Commissioning should therefore verify contact, flow, and per-chip load before relying on the rack's new thermal margin. If flow is low, fix flow first; if per-chip load exceeds the plate's 500–2000 W band, revisit plate selection rather than assuming the loop will compensate.
The evidence: where air stops working
Legacy enterprise data centers were designed for 3–5 kW per rack, so their air paths, raised floors, and CRAH sizing assumed a modest, predictable heat load. AI clusters now routinely exceed 20 kW and ultra-dense configurations exceed 50 kW per rack (Amaze, 2026-04-21). That is not a small overrun: it is several times the design basis, which is why the same air-cooled plant that was adequate for legacy storage racks becomes a thermal risk as object-storage nodes are upgraded.
Direct-to-chip fills the middle band of roughly 100–175 kW where rear-door heat exchangers fall short (Network World). The practical reading is that rear-door assistance does not extend the air-cooled range all the way to the highest densities; once a rack is in the gap between conventional air and the top density class, direct-to-chip is the technology that occupies that space.
The 30 kW tripwire is a decision rule set by this guide, not a measured air-cooling limit established by the cited sources. The sources place AI racks above 20 kW and ultra-dense configurations above 50 kW per rack (Amaze), and Network World puts direct-to-chip in the ~100–175 kW band, but none names a specific air-cooling ceiling for object-storage racks. Treat 30 kW as a conservative screening point, then confirm each rack with its own measured steady-state draw before acting.
For an object-storage rack, the test should be measured steady-state draw, not nameplate power. When that measurement exceeds 30 kW, the evidence supports converting the rack to direct-to-chip, provided the facility can supply 40–45 °C facility water or a CDU; at or below 30 kW, keeping air remains the supported choice.
The density bands above describe where air stops being reliable; they are not a construction guide, a cooling-option comparison, or a cost estimate. Use them as a screening rule, then confirm with the rack's own steady-state measurement.
Options compared: air vs DTC vs immersion
Air cooling is the lowest-capex option in this comparison and is described as workable below roughly 20 kW per rack (Amaze). No cited source establishes a measured reliability ceiling for air cooling in sustained object-storage duty; the 30 kW trigger is this guide's decision rule. Air therefore remains the default for lower-density racks, and a rack crossing 30 kW steady-state becomes a conversion candidate under the rule rather than a documented failure case.
Direct-to-chip cooling fills the middle band, handling roughly 100–175 kW per rack, which is the range where rear-door exchangers fall short (Network World). It is also retrofit-friendly for racks that must stay in place, but the retrofit depends on plant capability: it requires facility water at 40–45 °C or a CDU (Network World; ToneCooling). That water requirement is the main reason a rack-level conversion can stall even when the thermal case is clear.
| Option | Practical fit | Cost position | Binding constraint |
|---|---|---|---|
| Air | Below ~20 kW per rack | Lowest capex | Unreliable above 30 kW in sustained object-storage duty |
| Direct-to-chip | ~100–175 kW per rack | Retrofit-friendly | Needs 40–45 °C facility water or a CDU |
| Two-phase immersion | Rebuild scope, not rack-level retrofit | Lower 10-year TCO in present-day scenarios; higher capex | Higher capital cost than alternatives |
Two-phase immersion has the different cost profile. In present-day scenarios it showed a generally lower 10-year TCO, attributable to lower operating costs, while its capital costs were higher than those of the alternatives (Facilities Dive). That finding should be weighed against the higher upfront spend and rebuild scope; it is not by itself a reason to convert a rack that could be retrofitted with direct-to-chip. The cited finding is scenario-based, so it should be treated as a planning input rather than a universal ranking.
The comparison therefore turns on which constraint binds. Air wins on capex but loses above 30 kW; direct-to-chip wins on retrofit fit and density headroom but needs water or a CDU; two-phase immersion wins on 10-year TCO in present-day scenarios but loses on capital cost and rebuild scope. Apply the trigger directly: below 30 kW, keep air; above 30 kW with 40–45 °C water or a CDU available, convert to direct-to-chip; consider two-phase immersion only when a full rebuild is already planned.
Costs and numbers that matter
For object-storage racks, the retrofit decision should start with measured power, not nameplate ratings. The tripwire is a steady-state draw above 30 kW per rack: below it, air cooling remains the economical choice; above it, direct-to-chip is the in-place retrofit when the facility can supply 40–45 °C water or a CDU. Treat the 40–45 °C band as a supply condition, not a cost line—if it is unavailable, the retrofit path changes even when the rack exceeds 30 kW.
Capacity planning should use the 500–2000 W per-chip removal range cited for direct-to-chip cooling (ToneCooling) as a divisor, not as a promise. Divide a CDU loop's rated removal capacity by that range to get the cold-plate count envelope; the low end implies more plates per loop and the high end fewer. That calculation sizes a loop, but it is not a quote, because flow, coolant, and manifold limits are not specified in the available sources.
The cost-side opportunity that cooling comparisons often miss is heat reuse. Captured coolant heat can be reused for district heating and industrial applications, turning the data center from an energy sink into a supplier (LinkedIn, Data Centres Reduce Cost and Space). No per-rack revenue figure is given in the available sources, so any heat-sale value must be checked against local tariffs, interconnection costs, and actual delivered temperature.
To check your own numbers, log rack kW at the PDU for 7 days, then compare both the peak and the 95th percentile with the 30 kW tripwire. Sort the samples and take the value below which 95% of observations fall; if that value is at or above 30 kW, the rack is a steady-state exceedance. Use the peak to size transient headroom, not to declare a retrofit, and do not issue a retrofit quote from the log alone—the log establishes only whether the rack crosses the tripwire.
| Check | What to record | Decision |
|---|---|---|
| 7-day PDU log | Rack kW samples | Baseline only |
| Peak | Highest sample | Size transient headroom |
| 95th percentile | Value below 95% of samples | At or above 30 kW: convert; below 30 kW: keep air |
If a rack's 95th percentile is below 30 kW, keep air and revisit after the next hardware refresh or density change. If it is at or above 30 kW and 40–45 °C water or a CDU is available, convert that rack to direct-to-chip and carry the heat-reuse question into the same project scope.
What the evidence does NOT establish
No source in the available sources for this article gives a dollar-per-rack retrofit cost or a payback period for direct-to-chip conversion. Any figure you see quoted is therefore unverified, and it should be treated as a vendor or consultant estimate until it is tied to a written scope, utility impact, and commissioning plan.
The 100–175 kW direct-to-chip band is an AI compute figure from Network World, not an object-storage figure. Object-storage duty cycles are steadier and may sit lower, so that band should not be used to size, justify, or forecast an object-storage retrofit.
The reader rule also has a facility precondition: conversion above 30 kW assumes the site can supply 40–45 °C facility water or a CDU. A rack above 30 kW in a facility with no facility-water loop and no CDU cannot be converted without a plant upgrade; check for a CDU before committing.
What the evidence does establish is a thermal decision boundary, not project economics. It supports keeping air below 30 kW and converting above 30 kW when the water or CDU precondition is met, but it does not establish installation cost, downtime, or savings.
For planning, separate verified measurements from unverified claims: record measured steady-state rack draw, confirm CDU or 40–45 °C facility water, and require cost and payback figures to come from a site-specific quote rather than from the literature cited here.
If a proposal cites a per-rack cost or payback period, ask for the underlying assumptions—scope, labor, piping, controls, and utility work—because none of those are established by the sources in this article.
A 42U object-storage rack at 34 kW
Start with the measured load, not the nameplate. This rack is 42U, draws 34 kW steady-state at the PDU, is cooled by 3-year-old air-cooled CRACs, and has no facility water loop today. Record those four facts before any vendor discussion; they are the inputs for the decision.
Checkpoint 1 is density. The reader rule sets the tripwire at 30 kW measured steady-state. At 34 kW, this rack is above it, so air cooling is no longer the reliable path and the rack is a direct-to-chip candidate. The 4 kW margin is not a rounding error; treat it as a failed air-cooling gate rather than a future upgrade.
Checkpoint 2 is facility readiness. Direct-to-chip conversion is blocked until plant work is scoped because the site has neither a CDU nor a 40–45 °C facility water loop today. The rack can stay in place, but only after the cooling plant can deliver one of those two supply paths; until then, the decision is "convert after plant work," not "convert now."
Use the checklist below as a gate, not a wish list. A "no" on either checkpoint stops the project at that step; do not let a later step override an earlier failure.
| Step | Check | Pass condition | If it fails |
|---|---|---|---|
| 1 | Measure steady-state rack draw at the PDU | Record kW and compare with 30 kW | If 30 kW or less, keep air |
| 2 | Confirm the density tripwire | Greater than 30 kW | If 30 kW or less, keep air |
| 3 | Confirm facility supply | CDU or 40–45 °C water loop available or scoped | Block conversion and scope plant work |
| 4 | Confirm the rack stays in place | Retrofit path approved | Re-evaluate layout or immersion |
| 5 | Document the baseline | 42U, 34 kW, 3-year-old CRACs, no water loop | Do not proceed with an incomplete baseline |
For this specific rack, the result is a two-part answer: density says convert, facility says not yet. Keep the air-cooled CRACs running as interim cooling, scope the CDU or 40–45 °C loop, and re-run the checklist after plant work is defined. Do not treat the 34 kW reading as a reason to move the rack before the supply path exists.
Decision rules for the 30 kW tripwire
Use the 7-day peak rack draw, not a nameplate rating or a single spot reading, as the tripwire input. The canonical trigger is a measured steady-state draw above 30 kW per rack; the 7-day peak is the repeatable operational test for that condition. Compare the same rack, location, and workload mix across the full window, and keep the rack on air if the peak is at or below 30 kW.
If the 7-day peak exceeds 30 kW and the site already has a CDU or a 40–45 °C facility water loop, convert that rack to direct-to-chip. Confirm the loop or CDU has spare capacity for the added rack before scheduling the change; the cooling path must exist before the rack change is approved.
If the 7-day peak exceeds 30 kW but no CDU or suitable 40–45 °C facility water loop exists, scope the plant work first and do not order cold plates. Buying hardware ahead of the water path turns a cooling project into a stranded-asset risk, and the plant scope should include flow, temperature, and redundancy checks.
If the 7-day peak is 30 kW or less, keep air and re-check quarterly. Treat 30 kW as a moving threshold rather than a fixed rack property, because rising chip TDP can push a previously compliant rack over it between reviews; record the peak and workload in the capacity log each cycle.
If a rack crosses 30 kW only during a short test or benchmark window, record the event but do not convert on that evidence alone; require the 7-day peak to exceed 30 kW before changing cooling mode. This keeps one-off load spikes from driving permanent plant work.
| 7-day peak rack kW | Infrastructure available | Action |
|---|---|---|
| > 30 | CDU or 40–45 °C loop | Convert to direct-to-chip |
| > 30 | None | Scope plant work; no cold plates |
| ≤ 30 | Any | Keep air; re-check quarterly |
| Short test only | Any | Record; do not convert |
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Inventory object-storage racks and record each measured steady-state draw against the 30 kW threshold. | Identifies which racks are conversion candidates under the reader rule. |
| 2 | For racks above 30 kW, confirm the facility can supply 40–45 °C facility water or a CDU before converting. | Satisfies the required condition for direct-to-chip conversion. |
| 3 | Install direct-to-chip cold plates on processor dies, circulating coolant through internal micro-channels rated 500–2000 W per chip. | Delivers chip-level heat removal for converted racks. |
| 4 | Apply direct-to-chip in the ~100–175 kW density band where rear-door exchangers fall short, while keeping air cooling for racks at or below 30 kW. | Matches the cooling method to the density band and reader rule. |
| 5 | Before sign-off, confirm the 30 kW threshold, the 500–2000 W per-chip cold-plate range, and the 7-day / 95th-percentile logging method against your own rack measurements. | Prevents plan decisions from conflicting with figures already established in the guide. |
| 6 | Document which object-storage racks convert to direct-to-chip and which remain air-cooled. | Creates an auditable record of the 30 kW conversion decision. |
Frequently Asked Questions
What must be verified during commissioning before relying on a rack's new thermal margin?
Commissioning should verify contact, flow, and per-chip load before relying on the rack's new thermal margin.
What is the first corrective action if flow is low?
If flow is low, fix flow first.
What should be done if per-chip load exceeds the plate's 500–2000 W band?
If per-chip load exceeds the plate's 500–2000 W band, revisit plate selection rather than assuming the loop will compensate.
How much heat load were legacy enterprise data centers designed for per rack?
Legacy enterprise data centers were designed for 3–5 kW per rack.
What density gap does direct-to-chip cooling fill?
Direct-to-chip fills the middle band of roughly 100–175 kW where rear-door heat exchangers fall short.
Is the 30 kW threshold a measured air-cooling limit?
The 30 kW tripwire is a decision rule set by this guide, not a measured air-cooling limit established by the cited sources.
Quick answers
| What should commissioning verify before relying on the rack's new thermal margin? | Commissioning should verify contact, flow, and per-chip load before relying on the rack's new thermal margin. |
| What should be done first if flow is low? | If flow is low, fix flow first. |
| What should be done if per-chip load exceeds the plate's 500–2000 W band? | If per-chip load exceeds the plate's 500–2000 W band, revisit plate selection rather than assuming the loop will compensate. |
| Why does the same air-cooled plant become a thermal risk as object-storage nodes are upgraded? | AI clusters now routinely exceed 20 kW and ultra-dense configurations exceed 50 kW per rack, which is several times the legacy 3–5 kW design basis. |
| What is the 30 kW tripwire? | The 30 kW tripwire is a decision rule set by this guide, not a measured air-cooling limit established by the cited sources. |
Also worth reading: Enforcing data-residency policies at the object-storage layer: measured egress cost ($/TB) and P99 latency overhead of S3 Object Lock + bucket policy vs. gateway-side filtering across AWS, Azure Blob, and GCS: Enforcing data-residency policies at the · Object storage costs compared: 2026 500TB 12-drive Erasure Coding (EC:4) vs cloud: Object storage costs compared: 2026 · Object storage failover: Replication Time Control (RTC) 4,200 PUTs failover vs wait: Object storage failover: Replication Time