Short answer: maybe, and you can usually tell which way it will go in an afternoon. Floor space is almost never the problem. Power and cooling are, and in California the electricity rate and the rules around it make both matter more.
If you own or manage a building in California and someone has asked whether a room can host GPU servers, or you are wondering whether a colocation business is worth starting, this is the order I would check things in. Each check can end the conversation. Do the cheap ones first.
All dollar figures below are US dollars. Prices, tariffs and rules were checked on November 12, 2026, and the sources are linked with their dates. The hardware price and the load profile are assumptions to show the arithmetic, not quotes. Tariffs change at least yearly and building and fire codes are adopted on a cycle, so confirm the current version before you rely on any number here. This is not legal or engineering advice.
1. Power: find the tariff, the all-in rate and the spare capacity
Two numbers decide whether the power side works: what a kilowatt-hour really costs you, and how many kilowatts the building can still give. In California the first number depends heavily on which utility serves you, which rate schedule you are on, and whether a community choice aggregator (CCA) supplies your generation.
The tariffs. A commercial customer in the three large investor-owned utilities’ territories pays a customer charge, a demand charge per kW, and energy charges that change by season and time of day. The demand charge is the part that surprises people with a flat, around-the-clock load such as GPU servers, because it is billed on your highest 15-minute draw every month.
- PG&E. Schedule B-10 is PG&E’s medium general demand-metered schedule. PG&E’s tariff says a new customer expected to peak between 75 and 499 kW is assigned to it, and a customer above 499 kW for three consecutive months moves to Schedule B-19. In the sheets effective March 1, 2026, B-10 bundled secondary-voltage service shows a demand charge of $20.50 per kW per month and energy charges from $0.19139 per kWh (winter super off-peak) to $0.33947 per kWh (summer peak). The peak period is 4 p.m. to 9 p.m. every day, and summer runs June 1 to September 30. B-19 swaps a larger demand charge ($37.37 per kW of maximum demand on secondary service, plus $46.16 per kW of summer peak demand) for lower energy charges, $0.06442 to $0.18648 per kWh.
- SDG&E. The medium commercial summary table (effective June 1, 2026) covers Schedule AL-TOU-M, for customers between 20 kW and 200 kW. On secondary service it lists a monthly basic service fee of $209.77 for loads up to 100 kW, a non-coincident demand charge of $30.34 per kW, a summer maximum on-peak demand charge of $38.46 per kW (winter $15.88 per kW), and total energy charges from $0.1748 to $0.3469 per kWh depending on season and period. SDG&E also notes the non-coincident demand charge is based on the higher of the maximum monthly demand or 50 percent of the maximum annual demand, so a bad month keeps costing you.
- SCE. SCE’s summary of rate options (August 2025) describes TOU-GS-2 for demand of 20 kW to under 200 kW and TOU-GS-3 for 200 kW to 500 kW, both with the highest energy rates from 4 p.m. to 9 p.m. on summer weekdays. SCE’s TOU-GS-3 fact sheet explains that facilities-related demand is billed on your monthly maximum, time-related demand applies to on-peak or mid-peak hours, and that Option D trades higher demand charges for lower energy charges. I am not quoting SCE dollar figures here because the component rates sit in the tariff books rather than the summaries. Pull them from SCE’s tariff books for your schedule and option.
If a CCA supplies your generation, the generation part of these bundled numbers is replaced by the CCA’s rates, so your bill will differ. SDG&E’s table says as much: its generation rates do not apply to CCA customers.
What to pull from 12 months of bills. Get twelve consecutive months for the meter that would feed the room, and write down:
- total dollars and total kWh for each month;
- the billed peak demand in kW for each month, and which rate schedule and option you are on;
- the size of the electrical service (amps, voltage, phases) and what else is on it.
Divide total dollars by total kWh over the twelve months. That is your all-in rate.
An example. Take the five-rack room used later in this post: a flat 58.5 kW load, around the clock, on PG&E B-10 secondary service. That is 58.5 x 730 = 42,705 kWh a month. Weighting PG&E’s energy rates by hour (summer: 5 peak hours, 4 part-peak and 15 off-peak; winter: 5 peak hours, with a 5-hour super off-peak window from 9 a.m. to 2 p.m. in March through May), the energy comes to about $0.245 per kWh, or about $10,460 a month. Add the demand charge (58.5 kW x $20.50 = $1,199) and the customer charge ($11.36882 a day, about $346 a month) and the bill is about $12,000 a month, or about $0.28 per kWh all-in. That is my working figure for the rest of this post. For SDG&E’s AL-TOU-M the same load would pay roughly $3,460 a month in demand and fixed charges alone (about $0.08 per kWh) on top of $0.17 to $0.35 per kWh of energy, so I would expect an all-in rate somewhere between $0.26 and $0.43 per kWh. I did not weight those energy charges by hour, so treat that as bounds rather than an estimate. Your own bills beat all of this.
Subtract your highest monthly peak demand from the service capacity. What is left is the headroom, before any safety margin. Then ask your utility in writing how much capacity is available at your connection point, what an upgrade costs and how long it takes. The wait for a service upgrade is often longer than the wait for the hardware.
2. Power per rack: why a legacy rack is not a GPU rack
A legacy rack in an older room was often designed for roughly 5 to 10 kW. That is consistent with the Uptime Institute’s 2026 survey, as reported by Data Center Knowledge, which put the average rack at 7.8 kW and the most commonly reported density at 11 kW.
Modern GPU hardware does not fit that picture:
- One NVIDIA DGX B200 system is rated at about 14.3 kW maximum and takes 10 rack units, per NVIDIA’s product page (checked November 11, 2026). One box, with eight GPUs, is already more than a whole 10 kW legacy rack.
- An NVIDIA RTX PRO 6000 Blackwell card is rated at up to 600 W, per NVIDIA. A four-GPU server built on those cards is about 2.4 kW of GPU plus the rest of the machine, so roughly 3 kW at full load. I treat that as an estimate and check the vendor’s spec sheet for the actual server.
- At the top end, NVIDIA’s GB200 NVL72 is a liquid-cooled rack of about 120 kW, per NVIDIA’s OCP contribution and its DGX GB user guide. That is not a building-retrofit conversation. It is a purpose-built data hall.
For an ordinary older room, the realistic target is the small end: a few four-GPU servers per rack, at 6 to 12 kW per rack.
Here is the electrical arithmetic for a room of five racks at 9 kW each, which is three of those four-GPU servers per rack. This is an example to show the method:
- IT load: 5 racks x 9 kW = 45 kW.
- With cooling and distribution overhead of about 30 percent (a PUE of 1.3), the building feels 45 x 1.3 = 58.5 kW.
- On a 208 V three-phase supply, current is 58,500 W / (1.732 x 208 V) = about 162 A, before the margin your electrician is required to add for continuous loads.
If your spare capacity is 60 A, the answer is already no-go, or it is a service upgrade project. If it is 400 A, you are fine on this check.
3. Cooling: the hardest constraint, and a California code that applies to it
Every watt that goes into a server comes out as heat. One kW of IT load is about 3,412 BTU per hour. One ton of cooling is 12,000 BTU per hour, or about 3.5 kW.
So the 45 kW room above needs about 45 / 3.5 = 12.8 tons of cooling that runs all year, including in the cool months, at night and on a long weekend. The office air conditioner that handles the building’s occupants will not cover it.
The mistake I see is checking tons and stopping. Three other questions matter:
- Can the cooling reach the rack? Air has to move to the front of the servers and the hot exhaust has to go somewhere that is not the intake. A ceiling-return office room with a few floor diffusers mixes hot and cold air, and the servers at the top of the rack overheat first.
- Where does the heat go? A rooftop unit, a chilled water loop or a dry cooler outdoors all need space, structural support and, in many cases, a permit.
- What is the failure mode? If the cooling stops, a dense room heats quickly. Ask what happens at 3 a.m. on a Sunday and who gets the alarm.
Title 24 treats a dense IT room as a “computer room.” The California Energy Commission’s Computer Rooms and Data Centers fact sheet (August 2023, for the 2022 Energy Code) says the standards apply to space conditioning systems serving a room whose primary function is to house electronic equipment and whose design IT power density exceeds 20 watts per square foot of conditioned floor area. The requirements are in Title 24, Part 6, Sections 120.6(j) and 140.9(a). Your 45 kW room would pass that threshold unless it were larger than 2,250 square feet, and a five-rack room is nowhere near that. For a room like this, the fact sheet lists:
- Economizers. Each cooling system primarily serving the room must include an air or water economizer, with two exceptions: a room with an IT load under 5 tons (18 kW) in a building with no economizers, and a room under 20 tons (70 kW) served by a second fan system under conditions the sheet spells out. Our 45 kW room is above the first exception and below the 70 kW limit of the second.
- Air containment. A room with air-cooled racks and an IT load above 10 kW needs air barriers so discharge air cannot recirculate to the inlets.
- Other mandatory items. Controls that prevent reheating and simultaneous heating and cooling, adiabatic humidification only, variable airflow on larger fans, a fan power limit, and minimum efficiencies for AC-output UPS systems.
That fact sheet describes the 2022 code. The California Energy Commission says the 2025 Energy Code applies to permit applications filed on or after January 1, 2026, and I have not confirmed whether the computer room text changed. Ask your mechanical engineer to check Sections 120.6(j) and 140.9(a) of the version that applies to your permit, and whether a new or altered system triggers them. If your engineer says the existing system cannot carry the load or meet the code, price the new system before you price anything else. It is often the line that decides the project. At rack densities above what air can carry, you are into liquid cooling, which is a different building.
4. Fiber: two paths, not two bills
GPU inference and storage traffic need dependable connectivity, and tenants will ask about it first.
- How many carriers can serve the address, and what do they quote for a dedicated circuit? The FCC’s National Broadband Map is a free first look at which providers report service at an address, but carriers report availability, so confirm by asking for a quote.
- Do two carriers actually mean two physical paths? Two providers can share one conduit, one pole line or one building entrance. Ask for route maps and where each cable enters the building.
- Is there space in the building for a meet-me room or at least a demarcation point that a tenant can reach?
Diverse fiber at the street is a minimum. Diverse fiber into the building is the real answer. Quotes and install times vary widely by address, and some carriers will not build to a location without a commitment. Ask early.
5. Submetering and selling power to tenants
If you intend to bill tenants for the electricity they use, the meter and the billing structure become a regulatory question, not just a technical one. In California, the starting point is that a building owner who is the utility’s customer is usually not free to resell electricity at a markup, and the details live in the utility’s tariff.
Here is what I can point to from official sources, and where it stops:
- The utility rule. PG&E’s Electric Rule 18, “Supply to Separate Premises and Submetering of Electric Energy” (the sheet on nonresidential service is effective June 23, 2025), says PG&E will furnish and meter electricity to each individual nonresidential premises or space, with exceptions. Two matter here. One is a single meter where the cost of electricity is absorbed in the rent, there is no separate identifiable charge to the tenants for electricity, and the rent does not vary with consumption. The other is a high-rise building owned or managed by a single entity, where the owner’s submeters are used for cost allocation of dynamic pricing or conservation incentives, and tenants are billed at the same rate as the master meter is billed. Read those exceptions closely. A separate per-kWh line item at a markup does not fit either one, and I would not assume a separate per-kW “power fee” fits the first one either. That is a question for your lawyer and your utility.
- The statute and the CPUC. PG&E’s page for sub-metered tenants and landlords (checked November 12, 2026) says a landlord who bills tenants directly must follow Public Utilities Code Section 739.5 and PG&E’s tariffs, must charge the same rates PG&E charges, must itemize the bills, and must keep the submeters in working order, with meter accuracy handled by the county’s Department of Weights and Measures. It also cites CPUC Decision 05-05-026 for the rule that a landlord who converts to submetering must remove energy charges from rent. That page is written mainly around residential tenants, and I have not confirmed how far Section 739.5 reaches for commercial or colocation tenants.
- SCE and SDG&E. Each utility has its own tariff rules on supply to separate premises and submetering. I have not read them for this post, so I will not say they match PG&E’s. Find the equivalent rule in your utility’s tariff book.
The practical advice: before you quote a per-kWh price, a per-kW power fee or a “power included” rent, put the exact structure in front of your utility and a lawyer who has done this for commercial tenants. One safe-looking structure is flat rent with power absorbed in it and no separate identifiable charge, but that moves the power price risk onto you, so do the arithmetic in section 7 first. Meters that decide a tenant’s bill also need to be accurate and approved for billing, so ask the county sealer what applies to yours.
6. Generator and fire suppression
Neither is exotic, and both can cost more than people expect.
Backup power and air permits. Do your tenants need to stay up through an outage? If yes, you need a UPS for the ride-through and a standby generator for anything longer. Ask what the generator is sized for (the whole building or just the room), how long the fuel lasts, how often it is tested under load and who can install it. Then ask your local air district, because in California the generator is an air quality question first.
- State rule. CARB’s Airborne Toxic Control Measure for stationary compression-ignition (diesel) engines is at 17 CCR Section 93115 (a free mirror of the California Code of Regulations). For new emergency standby engines rated above 50 brake horsepower, Section 93115.6(a)(3) limits maintenance and testing to 50 hours per year, and the district may allow up to 100 hours at very low particulate emissions. In-use engines have their own limits (20 to 30 hours at the base level). An engine that runs for tenant SLAs or to cut your peak demand charges is not an emergency standby engine, so do not assume these limits cover that use. The state rules are enforced through your local air district.
- South Coast AQMD. SCAQMD’s emergency generator page says internal combustion engines above 50 bhp need a permit to construct before installation, an emergency generator is limited to 200 hours a year under Rule 1110.2, and testing and maintenance is limited, in most cases, to 20 to 50 hours a year under Rule 1470.
- Bay Area AQMD. The engine permit page (updated September 29, 2026) says a stationary engine over 50 bhp needs a permit, and tells applicants to submit at least six months before the desired installation date. For standby diesels it points to a Tier 4 engine, a health risk assessment, and a distance of more than 1,000 feet from any K-12 school as factors that speed or slow the permit.
- Elsewhere in California. Other districts have their own rules. I have only read the two above. Call yours before you buy a generator, and ask about natural gas, which the Bay Area page says generally permits faster than diesel.
I have not confirmed whether CARB or any district has amended these rules recently, and there are public petitions asking for tighter limits on data center generators. Treat the numbers above as a starting point for the conversation with the district, not as the final word.
Fire. A room full of high-density electrical equipment is a different fire risk than an office. Find out what suppression the room has today, and who the authority having jurisdiction is, normally your local fire department, plus your insurer. California adopts the International Fire Code with amendments as the California Fire Code, Title 24, Part 9. The Building Standards Commission says the 2025 edition was published July 1, 2025 and is effective January 1, 2026, and that the 2025 California Fire Code is based on the 2024 International Fire Code. NFPA publishes NFPA 75 on fire protection for IT equipment and NFPA 2001 on clean agent systems. Ask your fire protection engineer which edition and which local amendments apply. A conventional wet sprinkler over live servers is a common finding in an older room, and replacing it is a real project.
Tell your insurer before you sign anything. Some policies exclude or reprice this use, and it is much cheaper to learn that now.
7. The economics: hardware still leads, but power is not small here
Owners tend to ask about electricity first. In California it is a bigger line than it is in cheaper-power markets, though it is still smaller than hardware.
Take one four-GPU server, an RTX PRO 6000 class machine at about $82,000 all-in. That price is my assumption for November 2026, not a quote. Amortize it over 36 months:
- Hardware: $82,000 / 36 = about $2,278 per month.
- Power, at the all-in rate of $0.28 per kWh from section 1, and an average IT draw of about 2 kW at moderate utilization, which is 2 kW x 1.3 = 2.6 kW at the building’s meter with a PUE of 1.3: 2.6 kW x 730 hours x $0.28 = about $531 per month.
Power is about 19 percent of those two lines ($531 of $2,809). For the same server at full load of 3 kW around the clock:
- 3 kW x 1.3 x 730 x $0.28 = about $797 per month, or 26 percent of the total ($797 of $3,075).
Now scale it to a rack with three of those servers at 9 kW:
| Line (per month) | Calculation | Amount |
|---|---|---|
| Hardware amortization | 3 x $2,278 | $6,833 |
| Power at full load | 9 kW x 1.3 x 730 x $0.28 | $2,391 |
| Power share | $2,391 / ($6,833 + $2,391) | 26% |
This simple model uses a flat all-in rate. A real tariff adds the demand charge, which only falls if you cut your peak draw, so a flat rate is a fair approximation for a steady load and a poor one for a spiky load.
Three observations from that table:
- If you are the building owner and the tenant owns the hardware, your costs are power, cooling, space and your capital for the upgrades. Your revenue has to cover the power at the all-in rate first, then the cooling plant, then the rest. How you charge for power is constrained by the rules in section 5, so settle that before you settle the price.
- If you are the operator and own the hardware, your biggest risk is still utilization. Hardware amortization runs whether the GPUs are busy or not, so an idle server costs nearly the same as a busy one. Power is about a quarter of the cost at full load.
- A swing in the electricity rate of 25 percent moves the monthly total of a fully loaded server by about 6 percent. A swing in utilization moves the whole business.
For the same arithmetic applied to what a private AI deployment costs per user, see should we buy GPUs or keep paying for Copilot. For the larger, financed version, see what does it actually cost to run a GPU cluster.
Common questions
How much power does a rack of GPU servers need? For small deployments built on cards like the RTX PRO 6000 (up to 600 W each), plan on roughly 6 to 12 kW per rack, about 3 kW per four-GPU server at full load. An eight-GPU DGX B200 system is rated at about 14.3 kW by itself. A GB200 NVL72 rack is about 120 kW and liquid cooled. Check the exact server’s specification sheet.
Can an old office building in California host GPU servers? Sometimes. It passes if the electrical service has spare capacity, the room can be cooled around the clock to the Title 24 computer room rules, and the fire protection can be upgraded. Cooling and electrical capacity are what usually decide it, not floor area. Start with the twelve months of bills and the service size.
What is the all-in electricity rate for a commercial building in California? It depends on your utility, your schedule and whether a CCA supplies your generation, so calculate yours: total dollars divided by total kWh over twelve months. For context, I estimate about $0.28 per kWh for a flat 58.5 kW load on PG&E’s B-10 schedule using the March 2026 tariff sheets, and SDG&E’s AL-TOU-M sheets suggest a range of roughly $0.26 to $0.43. Those are estimates, not quotes.
Can I bill tenants for electricity in California? Possibly, but the rules are specific and differ by utility and tenant type. PG&E’s Rule 18 limits when a landlord may charge nonresidential tenants for electricity, and Public Utilities Code Section 739.5 requires that submetered tenants pay no more than the utility’s rate. Confirm with your utility and a lawyer before quoting a price.
Do I need an air permit for a backup generator? Almost certainly, if it is a diesel or gas engine above 50 brake horsepower. South Coast AQMD and Bay Area AQMD both require a permit above that size, and CARB’s rule limits emergency standby engines to a small number of maintenance and testing hours. Ask your local air district before you buy.
Is power the biggest cost of hosting GPUs? No, but in California it is not small. In the example above, hardware amortization is 74 to 81 percent of the monthly cost of a four-GPU server, and power is 19 to 26 percent. Utilization still matters more than the electricity rate.
How long does it take to find out if my room is viable? The first screen is a few hours: the bills, the service size, the cooling system’s capacity and a look at the room. That is the scope of the quick assessment described below.
What to do next
If you work through the seven checks and the answer is still maybe, the next step is a walk through the room with someone who has seen these rooms before. I do that as a fixed-fee GPU site readiness review for California property owners and operators: a one-hour quick assessment with a one-page go or no-go memo, or a full review with a cost model per rack and per GPU. The electrical, mechanical and fire work itself still belongs to licensed trades and engineers, and colocation agreement and SLA work can be referred to a data center contracts lawyer. What I give you is the decision, the numbers behind it and the risks to price before you spend money on drawings.