Summary
Could income from a small data center help turn an empty office tower into affordable homes? The Warm Tower tests that question on Seattle's Fourth & Blanchard Building. Its lower floors would hold computing equipment and the floors above would become apartments. Rent from the computing operator could help finance the conversion, and recovered heat could supply the homes' heating and hot water.
Under the current assumptions, three compute floors reduce the estimated financing shortfall from about $183 million to about $142 million and leave room for about 316 homes, including 126 income-restricted homes. That is a real reduction, about $41 million, but it does not make the project work on its own. It costs 60 homes compared with a housing-only conversion and adds about 9 MVA of electric demand. The results depend on assumptions about rents, construction costs, heat sales and electricity, and none of them has been confirmed by the utility, an operator or an engineer.
We propose a 90-day feasibility study to find out whether this approach can deliver lasting housing benefits at an acceptable cost to residents, the neighborhood and the power system, with clear conditions for stopping.
What is proposed?
A small data center on floors 2–4 of the half-empty Fourth & Blanchard tower in Belltown, with the floors above converted to about 316 homes. The data center operator pays rent to the building owner, and the computers' waste heat warms the apartments.
What do residents gain?
About 126 homes with rents capped for households earning up to 60% or 80% of area median income, plus about 190 market-rate homes, in a reused building heated by recovered heat.
What does it require?
About 10 MVA of electric service, roughly 9 MVA more than the housing alone. About $343 million in total project cost. And, under current assumptions, about $142 million more financing than the property's income can support, including a 10% developer margin.
What is uncertain?
Whether City Light can deliver the power, what an operator would pay, retrofit and facade costs, the building code path for mixing computers and homes, the size of the upper floors, and whether anyone would buy surplus heat.
What happens next?
A bounded 90-day feasibility study with the utility, the building's owner, operators and engineers, and a public report that recommends proceeding, redesigning or stopping.
The shortfall figures throughout are estimated differences between project value and cost, including the developer margin. They are not established requests for public funding.
How sure are we?
Not very, yet. The estimated shortfall swings by tens of millions of dollars with a single assumption. Every row below uses the model's default inputs except the one named.
| Scenario | Homes | Income-restricted | Est. demand | Est. shortfall |
|---|---|---|---|---|
| Housing only | 376 | 150 | 1.5 MVA | $183.1M |
| Three compute floors | 316 | 126 | 10.3 MVA | $142.4M |
| Three floors, half the surplus heat sold | 316 | 126 | 10.3 MVA | $129.7M |
| Three floors, compute rent 20% lower | 316 | 126 | 10.3 MVA | $168.2M |
| Three floors, existing curtain wall kept | 316 | 126 | 10.3 MVA | $115.1M |
| Three floors, conservative set | 316 | 126 | 10.3 MVA | $217.2M |
| Three floors, optimistic set | 316 | 126 | 10.3 MVA | $56.9M |
| Five compute floors | 276 | 110 | 16.2 MVA | $116.0M |
| Five floors, optimistic set | 276 | 110 | 16.2 MVA | $0.5M surplus |
Three things stand out. Compute narrows the shortfall in every case, by roughly $13 million per floor at default rents. Only the optimistic set closes it, at about five floors, and that needs about 16 MVA, well into City Light's large-load class. And the facade decision alone is worth about $27 million, as much as two compute floors.
Inputs in the model carry one of four labels: Published fact Derived estimate Unverified assumption Confirmed by investigation. Nothing has reached the fourth category yet. The traced floor plan and home count are derived estimates. Structural capacity, facade cost, utility availability, operator rent and the code path are unverified.
The problem
Downtown Seattle's office vacancy reached 35.8% in the second quarter of 2026, and 43.1% in the Denny Regrade submarket. Downtown office property values have fallen about $15 billion since 2020, a 46% decline that costs about $128 million a year in tax revenue. At pre-pandemic demand it would take about eight years to refill the space.
Towers are trading at prices that would have been unthinkable five years ago. Plaza 600 on Stewart Street sold for $12.5 million in September 2026, about $59 per square foot, after a $97 million sale in 2019. U.S. Bank Center sold for $280 million, roughly half its 2019 price, at 45% vacancy.
Converting offices to housing is the standard answer, and it runs into the same wall everywhere. Gutting a tower for apartments costs about as much as building new; New York's comptroller uses about $500 per gross square foot including financing. Deep floor plates leave dark interiors that can't legally be bedrooms. And rents in a weak downtown don't cover the cost. Moody's found only about 14% of Seattle's office stock is a good physical candidate. Even then, the deals that pencil tend to be luxury condos, not homes for working households.
In this building, affordability requirements are a smaller part of the problem than they look. Area median income (AMI) is the midpoint household income for the Seattle area, adjusted by household size. In 2026, 60% of AMI is $69,060 for one person and 80% is $92,080; for two people, $78,960 and $105,280 (City of Seattle limits; the tax-exemption program publishes its own schedule). A one-bedroom restricted to 80% of AMI rents for no more than $2,162 a month under the City's MFTE Program 6, including basic utilities. That is only about 15% below the downtown average of $2,534. For small units the cap can sit above what the market would pay, so the model uses whichever is lower. Even at market rents, the conversion does not pay for itself.
The idea
Treat the tower as a vertical piece of city infrastructure. The lower floors become a long-lived computing and energy plant, and the floors above become housing. One owner holds both, and the plant's rent helps finance the homes.
- Colocation
- A data center that rents space, power and cooling to many customers' servers, rather than serving one company.
- MW and MVA
- Megawatts measure real power used. Megavolt-amperes measure the capacity a utility must deliver; for this kind of load MVA runs a few percent higher. City Light's thresholds are in MVA.
- PUE
- Power usage effectiveness: total data center power divided by the power the computers use. 1.3 means 30% extra for cooling and losses.
Who earns the compute income, and who carries its costs
The model assumes one arrangement. The building owner pays to build the data hall's base infrastructure: structure, electrical distribution, cooling plant and heat recovery. A colocation operator leases that capacity on a long-term wholesale lease, priced per kilowatt per month. The operator pays for its own technology fit-out, pays the power bill and sells space to its customers. The owner's income is the lease rent, less upkeep and an equipment replacement reserve. That rent, capitalized at a higher rate than the housing to reflect its risk, is what narrows the housing shortfall.
Two caveats on the benchmarks. CBRE's published asking rates are for deployments of 250 kW and up at N+1, Tier III redundancy in purpose-built facilities; a retrofit tower may not achieve them, so the model uses the low end of the published wholesale range. JLL's $11.3 million per MW covers shell and core construction and excludes the tenant's technology fit-out; whether it fits a retrofit inside an existing steel frame is unverified. An affiliated operator would need a credible paying customer base behind its lease for lenders to count the rent.
Seattle already has a precedent for the parts. The Westin Building Exchange on Sixth Avenue is an office tower that became the region's main carrier hotel. Since 2019 its waste heat has warmed Amazon's Denny Triangle campus across the street: two 16-inch pipes carry about 65°F water to heat-recovery chillers that raise it to about 130°F. That shows the heat reuse works. It does not show this project would find a heat market.
The public bargain
Seattle has good reasons to be wary of data centers. In 2026 five proposals sought a combined 369 MW, about a third of the city's average use, and residents sent more than 54,000 messages of concern. Those concerns are about who gets scarce electricity, who pays for the grid, and what the neighbors live with. A proposal like this one has to answer them with commitments, not framing. The draft terms we would test:
- Homes. At least 40% of homes income-restricted at 60% and 80% of AMI, about 126 homes in the three-floor case, recorded on the property's title for at least 50 years.
- Who pays for power. The project pays for every service upgrade and connection, consistent with City Light's principle that new large loads should not shift costs to other customers. It accepts curtailment at system peaks and publishes its annual energy use.
- Residents are protected if compute income falls. Restricted rents and the housing's operations cannot depend on the compute lease once the building is stabilized. That means a funded operating reserve, a compute lease that is subordinate to the housing's regulatory agreement, and full backup heating so no apartment relies on the data center for warmth.
- Neighborhood impacts. No diesel storage beside homes, rooftop equipment held to residential noise limits, and batteries in their own fire-rated rooms below grade.
- Transparency. An annual public report on homes, rents, energy, heat delivered and compute occupancy.
Compared with the alternatives
| Option | Homes | Income-restricted | Est. shortfall | Electric demand | What it buys, and what it costs |
|---|---|---|---|---|---|
| Housing-only conversion with public assistance | 376 | 150 | $183M | ~1.5 MVA | The most homes and the least power. Needs the most outside financing: about $1.2M of shortfall per income-restricted home. |
| Compute-housing hybrid, three floors | 316 | 126 | $142M | ~10.3 MVA | About $41M less shortfall, about $1.1M per income-restricted home. Costs 60 homes, about 9 MVA of power, and adds operator and technology risk. |
| Continued office use | 0 | 0 | None | Existing | No conversion cost. The tower stays about half leased in a 43%-vacant submarket, and its value and tax base keep eroding. |
| Affordable housing built elsewhere | Varies | Varies | City capital of about $146,000 per new home in its 2025 awards | New load elsewhere | Likely far cheaper per home, since City capital is one of several funding layers. Does nothing for empty downtown towers. |
On these numbers, the hybrid is not a cheaper way to produce affordable homes than building them elsewhere. Its case rests on what else it buys: reuse of a steel tower instead of a vacant or demolished one, homes downtown, and income from an asset that is otherwise losing value. Whether that is worth the power and the remaining shortfall is a public choice, and the feasibility study should put the comparison in front of the people making it. We have not yet quantified the carbon benefit against a stated comparison, such as demolition and a new building of equal area.
Case study: Fourth & Blanchard
We screened distressed downtown towers by vintage, distress, floor plate, distance to fiber and heat networks, and availability. Fourth & Blanchard came out on top. This is an illustrative test fit from public data, not a proposal for the property.
Why this building. Martin Selig lost it to lenders, and the special servicer bought it at a courthouse auction in August 2025. Colliers now markets it as a "basis play" at less than half its assessed value. The assessor puts most of the property's value in the land, $58.3 million, against $11.7 million for the building. That does not by itself mean demolition is likely, but it weakens the economic case for keeping the tower as it is. Its full floors are about 21,000 rentable square feet. It sits about two blocks from the Westin Building's fiber hub, Denny Regrade has the highest office vacancy downtown, and locally it's known as the "Darth Vader building," a landmark people already know.
Why it's hard. The dark reflective curtain wall and angled massing are a real housing problem: daylight, ventilation and the sharp corners of the floor plate likely mean replacing the facade, which the model prices at about $21 million (unverified). Floor-to-floor heights, slab design loads and the column grid still need the original drawings. With about half the space leased, conversion would have to be phased around tenants. We could not confirm that CenTrio's steam mains reach this block.
Physical fit
Which kind of compute
Compute types differ by an order of magnitude in density. Office floors are built for about 5–10 watts per square foot. Colocation runs about 100–200 W/sf of floor used for servers. Liquid-cooled AI racks can draw 40–130 kW each. In this model more density doesn't help: AI halls cost more per megawatt to build, and two AI floors would put the building past the 10 MVA threshold. Edge and retail colocation give the best return per megawatt here, and they fit an office structure.
Where it goes
Compute goes on floors 2 through 4, the full parallelogram floors at the base. They are closest to the street, the loading dock and the utility vault, and they are the least desirable apartments in the building. The three below-grade levels can take switchgear, transformers, batteries and thermal storage, which keeps the heaviest equipment off the floors and away from homes.
Structure
The frame is steel. A 1980 U.S. Steel feature in Progressive Architecture describes about 2,650 tons of A-36 and A-572 steel engineered by KPFF for seismic Zone 3, two interconnected towers with a parallelogram floor plan, and a minimum of interior columns. A 1983 UC Berkeley study of moment connections also cites the building as a moment-resisting frame. Office floors are typically designed for a 50 psf live load; colocation wants 150–250 psf, and loaded racks put 2,500–3,500 lb on about 8 sf. Steel is the easier case for reinforcing, because new beams, plates and columns can be added to the existing frame, but the drawings must show the deck, the design loads and the column grid before any of this is more than a reasonable expectation.
Getting the equipment in
Racks now arrive pre-loaded and are about 8 ft tall, taller than most freight elevator doors and near most elevators' 4,000 lb rating. On the low floors the plan can use a dedicated hoistway, a facade knockout panel with a crane pick, or a widened freight car. On floor 20 none of those would be easy, which is one more reason the compute belongs low.
A typical floor, tested
The Colliers leasing plans show the real floor. Floors 2 through 14 are full parallelograms of about 21,000 rentable square feet, with an elongated central core of about 95 by 27 ft. Two offset towers meet at a notch on the east side. We traced the floor-2 plan and scaled it to its listed 21,037 RSF. That gives a gross floor of about 22,000 sf, roughly 224 ft along each long face and 97 to 110 ft deep. This is a derived estimate from a leasing plan, not a measured drawing.
Units are laid along the glass at about 29 ft deep, so living rooms and bedrooms sit on the window. A 5.5 ft corridor runs around the core, and spines run parallel to the slanted end walls. The two long corridors end at windows. The deep, dark pockets the parallelogram leaves at each end become support rooms: the heat pumps that use the compute floors' heat, plus storage, laundry and trash.
A full floor holds 20 homes in Scheme A and 25 in Scheme B, at about 66% efficiency. That is lower than a rectangle would give, because a parallelogram's sharp corners and deep ends are hard to use. Above floor 14 the floors shrink under the angled roofs; we count floors 15–25 as about 5.8 full floors, which is the least certain number in the home count. On that basis a housing-only conversion holds about 376 homes in Scheme A. Small units hit a ceiling: for 5 of Scheme A's 20 homes, and 17 of Scheme B's 25, the 80% AMI rent limit is above what the unit would likely rent for, so the model uses the lower market figure. Scheme B earns about 5% more rent per floor. Scheme A stays the base case because it keeps two-bedroom homes in the mix.
Heat
A full colocation floor at about 2.2 MW of computing load gives off about 1.5 MW of recoverable heat at 80–100°F, more with liquid cooling. The homes need much less: at about 2.5 kW of peak heat per home, half a compute floor covers the whole tower's peak. The three-floor case produces about 7.6 times the housing's peak need.
Heat produced is not heat delivered or paid for. The model credits only the heat the building itself uses, valued at what it saves after heat-pump electricity, distribution and upkeep: about $50,000 a year. That is small, so heat matters here for carbon, building-performance compliance and resident comfort more than for finance.
Selling surplus heat to neighbors is an upside, not part of the base case. If half the surplus sold at $35 per MWh, the three-floor shortfall would fall by about $13 million, because a modest income stream is worth a lot once capitalized. That is exactly why it stays out of the default until there are buyers, connection costs, seasonal demand figures and contracts. Downtown has two networks to approach:
- CenTrio steam (formerly Seattle Steam / Enwave). It runs about 20 miles of pipe serving 140 buildings and 28.5 million sf downtown and on First Hill, including three hospitals. Steam is far hotter than data center heat, so direct injection would need inefficient high-temperature heat pumps. The realistic near-term role is preheating boiler makeup water; the larger opening is CenTrio's own decarbonization.
- Amazon's Westin loop. The working template for low-temperature reuse, but a private, closed system to learn from rather than join.
The building was designed for heat sharing. The 1980 article describes an electric-hydronic heat pump system tied to a main circulating water loop. If that loop or its risers survive, they are a ready path for moving compute heat to the apartments. Confirming that is part of the building investigation.
Power
In the current model, power is the binding constraint; an engineering investigation could reveal another. Three quantities need to be kept apart:
- Estimated building demand. Each full compute floor adds about 2.2 MW of computing load, about 2.9 MW with cooling. All the housing adds about 1.2–1.4 MW. With three compute floors the building would draw about 9.8 MW, or about 10.3 MVA at an assumed power factor of 0.95.
- Confirmed service capacity. None. City Light has not said what it could deliver to this building, on what timeline, or at what cost.
- Regulatory thresholds. City Light's 2026 policy puts data centers requesting 10 MVA or more into a large-load class with cost-based rates, full responsibility for infrastructure costs, curtailment at peak and a supply queue. The city has also enacted a one-year moratorium on data centers above 20 MVA.
Being below a threshold is not the same as having power. On our estimate the three-floor case sits just above 10 MVA, so a design meant to stay out of the large-load class would need slightly less compute, lower density, or a ruling on how a mixed-use building is measured. Closing the shortfall under optimistic assumptions takes about five floors and about 16 MVA, inside the large-load class but under the moratorium. At the model's AI cost and rent assumptions, AI-density compute widens the shortfall instead of narrowing it, and passes 20 MVA within three floors.
The site is near new supply. City Light's Denny Substation opened in 2019 about 0.7 miles away, built to serve Denny Triangle and South Lake Union through an underground network and expandable from 50 to 405 MVA. Which feeder serves this block, and how much spare capacity it has, is the study's first question.
This section assumes grid power from City Light. A later round will look at where the electricity comes from and how the building could supply part of its own load:
- Off-system power purchase agreements, which City Light has floated for large loads
- On-site batteries and thermal storage for peak shaving and curtailment
- Backup strategy: network service plus batteries and a small renewable-diesel unit, versus full generator redundancy
- Whether a demand-flexible design could earn priority or better terms from the utility
- Rooftop solar: the original design angled the roofs at 45 degrees partly to take solar panels later