Where can robotaxi depots actually get power?

San Francisco: the busy part is the tight part

PG&E publishes, for every feeder and substation transformer bank in San Francisco, its rating and its forecast peak load for each year to 2040.¹ Rating minus the 2026 forecast is the spare capacity. We took the 126 surface parking lots of an acre or more in the city,² found the PG&E line each one sits next to, and kept the smaller of the feeder's and the substation's spare capacity. That is the room a depot there could plug into without an upgrade, on PG&E's own forecast.

41 lots have room for 1 MW. 22 have room for 2 MW and 14 for 5 MW. For 45 more, PG&E withholds the load figure. 29 are not near a PG&E San Francisco feeder at all: the Presidio, Treasure Island and lots just over the county line. The feeder, not the substation, was the tighter of the two for 49 of the 52 lots with numbers.

1 MW is not a lot of depot. Running flat out for 24 hours it delivers 24 MWh, enough for at most 286 robotaxi-days at 83.8 kWh each (the bridge section shows where that figure comes from).

The demand sits where the grid is tightest. PG&E lists a capacity need by 2027, or has a transformer bank already over its rating, at 5 of the 19 San Francisco substations in its data: Potrero, Embarcadero, Larkin, Mission and Hunters Point.¹ The only public count of ride-hail trips in San Francisco is the city transportation authority's: Uber and Lyft pickups and drop-offs by zone, for a typical week in late 2016.⁴ Of the 2,199,419 weekly pickups and drop-offs we could place on a PG&E feeder, 1,428,811 (65%) sit on those 5. Robotaxis serve the same trips, so a depot near the riders lands on the tightest part of the grid.

Austin: no headroom data, so watch the wires

Austin Energy is a city-owned utility and publishes no feeder or substation headroom. ERCOT, which runs the Texas grid, does publish every transmission limit that binds in its five-minute dispatch,⁵ the substations at each end,⁷ and the price at Austin's load zone.⁶

From Oct 12 to Oct 19, 2 lines into Austin Energy substations bound. The one between Magnesium Plant and Northland bound in 65 of 2,165 dispatch runs (3%), at an average shadow price of $462/MWh; the one between Burleson and Hi Cross bound in 6. Austin's load zone averaged $42.57/MWh against $33.55 for the ERCOT hub average, 27% more.

Austin Energy's own slides say the rest. Summer peak is 3,110 MW against about 800 MW of local generation, so the city leans on imports, and $500M of transmission upgrades are planned.¹⁰ On large new loads across ERCOT, it says transmission planning, approvals and construction "can take 3-5+ years".⁹ A depot is smaller than the data centers that sentence is about, but Austin publishes no depot-sized timeline.

The bridge: what a battery buys while you wait

PG&E's own reports put numbers on the wait. EV charging service jobs took 467 days end to end on average.¹² When the site needed a new circuit or a circuit upgrade, PG&E averaged 1,142 days, about 37.5 months. Substation upgrades averaged 1,307.¹¹

A bridge is energy that arrives some other way: a battery trucked in, or a big battery on site filled when the grid has room. Pick a fleet size and a wait below. The constants come from Waymo,¹⁴ the EPA,¹⁵ Xos¹⁶ and Tesla.¹⁷

The arithmetic points one way. Twenty robotaxis need about 2.7 mobile battery fills a day; a hundred need 13. Trucks can carry a small stop inside the zone through a 37.5-month wait. They can't carry a whole edge depot. PG&E also offers Flex Connect, which lets a site connect now under hourly limits while upgrades are built, for sites facing waits of more than 12 months.¹³

Five more cities: how fast can a depot open?

A site opens when it has both a permit and power. Cities publish their permit records, so we measured the days from application to issue for commercial EV charger permits since 2023. For power, California's utilities file their average time to energize an EV charging job with the CPUC. The other utilities publish none.

Where both are public, power sets the date, not the permit. Permits take days to about two months. SDG&E averaged 347 business days to energize an EV job and SCE 571, about 1.4 and 2.3 years.²⁰¹⁹

Where power exists matters as much as how fast. Of the 100 Los Angeles sites our depot look-alike screen ranks highest, 38 sit in SCE areas, and 28 of those have a circuit with room for 1 MW or more on SCE's grid-capacity map.²¹ In LADWP areas, 47 sit by a 34.5 kV line with more than 7.5 MW to spare, which usually means building your own transformer yard, and 13 only by local 4.8 kV lines that offer at most 600 kW.²²

Dallas, like Austin, shows up only in ERCOT's data. From Oct 14 to Oct 21, the North load zone that covers Dallas averaged $28.74/MWh against $30.46 for the hub average, 6% cheaper. Austin paid 22% more the same week.⁶

Method

• Candidate sites: OpenStreetMap parking areas of at least 1 acre (4,047 m²), excluding underground, multi-storey, rooftop and street parking. They are lots big enough to hold a depot, not lots for sale or ours. Shown only as counts per square.
• Demand: we count only lots inside a published robotaxi service area:³ all 126 in San Francisco, 582 of 1,053 in Austin. San Francisco also shows ride-hail demand: SFCTA's weekly Uber and Lyft pickups plus drop-offs per zone, placed at each zone's centre.⁴ Austin has no public ride-hail trip data, so its demand is the service area alone.
• SF headroom: each lot and each demand zone's centre is matched to the nearest PG&E line section within 300 m, which gives its feeder; the feeder id gives the substation. Feeder room is rating minus 2026 forecast peak. Substation room is the same summed over its banks. A lot's room is the smaller of the two.¹
• Tight grid (SF): a substation where PG&E's grid-needs data lists a capacity need in 2027 or earlier for any bank or feeder, or where a bank's 2026 forecast is over its rating.
• Austin: ERCOT binding constraints whose end stations are Austin Energy substations, from ERCOT's bus mapping. 61 of 66 substations matched to OpenStreetMap by name. Lines on the map are drawn straight between the two substations, not along the real route. Prices are 15-minute settlement prices for LZ_AEN and HB_HUBAVG over the same week.
• Days to open: permit days are calendar days from application to issue, median, for permits issued since 2023 whose description mentions EV charging and that the city flags as commercial (San Diego has no flag, so all property types count). Power is the utility's own filed average for EV charging jobs, in business days, read as years at 250 working days a year.
• LA power: for each of the 100 top-ranked candidate sites, the best circuit within about 150 m on SCE's load capacity map or LADWP's capacity map, taking the low end of LADWP's ranges.
• Bridge: 190 miles a day is Waymo's weekly miles divided by its fleet and seven days. EPA's 44 kWh per 100 miles is measured at the wall, so charging losses are in it.

Caveats

• Spare capacity on a forecast is a screening number, not a connection offer. Only PG&E or Austin Energy can say what a specific site can get, and hourly limits can allow more than the peak-hour figure.
• PG&E withholds load data for feeders with few customers, so many lots show no number. Downtown's networked grid works differently from a radial feeder.
• The ride-hail counts are from November and December 2016 and cover trips within San Francisco only. They are the latest public trip data; the CPUC and the companies have not released newer zone counts. Demand has grown and shifted since, so read them as where riders cluster, not how many there are now.
• One October week of ERCOT data. Constraints bind most in summer. ERCOT's public download keeps about a week, so we will re-pull it monthly.
• Utility averages include time the customer spends on design and paperwork. Most LA EV permits were issued without plan check, and the data has no application date for them, so we report their share, not their days.
• Bridge figures leave out cost, permits, and the time to charge each car, and the deploy time is a vendor's claim.
• PG&E data is PG&E's, shown here aggregated and credited. ERCOT allows its public data in analyses.⁸ OpenStreetMap data is available under the ODbL.