The third plug is worth more than the twelfth

Every CPO with a budget faces the same fork in the road. You have money for one upgrade at a busy site. Do you buy more power per bay, or more bays?

Kempower's August white paper, More plugs, more utilization, says plugs. We think they are pointing the right way. We also think the paper does not prove it — and that the real proof is more useful, because it tells you when to stop.

The telephone problem

A charging site is a queueing system: arrivals, servers, waiting. Telecoms solved this a century ago.

Agner Krarup Erlang, working at Copenhagen Telephone Company, published the model in 1917 to size switchboards. It is still how you size a call centre or a cell site. The result that matters here is trunking gain: pooling servers is super-linear. Three operators handle 2.3 times what two handle, not 1.5 times, because a bigger pool absorbs randomness a small pool cannot. That never stops being true — at every size, capacity grows faster than the server count. What changes is the size of the premium, and it shrinks fast: worth +80 percentage points over linear at three servers, +1.7 at seventeen, and essentially nothing by a hundred. That decay is what decides where a charging site should stop.

Run it for charging bays, holding service quality fixed at “5% of arrivals have to wait”:

Source: Erlang C, Chargalytics calculation. The dashed line is what you would get if capacity simply scaled with bay count.

Adding the third bay to a two-bay site buys you +130% capacity — more than double, for a 50% increase in hardware. Adding the twelfth bay to an eleven-bay site buys you +12%, against the +9% you paid for.

Does more power actually shorten a session?

Before spending money on either option, it is worth checking what power actually buys. And here is the awkward part: we measure charging time. We do not measure energy. So we lean on the evidence that needs no energy at all.

Sort each market’s sites by how much power is actually installed per bay, and ask what that buys in session time. Going from the 100–175 kW band to the 175 kW-plus band is roughly a doubling of nameplate:

Source: Chargalytics unified telemetry, June 2026. Every DC site with at least 50 sessions and 10 observed days — 5 100 sites, 4.0 million sessions. Measured charging time against measured nameplate, with no energy assumption.

In four of five markets, doubling the installed power per bay buys between nothing and thirteen per cent off a session. In Sweden it buys nothing at all. That is the charging curve doing what charging curves do: above roughly 150 kW the battery, not the charger, sets the pace.

Below that line it is a different story. Going from sub-60 kW hardware up to the 100–175 kW band — roughly a tripling — cuts sessions by 17% to 50%. At 50 kW the charger really is the bottleneck. The appendix works through why that leaves France and Switzerland with much longer average sessions than the Nordics.

That is not a failure of the hardware. It is physics. Charging curves taper hard above 60–70% state of charge, and a queue of real cars with real batteries never sustains nameplate power. The top half of a 350 kW rating spends most of its life doing nothing.

Which is the entire commercial case for load balancing across a site rather than hard-wiring power to an outlet. But it is only half the argument.

One more plug, or twice the power?

Erlang answers this cleanly. At a four-bay site running 50% busy, the chance an arriving driver has to queue is 17.4%. Double the power per bay and it falls to 10.9%. Double the bays and it falls to 0.1%.

Source: Erlang C, Chargalytics calculation.

The cleanest way to say it: at a four-bay site running 40% busy, one extra bay does as much for queueing as cutting every single session by 33%. At a two-bay site the equivalent is a 56% cut. No charging technology on the market delivers that. An extra dispenser is available today, from stock.

Does this actually happen? Yes — and almost only at small sites

So far this is theory. Our telemetry lets us check it. We took the busiest hour of the day at every site and asked the model how often a driver arriving then would find every bay full.

21.2% of European DC sites have more than one arrival in twenty queueing at peak. One site in twenty has more than one arrival in five queueing.

But look where the pain actually is.

Source: Chargalytics unified telemetry, June 2026, 8 672 DC sites with a full hour-of-day profile. Modelled with Erlang C at measured peak-hour load.

The split by market is just as lopsided — and it is not explained by how many bays a country has built.

Sources: Chargalytics unified telemetry, June 2026; BEV stock from Chargalytics market analysis, 2026 base scenario.

Compare Switzerland and Norway. They have almost exactly the same national provision — 14.6 and 14.1 DC bays per thousand electric cars. Swiss sites even run cooler at peak, 16.0% load against Norway's 18.3%.

And yet 38.8% of Swiss sites queue at peak against 7.1% of Norwegian ones. Five and a half times the pain, from the same number of bays per car, at lower load.

The difference is how those bays are arranged. Norway averages 9.4 bays per site. Switzerland averages 2.7. Same provision, opposite experience — because the Swiss network is chopped into pieces too small to benefit from pooling.

Which is a warning about the metric the whole industry plans with. Bays per electric car tells you almost nothing about whether drivers queue. Finland has twice Norway's ratio — 29.4 bays per thousand BEVs against 14.1 — and more than twice the share of sites queueing. How you group the bays matters more than how many you build.

None of this is new to us. In July we published Europe's charging network survived July. 145 sites didn't., which tracked 20 955 DC bays through the holiday peak and found 145 sites running above 80% load in their peak hours — seven times as many as in May. Same phenomenon, seen from the other end: a handful of sites absorbing the pain while the network average stays comfortable.

That piece put it well: portfolio averages hide the pain. A CPO reporting 16% July utilization can still be turning drivers away at a dozen sites while idling at two hundred. What June's hour-by-hour data adds is which sites — and the answer is overwhelmingly the small ones.

Half of all two-bay sites are in queueing territory at peak. Under 1% of eight-bay sites are. By ten bays it is essentially zero.

And this is not because big sites are quiet. Peak-hour load per bay is remarkably flat across site sizes — between 16% and 25% everywhere. The queues vanish because of trunking gain alone, exactly as Erlang predicts.

That is the whole argument in one chart. The queueing problem in European fast charging is a small-site problem. It is not a power problem, and above about eight bays it is not really a problem at all.

And here is where the argument runs out

This is the part that tends to go missing from vendor material, including Kempower's.

Look again at the first chart. The trunking bonus is enormous at two bays and almost gone by sixteen. But the blunter way to see it is in absolute queueing.

Take a site running 40% busy. At two bays, 23% of arrivals wait. At four bays, 9%. At twelve bays, 0.4% — and a thirteenth bay would take that to 0.15%. At that size you are spending real capital to fix a queue that has already stopped existing.

The plug argument is strongest exactly where sites are smallest, and it fades as they grow.

So “more plugs” is not a universal law of charging site design. It is a very strong argument for going from two bays to four, a decent one for going from six to eight, and a weak one for going from twelve to fourteen. Anyone selling you a twelfth plug on the strength of what the third one did is showing you the wrong chart.

So what should a big site do instead?

Balance the load across the whole site. And this is the neat part: the two levers run in opposite directions.

Adding a bay matters most when a site is small. Site-wide load balancing matters most when a site is large — because a large site split into two-outlet cabinets has more separate islands to strand power in. One car charging alone in a paired cabinet is capped by that cabinet, while the rest of the site sits idle.

Source: Erlang C and the Chargalytics power-allocation model. Small sites should add bays; large sites should balance load.

At a four-bay site, balancing site-wide instead of in pairs is worth about 50% more delivered power per car under light load. By six bays it is roughly 60%, and it stays there. Meanwhile the value of one more bay has collapsed from +130% to +26%.

Small site: add plugs. Large site: balance the load. Both are true, and neither generalises to the other.

What our own data actually shows

We looked at per-bay utilization across 8 966 European DC sites. Bigger sites are less utilized per bay, not more — and the pattern differs sharply by market.

Source: Chargalytics unified telemetry, June 2026. France is U-shaped, the Nordics flat, Switzerland rising. None of them decline.

Controlling for location quality, power per bay and country, doubling the number of bays is associated with 10.5% lower utilization per bay. Doubling kW per bay goes the other way: 7.7% higher. Location quality remains the single strongest predictor of them all.

Read quickly, that looks like a contradiction of the white paper. It is not — and the commercial reading is more interesting than either.

Utilization is the wrong single scoreboard

Not every operator sizes a site the same way, and it matters who is paying.

Networks where charging is a means to another end — Tesla protecting the ownership experience, IONITY delivering a corridor promise for its carmaker shareholders — can size for peak and accept the empty Tuesday. The headroom buys brand reliability, and the return shows up somewhere other than the charger.

A for-profit CPO does not have that luxury. It has to serve both numbers at once. Peak demand determines whether drivers come back. Average demand determines whether the site pays. Revenue is made on energy sold across the whole month — in our data, usage per bay per day — while queues at peak quietly destroy the repeat business that fills the rest of the week.

So the dilution is real but modest — and in the cleanest markets it is close to nothing. None of the five countries above shows utilization falling away as sites get bigger; France is U-shaped, the Nordics flat, Switzerland actually rising. Whatever a bigger site costs an operator, it is not a collapse in how hard each bay works.

Our July analysis found the same tension from the revenue side: the sites that saturate in the holiday peak are precisely the ones that spend the other eleven months half empty. Headroom is not free, and it is not evenly distributed.

Which is why the framing in the white paper is the wrong way round. Kempower asked whether more plugs raise utilization. On our data they barely move it either way. The question worth asking is what a marginal bay does to queueing — and there the answer is large, measurable, and concentrated almost entirely at small sites.

What this means for how you build

If plugs beat power at small sites and load balancing beats both at large ones, the architecture follows. A power block feeding multiple dispensers — PB/D — balances one rectifier bank across the whole site, so idle capacity flows to whoever is actually charging. Balancing inside a single two-outlet cabinet covers two bays. PB/D covers twelve.

Most CPOs already load balance in some form. The question is the size of the pool, and our model says the difference between balancing in pairs and balancing site-wide is worth roughly 60% more delivered power per car at any site above six bays.

Worth watching: EcoG's hardware-agnostic Powerblock/Dispenser platform, built with Jabil, which supports 32 charging architectures and 15 power converters. If PB/D becomes genuinely multi-vendor, CPOs stop being locked to one OEM's roadmap for a fifteen-year asset. For most networks that matters more than any single vendor's plug count.

And then there is what a bay costs

Everything above treats bays as if they were priced the same. They are not, and the difference cuts directly across the Erlang argument.

A standalone two-outlet cabinet — the classic Alpitronic pattern — is a self-contained box with its own rectifiers. Hardware and installation scale more or less linearly with the number of boxes. Cost per bay is flat: the tenth bay costs what the second did.

PB/D behaves completely differently. The expensive part is the power block, and it is shared. Each additional bay is a comparatively cheap satellite, so cost per bay falls the more bays you hang off one block — until you fill it, at which point the next bay drags in a whole new block and the curve jumps.

Indicative structural model, not Chargalytics price data. See the assumptions above. The jump at nine bays is a second power block.

On these assumptions PB/D is the worse buy at two or three bays — you are paying for a power block that almost nothing is sharing. It crosses over around four bays, and by eight it is roughly a third cheaper per bay. Then the ninth bay costs you a step.

Which sets up the tension that decides most site designs. The queueing value of a marginal bay collapses as sites grow. The cost of a marginal bay falls as sites grow. Those two curves run against each other, and where they cross is where the argument for a site actually lives.

For a for-profit CPO that is a genuinely useful asymmetry. At a large PB/D site the extra bay is cheap, so the dilution of average utilization stings less — you are buying peak coverage at a discount, and covering the four weeks a year that decide whether drivers come back. It is a much easier board paper than the same bay at a standalone site, where it costs full price and still does almost nothing for queueing.

The trade-off is real and worth stating plainly: a shared block divided across more bays means lower peak charging speed when the site is genuinely full. Off-peak, which is most of the time, drivers see the full satellite rating. At rush hour they see less. Given that the average session already draws well under half of nameplate, that is usually a cheap price — but it is a price, and it is paid on exactly the days the site is busiest.

The short version

• Doubling installed power above 150 kW per bay buys 0-13% off a session. Measured, no energy assumption needed. Above that line the battery sets the pace, not the charger.
• Going from 2 bays to 3 buys +130% capacity. Going from 11 to 12 buys +12%. The plug argument is a small-site argument.
• One extra bay at a four-bay site equals a 33% cut in session times. Nothing on the market delivers that.
• Large sites should balance load, not add plugs. Site-wide balancing is worth ~60% more delivered power per car than paired cabinets.
• A marginal bay cuts peak queuing and dilutes average utilization. For-profit CPOs are paid on the second and judged on the first.
• Half of all two-bay sites already queue at peak. Under 1% of eight-bay sites do. The queueing problem in European fast charging is a small-site problem.
• PB/D cost per bay falls as the site grows; queueing value per bay falls too. Where those two curves cross is the real site-design decision.

Kempower is directionally right, and right for a reason their own paper does not quite land. Plugs beat power because at under half of nameplate delivery, extra rated power has nowhere to go — not because plugs raise utilization. Our data says they lower it, and that is a cost a for-profit operator has to justify.

And the caveat that sells no hardware: the effect fades. The third plug is worth more than the twelfth. Build accordingly.

Appendix

Why session length varies so much between markets

Mean charging time per session ranges from 23.5 minutes in the Netherlands to 39.2 in France. Taken at face value that looks like it contradicts the argument above about power not mattering — so it is worth being precise about when power matters and when it stops.

The gap tracks one thing: the share of legacy sub-60 kW DC still in the ground. Switzerland is 42.7% and France 30.9%. Sweden is 8.2%.

Source: Chargalytics unified telemetry, June 2026. Sites with at least 50 sessions in the month.

Inside every single market, sessions on sub-60 kW sites run 1.35 to 2.1 times longer than on 175 kW-plus sites. And when you compare only the modern sites, the country gap largely closes: 24.5 minutes in Norway, 24.8 in Sweden, 29.0 in the Netherlands, 30.9 in Switzerland, 34.4 in France.

So the rule is not “power does not matter”. It is that power matters until the car becomes the constraint, and then it stops. Going from 50 kW to 150 kW roughly halves a session, because at 50 kW the charger is the bottleneck. Going from 150 kW to 350 kW changes almost nothing, because by then the battery is.

That is precisely why buying power is a weak move at a modern site and a strong one at a legacy site. Most of Europe's DC fleet is already past the knee in the curve.

A residual gap does survive — French sessions still run about 39% longer than Swedish ones on comparable hardware. We cannot settle that from our data. The likeliest candidates are vehicle mix (France skews to smaller cars with lower charge acceptance) and siting, with more French DC at destinations where drivers linger. Both are worth a separate look.

Deriving delivered power, and why we keep it out of the argument

We hold no kWh data. To talk about delivered power in kW at all, we have to borrow session energy from operators who publish it — Fastned at 26.9 kWh per session, IONITY at a stated 30 kWh — and apply the same figure to every market.

That is a real assumption, and it is why none of the main argument rests on what follows.

The sample is 5 300 160 DC charging sessions across five European markets, June 2026.

Delivered power is derived, not measured: it is published session energy divided by our measured charging time. Because the same energy figure is applied to every market, the delivered column is a restatement of charging time and carries no independent information about energy. Treat the levels as illustrative and the ranking as provisional.

On that basis the average session draws somewhere between 30% and 45% of the power the hardware is rated for — well under half, in every market. The health warning is that this range would collapse if session energy varies much by market. For France to be delivering the same power as Sweden, French sessions would have to carry 1.56 times the energy — about 42 kWh against 27. We think that unlikely, given France skews to smaller cars, but we cannot rule it out from our data. It is why the main argument is built on measured time, not on this.

Notes on the Kempower paper

We read More plugs, more utilization (Kempower, August 2026, form-gated) closely, including extracting the scatter-plot coordinates. Three things a reader should know.

The “3× stronger” claim compares two different samples. The plugs chart carries roughly 30 points with utilization capped at 11%. The power chart carries roughly 45 points reaching 23%. Correlation coefficients from different samples cannot be divided to rank two drivers.

Their own energy charts show plugs and power as near-equals — R = 0.47 versus R = 0.401 — printed beneath a headline saying plug density matters far more. Our data agrees with their charts: we measured 0.657 versus 0.667 on the same comparison.

Their utilization metric is defined as utilization of total installed power, which puts installed power in the denominator, and then plots it against installed power. That construction weakens the power relationship mechanically, independent of any real effect.

None of this makes their conclusion wrong. It makes it unproven by the evidence shown.