CRE investors who have spent four years chasing tenants for Class B office may have an entirely new way to generate asset NOI.
The first phase of the AI buildout has been focused on massive centralized data centers that assemble the giant compute necessary for frontier model training. As adoption and revenue climb, the scramble is now shifting closer to the user with inference workloads that serve queries and agents.
The bottleneck? Permitted electrical service in urban-adjacent locations with fiber. Guess who has that? Real estate investors.
Three different asset classes have read the same memo in the last six weeks.
In April, Span and NVIDIA introduced XFRA, a distributed data center network that puts RTX Pro 6000 nodes inside the electrical panels of new-construction homes. Target is 1 GW of inference capacity by 2027, with PulteGroup as the initial homebuilder partner. Span has concluded that the fastest way to add gigawatts of compute is to monetize service capacity that already exists on the customer side of the meter.
At GTC 2026, NVIDIA announced AI grids with Spectrum, T-Mobile, and Nokia, embedding GPUs at base stations and running inference across more than 1,000 fiber-served edge sites. Telecoms already own the urban-adjacent, fiber-rich, well-permitted footprint that inference needs. The cell tower is being repositioned as an inference node with a radio attached.
In the same window, Prologis disclosed an $8B development pipeline converting warehouses to data centers, with optionality on 100 more sites. CEO Hamid Moghadam pegs the value pickup at roughly $500/sf per conversion. One percent of the US Prologis portfolio at that uplift is $3B of pure repositioning gain.
The holy trinity: power, space, BUGs
The neocloud customer is shopping for three things. Permitted power on an existing meter. Physical space, fiber-served, in a metro. And a backup arrangement that can deliver 4 or 5 nines of uptime without waiting on a new air permit.
Half-vacant Class B office checks every box.
The power leg is the most overlooked. Permitted electrical service in 70s and 80s mid-rise office buildings was sized for the lighting, equipment, and HVAC loads of that era. Lighting power density has since fallen from 2 W/sf to under 0.5 W/sf with LED retrofits. The result is that most older Class B buildings are sitting on roughly half of their permitted service capacity, indefinitely.
Space is painfully abundant. RTO has plateaued well short of 2019, and flight to quality means Class B has simply not recovered.
The BUGs leg is the asymmetric one. Most major metro air districts have effectively frozen new diesel backup permits, which is what makes a grandfathered genset on a Class B building an asset no greenfield site can replicate.
Fiber is the fourth thing, and most older urban Class B has it or is close to it already, courtesy of decades of tenant data needs.
The math on a 4,000-amp service
Start with a generic 70s or 80s Class B mid-rise. Permitted electrical service in that era was typically 4,000 amps at 480V three-phase, sized for a 150,000 to 250,000 sf building under the design assumptions of the time. Convert that service to real power: 4,000A × 480V × √3 × 0.9 power factor is roughly 3.0 MW of permitted capacity. Half is in use today. The other 1.5 MW is headroom that nobody is using.
Map 1.5 MW of headroom onto a 200,000 sf reference building and three reliability tiers:
| Configuration | Rent ($/kW-month) | Annual revenue | Per sf of building |
|---|---|---|---|
| 3-9s (battery only) | $50 | $900K | $4.50 |
| 4-9s (battery + gas linear gen) | $100 | $1.8M | $9.00 |
| 5-9s (BUG permit) | $150 | $2.7M | $13.50 |
The pod itself fits in roughly 2,500 sf, or 1.25 percent of the floor plate. The other 197,500 sf is still leasable. The inference revenue is additive, not replacement.
For comparison, a half-leased Class B office in a soft metro is netting $3 to $6 per sf of NOI on the whole building. The 5-9s inference pod, on its own, generates more per-sf yield than the entire residual office NOI.
Translate the rent stream to value uplift at a 9 to 10 percent cap rate, which is the right range for behind-the-meter inference colo:
| Configuration | Value added | Per sf of building |
|---|---|---|
| 3-9s | $9M | $45 |
| 4-9s | $18M | $90 |
| 5-9s | $27M | $135 |
Of course there is capex. The GPU container is tenant-funded, but the landlord is on the hook for:
- Interconnect, switchgear, sub-distribution: $300K to $600K
- Cooling tie-in, rooftop chiller plant or CDU rooms: $500K to $1M
- Permit work, structural review, neighborhood approvals: $200K to $400K
- Total capex: roughly $1.0M to $2.0M
Payback against the rent stream runs 20 months in the 3-9s case and 7 months in the 5-9s case. That is a faster return on capital than any other use of half-vacant Class B floor plate, including resi conversion and demolition for ground-up redevelopment.
The startups already swarming
A small set of well-capitalized firms is already running this play on Class B office.
Perimeter Compute is the cleanest pure-play: wiring office service capacity into inference nodes, with the building owner as a counterparty rather than a passive landlord. Hammerhead AI, founded by AutoGrid alums, is the orchestration layer for stranded behind-the-meter power, including the GPU clusters that sit underutilized inside shrinking enterprise IT rooms.
The pitch will not look like a tenant inquiry. It will look like a power offtake conversation, with a 10 to 20 year take-or-pay structure and a counterparty that wants the panel, not the floor plate. CRE pros who have spent this cycle defending their cap rates against tenant turnover may be surprised at the deal structure when it lands.
How to assess your portfolio
The diligence checklist for an investor with a Class B portfolio reduces to four pieces of information per building.
The first is electrical. What is the rated service, what is the actual peak demand at the meter, what is the headroom, and what does the utility say about a 1 to 5 MW expansion path. Most of this can be pulled from interval data and utility filings in an afternoon.
The second is fiber. Which carriers are in the building or under the street, what is the latency to the nearest interconnect, and what is the cost to light a dedicated path to a major peering point.
The third is the backup permit. Does the building carry an existing diesel BUG permit, what is the permitted size, what is the annual hour cap, and what is the documented runtime against that cap. A clean, underused permit is worth multiples of the genset’s replacement cost in the right metro.
The fourth is competitive position. What is the shape of your utility’s large-load queue? How tight is demand and supply?
None of that is leasing math. All of it is grid math.
The bottom line
The same physics that makes Span put GPUs in residential panels, makes NVIDIA put inference at the cell tower, and makes Prologis convert warehouses also makes Class B office a viable inference host. The valuation conversation in CRE is still anchored on cap rates and rent rolls. The valuation conversation in the inference market is anchored on service capacity, fiber adjacency, backup-generation permits, and queue position.
The Class B owners who reprice the panel first will capture the asymmetric opportunity. Everyone else will wait for the pitch.
The fastest-appreciating things on the balance sheet are the panel and the permit.
