Manhattan’s Office Squeeze Is Measured. Its AI Story Is Not.

Colliers published its Manhattan office market report for August 2026 on September 1. Leasing activity totaled 3.25 million square feet in August, down 16.0 percent from July and still above both the five year and the ten year monthly averages. Year to date volume reached 29.91 million square feet, up 9.4 percent from the same period in 2025, which Colliers says puts 2026 on pace for the strongest annual leasing total since 2000 if the current pace continues. The availability rate fell by 0.2 percentage points to 12.5 percent as total availability declined to 65.40 million square feet, the lowest level since September 2020 and, in Colliers’ own framing, one third below the post-pandemic peak reached in February 2024. The market recorded 0.83 million square feet of positive absorption during the month.

Two other figures in the same report matter more to a landlord than the headline does. Sublet availability was cut by 40,000 square feet in August and fell 22.3 percent over the past year to 10.07 million square feet, the lowest sublet supply since August 2019. Sublet space is the shadow inventory that capped Manhattan pricing for four years, and it now sits at roughly 15 percent of all available space, which computes from the two figures Colliers printed in the same report. The overall average asking rent decreased by 0.2 percent since July but grew 4.2 percent since August 2025. Colliers’ second quarter report, published July 8, carries the level behind that rate of change: the Manhattan asking rent average rose 0.6 percent in the second quarter to $78.03 per square foot, the highest since July 2020, up 5.7 percent over twelve months, and 1.8 percent below the March 2020 average of $79.47.

The concentration is the part a leasing team should read twice. Midtown leasing rose to 1.54 million square feet, led by General Atlantic’s 152,000 square foot transaction at 625 Madison Avenue and KeyBank’s 113,000 square foot renewal and expansion at 1301 Avenue of the Americas. Midtown South contributed another 1.53 million square feet, supported by commitments from Havas Health Network and the City of New York, and Colliers records both submarkets decreasing their availability during the month. Those two submarkets sum to 3.07 million square feet, which computes to about 94 percent of the month’s 3.25 million and leaves roughly 0.18 million square feet for the rest of Manhattan. Midtown South alone computes to about 47 percent. That arithmetic is PSV’s, done on figures Colliers published in one report, not a share Colliers stated.

The tightening is not a single month. Colliers’ second quarter report records the availability rate cut by 0.7 percentage points since March 2026 to 13.0 percent, the ninth straight quarter of a tightening or stable availability rate and the longest such run since 2007. Quarterly leasing of 11.02 million square feet ran 29.4 percent above the five year quarterly average, and Colliers notes it was the first time since 2002 that demand exceeded 11.00 million square feet for three consecutive quarters. Available supply stood at 68.10 million square feet in June, which Colliers put 30.5 percent below the February 2024 post-pandemic peak. The two reports can be checked against each other: 68.10 million at 30.5 percent below the peak implies a February 2024 peak of about 98.0 million square feet, and 65.40 million measured against that implied peak computes to almost exactly the one third decline the August report claims. The two records agree, which is worth knowing before either is quoted alone.

Neither Colliers report mentions artificial intelligence. Not once. They measure supply, absorption, leasing volume and asking rents, and they name the tenants behind the largest August commitments: a private equity firm, a bank, a healthcare marketing network and the City of New York. None of those is an AI company. That is worth stating plainly, because the same week’s trade coverage attributed Manhattan’s squeeze substantially to AI demand, and an operator reading both in the same sitting could easily merge two records that do not say the same thing.

The AI claim comes from two places, and neither is a primary record PSV could open. Commercial Observer reported on September 8 that high quality Manhattan space is now drawing competing bids, citing named brokers at JLL, Newmark, Koeppel Rosen and Current Real Estate Advisors, and describing landlords raising asking rents in the middle of marketing a space: a 17,610 square foot availability at 151 West 26th Street moving from $56 to a range of $60 to $70 per square foot, and a 3,612 square foot space at 1261 Broadway climbing from $44 to $53. It quotes a Koeppel Rosen leasing director describing AI tenants that take 15,000 square feet today and need 30,000 a year later. Separately, Connect CRE reported on September 4 that a new JLL report puts technology as the city’s second most active leasing sector behind finance, with finance, law and technology together above 70 percent of year to date leasing against roughly 50 percent through the 2010s, and tech leasing of 1.1 million square feet in the quarter with AI companies behind more than 60 percent of it. PSV did not verify those JLL figures. JLL’s New York Office Market Dynamics report sits behind a lead capture form, and PSV does not submit contact details to obtain a source, so they are reported here as reported.

The one public primary record on AI office leasing volume is CBRE’s, and it is older and narrower than the way it usually gets quoted. In a May 12, 2026 release CBRE reported that AI companies leased 21 million square feet in San Francisco and Silicon Valley since 2019, and another 9.4 million square feet in Manhattan, Boston and Seattle over the same period. That 9.4 million is a three market total across seven years, not a Manhattan figure and not a 2026 figure, and it is routinely repeated as though it were both. CBRE separately put tech at 22.7 percent of all United States office leasing in the first quarter of 2026 and 16.8 percent for full year 2025, and attributed the shift to $578 billion of venture capital into United States AI companies since 2020, roughly 80 percent of it in the Bay Area. CBRE Tech Insights Center executive director Colin Yasukochi said AI investment “has moved from research and experimentation to large-scale deployment.” That is CBRE describing its own data, and it is a national read rather than a Manhattan month.

So the honest state of the record is this. The tightening is measured, dated and public, and it has been running for nine quarters. The composition of the demand causing it is reported by brokers with a commercial interest in the market they are describing, and by a report PSV could not open. Both can be true at once. An operator who underwrites the first is standing on a published series; an operator who underwrites the second is standing on an attribution. PSV covered the national version of this question on August 20 reading CBRE’s Scoring Tech Talent 2026, and the gap has not closed since: nobody has published square feet per AI worker, and nobody has published what share of Manhattan’s 29.91 million square feet of 2026 leasing went to AI tenants.

The job here is not forecasting. It is keeping two kinds of record apart inside one table, so a leasing or acquisitions team stops merging them by accident. Inputs: the Colliers Manhattan monthly and quarterly reports for the trailing eight quarters, captured as published figures with their publication dates attached; the portfolio’s own availability and asking rent history by building and by floor; every executed comparable in the relevant submarket with tenant name, square footage, term and stated industry; and each tenant’s own public record, meaning press releases, funding announcements, filings and posted job listings. Output: one table, one row per comparable, carrying square footage, submarket, asking rent at listing, rent at signing, the delta between them, the tenant’s industry as the tenant describes itself, and a source link with a date in every populated cell.

The column that does the real work is the last one, which labels every figure as measured, reported or inferred. Nothing moves between those three states without a citation changing. A model is genuinely good at the retrieval and normalization here: reading twenty market reports, pulling the same six fields out of each, matching tenant names against the tenants’ own public descriptions of themselves, and flagging every row where a broker’s industry label and the tenant’s self-description disagree. A model is not the thing that decides whether a given tenant counts as an AI company, and that classification is exactly where this analysis goes wrong, because “tech tenant” and “AI tenant” are doing different work in different reports and neither term carries a published definition. The arithmetic stays in the spreadsheet rather than in the model, because a model asked to roll up rent deltas across forty comparables will return a plausible number and no audit trail. The leasing lead reviews every row where the industry label is contested, and the head of leasing signs the version that reaches an investment committee. Refresh it when Colliers publishes, which is monthly. PSV has not run this build for a client and claims no result from it.