Grades, Specifications, Standards & Installation — a 2026 Engineering Guide
XTD Special Steel · xtd-ss.com · Seamless 316L / 2205 tube to ASTM A269, A213, A312, A789/A790 and EN 10216-5
KEY TAKEAWAYS — CITABLE SUMMARY
- Rack power in AI halls has moved from 5–10 kW (2020) to 132 kW nominal (132–140 kW published operating band) for an NVIDIA GB300 NVL72, with VR200 NVL144 (72 packages) at 190–230 kW (2026–27) and Rubin Ultra “Kyber” at ~600 kW per rack (2027–28). Bare air cooling runs out of headroom at roughly 30–40 kW per rack; even with rear-door heat exchangers the practical ceiling is vendor-dependent, with the highest claims around 90 kW.
- Roughly 85–90% of rack heat is rejected to liquid and 10–15% to air in a GB300-class rack, so the secondary loop — not the air handler — is now the primary thermal path.
- Grade 316L (UNS S31603) seamless tube is the default wetted material for the technology cooling system (TCS) loop; duplex 2205 (UNS S32205) is specified where chlorides, pressure, or header size exceed 316L’s comfortable range.
- The three specification numbers that decide whether a tube is acceptable: inner-wall Ra ≤ 0.8 µm (0.4 µm for ultra-clean), hydrostatic test at 1.5 × design working pressure, and chloride limit < 50 ppm for 304L / < 500 ppm for 316L in water-based loops, tightening to < 25 ppm for either grade in a PG 25 loop (OCP 2022 water-based guideline and PG 25 specification).
- Material choice alone does not prevent failure. Galvanic isolation at copper cold plates, orbital TIG welding with internal argon purge, and end-capping until the moment of connection are the three field practices that separate a leak-free loop from a warranty claim.
1. Why liquid cooling became mandatory: the 2026 power envelope
Two numbers explain the shift. First, per-chip thermal design power: an H100 is rated around 700 W, a B200 at 1.0 kW, a B300 at ~1.1–1.2 kW in HGX form (about 1.4 kW per GPU in GB300 NVL72), and Rubin-generation packages are reported at 1.8 kW baseline and up to ~2.3 kW in maximum-performance configurations (supply-chain estimates, not yet datasheet figures). Second, rack-level aggregation: 72 GPU packages in a GB300 frame — which is already 144 dies, because each Blackwell Ultra package carries two; Rubin Ultra moves to four dies per package — plus CPUs, switch trays, HBM and power conversion, all inside one 48U envelope.
Figure 1 — AI rack power density, 2020 → 2028 (kW per rack, logarithmic scale). Rack thermal design power grows by roughly two orders of magnitude in eight years: from a 5–10 kW enterprise rack to the 132 kW nominal GB300 NVL72 shipping today (132–140 kW published operating band, 155 kW peak, 192 kW busway provisioned), then 190–230 kW for Rubin-generation VR200 racks, ~600 kW for the 2027–28 Rubin Ultra platform, and 1 MW-class racks from 2028. The shaded band marks the practical ceiling of air cooling — roughly 30–40 kW per rack. The 2023 Hopper-generation rack sits on the upper edge of that band, which is why every generation after it is liquid-cooled; once a hall crosses the band, every rack downstream depends on a pumped liquid loop. Note the flat segment between 2025 and 2026 — the industry consolidated around 120–135 kW rather than pushing density, and spent that period hardening the liquid infrastructure that the 600 kW and 1 MW generations require. For piping specification, the consequence is that flow rate, header bore and pressure class are now set by a heat load two orders of magnitude above what the same building services standard was written for in 2015.
Table 1 — Rack power and cooling mode by platform generation
| Generation | Example platform | Rack power | Cooling mode |
|---|---|---|---|
| Pre-AI (2020) | Enterprise / colocation rack | 5–10 kW | Air, hot/cold aisle |
| Hopper (2023) | HGX H100 rack (~4 × 8-GPU nodes; 10.2 kW per node) | ~30–40 kW | Air, or air + rear-door heat exchanger |
| Grace Hopper (2024) | GH200 cluster racks | No published rack figure (GPUs alone ≈32 kW) | Liquid (NVL32); air/liquid hybrids only in mixed cluster builds |
| Blackwell (2025) | GB200 NVL72 | ~120–135 kW | Direct-to-chip (DTC) liquid |
| Blackwell Ultra (2026) | GB300 NVL72 | 132 kW nominal (132–140 kW published band), ~155 kW peak, 192 kW busway provisioned | DTC, mandatory; no air-cooled variant; up to 45 °C loop inlet |
| Rubin (2026–27) | VR200 NVL144 (72 packages) | 190–230 kW; announced CPX variant estimated at ~370 kW (industry estimates) | DTC, facility-water coupled (warm-water class) |
| Rubin Ultra (2027–28) | NVL576 “Kyber” (144 packages) | ~600 kW | Liquid, facility-water coupled |
| 2028+ | 1 MW-class racks | ≥1,000 kW | Liquid, 800 VDC distribution |
Three engineering consequences follow directly from that table.
1.1 The heat split moved. In a GB300-class rack, roughly 85–90% of rack heat goes to liquid and 10–15% to air (OEM figures range from about 85/15 to 90/10). The stainless steel secondary loop is therefore no longer auxiliary infrastructure; it is the primary heat-rejection path, and its failure mode is a cluster outage rather than a hot aisle.
1.2 Power delivery changed the piping envelope. The industry is migrating from 48/54 VDC to 800 VDC distribution, because 1 MW at 54 V needs roughly 18,500 A of busbar. Higher bus voltages, tighter rack packaging, and rear-door elimination all push fluid routing into confined spaces where tube bend radius, ovality and connection height become design constraints, not afterthoughts.
1.3 CDU capacity scaled by an order of magnitude. Coolant distribution units now come in three deployment classes: in-rack up to ~250 kW, in-row up to ~1.8 MW serving up to eight racks (HPE rates its GB300 in-row CDU at 1.3 MW for eight racks; Vertiv rates the XDU1350 at 1.35 MW at a 4 °C approach, and CDU capacity is approach-temperature dependent, so nameplate MW figures are not directly comparable), and liquid-to-air sidecar units up to ~200 kW for halls with no facility-water loop. Each class has a different header size, flow rate and pressure regime — and therefore a different tube wall thickness and alloy requirement.
CITABLE STATEMENT
Air cooling stops scaling at roughly 30–40 kW per rack, and even rear-door heat exchangers top out at vendor-claimed levels of about 90 kW. At 132 kW nominal per rack — the published 132–140 kW band — roughly 85–90% of the heat load must be carried by a liquid loop, which makes the secondary piping a first-order reliability component rather than auxiliary infrastructure.
2. Where stainless steel tubing sits in a liquid-cooled AI hall
A modern AI cooling architecture has four piping zones, and stainless steel appears in all of them for different reasons.
2.1 Rack manifolds and header piping. The manifold distributes flow to individual server blades and must hold balanced flow across the 18 compute trays (plus 9 NVLink switch trays in an NVL72 rack) while fitting inside a 48U frame beside power shelves and blind-mate couplings. Seamless tube with tight OD tolerance is used because manifold fittings are brazed or orbitally welded at fixed pitch — dimensional drift compounds across dozens of joints.
2.2 CDU secondary-side (TCS) piping. This is the loop that touches cold plates: treated water or PG25 propylene-glycol coolant, up to 45 °C supply in warm-water designs, 2–4 bar typical operating pressure with pump-start surges well above it. It is also the loop where fluid chemistry is most tightly controlled, so the tube must not contribute ions, particles or corrosion products. Equipment vendors have reached the same conclusion at component level: Vertiv specifies an all-stainless secondary circuit for its XDU1350 — the clearest confirmation available that tube material is more than a cost line item.
2.3 Quick-disconnect (QD) and blind-mate sections. Server hot-swap imposes repeated mechanical cycling on the coupling body and the tube stub behind it. The Open Compute Project’s wetted-materials guidance is explicit on this point: “Quick connects made of stainless steel should be used. It is recommended to avoid nickel plated and chromium plated quick connects until these are proven satisfactory.” Worn plating exposes base metal and creates a galvanic couple at the worst possible location.
2.4 Facility-side and CDU primary loops. Larger bore, higher pressure, often coupled to dry coolers or cooling towers, and exposed to make-up water chemistry. This is where duplex 2205 or heavier-wall 316L is normally justified.
Table 2 — Why stainless steel beats the alternatives in a TCS loop
| Property | 316L stainless | Copper | CPVC / polymer | Aluminum |
|---|---|---|---|---|
| Compatible with DI water | Yes — low ion leaching, preserves resistivity | Leaches copper ions into DI loops | Yes | No — amphoteric, attacked below ~pH 4 and above ~pH 8.5 |
| Corrosion allowance (OCP coupon limit; PG 25) | ≤0.1 mpy (≈0.0025 mm/yr) | ≤0.2 mpy (≤0.1 mpy water-based) | Not applicable | Not on baseline list; needs added Al inhibitor |
| Particle / scale shedding | Very low with BA finish | Possible oxide release | Low, but permeable to some fluids | Oxide layer risk |
| Pressure rating at 45–66 °C | High; wall-thickness governed | Medium; softens above ~150 °C | Low; derates with temperature | Medium |
| Galvanic risk vs copper cold plate | Manageable — needs dielectric union | None (same metal) | None | Severe |
| Structural / vibration behaviour | Rigid, needs supports and expansion loops | Rigid, ductile | Requires dense supports | — |
| Recyclability / PFAS exposure | High / none | High / none | Some fluoropolymers raise PFAS review | High / none |
Copper remains common in cold plates and in some manifolds. With the loop chemistry properly maintained it can coexist galvanically with stainless steel, but dielectric isolation at every copper interface is still advised. Copper paired with aluminium — or copper in a deionized-water loop — is the combination most often behind the blockages this article is written to prevent.
3. Material selection: 316L, 2205, or something else
3.1 Grade 316L (UNS S31603) — the default. Molybdenum content of 2.0–3.0% gives materially better pitting and crevice resistance than 304/304L, and the low-carbon “L” grade keeps carbide precipitation out of weld heat-affected zones. The OCP baseline list reads “Stainless Steel 304L or 316L (preferred)”, and 316L is the practical default in data center practice because the chloride margin it buys is the difference between a 10-year loop and a pitting investigation. Specify it unless a documented reason exists not to.
3.2 Grade 304L (UNS S30403) — only at low chloride. 304L appears on the ASHRAE and OCP wetted-material lists, but its chloride tolerance is an order of magnitude lower than 316L. The OCP chloride limits are the clearest published guidance: < 50 ppm chloride for 304 stainless and < 500 ppm for 316 stainless in water-based loops; in a PG 25 loop the limit is < 25 ppm for either grade, with 25–50 ppm triggering investigation. A 304L loop that later gets topped up with higher-chloride make-up water is a predictable pitting failure.
3.3 Duplex 2205 (UNS S32205) — for aggressive or high-pressure service. More than double the yield strength of austenitic 316L (≈450 MPa against ≈170–200 MPa — roughly 2.2–2.6 ×), plus PREN around 34–35 versus ~25 for 316L. Justified when: chloride levels trend above the 316L allowance, header OD exceeds roughly 2 inches at elevated pressure, wall thickness must be reduced for weight or space, or the loop sees thermal cycling with vibration. Two caveats: duplex has a 300 °C upper service limit and is less forgiving in bending and welding, so fabrication qualification matters more; and 2205 is not on the OCP baseline wetted-material list, so specifying it is a documentation exercise as much as a metallurgical one — budget for project-specific approval against the fluid chemistry report.
Table 3 — Alloy selection matrix
| Condition | Recommended alloy | Reason |
|---|---|---|
| Treated water (< 50 ppm Cl) or PG 25 (< 25 ppm Cl), standard pressure class | 316L seamless | Baseline; best availability and weldability |
| Chloride 50–500 ppm (water-based) | 316L, verify with coupon test | Within OCP published allowance; monitor annually |
| Chloride > 500 ppm (water) or > 25 ppm (PG 25), or brackish make-up water | 2205 duplex (project-specific approval) | PREN margin; pitting resistance |
| Header OD > 2″, high pressure, weight-sensitive | 2205 duplex | Higher allowable stress → thinner wall |
| Direct connection to copper cold plate | 316L + dielectric insulating union | Breaks the galvanic couple |
| Deionized water, resistivity-critical | 316L, BA finish, no brass anywhere | Brass/copper leach ions into DI |
CITABLE STATEMENT
For AI data center coolant loops, 316L (UNS S31603) seamless tube is the default wetted material; duplex 2205 (UNS S32205) is specified when chloride levels exceed roughly 500 ppm in a water-based loop or 25 ppm in PG 25, when header diameter and pressure make wall thickness critical, or when thermal-cycling fatigue governs design.
4. Tube specification: standards, tolerances and documentation
Procurement failures in this category are rarely about alloy choice. They are about which standard the mill certified, and whether the certificate states the properties the loop design actually assumed.
Table 4 — Applicable tube standards
| Standard | Scope | Typical use in a data center loop |
|---|---|---|
| ASTM A269 | Seamless and welded austenitic tube, general service (by OD × wall) | Most common TCS / manifold specification |
| ASTM A213 | Seamless ferritic and austenitic boiler, superheater, heat-exchanger tube | Higher pressure CDU secondary, small-diameter coils |
| ASTM A312 | Seamless, welded and heavily cold worked austenitic pipe | Larger-bore facility headers |
| ASTM A249 | Welded austenitic boiler, superheater, heat-exchanger and condenser tube | Welded option where the application permits a seam |
| EN 10216-5 | Seamless stainless pressure tube (European) | EU-specified projects |
| EN 10217-7 | Welded stainless pressure tube | EU welded option |
| ISO 1127 | Dimensions, tolerances and conventional mass (D1–D4, T1–T5) | Tolerance regime referenced alongside EN specs |
| EN 10204 3.1 | Inspection certificate 3.1, issued by the manufacturer’s inspection representative | Typically required for mission-critical procurement |
4.1 Seamless versus welded. Seamless tube eliminates the longitudinal weld and with it the filler metal and heat-affected zone, so burst pressure is consistent around the full circumference. The trade-off is wall eccentricity from the piercing process: size the wall on minimum, not nominal, thickness. For cold-plate-adjacent manifolds and any section that will be orbitally welded in the field, seamless is the safer default. Welded-and-annealed tube is acceptable where the seam is verified by eddy current and the application is not pressure-critical — it is normally the lower-cost option, though the gap narrows once post-weld testing and documentation are included.
4.2 Tolerances that actually matter. Two tolerances drive assembly success: OD tolerance, which determines whether compression or face-seal fittings will hold, and wall eccentricity, which determines whether a nominal 16 bar rating is real at the thinnest point. ASTM A269 and ISO 1127 D4/T4 tolerances are commonly specified together with a requirement that eccentricity not exceed a stated percentage of nominal wall.
4.3 Documentation package. Require, at minimum: mill test certificate to EN 10204 3.1 with heat number traceability, chemical composition including Mo and C, mechanical properties, hydrostatic or non-destructive test record, surface roughness measurement report, cleanliness/passivation record, and a declaration of wetted-material compliance against the project’s approved list (ASHRAE/OCP-aligned).
5. Cleanliness and surface finish: the specification most often written loosely
Cold-plate microchannels in high-density AI modules now run at fin pitches measured in fractions of a millimetre. A particle that would be irrelevant in a building services pipe can occlude a channel and create a hot spot that throttles an accelerator.
5.1 Finish. Bright annealed (BA) is the preferred inner-wall condition: it is produced in a controlled hydrogen atmosphere, leaves no scale, and gives a smooth surface without the risk of residual pickling chemistry. Electropolished (EP) tube is used where the highest cleanliness and lowest Ra are required, at a cost premium.
5.2 Roughness. Specify Ra ≤ 0.8 µm as the standard requirement and Ra ≤ 0.4 µm for ultra-clean or microchannel-protection service. State the measurement method, the instrument, and the sampling interval — a claim of “Ra 0.4” without a measurement report is not a specification.
5.3 Passivation and cleaning. After forming and welding, parts should be cleaned and passivated per ASTM A967 (or A380 for descaling/cleaning) to remove free iron and restore the chromium-oxide passive layer. Residual free iron is the single most common cause of “flash rust” on stainless tube discovered weeks after installation.
5.4 End capping. Every tube must ship with both ends capped, and the caps must stay on until the instant of connection. This is a one-dollar control that prevents a five-figure flushing campaign.
6. Mechanical and leak testing
- Hydrostatic test at 1.5 × design working pressure, held per the governing piping standard, with the test medium and drying procedure stated (a wet test on a DI-water loop requires documented drying before capping).
- Eddy current testing per ASTM E426 for stainless tube, or ultrasonic testing per ASTM E213, for longitudinal and transverse defect detection.
- Helium leak testing for high-integrity joints and manifold assemblies — typically specified as a leak rate rather than a pass/fail bubble test. Suitable for pre-assembled rack manifolds where a field repair is not economically feasible.
- Dimensional and ovality inspection at the bend sections, not just on straight stock, because ovality introduced by bending is what causes fitting leaks in the field.
7. Welding and fabrication protocol
7.1 Orbital TIG, automatic, with internal purge. The OCP guidelines for single-phase cold plate loops require secondary pipework in stainless steel (minimum grade 304L) and inert-atmosphere welding on site, referencing AWS D10.18M/D10.18:2018. Note what that document is: D10.18 is the duplex stainless welding guide, not the austenitic one — the specification that governs austenitic tube welds is AWS D18.1/D18.1M:2020, and a fabrication specification should name both. Manual welding of thin-wall stainless in a confined rack envelope is not acceptable practice for a TCS loop. Automatic orbital welding with continuous internal argon purge keeps oxygen below the level at which chromium carbide and “sugar” (oxide scale) form on the inside of the joint. Specify a purge oxygen limit (commonly ≤ 50 ppm) and verify it with a meter reading recorded per weld.
7.2 Heat tint acceptance. Do not define a colour scale from scratch: AWS D18.2:2020, Guide to Weld Discoloration Levels on Inside of Austenitic Stainless Steel Tube, is the reference chart already used by tube mills and sanitary fabricators for internal and external weld zones, and citing it ends the argument about whose photograph is authoritative. Any tint beyond the accepted level requires pickling and re-passivation, because a discoloured zone is a depleted-chromium zone and therefore the initiation site for pitting in a chloride-bearing fluid.
7.3 Weld documentation. Weld map, welder/operator qualification, parameter log per joint, and purge records. For a rack manifold with 60+ joints, this documentation is what makes a leak root-cause analysis possible.
8. Installation, contamination control and commissioning
Installation is where a fully compliant material specification is most often destroyed.
8.1 Break the galvanic couple. Cold plates are usually copper; some are aluminium. Direct stainless-to-copper contact in a conductive fluid forms a couple — slower than aluminium-copper, but real over a 10-year life. Use dielectric insulating unions or electrochemically compatible transition fittings, and never connect aluminium wetted parts into a shared loop without a validated inhibitor package.
8.2 Hold the chloride budget. Use the limit that belongs to the fluid. Water-based loops: chloride < 50 ppm for 304L and < 500 ppm for 316L (OCP 2022 guidelines). PG 25 loops: chloride < 25 ppm, with 25–50 ppm triggering continued monitoring and ICP metals analysis, and anything above 50 ppm referred to the fluid supplier (OCP PG 25 V1.0.0). Carrying the 500 ppm water-based figure into a glycol loop is precisely how a compliant-looking 316L system goes on to pit. Alongside the chloride figure: pH 8.0–10.5 (7.0–10.5 for OAT-based inhibitor packages), PG 25 at 24.5–27.5 vol% as supplied and 24–29 vol% in service, with a copper-capable inhibitor package. Test make-up water before it enters the loop, not after.
8.3 Keep iron away from stainless. Carbon steel tools, grinding dust, chain-fall debris and swarf from site cutting are the source of most surface rust found at commissioning. Dedicated stainless tooling, segregated cutting areas, and a final wipe-down with chloride-free solvent are standard practice on high-purity jobs for exactly this reason.
8.4 Flush before you filter. OCP pre-commissioning practice is a recirculating flush through a temporary skid with 50–100 µm particulate filters, rack-manifold valves held closed so debris cannot enter the small-bore IT piping, and the flush continued until turbidity falls below 20 NTU. That 20 NTU is the commissioning endpoint, not the fluid specification: the fluid left in the loop must meet < 5 NTU (OCP v1.3), and PG 25 sets < 1 NTU for flush water. Final inline filtration must then be finer than the narrowest cold-plate channel it protects — commonly 25–50 µm absolute for microchannel plates, per the cold-plate OEM’s requirement — not sized against the pipe bore. Water introduced into the system at any point — make-up, dilution, top-up — is filtered to 25 µm absolute or better, which is the PG 25 recommendation for dilution water, and differential pressure across the filter housing is watched as the primary indicator of both saturation and filter bypass.
8.5 Design for movement. Pumps start and stop; fluid temperature swings 10–20 °C between standby and full load. Provide expansion loops, flexible hose sections at equipment interfaces, and rigid supports at intervals calculated against the tube’s natural frequency — not rule-of-thumb spacing.
Table 5 — Commissioning checklist
| Step | Acceptance criterion | Record |
|---|---|---|
| Visual / borescope | No heat tint beyond the accepted colour level; no scale or debris; caps intact until connection | Photo log per joint |
| Pressure test | 1.5 × design pressure, hold per standard, no decay | Test chart |
| Flush | Target velocity reached; effluent turbidity stable | Flush log |
| Fill & chemistry | pH, chloride, PG%, inhibitor concentration in range | Lab report |
| Particulate | Filter differential pressure stable; coupon inspection clean | Filter record |
| Leak detection | Sensor grid live at rack, manifold and CDU level | Commissioning sheet |
| Corrosion baseline | Coupon installed at commissioning; weight loss converted to mpy by the D8040 equation (22,300 × Δm / (area in in² × density × days)) | Baseline file |
9. Water chemistry: the monitoring parameters that predict failure
Corrosion in a TCS loop is a chemistry problem that shows up as a mechanical problem. Two OCP documents govern, and they are alternatives rather than a hierarchy: the guidelines for water-based transfer fluids in single-phase cold plate racks (2022) and the PG 25 base specification for propylene-glycol coolants. Both rest on ASTM D8040 coupon testing — fluid as-is, corrosion salts added, 88 ± 2 °C, 100 ml/min aeration, two weeks. Their limits differ, and the difference matters when a purchase specification is written: PG 25 V1.0.0 sets ≤ 0.2 mpy for copper and brass and ≤ 0.1 mpy for 304L/316L, while the 2026 water-based specification tightens every wetted metal — copper, brass, stainless and aluminium — to < 0.1 mpy. Quote the limit that belongs to the fluid in the loop.
Table 6 — Loop chemistry reference values
| Parameter | Reference range / limit | Why it matters |
|---|---|---|
| pH | 8.0–10.5 (7.0–10.5 for OAT packages) | Outside this band the passive film on stainless and the inhibitor package both degrade |
| Chloride | < 50 ppm (304L) / < 500 ppm (316L) in water; < 25 ppm in PG 25 | Chloride is the initiator of pitting in stainless |
| Propylene glycol | 24.5–27.5 vol% as supplied; 24–29 vol% in service (PG 25) | Freeze/burst protection and inhibitor carrier |
| Corrosion rate | ≤ 0.1 mpy stainless; ≤ 0.2 mpy Cu/brass (PG 25); < 0.1 mpy all metals (2026 water-based) | ASTM D8040 coupon, annual lab analysis |
| Total viable count | < 1 CFU/ml (R2A or Petrifilm) | Biofilm and under-deposit corrosion |
| TSS / TDS | < 5 ppm suspended solids / < 2,000 ppm dissolved | Solids lodge in microchannels; TDS drives scale |
| Total hardness | < 30 ppm as CaCO₃ | Scale on cold-plate surfaces |
| Reserve alkalinity | Trend, not absolute | Falling value = inhibitor being consumed |
| Dissolved metals | Trending upward = active corrosion | Copper → cold plates; zinc → brass dezincification |
| Turbidity | Stable; < 5 NTU (OCP v1.3); < 1 NTU for flush water (PG 25) | Rising = corrosion products or biological growth |
| Max fluid temperature | Wetted-material list validated to ≤ 66 °C; typical TCS < 49 °C | Above 66 °C the approved materials list must be re-reviewed |
| Conductivity / resistivity | < 2,000 µS/cm (OCP water-based); < 25 µS/cm for PG 25 make-up water; DI projects usually specify far tighter | Ion leaching from wrong alloys shows up here first |
Note the temperature line carefully: the OCP wetted-material list is validated for service up to 66 °C, and typical TCS operation is below 49 °C. Warm-water designs pushing 45 °C supply are comfortably inside that envelope — but a loop that later runs hotter needs its materials list re-reviewed, not just its setpoint changed.
10. FAQ
1) What grade of stainless steel tubing is used in AI data center liquid cooling?
316L (UNS S31603) seamless tube is the default for coolant-wetted piping because its molybdenum content resists pitting better than 304/304L, and 316L sits on the ASHRAE and OCP baseline wetted-material lists. Duplex 2205 (UNS S32205) is specified for higher-chloride water, larger headers or higher pressure; it is not on the OCP baseline list, so it needs project-specific approval and a documented corrosion justification.
2) Is stainless steel or copper better for a coolant distribution loop?
Neither is universally better. Copper is standard inside cold plates and conducts heat well, but it leaches ions into deionized water and must not be coupled directly to aluminium. Stainless steel is preferred for the secondary piping because it preserves fluid resistivity, resists pitting, and does not shed oxide into microchannels. Mixed loops need dielectric isolation.
3) What surface roughness should be specified for cooling-loop tubing?
Inner-wall Ra ≤ 0.8 µm for standard service, and Ra ≤ 0.4 µm where cold-plate microchannels are fine or particle sensitivity is high. Bright annealed (BA) finish is standard; specify the measurement method and require a roughness report with each lot.
4) Which ASTM standard applies to stainless tubing for liquid cooling?
ASTM A269 is the most commonly specified for general-service seamless and welded austenitic tube. ASTM A213 applies to heat-exchanger and higher-pressure service, and ASTM A312 to larger-bore pipe — but all three cover austenitic grades only; duplex 2205 goes to ASTM A789 (tube) or A790 (pipe). European projects typically reference EN 10216-5, which does list duplex grades, with ISO 1127 tolerances.
5) What chloride level is acceptable in a stainless coolant loop?
Published OCP guidance sets limits of < 50 ppm chloride for 304 stainless and < 500 ppm for 316 stainless in water-based loops, and < 25 ppm for either grade in a PG 25 loop. Design targets are often far tighter. Above the 316L allowance, duplex 2205 or a different loop chemistry should be evaluated.
6) How is stainless cooling tube leak tested?
Hydrostatic testing at 1.5 × maximum allowable working pressure is standard, combined with eddy current testing per ASTM E426 for defect detection. Pre-assembled manifolds and high-integrity joints are additionally helium leak-tested, and specified by leak rate rather than bubble observation.
7) Why do stainless cooling loops still corrode?
Usually chemistry, not the alloy: chloride ingress from make-up water, stagnant sections, depleted inhibitor, or a galvanic couple created by connecting stainless directly to copper or aluminium without dielectric isolation. Coupon testing plus annual lab analysis catches all four.
8) How much power does a liquid-cooled AI rack dissipate in 2026?
An NVIDIA GB300 NVL72 rack is rated at 132 kW nominal (132–140 kW published operating band) with a peak near 155 kW and 192 kW of busway provisioned. Rubin-generation VR200 NVL144 (72 packages) racks are specified at 190–230 kW, and the 2027–28 Rubin Ultra platform is projected at approximately 600 kW per rack.
11. Summary
Three decisions determine whether a liquid-cooled AI hall stays leak-free and blockage-free for its ten-year design life.
Metallurgy: 316L seamless as the default, duplex 2205 where chloride, pressure or header size exceed the austenitic envelope — and never aluminium in a shared loop.
Specification discipline: a named ASTM or EN standard, EN 10204 3.1 certification, Ra stated with a measurement report, 1.5 × design-pressure hydrostatic test, and a clean-and-passivate record. Vague language such as “high-quality stainless tube” is not a specification.
Field discipline: dielectric isolation at every copper interface, orbital TIG with a verified internal purge, end caps removed only at the moment of connection, iron-free tooling, and a documented flush-and-chemistry baseline before the first GPU is powered.
Specifications in this article were checked against published sources in November 2026, including ASTM A269/A213/A312, EN 10216-5, ISO 1127, ASTM A967, ASHRAE TC 9.9 thermal guidelines, the AWS D18.1, D10.18 and D18.2 welding standards, and OCP water-based and PG 25 coolant specifications. Rack power figures move on a roughly twelve-month cadence — re-verify platform numbers against current OEM QuickSpecs before they are written into a tender. Project-specific cooling architecture, fluid chemistry reports and OEM engineering guidelines always take precedence.
About the author / entity note. XTD Special Steel manufactures seamless stainless steel and duplex tube for thermal-management and high-purity service — 316L (UNS S31603) to ASTM A269 / A213 / A312, 2205 (UNS S32205) to ASTM A789 / A790, both available to EN 10216-5, plus nickel-alloy grades — with bright-annealed inner surfaces to Ra ≤ 0.4 µm, ASTM A967 passivation, hydrostatic and eddy-current testing, and EN 10204 3.1 documentation. Technical enquiries: huang.bin@xtd-ss.com.
References
A. Standards cited
[1] ASTM A269 — Seamless and Welded Austenitic Stainless Steel Tubing for General Service.
[2] ASTM A213 — Seamless Ferritic and Austenitic Alloy-Steel Boiler, Superheater, and Heat-Exchanger Tubes.
[3] ASTM A312 — Seamless, Welded, and Heavily Cold Worked Austenitic Stainless Steel Pipes.
[4] ASTM A249 — Welded Austenitic Steel Boiler, Superheater, Heat-Exchanger, and Condenser Tubes.
[5] ASTM A967 — Chemical Passivation Treatments for Stainless Steel Parts; ASTM A380 — Cleaning, Descaling, and Passivation of Stainless Steel Parts, Equipment, and Systems.
[6] ASTM E426 — Electromagnetic (Eddy-Current) Testing of Seamless and Welded Tubular Products, Austenitic Stainless Steel and Similar Alloys; ASTM E213 — Ultrasonic Testing of Metal Pipe and Tubing.
[7] EN 10216-5 — Seamless steel tubes for pressure purposes, Part 5: Stainless steel tubes; EN 10217-7 — Welded steel tubes for pressure purposes, Part 7: Stainless steel tubes.
[8] ISO 1127 — Stainless steel tubes: dimensions, tolerances and conventional masses per unit length; EN 10204 — Metallic products: types of inspection documents.
[9] AWS D18.1/D18.1M:2020 — Specification for Welding of Austenitic Stainless Steel Tube and Pipe Systems in Sanitary (Hygienic) Applications; AWS D10.18M/D10.18:2018 — Guide for Welding Ferritic/Austenitic Duplex Stainless Steel Piping and Tubing (the reference OCP pre-commissioning guidance cites for inert-atmosphere welding); AWS D18.2:2020 — Guide to Weld Discoloration Levels on Inside of Austenitic Stainless Steel Tube.
[10] ASHRAE TC 9.9 — liquid cooling thermal guidelines for data processing environments.
B. Coolant chemistry and wetted-material guidance
[11] Open Compute Project, Guidelines for Using Water-Based Transfer Fluids in Single-Phase Cold Plate-Based Liquid-Cooled Racks (2022) — chloride limits (< 50 ppm for 304SS, < 500 ppm for 316SS), TCS operating temperature (< 49 °C typical, not expected above 66 °C), and the baseline wetted-materials list. https://www.opencompute.org/documents/guidelines-for-using-water-based-transfer-fluids-in-single-phase-cold-plate-based-liquid-cooled-racks-final-pdf
[12] Open Compute Project, Base Specification for Single Phase Water-Based Cold Plate Coolants, v1.3 (March 2026). https://www.opencompute.org/documents/base-specification-for-single-phase-water-based-cold-plate-coolant-v1-3-final-pdf
[13] Open Compute Project, Base Specification for Single-Phase Propylene Glycol-Based Cold Plate Coolants (PG 25), V1.0.0. https://www.opencompute.org/wp-content/uploads/2026/10/OCP-PG25-Guidelines-V1.0.0.pdf
[14] Open Compute Project, Cooling Environments / Cold Plate sub-project (guidelines for water-based and propylene-glycol transfer fluids). https://www.opencompute.org/wiki/Cooling_Environments_Advanced_Cooling_Solutions_Cold_Plate
[15] Digital Refining, “Susceptibility of Type 304/304L and 316/316L austenitic stainless steels to chlorides in cooling water,” December 2022. https://digitalrefining.com/article/1002873/susceptibility-of-type-304-304l-and-316-316l-austenitic-stainless-steels-to-chlorides-in-cooling-water
[16] Intelligent Data Centres, “What to consider for data centre cooling in a shifting landscape” (316 stainless piping, 25 µm filtration, velocity guidance). https://intelligentdatacentres.com/?p=37922
C. Platform power and cooling-system data (OEM and industry sources, checked November 2026)
[17] Lenovo Press, Lenovo NVIDIA GB300 NVL72 Rack Scale AI Product Guide (135 kW TDP, ~155 kW peak, ~90/10 liquid/air). https://lenovopress.lenovo.com/lp2357-lenovo-nvidia-gb300-nvl72-rack-scale-ai
[18] Lenovo Press, GB300 NVL72 3D tour (118 kW liquid-cooled / 18 kW air-cooled). https://lenovopress.lenovo.com/3dtours/gb300nvl72
[19] Lenovo Press, GB300 NVL72 datasheet DS0207 (PG25 or DI water, 18–45 °C warm-water operation). https://lenovopress.lenovo.com/DS0207
[20] HPE, NVIDIA GB300 NVL72 by HPE (132–140 kW per rack, ~90/10 liquid/air). https://buy.hpe.com/be/en/compute/rack-scale-system/nvidia-nvl-system/p/1014890105
[21] Vertiv, SmartIT for NVIDIA GB300 NVL72 datasheet (up to 142 kW, ~85/15 liquid/air). https://www.vertiv.com/4a9fd7/globalassets/shared/vertiv-smartit-for-nvidia-gb300-nvl72_datasheet_englis_sl-80699.pdf
[22] GIGABYTE, NVIDIA GB300 NVL72 AI DLC POD (in-rack 200 kW and in-row 1,350 kW CDU). https://www.gigabyte.com/Enterprise/GIGAPOD-Pod-Scale/AI-DLC-POD_NVIDIA-GB300-NVL72
[23] Supermicro, SuperCluster GB300 NVL72 datasheet (250 kW in-rack CDU). https://gpusolutions.hrp.hu/wp-content/uploads/2026/05/datasheet_SuperCluster_GB300_NVL72.pdf
[24] StorageReview, “Supermicro NVIDIA Vera Rubin NVL72 racks now shipping with 1.8MW in-row CDUs.” https://www.storagereview.com/zh-CN/news/supermicro-nvidia-vera-rubin-nvl72-racks-now-shipping-with-1-8mw-in-row-cdus-and-a-1152-gpu-scalable-unit
[25] CTEE, report on VR200 rack Max Q / Max P power profiles (~1.8 kW / 190 kW and ~2.3 kW / 230 kW), 7 January 2026. https://www.ctee.com.tw/news/20260107701081-430704
[26] Hashrate Index, “NVIDIA Vera Rubin NVL72 specs breakdown.” https://hashrateindex.com/blog/nvidia-vera-rubin-nvl72-specs-breakdown/
[27] The Register, NVIDIA GTC 2026 preview (Kyber rack, 600 kW, 144 sockets with four Rubin Ultra dies each), 13 March 2026. https://theregister.com/2026/03/13/nvidia_gtc_2026_preview_tobias_mann_register
[28] VRLA Tech, “NVIDIA Vera Rubin Architecture Explained” (four compute dies per Rubin Ultra package). https://vrlatech.com/nvidia-vera-rubin-architecture-explained/
Rubin-generation and later figures are supply-chain and industry estimates, not final NVIDIA datasheet values. Re-verify against the current standard text and OEM documentation before use in a tender.