Behring Wants to Turn an Oakland Office Asset Into an AI Data Center, but the Tenant Comes First

Behring Companies is pitching 415 20th Street as an AI data center with 4.5 megawatts available, despite having no publicly confirmed anchor tenant. The Oakland developer says the building can scale toward 20 megawatts and support high-performance computing. Yet that expansion remains a marketed opportunity, not a completed conversion.

The distinction matters because headlines circulating through Google News can make a proposed project sound more settled than it is. Behring has acquired a building with a genuine computing history and retained infrastructure. It is now seeking an AI user whose requirements would determine how much of that infrastructure gets rebuilt, expanded, or activated.

That makes the project a test of two competing stories about urban AI infrastructure. Behring sees a scarce, connected facility close to Bay Area companies and technical talent. The opposing view sees a small speculative conversion facing difficult power, cooling, financing, and leasing questions.

Behring Has a Real Building but Not Yet a Finished AI Facility

The concrete event is a change in how Behring is positioning 415 20th Street, not the opening of an operating AI data center.

Behring acquired the property in late 2025 and added it to a group of neighboring assets around Oakland’s 19th Street BART station. The company’s workspace affiliate described the purchase as part of a nearly four-acre “innovation corridor” connecting several buildings.

The facility sits at 415 20th Street, near Webster Street in Uptown Oakland. It is adjacent to Behring’s 1900 Broadway residential tower and close to its coworking and office property at 1950 Franklin Street.

The property is unusual because it previously served as a supercomputing data center for Lawrence Berkeley National Laboratory. That history gives Behring more than an empty office shell and an AI-themed marketing campaign.

According to the company’s property announcement, the site retains raised floors, electrical distribution equipment, cooling infrastructure, mechanical systems, and data center floor plates. Behring also identifies an active 4.5-megawatt chiller configuration and 6,000-amp switchgear.

Those components can shorten the path to reuse. They do not establish that the building meets the final requirements of a modern AI tenant.

AI inference is the process of running a trained model to produce an answer, prediction, or action. It can require less concentrated computing power than training a frontier model, but large deployments still need dense servers and dependable cooling.

Behring has promoted the site as a potential inference facility rather than a hyperscale training campus. That positioning fits the property’s urban footprint better than a plan based on enormous new electrical substations and warehouse-sized server halls.

The company says 4.5 megawatts is available immediately and describes a route toward 20 megawatts. Any prospective tenant would still need to validate how much power can reach its computing equipment, when additional capacity could arrive, and which upgrades would be required.

That validation process can change a project materially. A tenant using conventional enterprise servers has different requirements from one deploying high-density GPU racks. The latter may require heavier electrical distribution, liquid cooling, reinforced floors, and redesigned fire-protection systems.

The property’s earlier use is therefore evidence of potential, not proof of present readiness. Legacy infrastructure can reduce redevelopment work, but it can also conceal equipment that no longer matches current rack densities or efficiency standards.

Behring’s opportunity starts with this inherited technical foundation. Its risk begins with the gap between a reusable building and a commissioned AI facility.

Why an Oakland Data Center Looks Attractive Now

Behring is trying to match two distressed markets: underused urban property and accelerating demand for AI computing capacity.

Oakland’s office market has struggled since the pandemic. A 2025 report on Behring’s broader downtown strategy said the city’s office vacancy rate had reached 34.4 percent during the previous year.

Conditions later deteriorated further in the central business district. Cushman & Wakefield’s market data placed Oakland City Center’s overall office vacancy rate at 41.5 percent in the fourth quarter of 2025.

That weakness creates pressure on landlords to find uses beyond conventional office leasing. It also lets buyers acquire older properties at a lower basis and consider conversions that would be harder to justify in a strong office market.

Behring has already pursued that strategy nearby. It combined a new residential tower with coworking areas, offices, parking, fitness facilities, and event space. The company markets the collection as an urban campus rather than a group of unrelated buildings.

The Oakland campus strategy depends on attracting several kinds of users. Residents, small companies, independent workers, events, robotics teams, and AI infrastructure would generate different patterns of activity.

A data center complicates that vision. It occupies valuable space and consumes substantial power, but it usually brings fewer daily workers than an office filled with employees.

Behring’s more interesting argument is that the facility could support an AI cluster rather than operate as an isolated server building. A local inference tenant could sit near robotics labs, startups, offices, housing, and public transportation.

That physical proximity has practical value for some users. Robotics teams must move machines, sensors, and test equipment between computing systems and real environments. Developers testing new hardware may also value direct access to infrastructure staff.

Latency creates another possible advantage. Latency is the delay between sending a computing request and receiving the result. Locating inference systems near users can reduce that delay, although network design matters more than municipal boundaries alone.

Oakland also offers access to the wider Bay Area without requiring a central San Francisco address. The site sits directly above a BART station and near fiber routes associated with its former scientific use.

Still, geographic proximity does not automatically create a viable market. Cloud platforms already provide on-demand access to GPUs across regional data centers. A local facility must offer something better, including predictable capacity, specialized hardware, lower latency, or operational control.

The immediate pressure falls on Behring. It must convert AI interest into a binding lease before spending heavily on tenant-specific upgrades.

The longer-term pressure falls on Oakland’s conventional office owners. If Behring secures a credible computing tenant, other landlords may consider technical reuse instead of waiting for office demand to return.

The Google News Headline Hides a Tenant-First Strategy

The central reversal is simple: Behring is not building a facility and then finding a customer; it appears to be finding the customer who can define the facility.

Readers arriving through Google News may reasonably assume that a conversion plan represents a committed construction project. The available evidence supports a narrower conclusion.

Behring is marketing an existing technical building to AI users. It has described active capacity, expansion potential, and retained mechanical infrastructure. Public information does not identify a signed data center tenant, a construction schedule, or a completion date.

That tenant-first sequence is common in specialized real estate. Owners often avoid finalizing costly systems until they know a customer’s equipment density, redundancy standard, networking needs, and lease duration.

Redundancy refers to backup capacity that keeps a facility operating after equipment or utility failures. A tenant may request duplicated power feeds, backup generators, batteries, cooling loops, or network connections.

Those requirements can reshape the economics. A company operating internal research systems might accept one configuration. A commercial cloud provider promising continuous service to outside customers might demand far more redundancy.

The 4.5-megawatt figure also needs context. It is enough to support a meaningful urban computing installation, but it remains small beside hyperscale campuses measured in hundreds of megawatts.

A smaller facility can still serve a valuable niche. Inference systems, private enterprise workloads, robotics development, and regional cloud services do not always require hyperscale land.

The proposed 20-megawatt expansion would change the project’s position. However, Behring’s public materials present that figure as scalable potential, not operating capacity.

Power capacity has several meanings during data center development. Utility service may be available to a property, but not all of it necessarily reaches computing racks. Cooling, transformers, switchgear, and backup equipment consume space and electricity.

Operators therefore focus on usable information technology load, delivery timing, and contractual certainty. A headline number without those details cannot establish how many GPUs the site can support.

Hardware choices can also change before a lease closes. New AI accelerators draw more electricity per server, while liquid cooling moves heat differently from conventional air systems. A design based on one equipment generation can become inadequate for the next.

This uncertainty explains why the potential tenant matters more than the initial rendering or marketing description. The customer determines whether the facility becomes a modest inference site, a private computing center, or something closer to a regional colocation operation.

Colocation allows customers to place their own servers inside a facility that supplies power, cooling, connectivity, and security. It spreads infrastructure costs across multiple users but adds operating and service obligations.

Behring has not publicly specified which commercial model it intends to pursue. The company could lease most of the building to one operator, partner with an infrastructure provider, or support multiple technical users.

Each model shifts risk differently. A single tenant offers clearer demand but creates concentration risk. Multiple customers diversify revenue but require a more involved operating platform.

The tenant-first approach protects Behring from designing the wrong facility. It also reveals how early the proposal remains.

Existing Infrastructure Does Not Remove the Conversion Risk

The building’s data center history is Behring’s strongest advantage, but it does not settle the engineering, utility, or community questions.

Reusing an old computing facility sounds easier than converting an ordinary office building. The site already has features that many urban properties lack, including heavy electrical equipment, mechanical loops, raised flooring, and secure service areas.

The former laboratory use also suggests the location has supported demanding scientific workloads. That record gives potential tenants a reason to investigate the building rather than dismiss it as a speculative office conversion.

However, modern AI servers concentrate much more heat in each rack than many legacy systems. A floor that once supported supercomputers may still require new power distribution and cooling equipment.

Liquid cooling circulates a fluid near computing components to remove heat more efficiently. Deploying it inside an older building can require new pipes, heat exchangers, pumps, leak controls, and maintenance procedures.

The company’s claimed 4.5-megawatt chiller configuration helps its pitch, but a tenant must test the system’s age, efficiency, redundancy, and usable output. Equipment labels alone cannot answer those questions.

Structural capacity also matters. Dense racks, batteries, and cooling systems add substantial weight. Behring identifies a high-load shell, yet final equipment layouts would still require engineering review.

Power expansion presents the largest uncertainty. Moving from 4.5 megawatts toward 20 megawatts would likely require coordination among the property owner, utility, contractors, and local authorities.

The relevant questions include whether grid capacity is reserved, which interconnection work is necessary, and who pays for upgrades. Public materials reviewed for this article do not provide those answers.

Timing can be as important as total capacity. An AI company choosing between sites needs a credible energization date, not only a statement that more power is technically possible.

Permitting adds another layer. Backup generators, cooling systems, construction work, noise controls, and electrical equipment can trigger separate reviews. An operating plan may also draw scrutiny from nearby residents and businesses.

Urban data centers face a sharper land-use debate than remote campuses. They can reuse difficult buildings and keep technology investment inside a city. They can also consume scarce electrical capacity while producing limited street activity.

That tradeoff is especially visible in Oakland. Behring’s broader pitch emphasizes an active district where people live, work, meet, and build companies. A server facility contributes infrastructure but may place fewer people on surrounding sidewalks.

The project needs a credible explanation of how the data center connects to the rest of the campus. A nearby robotics lab would strengthen that case because physical testing and local computing can complement each other.

Behring says its STAK AI Labs program uses physical environments to gather rights-cleared data for embodied AI and robotics. Embodied AI links software intelligence to machines that perceive and act in the physical world.

That program remains separate from a confirmed data center lease. It supplies a possible use case, not guaranteed demand.

Community response will depend on details that have not been published. Residents will want to know about noise, backup generation, water use, construction, jobs, and grid effects.

Comparisons with hyperscale facilities can mislead both supporters and critics. A 4.5-megawatt urban site operates on a different scale from a remote campus built for frontier-model training.

The environmental and neighborhood effects could therefore be smaller. They are not zero, especially if the property expands toward the advertised 20-megawatt level.

Behring should receive credit for starting with an existing structure rather than proposing a new greenfield campus. Reuse can preserve embodied carbon, infrastructure, and land.

That benefit cannot substitute for operating data. The eventual tenant, equipment plan, energy source, and cooling design will determine the site’s actual impact.

Oakland’s Office Crisis Makes the Bet More Than a Data Center Story

The project tests whether AI infrastructure can become a credible reuse strategy for distressed downtown property without weakening the surrounding district.

Commercial real estate owners across the Bay Area face a difficult mismatch. Many companies need less traditional office space, while AI developers need more computing capacity.

Those needs do not map neatly onto each other. Most offices lack the power, cooling, floor strength, fiber connectivity, and service access required for a data center.

415 20th Street is a rare candidate because it began with technical infrastructure. That makes Behring’s proposal more credible than a generic claim that any vacant tower can become an AI facility.

The scarcity of suitable buildings also limits the precedent. If the project succeeds, it will not mean Oakland can convert its entire vacant office inventory into server space.

Instead, it would show that selected properties can support new technical uses when their physical history aligns with current demand. Owners would need to identify buildings with exceptional utility and structural characteristics.

Other downtown landlords may still watch closely. A signed AI tenant would validate a source of demand that barely existed in commercial property discussions before the current computing boom.

The comparison with San Francisco is unavoidable. AI companies have taken large office spaces there, helping parts of its leasing market recover. Oakland has not captured demand at the same scale.

Behring’s strategy approaches the opportunity from another direction. Rather than competing only for conventional headquarters leases, it is trying to host the machines and technical environments that AI companies require.

That approach could broaden Oakland’s role in the regional market. The city might support inference, robotics testing, research, or specialized infrastructure even when corporate headquarters remain elsewhere.

The strategy also carries an economic-development limitation. Data centers generally employ fewer people per square foot than offices, laboratories, or manufacturing sites.

A facility can generate construction work, technical positions, property activity, and tax revenue. Yet it should not be presented as a direct replacement for thousands of office workers.

Behring’s surrounding campus may soften that weakness. Housing, coworking, events, retail, and robotics facilities can bring people to the area while the data center supplies technical capacity.

That mixed-use combination is the developer’s primary differentiation. It is also an unproven operating model.

AI infrastructure tenants usually prioritize uptime, security, power certainty, cooling, and cost. Amenities and transit access matter less to servers than they do to employees.

The campus becomes relevant only if people working with those systems benefit from being nearby. Robotics teams, infrastructure engineers, and startup researchers provide plausible examples.

Google News exposure can attract attention to the concept, but attention does not validate this connection. Lease terms and actual occupancy will show whether AI users value the wider campus.

Behring therefore faces two sales tasks. It must persuade a technical tenant that the building can operate reliably. It must persuade Oakland that the facility contributes to downtown recovery.

Success on only one side would produce an incomplete result. A technically viable site with little local connection would weaken the urban-campus narrative.

A lively campus without a committed computing tenant would leave the data center plan as marketing. The project needs both infrastructure demand and a convincing place in the district.

What to Watch Before Calling the Conversion Real

Three signals will determine whether Behring’s proposal becomes an operating AI asset or remains an ambitious leasing pitch.

The first signal is a named tenant or infrastructure partner. A binding commitment would clarify the intended workload, commercial model, space requirements, and deployment scale.

The tenant’s identity would also reveal what “AI data center” means here. An enterprise running private inference has different needs from a cloud operator or frontier-model developer.

A credible announcement should include more than expressions of interest. Readers should look for a lease, partnership structure, target opening date, and defined initial capacity.

That signal would strengthen Behring’s argument immediately. Continued references to unnamed interest would weaken it as months pass.

The second signal is documented power delivery. Behring’s existing 4.5-megawatt configuration gives the project a starting point, while the 20-megawatt figure represents a larger ambition.

Watch for utility commitments, interconnection milestones, electrical permits, equipment orders, or a phased capacity schedule. These details would show that expansion is moving beyond property marketing.

A lower initial deployment would not necessarily indicate failure. It could demonstrate a disciplined tenant-first plan if the first customer needs less capacity.

The concern would be a large gap between advertised scale and deliverable power. AI infrastructure projects often compete for electrical capacity, transformers, and specialized construction resources.

The third signal is a public operating and community plan. Behring needs to explain cooling, backup power, noise, water, construction, staffing, and the facility’s relationship to neighboring uses.

That disclosure would let residents and prospective tenants evaluate the project on the same facts. It would also reveal whether the building can support modern equipment without undermining Behring’s mixed-use campus.

Permits offer an early window into those choices. Major mechanical, electrical, or generator applications would indicate that a detailed design is advancing.

The absence of visible work is not decisive because some negotiations remain confidential. However, a long period without a tenant, permitting activity, or power milestones would weaken the conversion thesis.

Readers should also separate the project’s existing advantages from its future claims. The building’s location, computing history, retained equipment, and campus connections are already documented.

A commissioned AI facility, 20 megawatts of available capacity, and a functioning local technology cluster are not. Those outcomes depend on agreements and construction that remain ahead.

This distinction is the most useful filter when the story resurfaces through Google News. Ask whether the update concerns marketing, leasing, utility capacity, permits, construction, or actual operations.

Each stage carries a different level of certainty. Treating them as interchangeable makes speculative projects look complete and hides the execution risks that matter most.

Behring has assembled a plausible site at a moment when Oakland needs new uses and AI companies need computing infrastructure. The building offers uncommon advantages, but the hardest work starts after the pitch.

The next meaningful update will not be another broad statement about AI demand. It will identify who is taking the space, how much power they can use, and when their systems will begin operating.

Until then, the right question is not whether AI can revive one Oakland property. It is whether Behring can turn inherited infrastructure and tenant interest into a financed, permitted, occupied facility. Track those three signals before treating the conversion as complete.