A modular data center is a building that was manufactured, not constructed — and the distinction is now the main lever on the AI industry's most scarce non-GPU resource: calendar time. In the traditional mode, a data center is a serial construction project: civil works finish, then power distribution is installed, then cooling, then fire and security, each trade waiting on the last. In the modular mode, the power skids, cooling units, and control systems are built and pre-commissioned in factories simultaneously, shipped in container-like modules, and stacked and connected on site. Alibaba Cloud's data center infrastructure chief Wang Chaoyang described the change to Chinese state media in exactly those terms: traditional schedules are serial — "install power first, then connect cooling" — while the modular approach runs every system's production in parallel in the factory, pre-commissions it, and assembles it on site "like building blocks." The question this piece answers with the published record: how much calendar does that actually buy, and — more importantly — which part of the calendar does it buy?

The baseline first, because every headline number needs one. JLL's 2026 Global Data Center Outlook (the 35-page report the firm's own research summaries and multiple independent write-ups — SemiAnalysis-adjacent trade commentary, Chinese semiconductor trade press, market-report aggregators — all quote consistently) puts the global average build time for a 50 MW data center at 18 months. Alibaba Cloud's own August 10, 2026 announcement states the regional split the same way: six to twelve months in China, twelve to eighteen in the US for comparable projects. Two independent sources, one from a brokerage-research firm and one from the vendor marketing the alternative, agree on the gap modular construction is selling against — which is exactly what makes the vendor claim checkable.

How broken is the traditional delivery calendar?

JLL's numbers describe an industry whose schedules stopped holding. 57% of data center projects slipped three months or longer in 2025. Equipment lead times average 33 weeks, up 50% versus pre-2020 levels; in the US the average is 42 weeks, up 83% versus 2019. By equipment class (per JLL's breakdown, as cited across independent trade coverage): transformers and switchgear about 43 weeks, generator sets about 51 weeks. Developers are pre-ordering critical materials up to 24 months ahead, and large operators hold six to twelve months of strategic component inventory to buffer the queue. And upstream of all of it sits the clock that dwarfs everything: in the US, grid interconnection for a new 50 MW facility averages more than four years — which is why JLL reports that power availability has displaced geography and cost as the industry's top site-selection criterion. Meanwhile the money keeps burning: JLL's cost series has global construction at $7.7M/MW in 2020, $10.7M/MW in 2025 (about 7% CAGR), with a 2026 forecast of $11.3M/MW (+6%).

That is the demand side of the pitch. The supply side is the factory. Alibaba Cloud's CUBE 5.0 architecture — announced August 10, 2026, vendor-reported, and corroborated by China Securities Journal, Xinhua's Economic Information Daily, the Hangzhou municipal government portal, and multiple English-language trade write-ups (IDCNOVA, Pandaily, BriefAsia) — compresses delivery of a large AI data center to 100 days: 30 days of factory production and site preparation, 50 days of on-site installation, 20 days of commissioning and testing. The architecture raises the modularization rate of its five core systems — power, cooling, security, intelligent management, and fire protection — from roughly 30% to 90%, cuts unit construction cost more than 10%, achieves a near-100% one-time test pass rate, runs air-cooled PUE at ≤1.15 and liquid-cooled at ≤1.10, supports air-to-liquid cooling conversion, and is designed to span at least three chip generations. The company says it will more than double its global modular data center capacity within 2026.

Alibaba is not the only name in the units game. Fluidstack — the New York infrastructure company selected in November 2025 to lead Anthropic's roughly $50 billion US data center buildout in New York and Texas — states on its own homepage that it builds "technology to enable delivery of gigawatts of compute in 6 months rather than the industry's 18–24 months," and reporting on the company (Forbes, Bloomberg via secondary coverage) describes its approach in the same factory terms: prefabricated modules assembled on site, with the work that normally happens in a construction lot moved into a factory. JLL's own 2026 modular compilation tabulates announced projects against the 2–3-year traditional baseline: a CIMC-built 60 MW facility in Johor, Malaysia at 10 months, a Fluidstack-built gigawatt-scale project for Anthropic in the US at 6 months, and 3-month immersion-cooling "compute pods" in Guizhou, China. Treat the per-project numbers as proponent claims compiled by one analyst firm — but the direction is corroborated across every source we checked, and the industry-norm baselines (18–24 months, 2–3 years) appear identically in JLL's research, Fluidstack's own positioning, and Alibaba's announcement.

Where do the 100 days actually start and end?

This is the reversal, and it is stated in Alibaba's own materials if you read the scope carefully: the 100-day window runs from a basic shell or stacked containers to hand-off — before servers and network equipment enter the building. Land acquisition, permitting, grid applications, the substation, the long-lead electrical equipment, and the entire IT build-out are outside it. BriefAsia's reporting of the same announcement says this explicitly: the 100-day timeline "assumes site readiness, including grid interconnection" — and notes that securing power allocations in tier-one Chinese cities can take six to nine months on its own. Compare the excluded items against JLL's lead-time table and the picture inverts: the part of the calendar modular construction compresses (site works, systems integration, commissioning — the 12–18-month segment) was never the longest pole. The longest poles — 4+ years of grid interconnection, 43–51 weeks of transformers and generators — sit upstream of the factory gate and downstream of the first server, respectively, and modularization touches neither.

So the honest frame is not "data centers in 100 days." It is: modularization deletes the construction phase as a schedule risk, converting on-site labor variability (weather, trade availability, serial-trade delays — the causes behind most of JLL's 57% delayed projects) into factory throughput, which is measurable, schedulable, and repeatable. JLL's forecast makes the industrialization explicit: new regional assembly hubs could cut logistics costs by 40% and push delivery cycles toward eight weeks, and the firm expects the modular systems market to reach $48 billion in annual sales by 2030. That is a manufacturing-industry trajectory, not a construction-industry one.

What is the arithmetic on speed versus cost?

Run the money (ours; assumptions stated). JLL's global average construction cost is $10.7M/MW (2025), so a 50 MW shell is roughly a $535M project. Assume an 8% cost of capital — conservative for current AI infrastructure financing. One month of delay on that project costs about $3.6M in pure carrying cost before a single rack earns revenue. JLL's modal delay — the ≥3-month slip that hit 57% of 2025 projects — costs about $10.7M. Now the modular side: 100 days against the US 12–18-month range saves 11 to 15 months of build calendar (call it ~14.5 against the midpoint); against JLL's 18-month global average it saves about 15 months. At 8%, those months are worth roughly $39–52M on the same 50 MW project — which is, within rounding, the entire 10% capex saving Alibaba claims (10% of $535M ≈ $53M). The conclusion we draw: the speed benefit and the cost benefit are the same size, and they stack — which is why "modular as premium product" misreads the offer. On a revenue-bearing AI facility, earlier delivery is not a convenience; it is the whole return profile pulled forward (see our piece on time-to-power for the queue side of the same ledger).

Two second-order numbers sharpen it. First, utilization-of-design: Alibaba claims 5–10× higher compute density per square meter and nearly halved transformer counts, which shrinks the physical plant that has to be built, permitted, and cooled in the first place — less building is itself schedule. Second, quality-as-schedule: the reported near-100% one-time test pass rate attacks the commissioning tail, the phase where traditional projects hide their worst overruns (JLL's own staged timeline puts commissioning and IT bring-up at 3–8 additional months in the conventional model).

What is the 2026 reality check?

The vendor claims are claims, and the caveats are concrete. Every Alibaba number above is vendor-reported — corroborated by multiple independent press accounts of the announcement, but not by third-party deployment audits; trade commentary (Inside AI's analysis of the launch) flags exactly what to watch: factory-built modules must match site-specific grid and water conditions that vary by region, and a factory line is only as reliable as its component flow — a global medium-voltage equipment shortage "could idle factory lines and erase time savings." Note that this is the modular thesis's real weakness: it does not eliminate supply-chain risk, it relocates it upstream — from a construction site (where a missing crew delays one project) to a factory (where a missing transformer batch delays every project on the line). JLL's own report documents operators stockpiling 6–12 months of components precisely because of this exposure.

The sources disagree at the margins, and we flag rather than harmonize. On Alibaba's capacity expansion: the company's announcement as carried by IDCNOVA and Chinese state media says it will increase global modular capacity "by more than double" (两倍以上); Pandaily's write-up renders the same statement as "triple." The two readings are the classic Chinese-numeral translation gap; we carry the conservative one. On baselines: BriefAsia's regional context puts traditional Southeast-Asian builds at 18–24 months, while the Alibaba/JLL-converging figures for China and the US are 6–12 and 12–18 — both can be true; region matters. And the boundary of the 100-day claim itself is scope-dependent by construction (powered shell in, IT equipment out), which is why we anchored the comparison to the build segment JLL measures at 18 months rather than to a door-to-door project calendar.

Who should care, and how. For AI labs and neoclouds signing 2027 delivery slots, the modular segment is where credible acceleration now lives — the JLL compilation's 3–10-month examples and Fluidstack's 6-month gigawatt pitch sit in the same band as Alibaba's 100 days once you normalize scope, and the differentiator between proponents is no longer architecture slides but factory throughput and slot availability. For everyone financing the buildout, the discipline this piece argues for is the same one our colocation-rent analysis argues for: decompose the calendar the way you decompose the bill. Factory time, queue time, and equipment lead time are three different clocks with three different owners — and a "100 days" headline that silently starts at a powered shell is telling you only about the first one.