Power and Cooling Budget for a Robotics + AI Compute Lab

The compute side of a robotics lab is the part that gets photographed for the website. The wall outlet is the part that decides whether any of it actually works. Most labs we have seen that "grew into AI" underestimate their sustained electrical load by roughly a factor of two, find out about it in month one when a breaker trips during a training run, and then spend the next quarter rewiring a room they thought was done.

This article is the budget version of that conversation, before you sign the lease or order the server. It covers the full load list — server, robots, workstations, network, lighting — the cooling math that determines whether the room can keep up, the electrical install you actually need, and a checklist to hand to the electrician before they put a drill into the wall.

Audience: lab managers, integrators, and buyers in Europe who are putting one or two AI servers and a couple of humanoid or quadruped robots into a single room. We will assume EU 230 V / 400 V electrical reality throughout. North American readers can divide by two on most voltages and arrive at roughly the same conclusions.

The full load list

A robotics + AI lab has more devices drawing continuous power than first-time builders count. The compute is the headline; the rest is the body of the bill.

A few things worth flagging before we work an example:

• The G1 charger spec (54 V, 5 A ≈ 270 W) is much smaller than people guess. Larger humanoid platforms in the 70–130 kg class can draw 1 kW or more from a fast-charger; the rule of thumb "humanoid charging = 1 kW" is conservative but not crazy. Check the actual charger spec for your unit.
• Workstations are not "free." A developer with a 5090 in their desktop, doing local inference and debug, can sustain 500 W. Four developers is 2 kW continuous, comparable to a 4-GPU server.
• Lighting and misc add up. A 30 m² lab with proper task lighting, three printers, and a soldering bench is easily 1 kW before any AI hardware shows up.

Worked example: one server, two humanoids, four desks

A realistic build-out: 1× K-AI 256/8x server, 2× humanoid robots on charging docks, 4× developer workstations with monitors, network gear, lighting, and a small workshop bench.

1× K-AI 8-GPU server            ~ 4.5 kW
2× humanoid chargers (~1 kW ea) ~ 2.0 kW   (assume larger platforms, conservative)
4× workstations (500 W ea)      ~ 2.0 kW
8× monitors (50 W ea)           ~ 0.4 kW
Network gear + APs              ~ 0.2 kW
Lighting + misc                 ~ 0.6 kW
                                --------
IT + lab load                   ~ 9.7 kW sustained

Round to ~10 kW sustained for planning. That number drives every other decision in this article. Note also: this 10 kW figure assumes both humanoids are charging simultaneously, which they will be if the team is bad at scheduling — which they will be. Plan for the case where everything draws at once.

For comparison: a smaller lab with a single 4-GPU server, one humanoid, two desks, and the rest scaled down lands around 5–6 kW. A larger lab with two 8-GPU servers and a fleet of four robots hits 18–20 kW comfortably. The "10 kW lab" is the median case we see in 2026.

Single-phase vs three-phase: when each is enough

Here is the same question compressed to what matters for a lab: which circuit holds the load?

The 10 kW lab sits just above the 400 V 16 A three-phase continuous limit, which means a 16 A three-phase feed is theoretically possible but leaves no headroom for the second server or a third humanoid. The right answer for a 10 kW lab in 2026 is a 400 V 32 A three-phase feed on an IEC 60309 (CEE) red 5-pin connector, split internally to single-phase 16 A branch circuits at the load panel.

The 32 A three-phase choice is not just about today's load. It is the cheapest insurance against "we added a server" surprise. The marginal install cost between 16 A and 32 A three-phase is small — the breaker, the connector, and a slightly heavier cable run. The cost to redo it twelve months later is several times higher because you are pulling new cable through finished walls.

The "I'll just plug it in" trap

Almost every first-time lab thinks the existing room wiring will do. It almost never does.

A typical European office room has one or two 16 A single-phase circuits feeding all the outlets, often shared with the lighting on the same panel feed. Plug in a 4-GPU server (2.4 kW), a humanoid charging (1 kW), two workstations (1 kW), monitors and lights (300 W), and you are at 4.7 kW on a 3.7 kW theoretical / 2.95 kW continuous circuit. The breaker trips on the first sustained training run. Sometimes it trips at 02:30 on a Tuesday when nobody is in the building, and you arrive Wednesday to find the server has been hard-rebooted three times overnight.

This is preventable in fifteen minutes of forethought. Count the load, look at the breaker, decide whether you need to call an electrician before you place the server order. The electrician site visit is in the low hundreds of euros; the rework is in the low thousands plus a month of project drag.

A particularly nasty variant: the circuit holds at average load but trips on transient peaks. An 8-GPU 5090 server (4.5 kW sustained) can spike to 6 kW for tens of milliseconds during a training step transition. A B-curve 32 A single-phase breaker (7.4 kW continuous) absorbs this fine. A C-curve 16 A breaker on the wrong side of a heavily-loaded circuit does not. You will see this as "the server randomly reboots when we start a training run," diagnosed eventually by an electrician with a clamp meter.

HVAC: every kW in becomes a kW out

The first physics rule of a compute lab: all the electrical power you put into the room ends up as heat in the room. A 4 kW server is a 4 kW space heater with extra steps. The HVAC must remove that heat at the same rate, or the room temperature climbs until either the equipment throttles, the people leave, or both.

1 kW of heat  =  3,412 BTU/h
1 ton of cooling  =  3.517 kW  =  12,000 BTU/h

For the 10 kW lab, you need to remove 10 kW of heat continuously. That is ~34,000 BTU/h, or about 2.85 tons of cooling. But equipment is rated at nominal output under specific conditions, and a room with poor airflow, sun on the south wall, and people inside loses cooling capacity. The standard headroom rule:

HVAC capacity  =  1.3 to 1.5 × IT + lab load

For 10 kW IT, install 13–15 kW of cooling capacity. Round to a 15 kW system. This sounds like overkill until the first 35 °C July afternoon when the building's ambient creeps up and your nominal-rated AC delivers 80% of its plate spec.

ASHRAE TC 9.9 sets the thermal envelope every datacenter design follows. The current recommended range is 18–27 °C dry-bulb at the server inlet, with humidity 8 °C dew point to 60% RH, ≤ 80% RH ceiling. The newer H1 class for high-density AI/HPC tightens this to 18–22 °C. For a small lab with prosumer-grade GPUs (5090, 4090), the wider 18–27 °C window is fine; aim for 22 °C as the design point and accept that summer afternoons may drift to 25 °C.

Humidity matters as much as temperature, for two reasons. Below 30% RH, electrostatic discharge risk rises — bad news for humans handling robot electronics and bad news for sensitive sensors. Above 80% RH, condensation risk on cold surfaces rises and corrosion accelerates on long timescales. Aim for 40–60% RH year-round. Most split AC units do not actively manage humidity; for a lab in a humid coastal climate, this is worth a dedicated dehumidifier or a CRAC-class unit that does.

Cooling options: portable vs split vs CRAC

The honest recommendation by lab size:

• Under 5 kW IT load: one good dual-hose portable AC works in a pinch, but a single wall-mounted split AC unit is cleaner, quieter, and roughly the same money once you account for the portable's worse efficiency over a couple of years.
• 5–15 kW IT load: wall-mounted or cassette split AC, sized at 1.3–1.5× the IT load. This is the sweet spot for a one-room lab. A 15 kW split system is a single outdoor compressor unit and one or two indoor heads.
• 15–40 kW IT load: multi-split system or move to a small CRAC. At this scale you start wanting humidity control and you genuinely need to think about redundancy (single compressor failure = no cooling = thermal shutdown in 20 minutes under load).
• Over 40 kW IT load: dedicated CRAC unit or chilled-water loop; you are running a small server room, not a lab. This is outside the scope of this article.

Portable AC deserves a specific warning. Single-hose portable units take cool air from inside the room, blow heat out through the hose, and pull replacement air from the rest of the building (or outdoors, through whatever leak path exists). This means they are net inefficient — for every 1 kW of cooling delivered, they leak ~0.3–0.5 kW of warm air back in. They are also noisy enough to make a working space unpleasant. Use them as a stopgap, not a plan.

Airflow: front-to-back rack discipline matters

The HVAC delivers cool air to the room; the rack converts it to useful work. The mechanical question is whether the cool air actually reaches the GPU intakes before it gets mixed with hot exhaust air.

The standard convention is front-to-back rack airflow: cool air enters the rack at the front, passes through the servers, exits at the rear. K-AI servers follow this convention with 3× 120 mm front intakes and 1× 120 mm rear exhaust per chassis, sized for sustained 24/7 load with industrial fans. This is the same convention every server hall uses, for the same reason: it lets you separate cold supply from hot return.

In a serious datacenter the cold supply and hot return are physically separated by aisle containment — panels, curtains, or full enclosures around the rack rows. In a small lab with one or two racks against a wall, you typically cannot build a full hot-aisle containment, but you can do the simple version:

• Server fronts face the cold-air source (the split AC outlet, or the doorway from a cooler room).
• Server rears face the hot-return path (a wall, ideally with the AC return high on the same wall).
• No equipment with intake on the side or top mixed into the rack. (This rules out office desktops piled on top of the rack, which we have seen more than once.)

Without separation, cold supply mixes with hot exhaust before reaching the rack, the effective cooling capacity drops by 15–25%, and the GPUs slowly heat-soak through the afternoon. Studies of datacenter containment show 20–25% fan energy reduction and 20% chiller reduction when proper containment is added. The lab-scale version of the same principle: do not let the air mix where it does not have to.

Battery and charging considerations

A robot charging is not a steady load. The charger pulls near-rated power for the first 70–80% of the charge cycle, then tapers. A humanoid that comes back from a training session with 20% battery and gets plugged in pulls full charger power for an hour or two, then trickles for another hour.

Two implications:

1. Charger draw competes with the server. If your circuit holds the server alone but is sized at exactly 80% of breaker, plugging in a robot at full draw pushes you over. The 10 kW lab calculation above assumed both robots charging simultaneously; if your reality is "charge one robot while training," you can shave a kW from the budget. If your reality is "charge both robots at the end of every working day while a long training job runs overnight," you cannot.
2. Stagger or oversize. The clean answer is to oversize the circuit. The cheap answer is to stagger charging via a switched PDU outlet schedule — robot A charges 18:00–22:00, robot B charges 22:00–02:00, training job has the full circuit between. The PDU-scheduling answer is fragile because humans forget to plug in robots; the oversize answer is the one we recommend.

A subtler point: robots have their own batteries, so an outage of a few minutes is not a robot problem. You almost never need to UPS-back robot chargers — the robot stays alive on its own battery, and the charging just resumes when power returns. The UPS budget is for the server, not the fleet.

UPS sizing

The server is the part you cannot afford to lose to a brownout or a brief outage. A graceful shutdown from a sustained training job takes a couple of minutes. A flapping mains feed (drops to 0 V for 200 ms, comes back, drops again 30 seconds later) takes more.

Always specify online double-conversion topology for AI compute. Line-interactive UPS units have a 4–10 ms transfer time when switching to battery — fine for office workstations, occasionally catastrophic for an inference server in the middle of a CUDA kernel. The 4 ms is not enough to crash all hardware, but it is enough that we have seen training jobs hang and require a hard reboot. Double-conversion has zero transfer time because the load is always on inverter; the battery just becomes the DC source.

A common mistake: under-sizing the UPS for the transient peak. A 3 kVA UPS in front of a 2.5 kW sustained server with 3 kW transient peaks will trip to bypass exactly when you needed it most. Size the UPS for the transient envelope, not the sustained average. This is the same logic as PSU sizing from W04.

The robots, as noted, do not need UPS. Save the kVA budget for compute.

Concrete electrical install for the 10 kW lab

What we recommend, end-to-end, for a 10 kW robotics + AI lab on a fresh install:

Three-phase 32 A lab sub-panel split into six balanced 16 A single-phase branch circuits — server PSUs on separate phases, HVAC isolated, robot chargers on their own breaker.

Notes on this layout:

• Three-phase 32 A at the lab sub-panel is the parent feed. From there, single-phase 16 A circuits drop off each leg, roughly balanced.
• Dedicated server circuits, one per PSU. The dual-PSU server has its two cords on two different phases — same logic as the split-delivery wiring on a three-phase PDU.
• Humanoid charger circuit is separate from the server. A robot getting plugged in does not perturb the server's electrical environment.
• HVAC on its own circuit. AC compressor inrush is significant; sharing with the server circuit is asking for nuisance trips.
• Type B RCD. Switch-mode PSUs with PFC have small DC leakage components that can fool Type AC RCDs. Type B is the right answer; Type A is acceptable on smaller circuits.

This is a one-day job for a certified electrician in a building that already has three-phase to the main panel. In a building with only single-phase service, the utility may need to upgrade the feed, which adds weeks. Check the panel before the lease.

The honest take

Most labs underestimate their power and cooling needs by roughly 2×. The pattern is consistent: someone designs the room around the server they are buying today, forgets the second server they will want in a year, undersizes the cooling because the salesman quoted nameplate kW, and skips the UPS because "we have utility power, it's fine."

The cost of doing it right the first time — pulling a 32 A three-phase feed, installing a 15 kW split AC, putting an 8 kVA double-conversion UPS in front of the server — lands in the €8–15k range for a 10 kW lab. The cost of doing it wrong is roughly the same number, plus a quarter of project delay, plus the credibility hit of telling the team "the server doesn't work today." We have seen the wrong-way version more often than the right-way version.

Plan for 1.5× your current load. Treat the current load as the floor, not the ceiling. The marginal cost of one breaker size up at install is small. The marginal cost of redoing the install is large.

This is part of the Kentino Wiki, a reference series on AI compute, robotics, and the systems that connect them. Comments and corrections welcome at info@kentino.com.