Walk the floor at Supercomputing 2026 (SC26) and you will see a lot of impressive silicon. Far less visible is what happens between the purchase order and the first production job. That gap is where we spend our days at AHEAD Foundry. 

The rack is the fast part of an AI deployment. The slow part is the facility around it. An AI-ready rack is integrated and proven before it ships, so it can be brought into production the day it arrives.  

That includes inventory synchronized into ServiceNow or whichever asset system you use. It also removes the work that eats months of many deployments, including racking, cabling, labeling, re-cabling, firmware updates, OS installs, PXE boot problems, East-West network faults, dead-on-arrival parts, and the wait for replacements. 

Say a team commits to a January 15 go-live. The colo contract is signed and billing starts that day. Equipment lands on the dock piecemeal, a GPU server one day and a PDU the next, with 20 POs to track across 10 vendors. Cables get reseated. GPU jobs fail to start. Several components arrive dead. Someone finds a part nobody ordered. The clock keeps running, and weeks or months pass before the system does what it was bought to do. Every one of those days is billed and produces nothing, and the nights go to finding the one expert who can diagnose 10,000 new critical UFM messages. 

An AI-ready rack skips that. It is built, tested, and proven at AHEAD Foundry before your go-live date, ships on schedule, and is ready to power on when your data center is ready.  

Ready means a running AI service 

I define ready as the whole path from the facility to a working AI service. The stack has five layers, and the first two belong to your facility. 

• Cooling. In many cases, direct liquid cooling (DLC), with the plumbing and liquid networks in place to carry the heat away. 
• Power. The right whips to the PDU and power shelf inputs, plus battery UPS where early utility draw can cause trouble for a new load. 
• Hardware. A full working reference architecture, racked and cabled. 
• Bare-metal software. OS patched and out-of-band (OOB) firmware updated so a Kubernetes cluster can run. 
• AI platform. Red Hat OpenShift AI, or an open-source orchestration stack on Kubernetes. 

Get all five right and you have dial tone. Your teams can start putting workloads on the platform. AHEAD Intelligent Operations can extend that, so the rack arrives with an observability stack ready. 

My test is simple. As racks rolls into your data center, ask these questions. 

• How much work is left, and how much of it is understood? 
• How much did your supplier push into your corner that you have not planned for? 
• Where do you get the expertise and the budget for the missing pieces? 
• What do you tell your end users when delays hit their projects? 

Liquid cooling works when your facilities team understands it 

Most of the objections I hear to DLC are about complexity. In practice the facilities team owns three things, the supply and return temperatures that set delta T, the flow rate in gallons per minute, and leak detection, which we prove during burn-in before a rack ships. 

The return is worth the effort. I put the energy saving at around 20 percent compared with air. A zero-evaporation closed loop is a reality today, and water usage effectiveness (WUE) improves about 40 percent with closed-loop DLC. 

Our Executive Playbook for Liquid Cooled Infrastructure for AI covers the terminology and the business case in more depth. 

Ask about power first 

The first question I hear from customers moving from air to direct-to-chip is whether they have the power. Racks above 100kW are routine in AI deployments. NVIDIA rates a GB200 NVL72 at up to 120kW, and a GB300 NVL72 runs about 135kW NVIDIA’s roadmap points to 600kW Kyber racks by the end of 2027. 

On 208V three-phase, a row of 250kW racks draws roughly 700 amps. At 415V it draws about 350, which is why 415V and 480V are common and why DC busbar distribution with power shelves is how NVL72-class racks are fed. 

NVIDIA’s next-generation racks make the case for 800V DC as density climbs toward 600kW. Hyperscalers are already pushing that edge, and in some cases they leave the rest of us in the enterprise white space with a narrower set of DC-enabled equipment. 

Retrofits carry a trap of their own. A rear-door heat exchanger (RDHx) can bring an existing data center into the game, but a rack and its door can only move so much air. Match the rack’s CFM to what the door can actually cool, because the rated airflow often does not match what you get on the floor. In many cases I would skip this step and go straight to DLC. 

Prove it before it ships 

AHEAD Foundry has been integrating and proving racks for years, and that experience shaped our new facility on AHEAD Foundry’s campus. Opened in April 2026, it was designed specifically for large-scale rack integration and liquid cooling, with 10 megawatts dedicated to rack-scale AI and HPC integration. It supports air and single-phase direct-to-chip cooling, with multiple loops connecting to multiple cooling vendors’ solutions for full DLC rack integration. 

The Innovation Lab, now five times the capacity of our previous lab, includes a digital twin where we simulate before we deploy. We can tie into NVIDIA DSX Air to simulate the network fabric and test provisioning and orchestration before any hardware moves. Customers can test and benchmark air, direct-to-chip, and hybrid designs on thermal efficiency and power draw before anything ships. 

Hatch keeps serial records, firmware state, and shipment history attached to every rack, so what we tested is what arrives. 

The slow part is the building 

Here is the schedule math behind that. Once it reaches a ready facility, a rack built and tested on our floor can be online in days. The facility is the long pole. Re-plumbing a data hall for several dense liquid-cooled racks can take 18 to 36 months, and the hardware waits on that work.  

That is why the red flag I watch for is a promise to deliver the whole path, from facility to running AI service, in two months. If the facility is not already built, nobody can honestly promise that. If it is, the promise usually means the testing moved to your floor, and you get the January story, with cables to reseat, dead parts, and a clock that keeps running. Avoiding that takes people who know the network, the storage, the environmentals, and the GPUs. 

Bring your specs to AHEAD booth 6720 at SC26 

AHEAD will be at SC26 at McCormick Place in Chicago, booth 6720. The exhibit hall is open November 15 to 19. 

If you are planning for a 100kW rack today or something near 250kW down the road, bring your rack density, voltage, and water specs. I will tell you what I would check first and where a schedule like yours is most likely to slip. I would also like to hear what you are building.