Intelligence as,
Infrastructure.
We are currently using the world's most vulnerable hardware, powered by the world's most volatile energy. Measurement is the only way to prove we don't need to burn the planet to power the future. Our data proves we can do world-class science — we have the capacity to do hard things and make them better than it was. Now we must build the infrastructure to secure that future.
AI infrastructure is locking in 30–40 years of fossil dependency through new gas plants while consuming billions of gallons of water for cooling. Evaporative systems use 3–5x more water than closed-loop alternatives, straining drought-prone communities that host data centers.
Data center demand is projected to consume up to 12% of total US electricity by 2028, already driving regional utility bill increases of $11–$16/month and capacity price spikes of 833% in critical markets like Virginia.
While the US faces domestic renewable rollbacks and energy supply chain risks (evidenced by the 2026 Strait of Hormuz crisis), China is executing a "Dual-Track" strategy — surpassing 2030 renewable targets six years early and decoupling AI from urban grids via the "Eastern Data, Western Computing" initiative.
We must transition from "Phase 1" (unregulated growth) to "Phase 2" (Measurement and Transparency). Mandating granular, per-job energy reporting and incentivizing "temporal flexibility" (shifting training to peak renewable hours) is required to prevent 30-year fossil fuel infrastructure lock-in.
AI is one of the most powerful tools humanity has built. But its physical infrastructure is growing unchecked — impacting electricity grids, water systems, and communities. This isn't new, but the scale is unprecedented.
Internet boom. Data centers expanded rapidly. First wave of energy concern.
Efficiency gains matched demand growth. Cloud consolidation, better chips, virtualization. The industry solved its energy problem — temporarily.
AI broke the efficiency curve. GPU-intensive training overwhelms the efficiency gains that kept demand flat for a decade.
The pattern: Data center energy was a concern in 2005, was managed through efficiency in the 2010s, and is now growing faster than efficiency can compensate — driven by AI training and inference at unprecedented scale. Source: Lawrence Berkeley National Lab, IEA Energy and AI 2026
Data centers projected to consume 6.7-12% of total US electricity by 2028.
Virginia: Bills up $11-16/month. PJM auction up 833%. 78% of voters blame data centers.
Ohio: Bills up ~$16/month. 12% above national average.
National: 21M households behind on utility bills. Data centers = 63% ($9.3B) of PJM capacity bill.
Sources: CNBC, Brookings, EESI, Virginia SB 253
Tech pledged Net Zero. AI broke those pledges.
Guardian analysis: actual emissions 7.6x higher than reported (2020-2022). Morgan Stanley: 2.5B tons CO2 by 2030.
Sources: Sierra Club, TechInformed, Morgan Stanley
Land: Hundreds of acres consumed, often zoned for agriculture/housing.
Noise: Industrial cooling systems affecting nearby residential areas.
Grid: Competing with hospitals, schools, homes for electricity. Who gets priority?
Jobs: A facility using as much power as a small city may employ 50-100 people.
Sources: Lincoln Institute, Consumer Reports
AI chips generate heat. Cooling them requires water. But the real concern isn't "per query" — it's where and when that water is consumed.
Per query: ~500ml per 20-50 ChatGPT prompts (UC Riverside estimate, methodology contested). More recent data: Google AI search ~10ml/query, Mistral ~3.5ml/response.
Google alone: 6.4 billion gallons in 2023. Council Bluffs, Iowa peaked at 2.7M gallons/day in summer 2024.
Projected US total: Could double or quadruple by 2028 to 150-280 billion liters/year.
Data centers use water when and where it's most scarce. Cooling demand spikes on the hottest days — exactly when farmers and residents need water most. In Arizona, Iowa, and Texas, this creates direct resource competition.
Policy should focus on "Peak Daily Withdrawal Limits" during droughts and heat waves, not just total annual gallons.
The "water bottle" headline applies to evaporative cooling — older tech that's cheap but consumes water by evaporating it. Like a giant swamp cooler.
Cheap, evaporates water. Source of "water bottle" headlines. Common in older facilities.
Like a car radiator. Near-zero water. More electricity for fans. The solution for water-stressed regions.
Most focus on water at the data center. But the majority happens at the power plant. Coal and gas plants consume water to generate electricity.
Solar interconnectivity solves both. Solar and wind use virtually zero water. Moving AI to solar grids addresses water AND carbon simultaneously.
"The 'one water bottle per chat' headline is a symptom of outdated cooling tech. We are using 20th-century evaporation to cool 21st-century chips. Mandate transparency, incentivize dry-cooling + solar — grow scientific capacity without draining water tables."
Sources: Undark, Brookings, EESI, ScienceDirect
We've seen this before. Every transformative technology goes through the same cycle:
Early cars: 8 MPG. No emissions controls. Leaded gasoline. Nobody measured the impact because the technology was too exciting.
CAFE standards (1975). EPA fuel economy labels. Catalytic converters. You can't optimize what you can't measure. Regulation created visibility.
Fuel injection. Aerodynamics. Weight reduction. Once measured, the industry optimized — 8 MPG became 30 MPG without sacrificing performance.
Hybrids. EVs. Regenerative braking. The technology that caused the problem became the solution — when policy pushed it there.
AI compute is in Phase 1. Nobody measures fleet utilization. Nobody reports per-job energy. Nobody knows if a GPU is right-sized to its workload.
We need to move to Phase 2 — measurement and transparency — before we can get to optimization. That's what our H200 data demonstrates: the measurement tools exist, the waste is real, and the path to efficiency starts with visibility.
We cannot fix what we do not measure. The energy cost of unoptimized AI compute is invisible across the industry.
Total across 6 tracked experiments (~13 hours of GPU time)
Carbon emissions at US average grid intensity
Annual energy wasted on idle GPUs across 100K researchers
Conservative estimate based on measured primary data
American homes powered for a year — on idle GPUs
One hyperparameter sweep = driving 265 miles
| Scale | Energy | CO2 | Equivalent |
|---|---|---|---|
| Our clinical research (6 runs) | 15.4 kWh | 5.7 kg | Half a day of home electricity |
| Full hyperparameter sweep | 288 kWh | 106 kg | Driving 265 miles |
| GPT-4 training (estimated) | 50,000,000 kWh | ~20M kg | Powering 4,500 homes for a year |
Logarithmic scale. Our data is measured. Industry estimates from published reports and researcher analysis.
Sources:
• Luccioni, S., Jernite, Y., & Strubell, E. "Power Hungry Processing: Watts Driving the Cost of AI Deployment." ACM FAccT 2024
• de Vries, A. "The Growing Energy Footprint of Artificial Intelligence." Joule, 2023
• Stanford HAI. "AI Index Report 2025 — Training Compute & Carbon Emissions." hai.stanford.edu — GPT-3: 588t CO2, GPT-4: 5,184t, Llama 3.1 405B: 8,930t
• IEA. "Electricity 2026 & Energy and AI." iea.org — Data centres: 415 TWh (2024), projected 945 TWh by 2030
• Epoch AI. "How Much Energy Does ChatGPT Use?" epoch.ai
Note: Foundation model estimates vary widely. Google TPU energy is per-chip more efficient than NVIDIA GPUs but total training uses thousands of chips. Anthropic does not disclose training compute. Bar widths are approximate log-scale representations.
We ran real clinical AI research on an NVIDIA H200 — the most powerful GPU in production — and tracked every watt, every byte, every cycle.
Every experiment tracked with real instrumentation. Not estimated. Not theoretical.
| Run | Time | Energy | CO2 | GPU % | Memory |
|---|---|---|---|---|---|
| LMM standalone | 104 min | 2.68 kWh | 0.99 kg | 11% | 1.72 / 150 GB |
| GRASP+LMM | 103 min | 2.66 kWh | 0.98 kg | 11% | 1.73 / 150 GB |
| GRASP+LMM+codemap (lr=5e-4) | 177 min | 4.60 kWh | 1.69 kg | 14% | 2.58 / 150 GB |
| GRASP+LMM+codemap (lr=1e-4) | 181 min | 4.67 kWh | 1.72 kg | 12% | 2.58 / 150 GB |
| GRASP+LMM+codemap (3 fixes) | 161 min | 0.61 kWh | 0.22 kg | 14% | 1.67 / 150 GB |
| LMM standalone + batch=256 | 42 min | 0.17 kWh | 0.06 kg | 16% | 3.39 / 150 GB |
The last row is the same model, same data, same architecture — with one parameter changed: batch_size=32 → 256. GPU utilization jumped from 11% to 16%. Memory usage doubled from 1.7 GB to 3.4 GB. Training time dropped from 104 minutes to 42 minutes. Energy fell from 2.68 kWh to 0.17 kWh — a 15x reduction.
Utilization isn't a fixed property of the hardware — it's a function of how the workload is configured. With larger datasets (MIMIC-IV, 200K+ patients), time series data, and full hyperparameter sweeps, utilization climbs further. This table captures early-stage research prototyping, not peak capacity.
codecarbon (energy/CO2 from actual sensor readings) + pynvml (GPU metrics from NVIDIA driver)
NVIDIA H200 NVL, 150 GB HBM3e, CUDA 12.6
MIMIC-III (46,520 patients, public clinical data, Beth Israel Deaconess)
All code open source via PyHealth. Any researcher with GPU access can replicate.
The chip drawing 93W instead of 700W is good engineering. NVIDIA designed the H200 to scale its power draw to the workload. When the model needs less compute, the chip draws less power. It doesn't waste 607W heating empty silicon. That's efficient hardware design working as intended.
This is a shared academic cluster — when our job finishes, other researchers can use the same GPU. The chip isn't sitting idle between jobs in the same way a dedicated enterprise allocation would. We've generally been among the most active users, but we have no way to confirm fleet-wide utilization because that data isn't collected.
The infrastructure provisioning. The datacenter built cooling, power delivery, and physical space rated for 700W per chip. Whether our job draws 93W or 700W, that cooling infrastructure exists, consumes baseline power, and occupies space that could serve smaller, more efficient hardware.
The allocation mismatch. Our model used 1.7 GB of 150 GB memory. We could have selected a smaller chip — an A100 (80 GB, ~400W TDP) — and achieved the same scientific result. The cluster offers H200s because they handle everything, but "handles everything" means most research jobs are over-provisioned by default.
The measurement gap. Cluster administrators likely have access to fleet utilization data through scheduling and monitoring tools. But that data isn't published to users or reported publicly. As a researcher, I can measure my own job's utilization — I cannot see the fleet average. If even academic clusters with transparent governance don't surface aggregate utilization to their users, enterprise datacenters with commercial incentives to obscure waste certainly won't without a reporting mandate.
On this cluster, researchers could select different hardware tiers. Here's what our workload actually needed vs what was allocated:
| Hardware | Memory | TDP | Approx. cost | Runs our model? | Right-sized? |
|---|---|---|---|---|---|
| NVIDIA H200 NVL (what we used) | 150 GB | 700 W | ~$40,000 | Yes | No — 88x memory over-provisioned |
| NVIDIA A100 (available on cluster) | 80 GB | 400 W | ~$15,000 | Yes | Closer — still 47x over |
| NVIDIA RTX 4090 | 24 GB | 450 W | ~$1,600 | Yes | Better — 14x over |
| NVIDIA RTX 4060 (right-sized) | 8 GB | 115 W | ~$300 | Yes | Yes — 4.7x headroom |
Our model used 1.7 GB peak memory. An 8 GB consumer GPU with 4.7x headroom runs the same experiment at ~6x lower power provisioning and ~130x lower hardware cost. The science is identical. The energy footprint is dramatically different.
Note: Training speed would be ~1.5-2x slower on the RTX 4060 due to lower memory bandwidth. For a 42-minute experiment, that means ~60-80 minutes. Acceptable for research iteration, not for production inference.
We didn't go to the moon just to walk on rocks. We went to prove we could master the physics of the impossible. Along the way we invented the cordless drill, the integrated circuit, and the water purifier. The waste isn't the electricity that wasn't used — it's the infrastructure ghost built around it.
The problem isn't the chip. The H200 scaling down to 93W is good engineering — NVIDIA designed it to draw only what the workload needs. That headroom gives researchers the freedom to scale. If the H200 gets you a mortality prediction in 42 minutes instead of 3 hours, that's the hard thing we want to keep doing.
The problem is the ghost target. When a utility planner in Northern Virginia or Ohio sees a new data center request, they look at the nameplate capacity. 10,000 H200s installed? The grid reserves 7 megawatts of capacity. But if those chips are doing research or inference or cloud workloads, scaling to 93W like our case, the actual load is 0.93 megawatts.
The utility builds a gas plant and raises residential electric rates for 6.1 megawatts of power that doesn't exist. There is a massive difference between a local circuit being ready for a surge and a national grid building 30-year gas plants for a surge that our data shows happens a fraction of the time.
This isn't hypothetical. PJM — the grid operator for Virginia, Ohio, and 11 other states — received 166 GW of forecast peak load growth, roughly 90 GW from data centers. Industry analysts say actual need is closer to 65 GW. Forecasts are overstating demand by ~38%. The "ask-versus-accepted" gap is 43% before forecasts are even published. By summer 2026, PJM will have "just enough power to keep the grid reliable" — partly because they're planning for phantom load that may never materialize.
Sources: Modo Energy (PJM forecast), Utility Dive, NRDC, WRI
If we just tell people to "use a smaller chip," we ignore the fact that speed saves lives — especially in clinical research. The H200 got us a mortality prediction in 42 minutes. On a consumer GPU, that's 80 minutes. In a hospital, that time difference matters.
The goal isn't to restrict compute. It's to measure the efficiency so we can pack more science into the same grid — rather than building new gas plants for "peaks" that rarely happen. We're too smart to waste the energy we're fighting for in the Strait of Hormuz.
Exploring new architectures. This is fine — it's discovery.
Full datasets, larger batches. Utilization climbs naturally.
Full hyperparameter sweeps, MIMIC-IV. The H200 earns its keep.
Some companies ARE 100% utilized — they are the rationale for the Stargate project in Argentina. But most enterprise AI at this level doesn't need 100% of an H200's TDP. There are moments it does. The utility companies treat the 93W research run and OpenAI's 700W training run as the same 700W ghost target on the grid.
The H200 scales from 93W to 700W dynamically. That's precision engineering. If we focus on infrastructure awareness — showing that chips can scale down — we prove that solar interconnectivity is possible. If chips can match the solar cycle, the grid can too.
Nameplate capacity. 10,000 H200s = 7 MW reserved. They build gas plants for the peak. Raise residential rates to fund it. The "ghost target" becomes a 30-year fossil fuel commitment.
Actual load: 0.93 MW. Peak surges happen but they're brief and predictable. We don't need a gas plant for the surge — we need batteries and solar interconnects to buffer it. Move the training surge to daytime solar. Run inference work in the background.
Data centers are lying by omission. They know that while some chips are surging, 80% of the fleet could be idling. They're hoarding grid capacity like a landlord hoards empty apartments — keeping the prices high and carbon-heavy while the rest of us just want to get the work done.
Data centers should be legally required to report actual peak vs provisioned load. In March 2026, the Strait of Hormuz blockade makes every watt a strategic asset. We can't let ghost targets hoard energy that hospitals and homes need.
Residential bills should not subsidize ghost infrastructure. If a data center's provisioned load is 10x higher than their actual usage, they should pay a surcharge. That money funds the solar interconnects and batteries needed to bridge the energy gap — so the local community isn't paying for industry inefficiency.
Sources: Utility Dive (off-grid risks), ENR (large load rules), S&P Global (22% demand rise)
DeepSeek isn't just "Chinese ChatGPT." It's a masterclass in algorithmic efficiency. Their Mixture of Experts architecture only wakes up about 5% of its parameters for any given query. It does world-class science while drawing a fraction of the power of a brute-force model like GPT-4. They don't need a 24/7 gas baseline because their models are smart enough to scale with the solar cycle.
Meta is building 7 new gas plants in Louisiana just to keep its ghost targets from crashing the local grid. Brute force. Fossil lock-in. Residential bills up $16/month in Virginia.
Under the 15th Five-Year Plan, China moved its intelligent computing hubs directly next to the largest wind and solar farms in Inner Mongolia and Gansu. 15+ UHV lines pipe green energy directly into AI. By 2026, China's new computing hubs target 80% green electricity. They didn't just build data centers — they redrew the map.
Reclaiming the ghost gap is how we fund the next generation of American solar and nuclear without slowing down the science. Transparency forces the hyperscalers — Google, Meta, Microsoft — to admit they don't need the Louisiana gas plants if they just optimized their fleet. If we had transparency, we'd see we don't need a new gas plant for the surge. We need batteries and solar interconnects to buffer it. We need the precision to move the training surge to daytime and the inference work to the background. We have the capacity to do hard things — but we're too smart to let them use the surge as a shield for fossil fuel lock-in.
AI models have Model Cards. Compute should have Climate Compute Cards.
Every transformative technology eventually gets a disclosure standard. AI compute is overdue.
Cars got fuel economy labels. Manufacturers had to disclose MPG. Efficiency went from 8 to 30 MPG.
Appliances got efficiency ratings. Consumers could compare. Market rewarded efficiency.
Mitchell et al. proposed standardized documentation for AI models — bias, training data, intended use, limitations.
Every AI workload discloses its energy footprint, ghost ratio, and grid impact. The missing standard.
Source: Mitchell et al. "Model Cards for Model Reporting." FAT* 2019
| AI Model Card discloses | Climate Compute Card discloses |
|---|---|
| What the model was trained on | What hardware it ran on |
| Known biases and limitations | GPU utilization and memory usage |
| Intended use cases | Energy consumed (kWh) |
| Performance metrics | CO2 emitted (kg) |
| Who built it and when | Grid carbon intensity at time of run |
| Ethical considerations | Ghost ratio: TDP vs actual power draw |
Ghost Ratio = TDP / actual power draw. Higher = more infrastructure built for capacity not used.
Measured with codecarbon + pynvml | Primary data, not estimates
"Right-sized?" is the wrong question. The H200 scaling to 93W IS right-sizing — the chip does it automatically. The problem isn't the hardware choice. It's the infrastructure built around the nameplate TDP that the chip rarely hits. Ghost Ratio exposes that gap. 7.5x means the grid provisioned 7.5x more capacity than was actually used.
This is already a known problem at the grid level. PJM received 60 GW of data center capacity requests but accepted only 34 GW — a 43% cut for "phantom load." Dominion Virginia reports an 82% load factor for data centers, but some submissions assume 100%. The Ghost Ratio brings this visibility down to the chip level, where the gap is even larger.
Sources: Utility Dive (PJM), WRI, Grid Strategies
codecarbon and pynvml are open source, free, and work on any GPU. We built this into our research pipeline in a day. The barrier isn't technical — nobody requires it.
A researcher at 11% during testing and a hyperscaler permanently at 11% across 10,000 GPUs are very different stories. The CCC provides context — phase of work, time of run, grid source. It rewards efficiency without punishing exploration.
Fleet-wide Ghost Ratios become visible and actionable. If Google's average Ghost Ratio is 5x across 100,000 GPUs, that's the data a utility planner needs to stop building gas plants for peak loads that never happen.
CAFE standards for cars (1975). Energy Star for appliances (1992). LEED for buildings (1998). Model Cards for AI models (2019). Climate Compute Cards for AI infrastructure (2026).
The demand is outrunning clean energy supply. The gap is being filled with natural gas, and Big Tech is moving to control energy production itself.
Largest single source powering US data centers in 2024. Growing as AI demand outpaces renewable buildout.
Still a major source globally. US grid mix includes coal in PJM region (Virginia, Ohio) where data centers concentrate.
Solar, wind, hydro, existing nuclear. Growing but not fast enough to match AI demand growth of 12%/year.
Despite Net Zero pledges, the AI boom is locking in decades of new fossil fuel infrastructure. Demand for gas turbines has pushed delivery timelines out to 2028+.
Meta is paying for 7 new natural gas plants (5.2 GW total) to power its Louisiana data center — the largest single data center facility planned in the US.
Across PJM: Almost all new gas-fired plants secured in fast-track procurement won't come online until 2030 or later. Once built, they operate for 30-40 years — locking in fossil fuel dependency through the 2060s.
Data center developers have asked the Trump administration for exemptions from pollution rules for new gas plants, arguing AI demand is a national security priority.
The renewable rollback compounds the problem.
US solar installations fell 14% in 2025 after Trump rollbacks hit the sector
Federal solar tax credit (25D) eliminated December 31, 2025 via the "One Big Beautiful Bill"
Interior Department paused offshore wind projects under construction
New "project density" rules disqualify many solar/wind projects from federal land permits
US pulled from international renewable energy and climate organizations (Jan 2026)
AI demand is growing at 12%/year. Renewable capacity is contracting. The gap is filled with gas.
Sources: Insurance Journal (Meta gas plants), Grist (pollution exemptions), Semafor (solar -14%), Third Way (rollback timeline), SEIA (OBBB)
Unable to secure enough clean energy from the grid, tech companies are moving to control energy production directly — primarily through nuclear.
| Company | Nuclear deal | Scale | Timeline |
|---|---|---|---|
| Microsoft | Restart Three Mile Island with Constellation Energy | 835 MW, $16B (20-year PPA) | Target 2028 |
| Backing Kairos Power SMRs | 500 MW development agreement | Hermes 2: 2030 | |
| Amazon | $500M investment in X-energy SMRs + Susquehanna campus | $20B+ total AI campus | 2030+ |
| Meta | RFP for new nuclear generation | 1-4 GW requested | TBD |
| Combined: Big Tech signed 10GW+ new US nuclear in the past year | SMRs are 5+ years from commercial operation | ||
Sources: IEEE Spectrum, Introl, Marketplace, IAEA
4,088 facilities across 26 tracked states. Solid = existing, dashed = planned. Data from Visual Capitalist, Axios, DataCenterMap (2025-2026).
Sources: Visual Capitalist, DataCenterMap.com (4,088 facilities), Axios
AI compute policy is energy policy is foreign policy. The demand signal from data centers flows through a chain that ends at shipping lanes and military operations.
AI compute policy is energy policy is foreign policy. The demand signal from data centers flows through a chain that ends at shipping lanes and military operations.
"Drill Baby Drill" isn't just a slogan — it's sustained by two converging demand signals.
New terminals: Corpus Christi expansion (Mar 2025), Plaquemines (Dec 2024), Golden Pass (early 2026).
Growing 12%/year. Fastest new source of US electricity demand. Natural gas fills 40%+ of data center power.
Growing 15-19%/year. US is world's largest exporter. Hormuz crisis accelerates demand from Japan, Korea, India.
Both create political support for continued drilling. Both require investment in gas infrastructure. Both lock in fossil dependence for decades.
The connection to AI policy: If AI data centers shift to renewables, the domestic demand signal for gas weakens. This doesn't eliminate LNG exports, but it removes the fastest-growing justification for new gas infrastructure. Policy that makes AI compute more efficient is policy that reduces the demand signal for fossil expansion.
Sources: EIA, Natural Gas Intel, Wolf Street
US and Israel conducted strikes on Iran (Feb 28, 2026). Iran closed the Strait of Hormuz to commercial shipping.
Tankers rerouting via Cape of Good Hope: +10-14 days per voyage. War-risk insurance: +$250K per transit.
Sources: Wikipedia, Al Jazeera, CNBC, Bloomberg
Japan imports 85%+ of all energy. The Hormuz closure directly threatens national survival.
Japan is pushed toward US LNG — which is available, at US-set prices. Energy vulnerability is geopolitical leverage.
Sources: S&P Global, Energy Tracker Asia, CSIS
With the Strait of Hormuz closed and Houthi attacks resuming in the Red Sea (announced Feb 28), tankers must reroute around Africa:
Persian Gulf → Strait of Hormuz → Indian Ocean → destination
Persian Gulf → around Africa → Cape of Good Hope → Atlantic → destination
+10-14 extra days per voyage
Longer route = more fuel burned per tanker, slower delivery cadence, fewer trips per year per ship. The same fleet delivers less oil at higher cost — and burns more fossil fuel doing it. The reroute itself increases global emissions while reducing supply.
Sources: Middle East Insider, Maritime News
Energy costs passed to consumers and industry. Economic contraction.
Available immediately. At US-set prices. Deepens energy dependence on the US.
Politically difficult post-Fukushima. Years to restart safely.
Right answer long-term. Not fast enough for immediate crisis.
The geopolitical leverage: A country that imports 95% of its oil from a region controlled by US military power is a country that follows US policy preferences. Japan's energy vulnerability is, from a realpolitik perspective, a feature not a bug of the current system. The Hormuz closure pushes Japan (and South Korea, Taiwan, India) toward US LNG exports — which are growing at record pace. The military action that disrupted traditional supply routes simultaneously creates demand for American gas.
AI compute efficiency alone doesn't end fossil fuel dependence. LNG exports, transportation, heating, and industrial demand continue regardless. But AI is the marginal demand driver — the fastest-growing new source of electricity consumption. Removing it from the fossil equation changes the marginal economics of new gas plants.
A gas plant that's justified by AI demand + LNG exports + residential growth might not be justified by LNG exports + residential growth alone. The AI demand is the tipping point in many capacity decisions. Policy that addresses it is policy that shifts those decisions.
The chain starts with demand. If AI data centers operate more efficiently, the demand signal for new gas infrastructure weakens:
"AI is a tool worth powering. But powering it with fossil fuels locks in 30-40 years of carbon infrastructure, requires military control of energy supply routes, and concentrates energy production in the hands of the same companies that control the AI. The alternative — renewables, right-sized compute, temporal flexibility — breaks the cycle. It's not anti-AI. It's anti-lock-in."
Not a threat narrative — a competitive reality check.
China isn't just outpacing the U.S. — they're playing a different game. While the U.S. spent 2025-2026 navigating policy rollbacks, China leveraged a state-led "Dual-Track" strategy and hit their climate targets six years ahead of schedule.
Surpassed their 2030 target of 1,200 GW in mid-2024 — six years early
US expected: ~44 GW in 2026. China deploys 7-8x more solar annually
More than half of global total. US: ~150 GW
While the US worries about data centers in Virginia raising local bills, China is executing a national strategy to solve the exact same problem.
Move energy-hungry AI clusters from power-constrained coastal cities (Shanghai, Shenzhen) to resource-rich western regions (Inner Mongolia, Gansu, Ningxia).
Projected 400 GW of spare renewable capacity in the west by 2030. Data centers built directly next to solar/wind farms — the "Solar Interconnects" this report advocates for.
15 new Ultra-High Voltage (UHV) transmission lines between 2026-2030. One Tibet-to-south project delivers 43 billion kWh/year of green power to megacities.
China is not purely green. They're running a "Dual-Track" system: projected 1,333 GW of coal by end of 2026. But unlike the US where coal is often primary fuel, China increasingly uses coal as a strategic reserve to balance the intermittency of their massive wind and solar fleets. Coal as backup, not as base load.
| Factor | United States (2026) | China (2026) |
|---|---|---|
| Policy | Uncertainty: OBBB rollbacks, solar tax credits eliminated | Centrally mandated 15th Five-Year Plan, $580B grid investment |
| Manufacturing | Reshoring slow, struggling to build domestic solar cells | Absolute dominance, price war crashing global panel costs |
| Permitting | 3-5 year interconnection queues for new projects | Fast-tracked UHV "Green Lanes" for state-linked projects |
| Solar deployed (2025) | ~44 GW | 317 GW (7-8x more) |
| AI compute strategy | Data centers wherever land is cheap, powered by local grid (gas) | National plan: move compute to renewable surplus regions |
China's progress isn't a threat narrative — it's a competitive benchmark. They figured out that to win at AI, you have to win at green energy interconnectivity first.
If the US doesn't adopt mandatory transparency and right-sized compute, we will lose the AI race not because our models are worse — but because our grid is too expensive and inefficient to power them competitively.
China solved the "Resource Squeeze" by geographically decoupling AI compute from coastal population hubs. The US can do the same — but it requires policy that the current administration is actively dismantling.
If the Strait of Hormuz is the energy chokepoint, Taiwan is the compute chokepoint. In 2026, these are no longer separate issues — they are the same single point of failure.
of Japan's oil imports pass through here. Closed by Iran, Feb 28, 2026.
Controls: oil, LNG, helium (critical for chip manufacturing)
Impact: energy prices, shipping costs, grid stability across Asia
Duration: indefinite — Iran controls the coastline
of high-end AI chips (2nm-5nm) are manufactured by TSMC in Taiwan.
Controls: every H200, B200, and next-gen AI accelerator
Impact: if TSMC stops, the global AI supply chain ceases to exist
2nm capacity: 100% pre-booked through 2028 (Apple, NVIDIA, OpenAI)
The most immediate Taiwan threat in March 2026 isn't a missile — it's a blackout.
Taiwan's LNG reserves as of March 2026
of Taiwan's gas from Qatar/UAE — cut off by Hormuz closure
Nuclear units operating — last shut down mid-2025
of 2nm chip production capacity at risk
"We are betting the entire AI revolution on an island that is 11 days away from energy collapse if the Middle East stays closed. This isn't just a military risk — it's an infrastructure oversight."
Sources: CommonWealth Magazine (Taiwan gas crisis), Atlantic Council (energy resilience), LiveUAMap/Politico (11-day figure), Domino Theory (energy dependence)
The energy blackout scenario is the immediate risk. But the deeper threat is China's long-standing territorial claim over Taiwan — and 2026 conditions that make action more likely than any time since 1996.
China has conducted record military exercises around Taiwan in 2025-2026. PLA Navy now has more warships than the US Navy. Amphibious landing capability has tripled since 2020. The 2026 National Defense Strategy pivoted toward "denial-based defense" of the First Island Chain — signaling the US is preparing for a world where Taiwan can't be defended conventionally.
The US/Israel strikes on Iran and Hormuz closure have shown China what a chokepoint disruption looks like — and that the global response is economic sanctions, not military intervention. If the US won't risk war over Hormuz oil, will it risk war over Taiwan chips? Beijing is watching the answer in real time.
Sources: The War Zone (blockade drills), Defense News, The Diplomat, Global Taiwan Institute (drone warfare), Brookings (gray zone), AEI (March 2026 update)
Taiwan's chip dominance was their "Silicon Shield" — China wouldn't invade because they needed the chips too. In 2026, that shield is thinning from both sides.
US reshoring push: Jan 2026 MOU pushes Taiwan firms to move 40% of supply chain to the States by 2029. CHIPS Act funding accelerates domestic fab construction.
The signal to Beijing: If the US is building its own fabs, it's preparing for a world where Taiwan is no longer the sole source. That makes Taiwan less valuable to protect.
China's domestic chip push: SMIC producing 7nm chips (sanctions workaround). Huawei's Ascend AI accelerators reducing dependence on TSMC for some workloads.
The closing window: If both the US and China reduce TSMC dependence, Taiwan loses its shield. China may feel pressured to act before 2028 — while the world is still too dependent to risk a full military response.
March 2026: intensified drone incursions and transponder-spoofing. Not provocations — rehearsals for a selective blockade that could filter energy shipments while letting other trade pass. Economic coercion driving up insurance rates and shipping costs for Taiwan.
China doesn't need to invade. A naval blockade cutting LNG shipments to Taiwan for 11 days collapses their grid. TSMC goes offline. The global AI supply chain stops. China wins without firing a shot at the fabs themselves.
Even without a blockade, constant gray zone activity raises shipping insurance premiums for Taiwan-bound vessels. This makes Taiwan's chip exports less competitive vs reshored alternatives — achieving economic coercion through risk pricing alone.
Sources: CNBC (US-Taiwan deal), Stimson Center (Silicon Shield erosion), CSIS (Taiwan importance), Wisconsin SoB (CHIPS Act cost)
The Hormuz closure doesn't just affect oil. It's halted helium shipments from Qatar — one of the world's largest producers. Helium is essential for cooling the lasers that etch AI chips.
Samsung and SK Hynix in South Korea are already rationing helium.
This has triggered a "Memory Supercycle" — High-Bandwidth Memory (HBM), the brain of the H200, is becoming physically scarce.
The HBM3e in the H200 we tested (150 GB) requires helium-cooled manufacturing. No helium → no HBM → no H200s → no frontier AI training.
The 2nm bottleneck compounds the problem.
TSMC's 2nm chips (mass production late 2026) promise 25-30% power reduction. But 100% of capacity is pre-booked through 2028. Researchers are stuck on older, less efficient nodes — throwing more "brute force" power (natural gas) at older hardware.
This is what's driving the carbon backsliding. We can't get efficient chips, so we burn more gas to compensate with less efficient hardware.
| Threat | Innovation solution | Climate solution |
|---|---|---|
| Grid fragility (Taiwan 11-day reserves) | Right-sized compute: reduce base load so TSMC can stay online longer on limited reserves | Solar interconnectivity: push for off-grid AI clusters not dependent on imported LNG |
| Supply chain concentration (100% TSMC) | Geographic decentralization: distributed clusters reduce need for single-source 2nm chips | Hardware longevity: incentivize legacy chips (8nm-12nm) manufacturable in US/Europe today |
| Blockade risk (gray zone escalation) | Transparency mandates: force cloud providers to disclose where physical compute is located | Circular economy: tax credits for refurbishing AI hardware, reducing demand from conflict zones |
| Helium shortage (chip manufacturing) | Domestic helium: invest in US helium extraction (BLM reserves in Texas, Kansas) | Chip efficiency: right-sized models don't need frontier chips — our 1.7 GB model proves it |
"We are currently using the world's most vulnerable hardware (H200s from Taiwan) powered by the world's most volatile energy (natural gas from Hormuz)."
"To protect the climate and the country, we need distributed, transparent, and domestic compute. Our data proves we can do world-class clinical AI on right-sized, American-made hardware today."
"The 'Taiwan Threat' is a mirror of the 'Hormuz Crisis.' Both are caused by centralization. We have centralized our energy in the Middle East and our compute in the Taiwan Strait. To protect American science and the climate, we must decentralize."
Nations are racing to build self-sustaining AI infrastructure outside conflict zones. The winner doesn't just own AI — they own foresight.
In October 2025, OpenAI and Sur Energy announced a $25 billion data center project in Patagonia — one of the largest AI infrastructure investments outside the US.
Neuquén province, Patagonia. Geographically isolated from Middle East and Taiwan Strait conflict zones. Naturally cold climate reduces cooling costs. First phase: 500 MW capacity.
Sits on Vaca Muerta shale gas deposits and Limay River hydroelectric dams. Energy partnerships with Central Puerto and Genneia (renewable provider). Uses closed-circuit cooling — no river or sea water needed.
President Milei's RIGI (Incentive Regime for Large Investments) created the regulatory framework. The Hormuz crisis and Taiwan vulnerability make a Southern Hemisphere "compute island" strategically critical.
Timeline: First phase ($7-10B) — construction starts 2026, operational late 2027. Full buildout: $25B.
Sources: Infobae (OpenAI official), Data Center Dynamics, Bloomberg Línea, Zarego
The Stargate Argentina project doesn't exist in a vacuum. South America's political landscape has shifted dramatically in 2025-2026, directly affecting the feasibility of large-scale Western AI infrastructure in the region.
Under Milei, Argentina has realigned sharply toward the US. The RIGI framework was designed specifically to attract projects like Stargate. Argentina has historically had territorial and political tensions with Venezuela — the two represent opposite poles of South American politics.
The January 2026 US intervention in Venezuela (Operation Absolute Resolve) removed a China-aligned government from the northern coast of South America. Venezuelan airspace had been restricted to US flights. The intervention opened the region's logistics and air corridors — the same corridors that connect Miami to Patagonia.
The strategic implication: the Western Hemisphere now has a clear north-south corridor from the US through a stabilized Caribbean to a US-aligned Argentina — with a $25B AI data center at the southern end and local energy independence from Vaca Muerta.
2025-2026 has seen the largest wave of subsea cable construction in history. These cables determine where data — and therefore AI — can physically operate.
| Cable | Route | Builder | Strategic significance |
|---|---|---|---|
| Humboldt | Chile → French Polynesia → Australia (14,800 km) | Google + Chile (50/50 JV) | First-ever direct South America → Asia-Pacific link. 144 Tbps. |
| Project Waterworth | US → South Africa → India → Asia-Pacific → Americas | Meta | Multi-ocean system creating redundant global backbone |
| IOEMA | UK → Netherlands → Germany → Denmark → Norway | Consortium | Northern Europe compute corridor |
| Fiber-in-Gulf (FIG) | Gulf region interconnect | Regional | Gulf AI sovereignty — connects Saudi Hexagon to UAE compute |
The theme: "connectivity with intent." These cables aren't about bandwidth — they're about strategic redundancy. Each one reduces dependence on a single chokepoint.
Sources: Interglobix Magazine, Google Cloud Blog (Humboldt), US State Department
Every major power is building "sovereign compute" — self-sustaining AI infrastructure insulated from adversaries.
| Player | Strategy | Key move |
|---|---|---|
| United States | Western Hemisphere corridor | Stargate Argentina ($25B), CHIPS Act reshoring ($280B+), domestic fabs in Arizona |
| China | "Eastern Data, Western Computing" + UHV grid | 1,400 GW renewables, 15 new UHV lines, SMIC 7nm workaround |
| Saudi Arabia | Sovereign AI spend | 480 MW Hexagon data center in Riyadh — largest government-owned in the world |
| UAE | "Neutral compute hub" | OpenAI/Microsoft partnership, but under fire from Iran — neutrality tested |
| Russia | Energy spoiler | Profiting from $100+ oil while funding battlefield AI from stripped open-weight models |
| Israel | AI proving ground | Testing autonomous systems in Operation Epic Fury — real-world AI-first warfare |
The game ends when one side achieves compute sovereignty — the ability to run their society, military, and economy on infrastructure that cannot be shut down by an adversary.
The nation with the most efficient compute can run models that predict market shifts, disease outbreaks, and supply chain disruptions months ahead. Foresight becomes the new currency of power.
By using solar interconnects and right-sized compute, the winner decouples AI growth from residential grid strain. No voter backlash from rising bills. Innovation without the political cost.
A nation that loses this race becomes a "compute colony" — forced to rent intelligence from foreign infrastructure. Their economic strategy, military planning, and scientific output become dependent on someone else's grid.
The 87% compute waste we measured on the H200 isn't just an efficiency problem. In a world where compute is the strategic resource, waste is a vulnerability.
The current state: We are running the world's most important AI workloads on the world's most vulnerable hardware (Taiwan), powered by the world's most volatile energy (Middle Eastern gas), at 11% efficiency.
The alternative: Right-sized, domestically manufactured hardware. Solar-interconnected, behind-the-meter power. Transparent utilization reporting. Our H200 data proves the science works on a $300 GPU. We don't need to wait for the next generation of chips from a conflict zone.
"Optimization isn't just about the climate — it's about making sure American science stays online when the rest of the world goes dark."
Decoupling AI from the residential grid. Not all AI work is urgent.
CAN shift to solar peak or off-peak hours. Our 3-hour run doesn't care if it starts at 2 PM or 2 AM.
CANNOT shift. Hospital mortality predictions need answers now. Patients can't wait for solar peak.
Same scientific result. 27x less energy. Speed and efficiency aren't always trade-offs — sometimes better engineering gives you both.
Data center "flexibility" could save the US electricity system $40–150 billion over the next decade.
By avoiding new gas plants, reducing fuel costs, and shifting capacity toward renewables. Ross & Ewing, Duke Nicholas Institute, Feb 2026
Training can run anytime — but the grid carbon cost depends on when. The incentive question: can rate structures make daytime solar cheaper than overnight gas?
Training can run at 2 AM just as well as 2 PM. Many researchers run overnight because the cluster is less contested. There's no technical reason to prefer daytime. The question is purely economic: does the rate structure make solar hours cheaper?
If utilities offered "green compute rates" — lower $/kWh during solar peak, higher during gas-heavy overnight — researchers and hyperscalers would naturally shift training to daytime. Not because they have to. Because it's cheaper. The grid gets cleaner without a mandate.
Instead of building gas plants for ghost peaks, what if data centers funded neighborhood batteries and solar that buffer the surge? That's a Virtual Power Plant — and data centers are the first customers willing to pay for it at scale.
Devices in homes — rooftop solar, batteries, EVs, smart thermostats, heat pumps — can act as mini-power plants or adjust energy usage in real time. Link thousands of them together to respond to grid spikes, and you have a Virtual Power Plant. No new land. No new gas turbines. Just smarter use of what's already installed.
— Mark Dyson, Managing Director, RMI (Rocky Mountain Institute)
The crisis is data center demand outpacing the grid. The cash is hyperscalers willing to pay for power "as quickly as possible" — and impatient enough to circumvent the 3-5 year grid interconnection queue. Data centers may be the first customers to subsidize VPP programs at scale, funding residential batteries and solar in exchange for grid capacity that comes online in months, not years.
Data center requests 7 MW of capacity.
Utility builds a gas plant for the peak.
Actual load: 0.93 MW. Ghost: 6.1 MW.
Gas plant operates for 30-40 years.
Residential bills go up to fund it.
Cost: $billions in stranded fossil assets.
Data center funds residential solar + batteries in its service territory.
1,000 homes with batteries = distributed surge capacity.
When the data center hits peak, homes flex load down.
When the data center idles, homes charge from solar.
No gas plant needed. Residential bills go down.
Cost: fraction of a gas plant. Online in months.
"Does a data center company ever want to say 'I won't run my training model for a couple hours on the hottest day of the year'? No. Instead, the opportunity is to pay other people to flex their load, or pay other people to adopt batteries that create headroom on the system."
— Mark Dyson, RMI
Data centers provide capital to help customers in their service territory buy residential batteries — or contracts that guarantee a return for VPP aggregators. Hyperscalers get capacity unlocked quickly. Homeowners get subsidized batteries. The grid gets flexible.
~70% of US electricity customers are served by investor-owned utilities that earn profits by building infrastructure, not by reducing demand. Utilities lack natural incentive to support VPPs that shrink their capital expenditure pipeline. Regulation needs to catch up.
EnergyHub's VPP Maturity Model defines five levels. Most advanced VPPs today are at Level 2. The goal is Level 4-5: indistinguishable from conventional power plants.
VPP benefits accrue to households that can afford solar + batteries. The cost of maintaining shared grid infrastructure falls on everyone. This echoes the rooftop solar equity debate. But unlike rooftop solar, VPPs also benefit non-participants — if the programs succeed in driving down system costs and improving reliability, everyone's bills decrease. The question is whether the benefits are distributed fairly, not whether they exist.
Some researchers argue grid-scale batteries at the utility level achieve the same goals at lower unit cost, with benefits shared equally across all customers. VPP advocates counter that with 3-5 year interconnection queues, distributed resources deploy faster. The answer may be both — grid-scale storage for baseload, VPPs for rapid-response surge buffering. Neither requires a new gas plant.
Sources: Heatmap News: "Will VPPs Ever Really Be a Thing?" (Feb 2026) — Mark Dyson (RMI), Apoorv Bhargava (WeaveGrid), Matthew Plante (Voltus), Ryan Hanna (UC San Diego), Ben Hertz-Shargel (WoodMackenzie). VPP Maturity Model via EnergyHub. DOE 80-160 GW target via Jigar Shah (former DOE Loan Programs Office).
The EU's 2026 Data Centre Energy Efficiency Package already requires disclosure. The US should match or exceed it.
Per-job utilization, power draw, memory — reported quarterly. EPA requires emissions reporting. AI compute should too.
Impact fees. Behind-the-meter power. Demand response. Residential bills should NOT subsidize corporate AI training.
Tax credits for matching hardware to workload. Our 1.7 GB model on a 150 GB chip is a school bus for one person.
Federal funding for flexible scheduling. Solar-peak priority access. Datacenters that store energy for the grid get credits.
Research exempt from utilization mandates. Reporting yes, penalties no. A cancer researcher at 2% is still doing important work.
Require standardized disclosure — GPU utilization, Ghost Ratio, energy, CO₂, grid carbon intensity — for every training run above a threshold. Cars got EPA stickers. AI compute needs the same. → Section 05
Allow datacenters to meet behind-the-meter requirements by funding neighborhood VPPs — residential solar + batteries that buffer grid demand. DOE target: 80–160 GW by 2030. → Section 10
Mandate WUE disclosure. Cap peak withdrawal during droughts. Require liquid/dry cooling for new builds in water-stressed regions. 720 billion gallons/year by 2028 can't be invisible.
| Technology | Policy Lever | Impact |
|---|---|---|
| Liquid / Immersion Cooling | Mandate for new facilities in water-stressed regions | 70% water reduction vs evaporative |
| Closed-Loop / Dry Cooling | "Water-positive" mandates for facilities in drought-prone regions (AZ, TX, IA) | Near-zero direct water consumption |
| Virtual Power Plants | Datacenter impact fees fund neighborhood solar + batteries. Grid capacity in months, not the 3-5yr interconnection queue | 80–160 GW by 2030 (DOE target) |
| Measurement Tooling (codecarbon, pynvml) | Open-source, free, works on any GPU today. The barrier is cultural, not technical — standardize and require | Enables CCC / Ghost Ratio disclosure |
| Efficient Architectures (MoE, Distillation) | R&D tax credits for compute-efficient model design. DeepSeek matched frontier performance at a fraction of the cost | Orders-of-magnitude compute reduction |
| UHV Transmission | Ultra-high voltage lines to move compute to where clean energy is. China's "Eastern Data, Western Computing" runs on this | Decouples location from load |
| Right-Sized Chip Allocation | Tax incentives for high fleet utilization | Direct idle energy reduction |
| Temporal Flexibility | Rate structures rewarding off-peak training | $40–150B grid savings / decade |
| On-Site SMRs (Small Modular Reactors) | Streamlined permitting for co-location | Decouples AI from public grid |
| Hardware Circularity | Extended producer responsibility for AI chips | Addresses 2-3yr chip e-waste |
"AI uses too much energy. Restrict it."
"AI pays for grid modernization — IF policy ensures it."
"We don't need less compute for science. We need to know what compute costs and give researchers the tools to make informed choices. Right now nobody measures, nobody reports, nobody knows. That's the policy gap."
The AI boom is being used to justify new gas plants, new LNG terminals, and fossil infrastructure locked in for 30–40 years. That infrastructure doesn't just emit carbon — it requires imperial maintenance. Controlling the Strait of Hormuz. Locking allies like Japan into US gas exports. Export controls to contain China. American families pay for all of it through higher energy bills and supply chain fragility.
China figured out the alternative: domestic renewables, UHV transmission, domestic chip fabs. They don't need Hormuz. They don't need LNG imports long-term. They're building self-sufficiency, not dependency.
The US has the same option. VPPs deploy in months, not years. Right-sizing cuts demand immediately. Temporal flexibility uses existing solar at zero new cost. Domestic fabs reduce the Taiwan chokepoint. Self-sufficiency turns energy and chips from a necessity — that must be controlled through force — into a want — that can be traded freely. That eases imperial pressure abroad and lowers costs for citizens at home.
Any administration is temporary — and this one will likely face deadlock once the House and Senate flip in the midterms. There will always be resistance to progress at the federal level. But that's not where the fastest action happens.
Communities are swifter. Cities and states are already going green because it makes financial sense — not because Washington told them to. Local politicians can approve VPP programs, pass behind-the-meter solar requirements, mandate WUE disclosure for new datacenter permits, and set green compute rate structures. None of that requires Congress. A county commissioner can move faster than a Senate subcommittee.
The speed mismatch works both ways. AI demand is growing faster than federal policy can respond — but it's also growing faster than fossil infrastructure can be built. Communities that act now on VPPs, solar interconnects, and ratepayer protections don't have to wait for the gas plant. They can be the grid the datacenter plugs into — on their terms, at their price.
Don't lock the AI boom into fossil infrastructure that requires imperial maintenance for 40 years. Build the domestic energy and domestic chips that make all of that unnecessary. China is already doing both. American communities are starting to. Federal policy should catch up.
"The climate movement needs to continue to elect climate champions to local offices across the country. More than that, we have to support those leaders once they are elected so that they can be successful in passing aggressive climate policy. We are fortunate to have Run On Climate leading the way in that effort."
— Bill McKibben
Author, Environmentalist, Journalist, and Co-founder of 350.org
| Run | Time | Energy | CO2 | GPU % | Memory |
|---|---|---|---|---|---|
| LMM standalone | 104 min | 2.68 kWh | 0.99 kg | 11% | 1.72 / 150 GB |
| GRASP+LMM | 103 min | 2.66 kWh | 0.98 kg | 11% | 1.73 / 150 GB |
| GRASP+LMM+codemap (lr=5e-4) | 177 min | 4.60 kWh | 1.69 kg | 14% | 2.58 / 150 GB |
| GRASP+LMM+codemap (lr=1e-4) | 181 min | 4.67 kWh | 1.72 kg | 12% | 2.58 / 150 GB |
| GRASP+LMM+codemap (3 fixes) | 161 min | 0.61 kWh | 0.22 kg | 14% | 1.67 / 150 GB |
| LMM standalone + batch=256 | 42 min | 0.17 kWh | 0.06 kg | 16% | 3.39 / 150 GB |
Methodology: codecarbon (energy/CO2 from actual sensor readings) + pynvml (GPU metrics from NVIDIA driver) + torch.cuda (memory tracking)
Hardware: NVIDIA H200 NVL, 150 GB HBM3e, CUDA 12.6
Dataset: MIMIC-III (46,520 patients, public clinical data, Beth Israel Deaconess Medical Center)
Reproducibility: All code open source via PyHealth (github.com/sunlabuiuc/PyHealth). Any researcher with GPU access can replicate.