Underwriting CRE deals used to mean hours (or days) of manual work in legacy software or Excel. Introducing Minervian AI — institutional-grade cash flow modeling with full transparency and precision. Upload a rent roll or PDF → refine assumptions in plain language → full deterministic model (lease-by-lease, recoveries, debt, IRR, DSCR, equity multiples, sources & uses) in minutes, if not seconds at https://t.co/QwOCRKcet1
This series explores the potential impact of orbital launches on data centers and the real estate industry.
Starship reached orbit on September 28. Flight 14 put the ship into low Earth orbit for the first time and released 26 operational Starlink V3 satellites, its first revenue payload.
The compute implication is straightforward. Terrestrial AI is already power-constrained. Northern Virginia’s colocation vacancy was about 0.2 percent in the first half of 2026, with 10.8 MW available. Google’s own orbital-compute work treats roughly 1,800 Starship flights as the mass required to push launch cost toward $200 per kilogram, the band where space-based power might become comparable to ground power. The first Google TPU prototype flew on a Falcon 9 this month, in short bursts.
For real estate, the base case does not yet move.
Data centers still clear on deliverable megawatts, interconnect date, water, and offtake credit. Orbital compute belongs in a year-ten exit-cap sensitivity, not in year-one NOI. Sites with power keep the scarcity premium until a launch cadence actually exists.
Other asset classes split cleanly. Industrial land, gas laterals, and workforce housing near Starbase and supplier corridors are the direct trade. Cameron County is already stacking launch infrastructure against LNG and a terrestrial AI campus, competing for the same water. Office demand is narrow. Hospitality at the pad is noise.
What change:
• Split the memo. Terrestrial campus: contracted megawatts, ramp, power cost, vacancy after the first term, obsolescence of the shell. Orbital option: a sensitivity on exit cap rate in year ten, not a line in year-one NOI.
• Power is the rent roll. A data-center pro forma that treats electricity as a pass-through and ignores delivery date will mis-size debt the same way a multifamily pro forma mis-sizes it by ignoring loss-to-lease.
• Technology obsolescence gets a real case. Training clusters and orbital buses both shorten the useful life of today’s shell and today’s chip. Hold periods and residual assumptions should say so.
• Local fiscal deals are part of basis. Brownsville’s $220 million water MOU is a reminder that the public balance sheet is now inside the sources and uses.
Follow #MinervianAI for a 5-year window analysis.
#CRE #AI #OrbitalAI #CommercialRealEstate #Datacenters #SpaceX
Cushman & Wakefield’s note on healthy shopping centers is right about frequency. The part worth underwriting is what the frequency numbers are actually saying.
The center stopped being a place to buy goods. From 2019 to 2025, U.S. apparel visits fell about 25%. Fitness rose about 23%, med-spa and wellness about 32%, off-price about 29%. Service tenants took more than half of new retail leasing in 2025, the first time they outleased goods.
Frequency and dwell are different products. A gym at four visits a month, with 90% of those visits lasting more than 30 minutes, buys the trip. A med spa at roughly one visit a month buys duration and spillover, and has no e-commerce substitute. Coffee is the other habit node, and it barely shows up in the write-up. One without the other is a half-built center.
The goods that survived are a hunt, not a category. Off-price still pulls repeat visits because the inventory is not a SKU you can reorder on a phone. That is why subdivided department-store boxes often refill with a Dick’s, a Burlington, and a fitness operator, at two to four times the old anchor rent.
The supply event of the decade is the anchor vacancy, not a new format. Sears, Macy’s, Saks, and in Canada Hudson’s Bay, Nordstrom, and Target left the boxes. Almost no retail has been built, so the remix happens inside a fixed shell. Capital has already paid up for the clean version: grocery-anchored centers in the high-5s to mid-6s. The spread is the former anchor whose new mix is not yet stabilized.
One warning on the European receipt. Spend is up about a quarter since 2021 and units only about 5%. Nominal sales can look fine while the visit is not actually bigger. Percentage rent that does not deflate for that will lie.
Frank Wilczek’s WSJ op-ed, “Let Physics Regulate the AI Industry,” is half right. Vague “develop AI responsibly” rules get gamed, and gigawatts do not. The remedy does not follow from the physics.
His exhibit is Stargate in Abilene. The campus is planned at 1.2 GW, with 450,000-plus Nvidia GB200s at full build. About 200 MW of IT load was already operational by April 2026. On site: 10 gas turbines, roughly 360 MW, and 62 diesel generators, with a pending permit for 41 more turbines and 18 more generators. That is not a pause.
The constraint he wants to invent is already binding. Elon Musk has put chip output up 40–50% a year against power growth of 10–20% outside China. Morgan Stanley has circulated a figure of about 100 GW of US data-center power needed through 2028, on the order of 100 nuclear plants, against a new-build nuclear timeline of about a decade. Bank of America has pointed to 230-plus GW of new US generation within five years, more than half of it data-center driven. Eaton has counted about 307 GW of US data-center projects in the pipeline. Large transformers take years. GE Vernova’s turbine and grid backlog is about $176 billion, with slots stretching toward 2030. Oracle has already declared force majeure on a New Mexico campus for lack of power.
The industry response is a second grid, not a loophole. Analysts have projected nearly 30 GW of new behind-the-meter gas by 2030, with data centers about 88% of it. Bloom’s fuel-cell backlog is about $20 billion, including 2.8 GW for Oracle, and one Oracle site was energized in 55 days. The carbon-free deals that actually exist are restarts and existing plants: Microsoft at about 835 MW on Three Mile Island, Meta at 1.1 GW on Clinton, Vistra with roughly 3.8 GW contracted, including 2.6 GW for Meta and 1.2 GW for AWS at Comanche Peak. A rule that no large data center gets a connection unless the developer first delivers new carbon-free generation, storage, and transmission does not add physics. It bans the only megawatts that arrive before the chips depreciate.
It is also unilateral. SemiAnalysis puts China at about 24 GW of operational data-center capacity against roughly 56 GW in the US. Inner Mongolia’s Wulanchabu hub alone is building on the order of 15 GW across 89 sites, in 12–18 months versus 18–24 in the US, at about 20% lower cost. Test bans and the Montreal Protocol worked because they were multilateral. A US-only clean-power gate moves the training run.
“Pay the full cost” is the part worth keeping. The DOE has data centers near 12% of US electricity by the end of the decade. Do not socialize that onto residential ratepayers. Do not pretend a carbon-free precondition is the same thing as physics.
The rebound is real for a subset of properties, but it is a sorting process rather than a broad revival of the classic mall.
Green Street’s Commercial Property Price Index shows mall values up roughly 13–14% over the past year, the strongest gain among the sectors it tracks, and now about 6% above their 2022 peak, while the broader commercial property index remains around 13% below its prior high. Roughly 200 U.S. malls have closed since 2008, leaving about 900 operating. Analysts at the firm estimate only around 250 higher-quality (A- or better) properties are capturing most of the gains. Limited new supply has concentrated surviving tenants into those centers, supporting occupancy (mall REITs often near or above 96%) and rent growth on new leases.
The tenant mix has shifted away from traditional department-store anchors toward restaurants, entertainment, fitness, off-price, and luxury. https://t.co/sefSK48BH8 data show indoor-mall visits up about 5% year-over-year in August 2026, with longer average dwell times than open-air centers. Younger shoppers contribute: Coresight surveys found higher recent mall visitation rates among 18–29-year-olds than older cohorts. Flagship examples such as Mall of America and American Dream continue to draw on the order of 40 million visitors annually by pairing retail with large attractions.
Skeptics, including some portfolio managers quoted in the original reporting, argue the durable value often sits in the land more than the retail operation itself, and that restaurants or entertainment backfills can be hard to replace if they falter. The recovery also remains sensitive to consumer spending and broader capital-market conditions; Green Street’s overall index dipped slightly in September amid higher yields, even as high-quality retail fundamentals held up better than weaker sectors. Public markets have priced in the improvement: mall REITs have outperformed the broader equity REIT index over recent years, but the gains remain concentrated among the better-located, better-capitalized assets.