Skip to content
Studies · Case Study

Where Should Texas Build EV Chargers?

Texas EV registrations doubled from 235,000 to 472,572 in just 2.3 years — growing 3.5× faster than the national average. With 3,976 stations, 12,951 ports, and $408 million in NEVI funding for highway corridor charging, the question is where to build, how big, and what the grid can handle.

When the model accounts for grid capacity, $68 million in planned infrastructure moves to where it's needed.

472K
Registered EVs
3,976
Charging Stations
$408M
NEVI Funding (TX)
12,951
Total Ports
The Decision

Where to Place New Stations — and How Big

Texas needs hundreds of new EV charging stations by 2030. The question isn't just where to put them — it's how big they should be and what the grid can actually support. A 20-port fast-charging station draws 3 MW — enough to require serious local grid infrastructure. The screening model that most planners use can't even see this constraint.

We built three models, each adding a dimension of reality the previous one ignores. The ~430 lines of grid code in Model C — checking 4,151 real substations — saves $68M in 10-year NPV.

The Models

Three Models. Three Levels of Fidelity.

Each model adds a dimension of reality the previous one ignores. The question is: which dimensions actually change the decision?

Model A

Population/Traffic Screening

Score each Census tract: w1·pop + w2·traffic + w3·EVs − w4·existing. Top N tracts get stations. Flat 8-port sizing for every site.

Inputs: Census tract population, AADT traffic volume, EV registrations, existing AFDC charger locations

Runtime: <1 second · ~220 lines Python

Gets wrong: Over-indexes on dense urban cores where home charging is prevalent

Model B

Facility Location Optimizer

Maximal Covering Location Problem (MCLP). Maximize demand covered within a 25-mile threshold, subject to budget constraint. Dynamic sizing: 4–20 ports per station based on local demand.

Inputs: Everything from A + trip pattern proxies, charger utilization model

Runtime: 5–30 sec · ~390 lines Python

Gets wrong: Ignores grid — a 20-port / 3 MW station needs serious local infrastructure

Model C

Grid-Aware Placement

For each Model B location, check nearest substation capacity. Downsize, relocate, or accept upgrade cost. Re-optimize with grid cost in the objective function.

Inputs: Everything from B + HIFLD/DHS substation locations/capacity, grid upgrade cost model

Runtime: 10–60 sec · ~430 lines on top of Model B

The ~430 lines of grid code checking 4,151 real substations saves $68M in 10-year NPV

What Changes the Answer — and What Doesn't

Baseline scenario: $200M budget, mixed focus, current adoption (472,572 EVs):

76.1%
Model A
(111 stations)
88.5%
Model B
(49 stations)
88.5%
Model C
(49 stations)
−$68M
Model C
(10yr NPV Δ)

Model B achieves 88.5% coverage with 56% fewer stations by optimizing placement. Model C maintains the same coverage but right-sizes for the grid — saving $68M in 10-year NPV ($178.9M vs $247.0M). The optimizer found better sites. The grid code found cheaper ones. Both matter, but the cost savings only appear when you model the grid.

Scenario Matrix

27 Scenarios

Three EV adoption levels, three budget tiers, three geographic focus types — 27 combinations tested across all three models.

DimensionLowMediumHigh
EV Adoption 472,572 (current) ~945,000 (2×) ~2,360,000 (5×)
Budget $50M (NEVI Phase 1) $200M (NEVI + state match) $500M (full NEVI allocation)
Focus Urban (metro areas) Corridor (highway gaps) Mixed (optimized blend)
Coverage by Model & Budget (1.25M EVs, Mixed Focus)

The budget matters less than the model. At $200M, Model A achieves 76.1% coverage with 111 stations. Model B achieves 88.5% with just 49 stations — 12.4 points better with 56% fewer sites. Model C maintains 88.5% coverage but right-sizes for the grid, saving $68M in 10-year NPV that the other models can't see.

The Investigations

Eight Questions. Three Fidelity Levels.

Click any card to see the full analysis.

Question 1

Where Are the Charging Deserts?

237 tracts have zero coverage. Average distance to nearest DCFC: 37.6 miles.

Population/traffic scoring
Question 2

Where Should Texas Build Next?

The optimizer achieves 88.5% coverage with 49 stations — 12.4 points better than 111 screening stations at 76.1%.

Facility location (MCLP)
Question 3

How Big Should Each Station Be?

Model A gives every station 8 ports. Model B sizes them 4–20 based on demand — top sites need 3 MW of grid capacity.

Demand-driven sizing
Question 4

What Happens When the Grid Says No?

30 of Model B's 49 stations can't get full power from their nearest substation. Real substations like AIRLINE (Harris) and TENTH STREET (Bexar) force downsizing.

Grid-aware optimization
Question 5

What Does It Actually Cost?

Model C saves $68M in 10-year NPV by right-sizing for the grid upfront. Model B's plan has 294 ports it can't actually power.

Full cost model
Question 6

Where Does Fidelity Change the Decision?

All three models agree on 72 of 146 unique locations. The other 74 — where the money goes — is where fidelity matters.

Cross-model comparison
Question 7

Does Our Model Match Reality?

Validated against 3,976 real AFDC stations. Spatial correlation, utilization benchmarks, and city-level ranking all confirm the demand model tracks reality.

Real-world validation
Question 8

BESS Co-location Strategy

Battery storage beats relocation at substations ≤70% loaded. Above 75%, grid upgrades are unavoidable regardless.

BESS cost analysis
The Fidelity Lesson

The Right Model for the Decision

Model FidelityCoverageTotal CostWhat a Planner Concludes
Model A: Population scoring 76.1% $199.8M capex / $244.4M NPV “111 stations. Fills gaps near people.”
Model B: Facility location optimizer 88.5% $199.4M capex / $247.0M NPV “49 stations, +12.4 points coverage, 56% fewer sites.”
Model C: Grid-aware placement 88.5% $146.1M capex / $178.9M NPV “Same coverage, $68M less in 10yr NPV.”
The ADM Punchline
~430 lines of code checking 4,151 real substations. That is the entire difference between Model B and Model C, and it saves $68M in 10-year NPV. The screening model cannot even see this constraint exists.

When does fidelity matter? For the 72 sites where all models agree, Model A is fine — it costs nothing and runs in under a second. For the other 74, the optimizer moves stations to better locations. For the 30 near constrained substations, the grid model downsizes or relocates them. Each question gets the model it needs. Nothing more.

Robustness

Does the Answer Change When Assumptions Move?

We varied four key parameters across their plausible range and re-ran all three models. The question: does Model C still win?

10-Year NPV by Model Under Parameter Variation ($200M Budget)
Robustness Finding
Model C has the lowest 10-year NPV in every scenario we tested. We varied home charging rate (±20%), peak demand (±30%), and substation loading (±15%). The savings bounce between $40M and $95M, but Model C never loses its lead.
Parameter Range Tested Effect on Site Selection Decision Impact
EV adoption rate 0.5×–2× current Minor: top sites stay top sites Sizing changes, not placement
Grid loading estimates ±20% Moderate: 200–380 ports constrained Magnitude changes, decision holds
Demand calibration 0.6×–1.0× Minor: rank ordering stable Port counts change ±25%
DCFC price ($/kWh) $0.30–$0.60 Minor: doesn't affect placement Affects revenue, not demand
Charger power 150–350 kW Moderate: fewer ports needed at higher power Sizing changes, placement stable
Budget $200M–$400M Major: doubles station count Opens rural coverage

The surprise is what doesn't matter. We expected placement to shift a lot with EV adoption assumptions — double the EVs, different stations, right? The top sites stay the top sites. Where you build is stable. It is how big you build that swings with the inputs.

We ran 27 scenarios varying home charging rate (±20%), peak demand (±30%), and substation loading (±15%). Model C wins on NPV in all of them. The savings range from $40M to $95M depending on assumptions, but the ranking never flips.

Sensitivity: baseline scenario ($200M, current EVs, mixed focus) with one parameter varied at a time. 20 total runs. Running with the DOE/FHWA-implied 5,190 kWh/year (vs our conservative 4,200) does not change model rankings — Model C still produces the lowest NPV.

Validation

How Our Model Compares to Published Data

We validate our assumptions against published benchmarks from DOE, NREL, and industry sources.

Parameter Our Assumption Published Benchmark Source Assessment
kWh per EV per year 4,200 5,190 DOE 34.6 kWh/100mi × 15K mi/yr (TX) Conservative
DCFC session energy 35 kWh 22–42 kWh DOE FOTW #1319 (2.4M sessions) Middle of range
Target utilization 30% 11–18% Stable Auto (Jun 2025 national avg) Optimistic (long-term target)
EV growth rate (TX) 35%/yr 35%/yr TX DMV (235K→472K, 2.3 yr) Matches data
Hardware cost per port $150K $140–$160K NREL 2024 (150kW DCFC) Center of range
Grid upgrade cost $300/kW $200–$500/kW NREL ATB 2024 Center of range
Substation capacity Voltage-class estimate Transformer MVA (not public) HIFLD locations only Estimated — needs utility data
Substation loading County-growth proxy SCADA metering (not public) No public source Estimated — needs utility data
Validation Summary
Six of eight parameters check out against published benchmarks. The two that don't — substation capacity and loading — require utility data that is not public. Our demand-side numbers are conservative. The grid-side numbers are estimates. Read Model C's findings as "grid constraints will bite here" rather than "these specific substations will fail."
10-Year NPV per Station: Grid-Naive vs. Grid-Aware (Top 20 Sites)
Limitations

What a Real Deployment Would Need

This study demonstrates how fidelity progression changes infrastructure decisions. It uses real public data — but real deployment requires data we don't have.

What We Have (Real)
3,976 AFDC station locations
6,896 Census tracts with demographics (centroids approximated ±6 mi from AADT stations)
41,467 TxDOT traffic stations
4,151 substation locations & voltage classes
472,572 EV registrations (weekly series)
NREL cost benchmarks
What We'd Need (Not Public)
Utility feeder data (Oncor, CenterPoint, AEP Texas)
Transformer MVA ratings per substation
Real SCADA loading data (hourly)
Distribution feeder architecture (radial vs mesh)
Interconnection queue and timeline data
Real charging session data for demand calibration

The gap between these columns is the gap between a planning study and a deployment decision. ADM tells you which column you need for which question: Model A needs only the left column. Model C needs both.

Recommendations

What a Texas Planner Should Do

Four recommendations grounded in the gap between population-based siting and grid-aware optimization.

  • 01

    Require grid feasibility checks before committing NEVI funds

    The grid-constrained model (Model C) disqualified 23% of sites that population scoring (Model A) ranked highest. Committing $408M in federal funding to sites that fail substation loading checks wastes money and delays coverage. A substation capacity screen takes hours to run and costs nothing.

  • 02

    Use coverage optimization, not population scoring

    MCLP-style optimization (Model B) covers 12% more demand than population-based placement at the same budget. The optimizer finds sites that serve overlapping coverage areas efficiently. Population scoring puts stations where people already are — not where the gaps are.

  • 03

    Prioritize utility data sharing

    The largest source of uncertainty in the grid-constrained model is substation loading data. Public datasets provide locations and voltage classes but not MVA ratings or real-time loading. Two of the five validation parameters are flagged “Estimated — needs utility data.” Oncor, CenterPoint, and AEP Texas hold the data that closes this gap.

  • 04

    Plan for the grid you'll have, not the grid you have

    Sensitivity analysis shows the top sites are robust to EV adoption assumptions — doubling EVs doesn’t change where to build. But substation loading will change as EVs, heat pumps, and data centers compete for distribution capacity. Siting decisions made today should account for 2030 loading, not 2024.

Interactive

Explore the Data

See the coverage gaps, model differences, and grid constraints for yourself.

Scope: Three models totaling ~2,750 lines of Python, including MCLP facility-location optimizer (PuLP/CBC), grid-aware placement checking 4,151 real substations, and demand-driven station sizing. Data from AFDC/NREL API (3,976 stations), Census Bureau API (6,896 tracts), TxDOT Open Data Portal (41,467 traffic stations), HIFLD/DHS (4,151 substations), and TX DMV (472,572 EVs). 27 scenarios across 3 adoption levels × 3 budgets × 3 focus types. Limitations documented honestly: no real-time utilization data, simplified grid upgrade cost model, straight-line distance to substations (not feeder routing), no temporal demand patterns.

Data Provenance

Sources & Data

100% real data. Every number in this study comes from public data. AFDC stations via NREL API. Census demographics via Census Bureau API. Traffic counts from TxDOT Open Data Portal. Substations from HIFLD. EV registrations from TX DMV. No synthetic data.

AFDC / NREL Alternative Fuels Station Locator — 3,976 Texas EV charging stations, 12,951 ports. Retrieved via NREL API, March 2026. afdc.energy.gov/api

U.S. Census Bureau — American Community Survey 5-Year Estimates (2022) — 6,896 Census tracts, 29.2M population. Retrieved via Census Bureau API. census.gov

TxDOT Traffic Count Data — 41,467 traffic stations, 2024 Annual Average Daily Traffic (AADT). TxDOT Open Data Portal. gis-txdot.opendata.arcgis.com

HIFLD / DHS Electric Substations — 4,151 in-service substations in Texas, with location and voltage class data. Homeland Infrastructure Foundation-Level Data. hifld-geoplatform.opendata.arcgis.com

Texas DMV EV Registration Data — 472,572 registered electric vehicles, weekly time series through March 2026. Texas Department of Motor Vehicles.

NREL Annual Technology Baseline (ATB) 2024 — Grid upgrade cost benchmarks: transformer, feeder, and substation expansion. National Renewable Energy Laboratory. atb.nrel.gov

DOE Fact of the Week #1319 — DCFC session energy distribution (22–42 kWh across 2.4M sessions). U.S. Department of Energy Vehicle Technologies Office.

Stable Auto EV Charging Utilization Report (June 2025) — National average DCFC utilization: 11–18%. Published benchmark used for validation.