AI companies occupy 8.5M square feet of San Francisco office space. Leases with a published size account for 5.06M

8 chapters. The map follows as you scroll, and every count ties to the register below.

• AI companies occupy 8.5M square feet, by JLL's count. 5,057,231 of it is in leases with a published size; the rest is probably space whose lease size was never made public.
• Documented AI space nearly tripled in 10 months, from 1,722,977 square feet in December 2025 to 5,057,231 in October 2026. Frontier and model labs hold 2,468,329, AI infrastructure and tooling 1,290,798, AI-native applications 1,298,104.
• 207 of 271 locations have no public square footage. Most are small offices, by our headcount model.
• Two labs dominate. OpenAI (1,181,600 SF) and Anthropic (944,955 SF) hold more than the next 14 companies combined. Only 47 companies have any observed square feet.
• Financial District/South Beach has 141 current AI locations, more than the next five neighborhoods together. South of Market has 29.
• 15 buildings are at least half AI by observed square feet.
• AI is backfilling tech's old space. 14 AI leases name the prior occupant, from Slack at 500 Howard Street to Cruise at 333 Brannan Street.

JLL counts 413 companies to our 223. The broker comparison explains both gaps and why the brokers' own figures differ.

330 locations of 241 companies since 2019. Drag the timeline to any month or click a building for the records behind it. Filters cover tier and neighborhood.

73 deals with a published size, on one line in date order, with the logo sized by the deal's square feet. Each year's stretch of line grows with its number of deals, so 2026 takes the most room. 40 came in 2026 alone, against 17 in all of 2025. The largest is still OpenAI’s sublease at 1455 3rd Street: 486,600 square feet, October 2023. When an expansion restates a location’s total, it shows as a second logo for the same company. Each logo links to the company's page; the table has every address.

The series counts square feet a lease, sublease or filing states, held by companies with a recorded San Francisco location. Unsized locations add zero, which is why the line sits flat until 2022: in December 2019 no AI lease in the register had a published size.

The 5,057,231 observed square feet are 5.8% of the 87,080,074 square feet of office inventory in Cushman & Wakefield's third quarter 2026 report. JLL puts AI occupancy at 8.5M. The hatched band is the difference, 3.4M square feet that sits outside published leases. The rings mark the totals JLL and Avison Young publish.

OpenAI holds 1,181,600 square feet and Anthropic 944,955. Third place, Sierra, has 392,239. All eight panels share one scale, so the gap is drawn to size. Lines step up in the month a deal was reported.

Every AI company with a current San Francisco office location on the public record: frontier and model labs, AI infrastructure and tooling, and AI-native applications. 47 of them have published square feet. Sort by any column or search a name. Companies with square feet link to their own page with addresses and sources.

Financial District/South Beach has 141 current AI locations. Next come South of Market (29), Mission (29), Mission Bay (20), Chinatown (20), Potrero Hill (10). These are location counts, offices of every size.

For a landlord, AI concentration in one building is the exposure that matters. 3180 18th Street (Exa, 100%), 575 Florida Street (OpenAI, 100%), 295 Bay Street (Cursor, 100%), 550 3rd Street (Cognition, 98%) lead. The share is observed AI square feet over the building's square feet in the DataSF property record; where the two sources disagree upward (two buildings) the share is capped at 100%. Only 52 buildings can be measured at all. The strip shows the top 20; the 15 in bold, right of the half-building line, are the majority-AI buildings, and they are also the first 15 rows of the data table.

Every deal here names the previous occupant in a press report or filing. Most records name none, so treat this as a sample of turnover, with square feet as each source states them.

92 AI locations started in 2025 and 22 ended. Through October 2026, 2026 has 70 started and 9 ended. Starts are registration or signing dates and ends are registration closures, so neither marks a physical move.

JLL counts 413 AI companies occupying 8.5 million square feet. Avison Young says 8.75 million. This register has 223 companies and 5.06M square feet.

The square-foot gap is probably unpublished leases. We count a lease only when its size is public, and 207 of 271 current locations have none. A targeted search of every one of them (press, filings and city permits) turned up sizes for only a dozen. Most are small offices by our headcount model, and several addresses hold four or five of them, the pattern of shared and co-working space; leases that size are rarely announced. Brokers see those deals in their own lease databases, which is how JLL reaches 8.5M.

The company gap is a definition gap. The register leaves out 192 current records at home addresses and mailboxes, 109 current records whose company identity we could not verify, 141 locations of AI groups inside large non-AI companies, and AI-first companies founded before 2017 (outside our definition). No broker publishes its definition of an AI company, so the counts cannot be matched company by company.

Reconciliation of company counts. 223 verified companies + 83 unconfirmed office registrations + 116 registered at a home or mailbox only = 422 companies with a San Francisco registration. JLL counts 413. The 83 and 116 are businesses registered in San Francisco under the name of an AI company whose identity we could not confirm. They are unconfirmed, are not in the 223, and add no square feet to the observed total. The names are in tier2.csv.

A check on that explanation. Sizing the 116 of those 207 locations that can be sized from headcount (each company's employee range times its San Francisco share, 100% when it is headquartered here and 10% to 40% otherwise, at 125 to 250 square feet per person) gives 0.5M to 3.6M square feet. Published plus modeled is 5.5M to 8.6M, a range that takes in JLL's 8.5M. Modeled figures are estimates, not records; outside this paragraph they appear only in the data file's labelled columns.

5 brokers publish figures on AI tenants in San Francisco. None publishes its list of companies or its definition of an AI company, and none puts its figures in a data file. Each row below gives the figure, what it counts, and how we read it.

• Space held and space leased are different measures. JLL (8.5M SF) and Avison Young (8.75M SF) count space AI companies occupy at one point in time. Cushman & Wakefield (2.1M SF in 2025, 4.4M SF from January to September 2026), Savills (2.9M SF in the first half of 2026) and CBRE (10.6M SF since 2019) count space leased during a period. A lease signed this year may not be occupied yet, a renewal adds leasing without adding space, and space given back never shows in a leasing total, so the two kinds of figure cannot be added or netted. JLL's 2.4M SF committed since January is a leasing figure too.
• Shares rest on different totals. For the first half of 2026, Cushman & Wakefield (2.9M SF) and Savills (2.9M SF) report the same AI leasing, yet Cushman & Wakefield puts it at 47% of new leasing and Savills at 31.3%. Both can hold only if the new-leasing totals underneath differ, and neither publishes its deal list. Avison Young's 13.4% is a share of occupied space, not of all space; on Cushman & Wakefield's inventory of 87,080,074 SF, JLL's and Avison Young's totals both come to about one tenth (9.8% and 10.0%).
• The size of the market differs. Cushman & Wakefield counts 87,080,074 SF of office inventory in San Francisco; Kidder Mathews counts 103,515,948 SF. Any "share of the market" depends on whose inventory sits under it.
• Subleases and renewals are treated differently. Savills includes subleases and puts them at 23% of AI leasing from 2025 through mid-2026. Cushman & Wakefield's 47% counts new leases only. The other figures do not say.
• Nobody defines an AI company. Without a published definition or list, a company one broker counts can be missing from another's total, and no two counts can be matched company by company. This register publishes both: the definition and every record in tenants.csv.
• Brokers see deals that are never announced. Their figures draw on their own lease databases. This register counts a lease only when its size is public, which is why its 5.06M SF sits below JLL's and Avison Young's totals.

tenants.csv lists every company and location since 2019: company, tier, building and address, neighborhood, coordinates, start and end dates with their basis, and observed square feet with a label for lease signed or sublease. Its square-feet column holds observed figures only, and a blank cell means no public figure. Three further columns (sf_est_low, sf_est_high, sf_est_basis) carry a modeled range for the 116 blank-cell locations that can be sized from headcount; they are modeled, not observed.

tier2.csv lists the 199 unconfirmed registrations by company, tier and status. It has no addresses.

Map data files: series.json, replaced.json, stats.json, lease_tape.json (every dated lease and sublease behind Exhibit 1) and footprints.json (each company's observed square feet by month; the companies sum to series.json).

Embed the map. Paste this into any page; it shows the same data and links back here.

<iframe src="https://morainecre.com/resources/research/sf-ai-office-map/index.html#embed=1" title="San Francisco AI tenants map (Moraine Research)" width="100%" height="620" style="border:0" loading="lazy" allowfullscreen></iframe>

Corrections and updates. The register is refreshed monthly from the same public sources. If a record is wrong or missing, send a correction with a public source; fixes go into the next refresh. To hear when it is refreshed, ask to be notified.