Can the Texas Grid Handle AI?
ERCOT is America's only major isolated grid — 27 million people, no backup from neighboring states. In February 2021, Winter Storm Uri brought it within 4 minutes and 37 seconds of total collapse, leaving 246 dead and $130 billion in damage.
Now AI data centers have filed for 226 GW of new interconnections — 73% of the queue, nearly triple ERCOT's current peak demand. Can the grid handle it?
The answer depends on whether you model the grid as a whole — or model where the load actually lands.
Five Models. Two Dimensions of Fidelity.
This study uses five distinct simulation models. But model complexity is only one dimension of fidelity. The other — and for this study, the more important one — is how honestly you treat uncertainty in the inputs.
The Models
Hourly Dispatch
Merit-order simulation of all 8,760 hours. Nuclear first, then renewables, batteries, gas, coal. Shortfall = blackout. The workhorse behind the first four questions.
Spatial Dispatch
Splits ERCOT into 2 or 6 zones with capacity-limited transmission corridors. Same dispatch algorithm, now topology-aware. Reveals bottlenecks the single-node model can't see.
SCUC Optimizer
Mixed-integer programming. Gas plants cost $80K to start, can't run below 40%, must stay on 8+ hours. Tests whether real machine constraints change the answer.
Cascade Solver
Iterative coupled solve. Gas shock → dispatch → load shedding → compressor failure → more gas offline → repeat until convergence or collapse.
Backward DP
Dynamic programming for optimal battery dispatch. Perfect foresight over the full year, plus receding-horizon variants with limited forecast windows.
Deterministic vs. Stochastic
Several questions were analyzed both ways. Deterministic analysis uses one weather year, or sweeps across ten years (2015–2024, demand-normalized to 2024 peak), with all plants working perfectly. Stochastic Monte Carlo runs 200 draws per scenario with random weather year selection, 7–10% forced outage rates, and ±3% demand noise — this is how real resource adequacy studies work.
The biggest fidelity gap in this study isn't between models. It's between deterministic and stochastic analysis of the same model. The dispatch algorithm didn't change. What changed is whether we pretended every plant works every hour. That single assumption — all plants available — is what makes the grid look reliable when it isn't.
What Changes the Answer — and What Doesn't
60 GW gas + 300 GWh storage at 70% renewables + 15 GW data center load:
(best year)
(worst year)
(best year)
(P90)
The escalation converges. Weather variability matters — it reveals failures a single good year hides (6→49). A more complex model (SCUC) doesn't change the answer — the optimizer confirmed the simpler dispatch was right (49→6). But adding forced outages to the same dispatch model changes everything (6→476). The model was fine. The inputs were a lie.
Eight Questions. Eight Deep Dives.
Each question was answered at the fidelity it required. Click any card to see the full analysis, charts, and methodology.
Where's the Breaking Point?
ERCOT absorbs 3–4 GW of DC load before exceeding reliability standards in a good year. Less than 2% of the 226 GW queue.
What to Build — and What It Costs
70 GW gas + 50 GWh storage ($24B incremental CAPEX, undiscounted) is the cheapest reliable path. Gas dominates; batteries have sharp diminishing returns.
Can the Grid Survive Another Uri?
25% gas outage is manageable. 50% is catastrophic. Weatherization is the cheapest insurance.
Can Data Centers Help?
30% load flexibility cuts blackouts by a third at zero cost. 5% demand response doubles grid DC capacity.
Does It Matter Where You Build Them?
West Texas wins at small scale, loses at 15+ GW. A 6-zone model reveals bottlenecks the 2-zone model misses.
What If Gas Plants Are Real Machines?
A proper SCUC optimizer confirms gas headroom matters — but the heuristic overstated the penalty by 10×. The directional finding holds; the specifics change dramatically.
When Does a Gas Shock Become Self-Sustaining?
At ERCOT's estimated 1–2% gas-electric coupling, a 25% shock amplifies to 35–52%. At 3%+, even a small shock cascades to total collapse.
Does Battery Intelligence Matter?
Perfect-foresight dispatch cuts unserved energy by 87%. A 6-hour weather forecast captures 97% of that value. You don't need AI — you need a forecast.
Scope: Five models totaling ~4,500 lines of Python, including MIP-based SCUC (PuLP/CBC), iterative cascade solver, 6-zone spatial dispatch, stochastic Monte Carlo (200 draws with forced outages and demand noise), and backward DP dispatch optimization. Ten years of ERCOT weather data (2015–2024), demand-normalized to 2024 peak. Eighteen years parsed (2007–2024); years before 2015 excluded after testing showed demand dilution without additional weather diversity. Limitations documented honestly: simplified demand response, no new nuclear/SMR, no sub-hourly dynamics, greedy (not LP) zonal transfers. Each question got the model it needed — nothing more.
Finding the Right Model for the Decision
Each question above was answered at the fidelity it required. But scattered across eight questions, it's easy to miss the pattern: which dimensions of reality actually change the answer? Here it is in one place — the same investment evaluated at increasing fidelity, showing where the answer converges and where it breaks.
60 GW gas + 300 GWh storage — $66B (incremental CAPEX, undiscounted)
| Model Fidelity | Blackout Hours | What a Planner Concludes |
|---|---|---|
| Hourly dispatch, best weather year | 6 | “Meets the reliability standard.” |
| + Monte Carlo (10 weather years) | 49 | “Fails in 7 of 10 years. Not reliable.” |
| + SCUC optimizer (24h MIP) | 6 | “Still reliable. UC constraints don't bind with 300 GWh.” |
| + Stochastic MC (200 draws, forced outages) | P90: 476 | “Catastrophically unreliable when you model real plant availability.” |
| + Gas-electric cascade (Q7) | amplifies shock 2× | “A 25% gas outage becomes 52% via feedback.” |
Every model up to SCUC said 6 hours — reliable. Then we added forced outages and demand uncertainty: P90 jumps to 476 hours. The gap isn't between dispatch models anymore — it's between deterministic planning and stochastic reality. Real power plants break. Real demand deviates from forecast. A planning process that ignores both is planning for the best case, not the likely one.
70 GW gas + 50 GWh storage — $24B (incremental CAPEX, undiscounted)
| Model Fidelity | Blackout Hours | Verdict |
|---|---|---|
| Hourly dispatch, best weather year | 2 | Reliable |
| + Monte Carlo (worst weather year) | 8 | Reliable |
| + SCUC optimizer (24h MIP) | 0 | Reliable (optimizer uses batteries better than greedy dispatch) |
| + Stochastic MC (200 draws, forced outages) | P90: 270 | Not reliable. Need 85 GW gas to pass P90. |
| 6-zone spatial model (Q5) | — | Robust if load is distributed across zones |
| + Gas-electric cascade (Q7) | amplifies shock 1.7× | Manageable at 1.5% coupling; collapses at 3%+ |
Methodology note: This analysis uses 10 years of weather data (2015–2024) with demand normalized to 2024 peak levels (85.5 GW). We parsed 18 years of ERCOT generation data (2007–2024), but years before 2015 add weather diversity already captured by the 10-year window—the worst February wind drought in the 18-year record (2018) is nearly identical to one in the 10-year sample (2023). Demand normalization ensures every Monte Carlo draw faces current demand levels, not the 15–30% lower demand of earlier years. Without normalization, draws from low-demand years make the grid look artificially reliable.
Understanding P90: Why Reliability Matters
P90 means the 90th percentile of stress hours. Out of 10 weather years, 9 are "easier" than this. Only 1 year is worse. This is the reliability standard that grid operators must meet.
50h
200h
476h
~600h
What it means: If you design the grid for P50 (200 hours of stress), you'll fail 50% of the time. If you design for P90 (476 hours), you're safe 90% of the time — but you need way more gas capacity and battery storage.
The decision: ERCOT must choose between (a) frequent blackouts, (b) massive overbuilding, or (c) acceptance of 1-in-10 stress years. That's why P90 is the industry standard — it's the reliability level most societies accept.
What Drives the Answer?
Across eight investigations, three inputs dominate the capacity question. Two widely-discussed factors barely move the needle.
Weather & Outages Run the Show
Pick a different weather year and the breaking point swings from 6 GW down to 2 GW. Layer in realistic forced outages (7–10%) and it drops to zero. Nothing else in the model moves the answer that much.
Location Changes Everything
67 blackout hours or 528? Same 10 GW of data centers, same year. The only difference: one model treats ERCOT as a single node, the other splits it into two zones with transmission limits. That 8x gap is entirely about where you build, not how much.
Dispatch and Gas Prices? Barely Register
A 6-hour battery forecast captures 97% of perfect-foresight value. Gas price swings shuffle the dispatch order but don't change whether the grid holds. These are operational dials, not structural ones.
Model Validation
Before projecting future scenarios, we ran the dispatch model against 2024 actuals. Same model, same algorithm — but compared to what EIA recorded actually happened on the grid.
| Metric | Model | EIA Actual | Error | Explanation |
|---|---|---|---|---|
| Wind generation | 111.8 TWh | 111.8 TWh | +0.0% | Same underlying profiles |
| Solar generation | 47.8 TWh | 47.8 TWh | +0.0% | Same underlying profiles |
| Nuclear generation | 39.0 TWh | 38.9 TWh | +0.2% | Constant 87% CF assumption |
| Gas generation | 281.8 TWh | 204.3 TWh | +37.9% | Model dispatches gas before coal (see below) |
| Coal generation | 0.2 TWh | 58.9 TWh | −99.7% | Model barely dispatches coal |
| Gas + Coal combined | 282.0 TWh | 263.2 TWh | +7.2% | Total thermal need is close |
| Peak demand | 85.5 GW | 85.5 GW | +0.0% | Same input profile |
The honest reading: Our merit-order model dispatches gas before coal, so it over-predicts gas by 38% and under-predicts coal by 100%. But total thermal generation (gas + coal combined) is within 7.2% of reality. For the future scenarios in this study — which assume coal retires to zero — this error disappears: all thermal generation is gas, and the model's total thermal estimate is what matters. The validation confirms the model correctly predicts how much dispatchable generation the grid needs, even if it misallocates between fuel types in the current fleet.
Validation uses 2024 installed capacity: wind 34 GW, solar 30 GW, gas 56 GW, nuclear 5.1 GW, coal 11 GW, battery 13 GW / 52 GWh. EIA-930 hourly actuals. Battery discharge is over-predicted (model assumes full 13 GW fleet; actual 2024 operational fleet was ~3 GW still ramping).
What This Model Doesn't Capture
Every simplification was a deliberate choice. These are the ones that matter most.
Simplified Demand Response
DR is modeled as a fixed percentage of peak load that can be shed. Real demand response involves complex contract structures, participation rates, and fatigue effects over multi-day events. The model likely overstates DR availability during sustained stress.
No New Nuclear or SMR
The generation fleet is based on current installed capacity plus announced additions. Small modular reactors and new conventional nuclear are excluded. If SMR deployment accelerates, the capacity gap narrows — but not on the timeline of the data center queue.
No Sub-Hourly Dynamics
Dispatch runs at hourly resolution. Frequency regulation, ramping constraints, and intra-hour renewable variability are not captured. This matters most for battery valuation — batteries provide sub-hourly services that the model cannot see.
Greedy Zonal Transfers
Inter-zone power flows use a greedy algorithm, not linear programming. The model routes power to the nearest deficit zone without optimizing system-wide transfer efficiency. Real ERCOT dispatch is more efficient than what the model computes.
CAPEX Only, No Lifecycle Costs
Investment estimates use NREL ATB 2024 overnight capital costs. Fuel costs, O&M, financing, and capacity factor degradation are excluded. These are order-of-magnitude comparisons between options, not project-level investment cases.
No Market Feedback
Generators and load-serving entities are price-takers. In reality, scarcity pricing triggers investment responses and demand shifts that the model does not capture. This is a structural stress test, not an equilibrium forecast.
What's Actually Happening
The analysis above answers “what if.” This section answers “what is.” Real data center investments have been announced. Real generation is under construction. Real legislation has been signed. Here's how reality maps onto our model.
What's Coming: Data Center Demand
ERCOT's large-load interconnection queue hit 226 GW in 2025 — up 259% in a single year, with 73% of it data centers. But historically only 13% of queued generation projects reach commercial operation (Lawrence Berkeley National Lab). ERCOT itself applies a 49.8% haircut to requested amounts. Here's what's actually under way.
| Project | Location | Capacity | Timeline |
|---|---|---|---|
| OpenAI / Oracle (Stargate) | Abilene | 1.2 GW | First buildings operational Sep 2025 |
| Google (3 campuses) | West TX / Panhandle | Multi-GW | $40B investment, 2025–2028 |
| Meta | El Paso | GW-scale | Expected 2028 |
| Amazon / AWS | Round Rock, DeSoto, Amarillo | Various | In planning / zoning |
Analyst estimates converge on ~10 GW of data center load by 2028, with ~6 GW/year additions possible thereafter, constrained by gas turbine manufacturing (Modo Energy). ERCOT projects 55 GW of DC load by 2030; independent analysts call this unrealistic, estimating 20–30 GW based on supply-chain and construction limits (Ascend Analytics).
In our model, 10 GW of data center load at the current generation mix produces 71 blackout hours per year — 8× the reliability standard. At 20 GW, it's 700+ hours. These aren't hypotheticals anymore.
It's Not Just Data Centers
Data centers dominate the growth narrative — roughly 60–70% of ERCOT's incremental demand through 2030. But they're not the whole story.
| Demand Driver | Additional Load by 2030 | Context |
|---|---|---|
| Data centers | 40–55 GW (queue-adjusted) | 73% of the interconnection queue. The dominant driver. |
| Oil & gas electrification | ~8.5 GW | Permian Basin drilling, pumps, compressors shifting from diesel to grid power. |
| Cryptocurrency mining | 4.3–5.3 GW | 9% of queue. ~3 GW is price-responsive (curtails during scarcity). |
| Population growth | ~3–4 GW (organic) | 391,000 new Texans per year. 1.2% growth, #1 state in absolute additions. |
| Electric vehicles | ~2 GW peak | 456K EVs registered (2025), growing ~1,500/week. Small relative to grid scale. |
Before the data center boom, ERCOT demand grew at roughly 3% per year — driven by population, Permian Basin electrification, and increasing cooling load. That baseline alone would stress the grid over a decade. Data centers accelerated the timeline from “gradual challenge” to “imminent crisis.”
What's Being Built: New Generation
| Resource | Capacity | Timeline | Status |
|---|---|---|---|
| Texas Energy Fund (new gas) | 10 GW (17 projects) | 2028–2029 | $5.4B in loans approved; 3.3 GW finalized |
| Gas (under construction now) | 1 GW | Summer 2026 | Under construction |
| Battery storage | 14 → 19 → 40–55 GW | Now → mid-2026 → 2029 | 14 GW operational, on track |
| Solar (under construction) | 23 GW | Summer 2026 | Under construction |
The pattern: Texas is building renewables and batteries at scale, but almost no new dispatchable gas arrives before 2028. The TEF pipeline addresses this, but 10 GW of new gas brings the total to ~66 GW — and our stochastic analysis shows you need 85 GW (P90) to 90 GW (P99) to survive forced outages and weather variability at 15 GW of data center load.
Reserve margins go negative by summer 2028. ERCOT's own December 2025 Capacity, Demand and Reserves report projects the grid drops below its 13.75% target in 2027 and enters shortfall territory by 2028. This is not a model projection — it's ERCOT's published assessment based on what's under construction versus what's been requested.
The Gap
Mapping real-world supply plans onto our model results:
| Timeframe | Likely DC Load | Gas + Storage Available | Our Model Says |
|---|---|---|---|
| 2026–2027 | 6–10 GW | 56 GW gas, ~76 GWh battery | At the edge. SB 6 curtailment is the safety valve. |
| 2028–2029 | 10–20 GW | ~66 GW gas (with TEF), ~200 GWh battery | Falls short. Stochastic MC says you need 85 GW (P90) at 15 GW DC. |
| 2030+ | 20–30 GW | ~66 GW gas, ~220 GWh battery | Deeply insufficient. 700+ blackout hours at 20 GW without major new gas. |
SB 6: The Legislature's Answer
Senate Bill 6, signed June 2025, directly addresses several risks our analysis identified.
| SB 6 Provision | What Our Model Found |
|---|---|
| 50% on-site backup for loads ≥ 75 MW | Q4 found 30% flexibility cuts blackouts by a third. 50% backup exceeds this. |
| ERCOT curtailment authority during emergencies | DR reassessment: 5% demand response doubles grid DC capacity from 4 to 8 GW. |
| Remote disconnect for new large loads | Enables rapid load shedding our Q3 stress test showed is critical during Uri-type events. |
| Interconnection cost-sharing | Q5: Location matters. Cost-sharing discourages concentration in constrained corridors. |
SB 6 is well-designed for the near term. Its provisions map directly onto the mechanisms our analysis identified as most effective. The question is whether it's sufficient beyond 15–20 GW, where the gas capacity gap becomes dominant and on-site backup generators can't substitute for grid-scale dispatchable power.
What We'd Recommend
Eight questions, four model fidelities, ten years of weather data, hundreds of scenario combinations — and now grounded in what's actually being built, bought, and legislated.
How Sensitive Are These Numbers?
Every model depends on assumptions. We tested how the stochastic reliability of our recommended configuration (85 GW gas + 50 GWh storage, 70% RE + 15 GW DC) changes when we vary three key assumptions. The baseline passes LOLE at P90 with 6 blackout hours. Here is what moves it.
| Assumption | Value | P90 Hours | Verdict |
|---|---|---|---|
| Demand growth | 0% / yr | 6 | PASS — baseline |
| Demand growth | 2% / yr | 77 | FAIL — +8% demand by 2028 overwhelms supply |
| Demand growth | 4% / yr | 275 | FAIL — catastrophic; 275 blackout hours at P90 |
| Battery RTE | 80% | 9 | BORDERLINE — just at the LOLE threshold |
| Battery RTE | 85% | 6 | PASS — baseline |
| Battery RTE | 90% | 5 | PASS — modest improvement |
| Fleet EFOR | 5% | 2 | PASS — well-maintained fleet, plenty of margin |
| Fleet EFOR | 8% | 6 | PASS — baseline |
| Fleet EFOR | 12% | 32 | FAIL — aging or stressed fleet breaks reliability |
What matters most: Demand growth dominates. Even 2% annual growth (+8% by 2028) pushes P90 blackout hours from 6 to 77 — more than 12x the LOLE standard. Battery efficiency barely moves the needle (P90 range: 5–9). Fleet reliability matters at the margin: if forced outage rates climb from 8% to 12% (aging fleet, stressed conditions), the configuration fails. The implication: ERCOT's biggest risk isn't the supply plan — it's whether demand growth outpaces it.
What Texas Is Getting Right
| Action | Why It Works (Per Our Analysis) |
|---|---|
| Texas Energy Fund ($9B for gas) | Gas is the dominant reliability lever (Q2). 10 GW of new gas is the single most impactful investment possible. |
| SB 6 curtailment + backup mandates | Demand response is the cheapest tool (Q4). Legislating it removes voluntary participation risk. |
| Battery buildout (14 → 55 GW by 2029) | Short-duration storage handles daily peaks. Texas is building this faster than anywhere else. |
| SB 6 interconnection cost-sharing | Discourages corridor concentration (Q5). Market signal for distributed siting. |
What's Missing
| Gap | Why It Matters | What To Do |
|---|---|---|
| No gas plan beyond TEF | TEF delivers ~66 GW total. Stochastic MC needs 85 GW (P90) for 15 GW DC — that's 19 GW short. The cascade model (Q7) shows 66 GW is especially fragile to gas-electric feedback. | TEF Round 2 targeting 19–24 GW additional gas by 2030. |
| Gas turbine supply bottleneck | Only ~6 GW/year available for ERCOT. Manufacturing, not money, is the constraint. | Long-lead procurement now. Prioritize combined-cycle for reliability (Q6). |
| No long-duration storage strategy | 4h batteries can't address multi-day deficits. Uri was a 72-hour event. | Incentivize 8–12h storage. Duration matters more than capacity. |
| Gas weatherization + cascade resilience | Q3: 25% gas outage is manageable, 50% is catastrophic. Q7: the cascade amplifies a 25% shock to 52% at 2% coupling. Weatherization AND compressor backup power are the dividing lines. | Mandatory standards with compliance verification. Backup power for critical compressor stations. Cheapest insurance available. |
| No plan for 20+ GW DC load | At 20 GW with current plans, our model shows 700+ blackout hours in good weather. | Conditional interconnection: DCs beyond 15 GW total require demonstrated generation adequacy. |
Where the Next Dollar Should Go
Based on our analysis and the current supply pipeline, in priority order:
| Priority | Investment | Est. Cost | Reliability Impact |
|---|---|---|---|
| 1 | Gas weatherization (existing fleet) | $2–4B | Keeps outage rate below 25% in winter storms. Prevents 90% of Uri-scale damage (Q3). |
| 2 | 19–24 GW additional gas (beyond TEF) | $23–29B | Closes gap from 66 to 85–90 GW. Stochastic MC says 85 GW for P90 reliability, 90 GW for P99 at 15 GW DC. |
| 3 | Long-duration storage (8–12h, 50+ GWh) | $12–15B | Addresses multi-day deficits that 4h batteries cannot. Complements gas for winter resilience. |
| 4 | West TX transmission expansion | $3–6B | Raises the 20 GW transfer corridor limit. Enables West TX DC siting beyond 10 GW (Q5). |
Cost estimates are incremental CAPEX only (NREL ATB 2024 assumptions, undiscounted). They exclude fuel costs, O&M, financing, and capacity factor degradation. These are order-of-magnitude figures for relative comparison between options — not project-level investment estimates. A proper cost-benefit analysis would require discounted cash flow modeling with fuel price scenarios.
Analysis current as of March 2026. Data center announcements, generation construction, and policy are evolving rapidly. ERCOT queue data, CDR reports, and EIA generation data are updated quarterly. This analysis will be updated as material changes occur.
Explore the Data
Drag the sliders and see how the Texas grid responds.
Sources
Grid Operations Data
EIA-930 Hourly Grid Monitor — Generation by fuel type + demand, 2019–2024 (52,608 hours).
ERCOT Capacity, Demand and Reserves Report (Dec 2025) — Reserve margin projections, installed capacity.
NREL Annual Technology Baseline 2024 — Battery parameters and capital costs.
FERC/NERC Joint Inquiry into February 2021 — Uri failure analysis, gas supply chain data.
Demand Forecasts
ERCOT 2025 Long-Term Load Forecast (Apr 2025) — Peak demand projections, EV forecast (Table 1), large load queue methodology.
ERCOT 2026 Long-Term System Assessment — 207 GW new resource need.
Ascend Analytics — Independent critique of ERCOT load projections (119 GW realistic by 2030).
Data Center Investments
Utility Dive — ERCOT large load queue 226 GW, 73% data centers.
Modo Energy — DC buildout projections, gas turbine supply constraint (~6 GW/yr).
Lawrence Berkeley National Lab — Interconnection queue completion rates (13% historical).
JLL Year-End 2025 — 6.5 GW under construction in Texas.
Non-DC Demand Growth
EIA — Rapid electricity demand growth in Texas.
Yale Clean Energy Forum — O&G electrification, 3.4→11.9 GW by 2032.
EIA — Cryptocurrency mining electricity consumption.
TxETRA — Texas EV registration data.
Texas Tribune / Census — Texas population growth (31.7M, +391K/yr).
Policy & Legislation
McGuireWoods — SB 6 overview (75 MW threshold, 50% backup, curtailment authority).
Power Magazine — Texas Energy Fund, $5.4B for 10 GW.
Texas Tribune — $7.2B gas plant loan program details.
Battery & Generation Pipeline
Modo Energy — ERCOT battery buildout (14 GW operational, 40–55 GW projected by 2029).
EIA — ERCOT meeting demand with solar, wind, and batteries.
Texas Comptroller — The Future of Texas Power.