Straight answers.

A data center is a building engineered to keep computers running 24/7. Four physical systems: power (from the grid + backups), cooling (to remove heat), networking (fiber connectivity), and the IT itself (servers, storage, switches). Everything else is accessory.

Power density is roughly 10× higher. A traditional rack draws 5–10 kW; an AI rack with GPUs draws 70–130 kW. This single fact forces every other design decision: liquid cooling instead of air, InfiniBand instead of Ethernet, dedicated substations instead of grid taps, and campuses sized in gigawatts instead of megawatts.

A reliability rating from Uptime Institute. Tier III means 'concurrently maintainable' — you can take any one component offline for repair without losing the facility. Target uptime: 99.982% (≈1.6 hours of downtime per year). Most AI facilities design to Tier III; hyperscalers often skip certification and build to their own equivalent specs.

1U = 1.75 inches (44.45 mm) of vertical space. A standard cabinet is 42U or 48U tall. A 1U server is one slot; an 8U GPU server takes 8 slots.

Scale-up = making a bunch of GPUs act like one big GPU (NVLink within a rack, ~1.8 TB/s per GPU). Scale-out = connecting thousands of those units together across the building (InfiniBand at 400–800 Gbps per port). Both are needed to train frontier models.

One modern rack (GB200 NVL72) draws ~120 kW. A 100 MW campus = ~800 of those racks. A 1 GW campus = ~8,000. Frontier training runs need 30,000–500,000+ GPUs on a single network fabric, which drives power into the gigawatt range.

Utilities don't build new transmission + generation quickly. Adding 500+ MW of new industrial load requires new substations, transmission lines, and often new power plants. In tier-1 markets (Northern Virginia, Dublin, Singapore) the queue is 5+ years. That's why xAI built gas turbines in Memphis and OpenAI chose Texas.

AI needs 24/7 firm clean power. Solar stops at sunset. Wind is intermittent. Gas emits carbon. Nuclear is the only source that is firm AND low-carbon AND scalable. Microsoft-Three Mile Island, AWS-Talen Susquehanna, Google-Kairos Power (SMR) — all signed in 2024.

Shelby County, TN only requires air permits for generators in place more than 364 days. xAI declared them 'portable' and ran them unpermitted for ~18 months. The NAACP + SELC filed Clean Air Act notices; xAI removed them and got permits for the rest. The EPA closed the 'portable' loophole in January 2026.

UPS = Uninterruptible Power Supply. It's the bridge between the grid cutting out and the generators starting. UPS responds in <10 milliseconds; generators take 10+ seconds to come online. Without the UPS bridge, everything reboots every time the grid flickers.

Power Usage Effectiveness. Total facility power divided by IT power. 1.0 = perfect (all energy goes to computers). 1.10 = hyperscale target (10% overhead on cooling+lights). 1.5+ = legacy enterprise. A 100 MW facility with PUE 1.10 means 90 MW to the IT + 10 MW to cooling/lights/losses.

Physics. Air carries ~1.2 kJ of heat per cubic meter per degree. Water carries ~4,180 kJ per cubic meter per degree — 3,000× more. Above ~30-40 kW per rack, air flow can't keep up. B200 chips run at 1000W TDP; only direct liquid contact via cold plates can move that heat away fast enough.

Immersion = the servers are submerged in a dielectric (non-conductive) fluid. Best thermal performance, lowest PUE, but operationally complex — you can't just yank a drive out. Used in niche deployments. Two-phase immersion (boiling fluorocarbons) had momentum but was hit by 3M's PFAS phase-out announcement.

Water Usage Effectiveness — liters of water consumed per kWh of IT energy. Evaporative cooling ~1.8 L/kWh; closed-loop dry cooling ~0. In water-stressed regions (Phoenix, Madrid) operators switch to dry cooling and accept worse PUE. Microsoft pledged 'water positive by 2030' — replenishing more than they consume.

Technically no. Anthropic's Project Rainier runs on ~500,000 AWS Trainium2 chips. Google trains Gemini on TPUs. AMD MI300X is hardware-competitive. But in 2026 NVIDIA still has the most mature software stack (CUDA + NCCL + NeMo + TensorRT) so switching costs years of engineering.

NVIDIA's Blackwell rack-scale system: 72 B200 GPUs + 36 Grace CPUs in one liquid-cooled rack, connected as a single NVLink domain via 9 NVSwitch trays. Total: ~14 TB HBM3e, 1.4 EFLOPS FP4, ~120 kW. Effectively one very big GPU.

Model FLOPs Utilization — the % of theoretical peak FLOPs your training job actually achieves. 30-55% is normal. The gap is eaten by communication overhead, memory stalls, restarts, and sub-optimal kernels. Peak FLOPs are marketing; MFU is reality.

Number formats used in AI math. Lower precision = more throughput, less memory, slightly less accuracy. FP32 is classic. BF16 is standard training. FP8 is used by Blackwell for modern training runs. FP4 is used for inference. A single NVL72 rack does 1.4 EFLOPS at FP4 but only 360 PFLOPS at BF16.

InfiniBand was designed for HPC — lossless flow control, hardware RDMA, sub-microsecond latency. Ethernet was designed for internet traffic. For gradient synchronization across 10,000+ GPUs, InfiniBand's latency and packet-loss behavior matters. Ultra Ethernet (2024 consortium) is the open alternative catching up.

NVIDIA's proprietary GPU-to-GPU interconnect. NVLink 5 on Blackwell = 1.8 TB/s bidirectional per GPU. Within a rack (NVL72), 72 GPUs share one NVLink domain. Cross-rack, you use InfiniBand. NVLink is fast but short; InfiniBand is slower but scales.

Turnkey with modern GPUs: $2.5-5 billion. Shell-only (power + cooling + networking + building): $800M-1.2B. IT equipment (the 30,000 GPUs): $900M+. GPUs alone = roughly half of total capex.

Capital structure. Meta can't shoulder $27B of capex directly on its balance sheet. Private equity (Blue Owl) provides the capital. Meta becomes a tenant with an exit option every four years. This structure is now standard for 2+ GW builds — Stargate uses a similar four-party JV.

A GPU-as-a-service company built around AI compute. CoreWeave, Lambda, Crusoe, Nscale. They buy GPUs (often debt-financed), rent them out by the hour, and compete with hyperscaler clouds on price and availability. CoreWeave IPO'd in March 2025 at a $23B valuation.

Complicated. Grid intensity varies wildly: 50 g CO₂/kWh in France, 5 g in Iceland, ~400 g in Texas. A 1 GW campus on a coal grid is a disaster; on a hydro grid, it's not. Current AI buildout is happening mostly on US grids with significant gas + coal, so emissions are rising even as renewable share grows. Hyperscalers have 2030 net-zero pledges — their 2024 reports show emissions UP vs 2020.

Meta's Hyperion will ultimately consume ~5 GW — more than half of Entergy Louisiana's total current generation. The state couldn't provide that from existing plants. 10 gas plants + 240 miles of transmission + battery storage are the fastest-to-build option. An MOU on future nuclear is part of the long-term plan.

Three paths. (1) Technician: start with EPI CDCP certification ($1,400) → apply at Equinix/Digital Realty/hyperscalers. (2) Engineer: EE or ME degree → MEP consulting firm (JBA, Bala, Syska) → Uptime ATD cert. (3) AI cluster engineer: software/SRE background → NVIDIA NCA-AIIO cert → target Anthropic/OpenAI/CoreWeave.

US figures (median). Technician: $65-95k. Critical Facilities Technician: $75-120k. Data Center Engineer: $95-160k. Facilities Manager: $140-220k. AI Cluster Engineer at frontier labs: $200-400k+ total comp. France: generally 30-40% lower in base salary, compensated partly by 13th month, bonuses, and stronger benefits.

A Delaware joint venture of OpenAI, Oracle, SoftBank, and MGX committing up to $500B by 2029 for US AI infrastructure. Announced by Trump on Jan 21, 2025. Flagship site is a 1.2 GW campus in Abilene, Texas built by Crusoe on the Lancium Clean Campus. Five more sites now announced; total target 10 GW.

Oracle and OpenAI had been negotiating to expand to 2 GW. Additional grid capacity wouldn't be available until 2027+. Rather than wait, both walked. Microsoft leased the adjacent 900 MW. OpenAI pivoted new Stargate capacity to NVIDIA's next-generation 'Vera Rubin' silicon on five other sites.

Disputed. Bull case: AI demand compounds → every GPU finds a buyer → capex creates real assets. Bear case: 2024-2027 hyperscaler capex will exceed $1T, the closest analog is the 2000 telecom buildout which preceded ~90% peak-to-trough drops in carrier valuations. The honest answer: watch GPU utilization + training workload growth vs capex pace through 2027.