Quick answer
Direct-to-chip vs immersion cooling, in brief: Direct-to-chip pumps a water/glycol coolant (often PG25) through cold plates on the hottest chips, while other components remain air-cooled. Immersion submerges whole servers in a dielectric fluid, either single-phase (it stays liquid) or two-phase (it boils and condenses). Direct-to-chip fits existing rack designs more easily. Immersion captures nearly all heat but changes servicing and hardware choices.
Key takeaways
- Direct-to-chip: conductive coolant, so it must stay inside the loop.
- Immersion: non-conductive fluid, in contact with the hardware.
- Two-phase immersion uses engineered fluids that boil at low temperatures.
- Hardware warranty and serviceability decide most projects.
- Many AI deployments today use direct-to-chip.
How each works
| Direct-to-chip | Single-phase immersion | Two-phase immersion | |
|---|---|---|---|
| Fluid | Water/glycol (PG25) | Dielectric oil/synthetic | Low-boiling dielectric |
| Contact | Cold plates only | Whole server | Whole server |
| Heat removal | Chips (rest air-cooled) | Nearly all | Nearly all |
| Examples | Shell DLC S3, DOWFROST LC | Submer SmartCoolant, ElectroCool | Chemours Opteon 2P50 (developmental) |
Pros and cons
| Pros | Cons | |
|---|---|---|
| Direct-to-chip | Fits standard racks; mature supply of PG25 fluids | Leak risk; still needs some air cooling |
| Immersion | Very high heat capture; quiet | Tank form factor; servicing and hardware compatibility |
Decision checklist
- Rack density and future roadmap
- Hardware vendor support and warranty
- Facility water temperature
- Serviceability and staff training
- Fluid supply and lifecycle
Single-phase vs two-phase immersion
In single-phase immersion, servers sit in a tank of dielectric fluid that stays liquid. It's pumped through a heat exchanger to remove heat. In two-phase immersion, the fluid has a low boiling point. It boils at the hot components, the vapour condenses on a cooled coil above, and it drips back. Two-phase can move very high heat flux, but its fluids are specialized. Some historical two-phase fluid families have faced regulatory and supply changes, which is one reason newer HFO-based options are being developed (see the Chemours profile).
Hybrid approaches
Many deployments combine methods: cold plates on GPUs and CPUs with rear-door heat exchangers or air for the remaining components, or immersion for some racks and direct-to-chip for others. The right mix depends on the hardware roadmap and the facility.
Where each fits in the ASHRAE loop model
Direct-to-chip systems map directly onto ASHRAE's FWS/TCS model: the cold plates sit on the TCS, a CDU transfers heat to facility water, and the TCS fluid must meet strict quality requirements (see the table on the hub page). Immersion systems also use a heat exchanger to move heat from the dielectric fluid to facility water, but the fluid itself surrounds the electronics, so its electrical properties, material compatibility with cables, labels and thermal pastes, and long-term stability become the key questions.
Fluid questions for each approach
| Question | Direct-to-chip (PG25) | Immersion (dielectric) |
|---|---|---|
| Main reference documents | ASHRAE TC9.9 TCS guidance; OCP PG25 guidelines and base spec | Vendor and certification programs (e.g. Intel immersion certification cited by Shell) |
| Key fluid checks | pH, conductivity, glycol %, inhibitor reserve, microbes | Supplier-specified properties: not published here |
| Leak consequence | Conductive fluid near electronics | Fluid is already in contact, by design |
| Top-up | Same pre-diluted fluid | Same fluid, per the supplier |
| Serviceability | Standard rack access with quick-disconnects | Servers lifted from tanks |
Facility water temperature and heat reuse
Both approaches can work with warm facility water. ASHRAE's renamed W classes run from W17 to W45 and W+ (above 45 °C). Warmer water increases free-cooling hours and opens heat-reuse options, but the IT equipment's thermal requirements depend on the fluid temperature at its inlet, which is typically higher than the facility supply after the CDU approach. Confirm the supported inlet temperatures with your hardware vendor.
| ASHRAE class | Facility water supply temperature (upper limit) | Typical heat rejection (per ASHRAE) |
|---|---|---|
| W17 | 17 °C | Chiller / cooling tower |
| W27 | 27 °C | Cooling tower, water-side economizer |
| W32 | 32 °C | Cooling tower |
| W40 | 40 °C | Cooling tower or dry cooler |
| W45 | 45 °C | Cooling tower or dry cooler; district heating possible |
| W+ | >45 °C | District heating |
All classes carry a 2 °C minimum. W classes describe the facility water supply (FWS), not the fluid temperature at the IT equipment, which is typically several degrees warmer after the CDU approach.
How to run a pilot
- Pick a representative workload and rack density.
- Agree on approved fluids and materials with the hardware vendor.
- Instrument power, temperatures, flow and fluid chemistry.
- Run long enough to see fluid trends, not just thermal results.
- Document the service procedures (swap, drain, refill) and time them.
Common misconceptions
- "Immersion fluid never needs maintenance." Suppliers specify monitoring and filtration. Follow their program.
- "Direct-to-chip removes all heat." Remaining components are usually still air-cooled.
- "Any PG25 will do." Hardware approvals name specific products.
See who makes data-center and car coolants
Browse suppliers →Bottom line
Choose by hardware support first, then facility constraints, then fluid program.
Operations: what changes for staff
| Task | Direct-to-chip | Immersion |
|---|---|---|
| Server swap | Disconnect quick-disconnects; standard rack work | Lift the server from the tank; drip and clean |
| Fluid checks | Sample the TCS loop | Sample tank fluid per supplier |
| Spill response | Glycol-water cleanup | Dielectric fluid cleanup per SDS |
| Training | Familiar to HVAC/IT teams | New procedures for most teams |
Heat reuse potential
Both approaches can return warm fluid to the facility loop. Higher return temperatures improve the case for reusing the heat, for example in district heating, which ASHRAE associates with its warmest W classes.
Decision summary
| If… | Lean toward |
|---|---|
| Standard racks, cold-plate servers available | Direct-to-chip |
| Very high density, tank-ready hardware | Immersion |
| Mixed estate | Hybrid approach |
More questions answered
Can I retrofit direct-to-chip into existing racks?
Often, with liquid-ready servers and a CDU, but check the hardware and facility constraints.
Reader questions
Is immersion noisier?
Immersion removes most server fans, which typically reduces noise.
Key point to remember
Pilot before committing at scale.
In practice
In practice, involve facilities, IT and fluid suppliers from day one.
Next step
Shortlist fluids in the comparator for your chosen approach.
One more tip
Ask vendors for reference sites.
Frequently asked questions
Is immersion cooling better than direct-to-chip?
Neither is universally better. It depends on density, hardware support and operations.
What fluid is used in immersion cooling?
Dielectric fluids, synthetic or hydrocarbon (single-phase), or engineered low-boiling fluids (two-phase).
Related guides
Sources
- Intel: Shell customer story
- Opteon 2P50 product page
- Engineered Fluids: ElectroCool
- Submer: about us
- ASHRAE TC9.9 white paper: Water-Cooled Servers: Common Designs, Components, and Processes
- ASHRAE TC9.9 white paper: Emergence and Expansion of Liquid Cooling in Mainstream Data Centers
- OCP: Guidelines for Using Propylene Glycol-Based Heat Transfer Fluids in Liquid Cooled Racks (Rev 2.0, Sep 2026)