Direct-to-chip liquid cooling vs immersion: the real ROI for data centers in LatAm

When your data center goes from 15 kW per rack to 50 kW per rack, conventional air stops working. The choice is between two liquid cooling technologies: direct-to-chip (DLC, where fluid runs through cold plates mounted on CPUs and GPUs) or immersion (where the entire server is submerged in a bath of dielectric fluid). Both solve the density problem. The difference is in the real ROI, which depends on three factors: density per rack, local energy cost, and workload type.

What each technology solves

DLC (Direct Liquid Cooling) keeps the server in its standard form factor and adds a cold plate over the high-heat components (CPU, GPU, HBM). Coolant circulates between the cold plate and an external CDU. The rest of the server continues to dissipate heat through air via internal fans. Immersion removes air entirely: the server (no chassis, no fans) is submerged in a tank of dielectric fluid. The fluid boils gently on the chip surface and condenses in an external coil.

When DLC makes sense

DLC is the right option when your rack density sits between 30 and 80 kW and you want to preserve the physical form factor of the server. Operational advantages: you use the same server you already know, the rack stays standard 42U, the change is done in stages (you can start with one DLC rack and leave the rest on air). Limitations: the cold plate only cools the components it touches — RAM, SSDs, and VRMs still rely on air, so you still need residual ventilation. And it requires one CDU per every 3–5 racks, with its own water-glycol or refrigerant loop.

When immersion makes sense

Immersion is the right option when your density exceeds 80 kW per rack, or when you want to eliminate fans entirely (reduce noise and mechanical failure points). The dielectric fluid touches the ENTIRE PCB surface, so cooling is uniform. Energy efficiency is 5–10% better than DLC because there are no fans. Operational limitations: the server must be specifically designed for immersion (connectors, cables, and optical drives are fluid-sensitive); the fluid costs 5–10x more than water-glycol; it requires specialized fluid handling (seals, drainage, periodic degradation analysis).

How to calculate the real ROI in LatAm

The ROI of migrating to DLC or immersion depends on three local variables:

1. Energy cost. In Mexico, CFE tariffs for medium-voltage data centers range between MX$1.20 and MX$2.50 per kWh depending on schedule and region. The higher the energy cost, the greater the savings from lower PUE, and the faster the liquid cooling capex pays back.
2. Density tariff. If you are paying per square foot (rent) or per contracted kW, liquid cooling lets you pack more load per square meter. The rent savings can be the strongest argument.
3. Expected useful life. If the IT load will scale 2–3x over the next 3 years, liquid cooling is justified sooner. If the load will remain stable, the ROI extends.

What almost nobody includes in the calculation

There are three hidden costs that change the ROI:

• Property insurance. Standard data center policies contain clauses about flammable or toxic dielectric fluids. Some insurers raise the premium 15–25% if you use immersion with mineral fluid. Check with your insurer before purchasing.
• Staff training. DLC requires technicians who understand water-glycol loops, pressures, and fluid chemistry. Immersion requires technicians who understand dielectric fluid chemistry, vapor handling, and seals. Training takes 2–4 weeks per technician.
• Opportunity cost of space. A DLC rack occupies the same footprint as an air rack. An immersion tank occupies twice as much. If your DC space is limited, DLC gives you more density without sacrificing layout.

The capex detail that actually changes

For a typical DLC rack (40 kW): the cold plate per server costs USD $200–500, the rack-level manifold USD $1,500–3,000, the 200–400 kW CDU costs USD $25–60k, the external dry cooler or cooling tower USD $15–40k. The total per rack runs USD $60–120k depending on density. For immersion: the tank with dielectric fluid costs USD $40–80k per rack, the condensation system and fluid handling add another USD $30–50k. Capex per cooled kW is typically 1.5–2x higher in immersion than in DLC, but energy efficiency compensates over time in sites with expensive power.

Cases where neither makes sense

There are three scenarios where neither DLC nor immersion is justified, and they are worth making explicit:

1. DC with fewer than 50 racks total. Capex per rack is too high to amortize without a large fleet spreading the cost of CDUs, loops, and tanks.
2. Stable IT load with no growth plan. If your load will stay at 7–15 kW/rack for the next 5 years, conventional air with good maintenance remains the most financially efficient option.
3. Leased site with no major civil works possible. Both DLC and immersion require installing the external loop (dry cooler, condenser, etc.). If you cannot do that work, migrate the workload to a colocation with liquid cooling available.

Practical recommendation

If you are in the 30–50 kW per rack range, start with DLC — it is cheaper, easier to deploy in stages, and most of your current servers can be retrofitted with cold plates. If you are above 80 kW or you are building a greenfield, immersion gives better energy efficiency at the cost of higher capex and operational complexity. For DCs in LatAm in the 50–80 kW range, the decision comes down to energy cost: if your kWh is above MX$2.00, immersion pays for itself; if it is below MX$1.50, DLC is the financially correct decision.

Sources

1. Vertiv — homepage — https://www.vertiv.com/
2. ASET — homepage — https://www.aset.org/
3. Uptime Institute — AI Infrastructure Advisory (cooling considerations) — https://uptimeinstitute.com/ai-services/ai-infrastructure-advisory
4. TIA-942 — Telecommunications Infrastructure Standard for Data Centers — https://www.tiaonline.org/