What is the Dell NVIDIA AI Factory 2.0? Dell NVIDIA AI Factory 2.0 is a joint Dell and NVIDIA reference architecture launched May 18, 2026 at Dell Technologies World in Las Vegas, combining PowerEdge XE9780 and XE9785 servers, NVIDIA Blackwell Ultra GPUs, and liquid-cooled IR7000 rack designs to support up to 256 GPUs per rack for enterprise AI workloads. Michael Dell and Jensen Huang delivered the joint keynote at 10:00 AM Pacific. For CRE data center investors, the headline is not the silicon, it is what 200-plus kilowatt rack densities and mandatory liquid cooling mean for the buildings being built today and the leases being signed for delivery in 2027 and 2028. For background on broader AI tools for CRE investors, see our complete guide on AI tools for real estate investors.
Key Takeaways
- Dell and NVIDIA launched AI Factory 2.0 on May 18, 2026 with PowerEdge XE9780 and XE9785 servers (including liquid-cooled XE9780L and XE9785L variants), supporting up to 192 Blackwell Ultra GPUs per system and 256 GPUs per IR7000 rack.
- The rack density shift pushes typical AI hyperscale rack power requirements from 50 to 150 kilowatts toward 200 kilowatts and above, redefining the buildable footprint per megawatt for hyperscale and colocation operators.
- Liquid cooling moves from optional to required at these densities, changing data center construction specs: cooling distribution units, raised slab or rear-door heat exchangers, and water access become primary site-selection variables.
- The announcement extends Dell's existing $11.5 billion AI server backlog and reinforces the multi-hundred billion dollar 2026 AI capex commitments from hyperscalers, which translates directly into CRE land, power, and water demand.
- CRE investors with data center exposure should re-evaluate underwriting assumptions on rack density, cooling capex, and tenant power requirements; pro formas built on 30 kilowatt rack assumptions are now structurally stale.
What Was Announced and Why It Matters
The Dell AI Factory 2.0 keynote walked through three layers: hardware (PowerEdge XE9780/XE9785 with Blackwell Ultra GPUs), software (NVIDIA AI Enterprise distributed directly through Dell, plus managed services for the NVIDIA AI stack), and reference architecture (the IR7000 rack with integrated liquid cooling and up to 256 GPUs per rack). The hardware tier is what reshapes CRE underwriting. A single 256-GPU rack at full TDP draws 200 to 250 kilowatts, which is roughly 5x to 10x the standard rack density colocation operators underwrote five years ago. The buildings, the power capacity, and the cooling systems built for 30 to 50 kilowatt racks are not suitable for tenant deployments at these densities without retrofit.
The Rack Density Shift CRE Underwriting Hasn't Caught Up With
Traditional data center underwriting models rack density at 5 to 15 kilowatts per rack for enterprise colocation and 30 to 50 kilowatts per rack for cloud and AI workloads. AI Factory 2.0 normalizes 200 kilowatts and above per rack. The consequence is a fundamental reshaping of the buildable megawatt per square foot ratio. A 100,000 square foot data center building that historically supported 20 megawatts of IT load may, with liquid cooling and 200 kilowatt racks, support 40 to 60 megawatts in the same shell. Operators willing to retrofit for liquid cooling can effectively double the revenue per square foot. Operators unwilling to retrofit face stranded inventory.
According to CBRE Research, vacancy in primary North American data center markets has remained below 3 percent through 2026, and the pricing power that has produced a 12 to 18 percent annual increase in colocation rates depends on operators being able to deliver the rack densities AI tenants now require. For personalized guidance on stress-testing data center underwriting models against AI Factory 2.0 density assumptions, connect with The AI Consulting Network.
Liquid Cooling Is Now Table Stakes
At 200 kilowatts per rack, air cooling is not physically capable of removing the heat at a reasonable energy efficiency. Liquid cooling, whether direct-to-chip, rear-door heat exchangers, or full immersion, becomes mandatory rather than optional. This changes the data center building program in three ways: (1) cooling distribution units (CDUs) require additional floor space, (2) water access becomes a primary site-selection variable, similar to power, and (3) construction specs shift toward raised slab designs with under-floor plumbing, away from the slab-on-grade designs common in earlier cloud-era buildouts.
Power and Water: The New Site Selection Math
Hyperscale site selection used to optimize on fiber, latency to population centers, and power transmission availability. AI Factory 2.0 era site selection adds water access as a major variable, both for liquid cooling makeup water and for adiabatic cooling tower operation. Inland regions with abundant power but limited water (parts of Arizona, Nevada, west Texas) face new constraints, while regions with both power and water (Ohio, Virginia, Washington, parts of the Mid-Atlantic) become more competitive. For more on the broader CRE impact of these shifts, see our coverage of data center delays and the 2026 power crunch and PJM power price surges.
What This Means for Different CRE Investor Profiles
- Hyperscale developers: Underwriting specs need refresh. The 30 to 50 megawatt building is being replaced by the 100 megawatt and above building with liquid cooling baked in. CapEx per IT megawatt has shifted, and stabilized cash-on-cash returns need to be recalculated at the new density.
- Colocation operators: Existing buildings face a binary decision: retrofit for liquid cooling and AI tenant density, or accept that AI tenants will go elsewhere. The retrofit economics depend on existing power capacity, ceiling height, and water access.
- Land bankers and developers: The 100 to 500 acre parcel with both adjacent transmission and water access has appreciated significantly. AI Factory 2.0 reinforces the thesis that the limiting input for the next five years is power-plus-water sites, not buildings.
- REIT investors: Public data center REITs (Digital Realty, Equinix) and their private market peers are competing for the same AI tenant deployments. Investors should pay close attention to portfolio-level rack density capability and capex plans for liquid cooling conversion.
The Broader Capex Context
Dell AI Factory 2.0 lands at a moment when Meta has guided to $125 to $145 billion of 2026 capex, much of it AI infrastructure, and the top nine North American cloud service providers are forecast to spend approximately $830 billion of combined 2026 capex per TrendForce data. The aggregate demand for power, water, land, and buildings is unprecedented. The vast majority of that capex ends up in either real estate (the buildings and the land), power infrastructure, or compute hardware, with real estate capturing roughly 8 to 12 percent of total AI infrastructure capex by most analyst estimates. That percentage applied to current capex levels implies $60 to $100 billion of annual CRE-relevant AI infrastructure spending into 2027.
How CRE Operators Should Use AI to Underwrite These Shifts
The same AI tools driving demand for data centers are also the right tools for underwriting data center deals. Sponsors can use Claude, ChatGPT, and Gemini to: (1) build sensitivity tables on rack density assumptions and resulting revenue per square foot, (2) model liquid cooling retrofit capex and payback, (3) cross-reference site-by-site power and water availability against forward demand projections, and (4) stress-test tenant credit profiles in an environment where many AI workloads are still pre-revenue. CRE investors looking for hands-on AI implementation support on data center underwriting can reach out to Avi Hacker, J.D. at The AI Consulting Network.
What to Watch Over the Next 90 Days
The post-keynote signals worth monitoring include: (1) Dell and NVIDIA pipeline announcements and revenue guidance updates, (2) hyperscaler capex follow-on announcements in Q2 2026 earnings cycles, (3) data center REIT comments on rack density and liquid cooling readiness in earnings calls, (4) Google I/O (scheduled May 19, 2026) and what Gemini 4.0 implies for compute demand, and (5) regulatory and grid signals around power transmission and water access in primary data center markets.
Frequently Asked Questions
Q: How much power does a 256-GPU rack draw?
A: At full TDP, a fully populated 256-Blackwell-Ultra-GPU IR7000 rack draws roughly 200 to 250 kilowatts of IT power, depending on workload mix and supporting CPU configuration. This is approximately 5x to 10x the density of pre-AI hyperscale racks.
Q: Does this mean existing data center buildings are obsolete?
A: Not obsolete, but constrained. Existing buildings with sufficient power capacity, ceiling height, and structural loading can be retrofit for liquid cooling. Retrofit costs vary widely. Buildings without those underlying capacities face partial repositioning or non-AI tenant strategies.
Q: How does liquid cooling change the construction cost per megawatt?
A: Liquid cooling adds 10 to 25 percent to construction cost per IT megawatt, but it enables 2x to 4x higher density in the same shell, which improves revenue per square foot meaningfully. The economics typically favor liquid cooling for AI deployments.
Q: Which regions benefit most from the AI Factory 2.0 shift?
A: Regions with both available power and water gain the most. Ohio, Virginia, the Mid-Atlantic, parts of Washington and Oregon, and emerging markets like northern Indiana and central Pennsylvania are well positioned. Inland regions with strong power but limited water face more constraints.
Q: How should CRE investors update their data center underwriting models?
A: Move rack density assumptions to a range of 50 to 200 kilowatts depending on tenant mix, include liquid cooling capex as a base-case line item rather than an upside, treat water access as a hard input alongside power, and model retrofit options separately for existing buildings versus greenfield development.