the industry’s history, that heat was simply rejected to the atmosphere as a waste product. A growing number of facilities, particularly in Europe, have turned this liability into an asset by recovering waste heat and integrating it into district heating systems.

• Google’s data center in Hamina, Finland pipes waste heat to the municipal heating network, warming homes and businesses in the surrounding city.
• Equinix and Fortum have partnered in Stockholm to recover heat from data centers and supply it to approximately 10,000 homes annually through the city’s district heating grid.
• In Denmark, data center waste heat now contributes meaningfully to district heating networks that serve a significant share of the country’s residential heating load.

The economics are not always favorable — waste heat recovery requires proximity to a district heating network, which exists in most northern European cities but is rare in the U.S. American cities are built around individual HVAC systems rather than shared thermal infrastructure, so the opportunity is largely untapped domestically. But it represents a genuine model for reducing the net energy cost of data center operations where the infrastructure exists.

The Bad: Real Problems That Deserve Honest Accounting

Water Consumption: The Hidden Cost

Electricity consumption is the data center impact that gets attention. Water consumption is the one that is systematically underreported and, in water-stressed regions, potentially more consequential.

Most large data centers use evaporative cooling: water is evaporated to remove heat from the facility, and that water is consumed — it does not return to the source. A 100 MW data center using evaporative cooling consumes roughly 1–5 million gallons of water per day, depending on climate, cooling design, and outdoor temperature. That is comparable to the daily water use of a city of 30,000–150,000 people, drawn continuously from the local water supply.

Water Usage Effectiveness (WUE), measured in liters of water per kWh of IT load, is the industry metric for cooling water intensity:

• Industry average WUE: 1.5–2.0 L/kWh
• Hyperscale best practice: 0.2–0.8 L/kWh
• Air-cooled facilities in cool climates: near zero
• Older enterprise data centers: 2.0–4.0 L/kWh

Google disclosed that its global data centers consumed 5.6 billion gallons of water in 2022. Microsoft’s Iowa data centers attracted significant local criticism for their water withdrawals during drought conditions — drawing millions of gallons per day from aquifers and municipal systems that agricultural and residential users were already competing for. A 2023 study from researchers at UC Riverside estimated that training GPT-3 consumed approximately 700,000 liters of fresh water. Inference at scale — running ChatGPT for hundreds of millions of queries — adds substantially more.

The location problem compounds the consumption problem. Many of the largest U.S. data center clusters are in the Southwest — Phoenix, Las Vegas, northern Virginia — precisely the regions where water is already the binding constraint on growth. Data centers locate there for cheap power, flat buildable land, and favorable tax treatment, with water consumption treated as a solvable engineering problem rather than a genuine constraint. In a region facing multi-decade drought and declining Colorado River flows, that calculus is increasingly untenable.

Backup Diesel Generators

Every significant data center maintains banks of diesel generators sized to power the entire facility in a grid outage. For a 100 MW data center, that means diesel generator capacity of 100 MW or more — equivalent to the standby capacity of a small power plant. These generators are not primarily an emissions problem during normal operations; they sit idle the vast majority of the time. The problems arise in two specific ways:

• Required testing. Diesel generators must be run under load periodically to verify they will start and operate reliably when needed. Most facilities run generators monthly for 30–60 minutes under full load. In a data center cluster with dozens of facilities — like the massive concentration in northern Virginia ’s “Data Center Alley” — this generates localized diesel exhaust events that are visible in air quality monitoring data and disproportionately affect nearby communities, which are frequently lower-income and majority-minority.
• Extended outages. During major grid disruptions (ice storms, hurricanes, prolonged outages), generators run continuously for hours or days. The 2021 Texas winter storm drove extended diesel operation across the state’s data center base. A 100 MW facility running on diesel for 24 hours consumes approximately 70,000–80,000 gallons of diesel fuel and emits roughly 700–800 metric tons of CO₂.

Battery storage and fuel cell technology are gradually replacing diesel generators in some facilities, but the capital cost and energy density of alternatives means diesel will remain the dominant backup power source for large-scale data centers for at least another decade.

Hardware Manufacturing and E-Waste

The embodied carbon in data center hardware — the emissions associated with manufacturing servers, networking equipment, and storage systems — is a significant fraction of the total lifecycle carbon footprint that most operator sustainability reports undercount or omit entirely.

Manufacturing a single server generates approximately 1,000–4,000 kg of CO₂e depending on configuration and manufacturing location. A 100 MW hyperscale data center contains tens of thousands of servers. The total embodied carbon in the hardware can equal or exceed several years of operational electricity emissions. For AI training clusters built around NVIDIA H100 GPUs — which require advanced semiconductor manufacturing processes with high energy and chemical intensity — the manufacturing carbon per unit of computing is substantially higher than for commodity servers.

Hardware refresh cycles drive e-waste at scale. Hyperscale operators typically retire servers after 3–5 years, not because they have failed but because newer hardware is more energy-efficient and the power savings justify the replacement. This creates a continuous flow of relatively young, functional equipment. The best operators refurbish and resell end-of-life equipment or partner with certified recyclers. The worst ship it to developing-world e-waste facilities where informal processing releases toxic materials — lead, mercury, cadmium, and brominated flame retardants — into the local environment.

The Ugly: Where the Problems Are Getting Worse

The AI Energy Shock

The shift to AI workloads is not a gradual evolution of data center power density — it is a step change that is straining grid infrastructure, exhausting transmission capacity, and compressing development timelines in ways the industry has not experienced before.

The power density difference between traditional computing and AI is stark:

• Traditional server rack: 5–8 kW per rack
• Standard cloud compute rack: 8–15 kW per rack
• AI training rack (8x NVIDIA H100): 40–50 kW per rack
• Next-generation AI rack (GB200 NVL72): 100–120 kW per rack

A hyperscale AI training cluster is not just a bigger version of a cloud data center — it requires fundamentally different cooling infrastructure, structural floor loading, and electrical distribution. The liquid cooling systems required for 100+ kW racks cannot be retrofitted into facilities designed for 10 kW air-cooled racks. New AI data centers are essentially being designed and built from scratch.

The energy cost of individual AI operations is also significantly higher than their conventional equivalents:

• A standard Google web search: approximately 0.3 Wh of energy
• A ChatGPT query (inference): approximately 2.9–10 Wh — 10–30 times a web search
• Training GPT-3 (one time): approximately 1,300 MWh — equivalent to the annual electricity use of 120 U.S. households
• Training a GPT-4 class model: estimated 50,000–100,000 MWh — the annual consumption of a small city

Training is a one-time cost per model. Inference — actually running the model for billions of queries per day — is where the ongoing energy demand accumulates. As AI assistants are integrated into search engines, productivity tools, and consumer applications at scale, inference energy becomes a continuous and growing load. The IEA projected in 2024 that a single year’s worth of AI query growth could add 10 TWh to global electricity demand — roughly the annual consumption of Greece.

The grid impact is landing unevenly. Northern Virginia, which hosts approximately 70% of the world’s internet traffic through its data center concentration, is facing transmission capacity constraints that are delaying new data center connections by 5–10 years. Dominion Energy has publicly stated that planned data center growth in its service territory requires a grid expansion it cannot complete fast enough to meet demand. Several counties in the region have imposed moratoria on new data center permitting. The same dynamic is playing out in Dublin, Singapore, and Amsterdam, where grid operators have paused new large-load connections.

The Renewable Energy Accounting Problem

When a major cloud provider announces it is powered by “100% renewable energy,” the statement is technically defensible but practically misleading in most cases. Understanding the gap requires distinguishing between three different standards:

• Annual Renewable Energy Credit (REC) matching. The company buys RECs equal to its total annual electricity consumption. Each REC represents 1 MWh of renewable generation somewhere on the grid. The data center runs on whatever mix of generation the local grid provides — which may be 70% coal at 2 AM on a Tuesday — while RECs purchased from a wind farm in Texas provide the accounting offset. This is the standard most “100% renewable” corporate claims are based on. It does not mean the data center is running on renewable energy at any particular moment.
• Locational and temporal matching (PPAs). The company signs long-term power purchase agreements with specific renewable generators, ideally in the same grid region as the data center. This is meaningfully better than REC purchasing: new generation capacity is actually built, and the physical supply relationship is more direct. Most large operators use this approach for a significant portion of their renewable procurement. It still does not guarantee that renewable energy is flowing to the data center at every hour.
• 24/7 Carbon-Free Energy (CFE). The company matches its electricity consumption with carbon-free generation on an hourly basis, in the same grid region, every hour of every day. This requires dispatchable carbon-free generation (nuclear, hydro, geothermal, or storage-backed renewables) to cover the hours when solar and wind are unavailable. It is the only standard that genuinely means the data center runs on clean power at all times. Google is the only major operator that has committed to this standard by 2030, and it has not yet achieved it.

The practical gap between annual REC matching and 24/7 CFE is largest at night and during winter, when solar production is zero and wind is variable, and the grid falls back on whatever dispatchable generation is available — which in most U.S. markets includes natural gas and, in some regions, coal. A data center operating on “100% renewable” RECs is drawing from the same grid as everyone else; the renewable claim is an accounting convention, not a physical reality.

This is not to say the REC and PPA approaches are worthless — they fund real renewable capacity and represent genuine progress over doing nothing. But the gap between the marketing claim and the physical reality is large enough to matter, and it is important context when evaluating corporate climate commitments in the data center sector.

Location Optimization for Cost, Not Climate

Data center siting decisions are driven by a hierarchy that rarely puts environmental impact at the top. The factors that actually determine where large facilities are built:

• Electricity price. Power is typically the largest operating cost for a data center, often 40–60% of total operating expense. Operators pursue the lowest available rates aggressively. This frequently means siting in regions with high fossil fuel generation, because coal and gas plants tend to produce cheap baseload power. West Virginia, Kentucky, and Wyoming have attracted data center interest specifically because their electricity rates are low — powered largely by coal generation.
• Tax incentives. States and localities compete aggressively for data center investment with property tax abatements, sales tax exemptions on equipment purchases, and income tax credits. Virginia’s aggressive incentive structure is a primary driver of Data Center Alley’s concentration. These incentives reduce the effective cost of siting in a given jurisdiction but provide no environmental benefit and frequently represent a significant transfer of tax burden to other local taxpayers.
• Land and permitting speed. Flat, developable land with permissive zoning and rapid permitting processes attracts data centers regardless of other environmental factors. Rural communities eager for economic development often provide both, without fully accounting for the water, power, and infrastructure demands that accompany a large facility.
• Fiber connectivity. Latency-sensitive workloads require proximity to network exchange points, which concentrates data centers in specific geographic clusters regardless of local grid cleanliness.

The result is a geographic distribution of data centers that reflects the cheapest land, power, and tax environment rather than the cleanest grid or most sustainable water supply. Operators who genuinely prioritize clean energy in siting decisions — locating near hydro in the Pacific Northwest or geothermal in Iceland — exist, but they are the exception rather than the rule in the hyperscale buildout underway right now.

What Can Actually Be Done

The technology to substantially reduce data center environmental impact exists. The gap between what is technically possible and what is widely deployed is large — and it is primarily a gap of economics, incentives, and institutional inertia rather than a gap in available solutions.

Advanced Cooling Technologies

Cooling accounts for 30–50% of total data center energy consumption in conventionally designed facilities. Newer cooling approaches cut that fraction dramatically:

• Direct liquid cooling (DLC). Cold plates attached directly to CPUs and GPUs remove heat at the source via circulating water or refrigerant, eliminating the need to cool the entire room to cool the chips. DLC systems achieve PUE of 1.03–1.1, compared to 1.4–1.6 for air-cooled equivalents. For AI GPU clusters generating 50+ kW per rack, DLC is no longer optional — air cooling simply cannot remove heat at those densities. The industry is transitioning to DLC for AI workloads by necessity.
• Immersion cooling. Servers are submerged directly in tanks of thermally conductive dielectric fluid that carries heat away without the energy cost of fans, compressors, or chillers. Single-phase immersion systems achieve PUE of 1.03–1.05; two-phase systems (where the fluid boils and condenses) achieve near 1.0 — essentially no overhead for cooling. Immersion cooling also eliminates dust ingestion, reduces fan-related component failures, and can extend hardware life. The barriers are upfront cost, specialized facilities, and the operational transition from rack-based to tank-based maintenance workflows.
• Air-side economization. In climates with sufficiently cool and dry ambient air, outside air can be used directly for cooling without mechanical refrigeration for a significant fraction of annual hours. Google’s data center in St. Ghislain, Belgium operates using outside air for cooling roughly 100% of the time, drawing no mechanical cooling at all. This approach works in temperate climates; it does not translate to Phoenix or Singapore.

Water Reduction and Closed-Loop Systems

The water problem is more tractable than it appears, because the primary driver — evaporative cooling — is not the only way to remove heat. Direct liquid cooling and air-side economization both achieve near-zero water consumption in appropriate climates. For facilities that must use evaporative cooling, several approaches reduce consumption substantially:

• Closed-loop cooling towers recirculate water through a heat exchanger rather than evaporating it directly, reducing consumption by 60–80%.
• Non-potable water sources — treated wastewater, recycled process water, or non-potable well water — substitute for municipal supply without drawing from drinking water sources. Microsoft has committed to using 100% recycled or non-potable water for data center cooling by 2030.
• Climate-appropriate siting — locating facilities in cooler, more humid climates where air-side economization is practical for most of the year eliminates or minimizes cooling water requirements entirely.

Waste Heat Recovery at Scale

The European model of integrating data center waste heat into district heating networks is the most underutilized low-cost decarbonization option available. The barriers in the U.S. are infrastructure (lack of district heating networks in most cities), geography (data centers tend to locate in suburbs or rural areas distant from residential heating demand), and commercial complexity (matching a continuous heat supply with variable residential demand requires thermal storage or flexible offtake agreements).

Where proximity and infrastructure exist — urban data centers near hospitals, universities, or industrial facilities with constant heat demand — waste heat recovery is commercially viable and environmentally compelling. A 20 MW data center providing waste heat to a hospital complex eliminates the hospital’s gas-fired boiler load while recovering value from what was otherwise a waste stream. The economics require a shared infrastructure investment that neither party can justify alone, which is exactly the kind of project that benefits from municipal coordination or utility program support.

24/7 Carbon-Free Energy and the Nuclear Option

The renewable energy accounting problem described above has a solution: commit to hourly-matched, grid-regional carbon-free energy rather than annual RECs. Achieving this requires dispatchable carbon-free generation that operates when solar and wind cannot — which means, in practical terms, nuclear power, geothermal, or storage-backed renewables.

The nuclear option is being pursued with increasing urgency by the major operators:

• Microsoft signed a 20-year agreement in 2023 to purchase power from the restarted Three Mile Island Unit 1 reactor in Pennsylvania (renamed Crane Clean Energy Center), providing baseload carbon-free power for its Mid-Atlantic data center operations. The deal is directly motivated by the 24/7 clean power requirement that RECs cannot satisfy.
• Google signed agreements in 2023 with Kairos Power and other small modular reactor (SMR) developers for future nuclear capacity, targeting deployment in the 2030s.
• Amazon purchased the Susquehanna Steam Electric Station (nuclear) in Pennsylvania from Talen Energy in 2023 — a direct acquisition of nuclear generating capacity for data center power, the first such deal in the industry.

Small modular reactors (SMRs) are the longer-term technology that the data center industry is watching most closely. SMRs in the 50–300 MW range are well-matched to single large data center campuses: they provide continuous carbon-free power at a scale that one facility can consume, without the transmission infrastructure required for a large conventional nuclear plant. Commercial SMR deployment in the U.S. is still 5–10 years away for most designs, but the alignment between what the technology provides and what data centers need is the clearest demand signal the nuclear industry has had in decades.

Hardware Efficiency and Lifecycle Extension

The embodied carbon problem is addressable through two levers that are underutilized in most operator sustainability programs:

• Longer hardware lifecycles. Extending server refresh cycles from the typical 3–5 years to 6–8 years cuts manufacturing carbon roughly in half on a per-year-of-use basis. The energy efficiency penalty for running older hardware exists but is often smaller than the embodied carbon savings from deferring replacement. Google and Meta have both extended server lifecycles as part of their sustainability programs, with published data showing the net carbon benefit is positive in most scenarios.
• Circular economy programs. Equipment retired from hyperscale facilities is often functional and can serve a second life in less demanding computing environments. Certified refurbishment programs that test, repair, and resell end-of-life servers displace new manufacturing and prevent electronics from entering the waste stream prematurely. The market for refurbished enterprise hardware is substantial and growing; the barrier is that operators must invest in structured decommissioning processes rather than bulk disposal.

Bottom Line

Are data centers as bad as claimed?

By the most-cited measure — electricity consumption — data centers are a real but not dominant contributor to global energy demand, representing roughly 1–1.5% of global electricity. The claim that “the internet uses more power than the airline industry” depends heavily on what you include and how you measure it; on a direct electricity basis, it is not accurate. The efficiency improvements achieved through hyperscale consolidation are a genuine and underappreciated environmental success story.

The honest qualifications:

• The AI inflection point is real and is accelerating power demand growth faster than at any point in the industry’s history. The trajectory to 3–4% of global electricity by 2030 is plausible, and the grid infrastructure implications are already visible in transmission-constrained markets.
• Water consumption is a more serious impact than electricity in many locations, and it receives far less scrutiny. Siting 100 MW data centers in water-stressed desert regions is a genuinely poor environmental decision that economic incentives currently reward.
• Corporate “100% renewable” claims are mostly accurate as annual accounting exercises and mostly misleading as descriptions of physical reality. The gap between what the press release says and what is actually happening at the grid level is large and matters.
• Cryptocurrency mining is the worst environmental actor in the data center space by a substantial margin — consuming electricity at a scale approaching half of all productive data centers for computations that have no utility except the consumption itself.

What would actually help?

• Honest renewable accounting. Move the industry standard from annual REC matching to hourly locational matching. Google’s 24/7 CFE framework is the right target; the rest of the industry should adopt it rather than sheltering behind annual averaging.
• Water disclosure and standards. WUE (Water Usage Effectiveness) should be a mandatory public disclosure metric alongside PUE. Many operators do not disclose it at all. Siting guidelines for new data centers in water-stressed regions should require closed-loop or near-zero water cooling systems.
• Nuclear integration. The Microsoft TMI deal and Amazon’s Susquehanna acquisition are the right model for large operators with 24/7 carbon-free commitments. SMR development should be accelerated specifically for the co-location use case.
• Diesel generator transition. Battery energy storage systems (BESS) and fuel cell technology should replace diesel backup power on an accelerated schedule, prioritized in communities with air quality concerns near data center clusters.
• Waste heat mandates in new construction. Cities granting data center permits should require waste heat recovery infrastructure where district heating networks exist or are planned. The European model works; it requires municipal coordination to implement.
• Hardware lifecycle standards. Embodied carbon in server manufacturing should be included in data center sustainability reporting. Voluntary extension of hardware lifecycles to 6+ years should be recognized as a meaningful emissions reduction strategy.

Data centers are not uniquely villainous. They are a large and growing industrial sector with real environmental impacts, real solutions available, and a gap between the two that is mostly explained by economics and incentives rather than a lack of technology.