Thermal Sensing · AI Data Centers
Part of an ongoing series on temperature sensing across the physical-AI supply chain. This one covers the loudest line of all — and the components nobody counts.
In 2026, liquid cooling is the hottest word in AI infrastructure. Broker reports give a 94.2 billion RMB market; exhibitions queue up in Suzhou, Shanghai and the Greater Bay Area; NVIDIA's Rubin-generation volume production has pushed the thermal-vendor selection process into its final stretch.
But open the tables of those reports and you find a strange phenomenon: cold plates, CDUs and quick disconnects — the three components repeatedly dissected (some media calling them ~85% of system cost combined) — while the component that actually performs the "temperature measurement" action inside the liquid loop is almost never valued separately.
No vendor report prices the liquid-cooling temperature sensors as a line item; the filings of domestic listed sensor factories contain no design-win record for data-center liquid cooling. Not because the sensors don't matter — because they are small: small enough to be buried inside the BOM of cold plates and CDUs as a low-price, high-stickiness follower class.
This article opens that corner: where the temperature sensors sit in a liquid loop, how many there are, how the specifications are changing, who supplies, and where the gaps are for second-tier vendors.
Cold-plate level. After GPU power reached the 1,400 W class, the thermal-control window narrowed to single-digit degrees Celsius. Taiwan solution providers put it this way: a liquid loop's heat capacity is far smaller than air cooling's — a high-power GPU's temperature can spike within 5 seconds, so monitoring has to follow at millisecond scale. That forces the cold-plate sensing element into a trade between "accurate" and "fast": PT100 platinum RTDs are accurate (±0.1 °C class) but respond in seconds; thermocouples respond in milliseconds but with lower accuracy. Industry common sense so far — exact specs still need confirmation against OCP/ODCC source documents.
CDU level (the densest sensing point). The coolant distribution unit is the "heart" of the system and the densest place for sensors: primary-to-secondary temperature differential, flow, pressure, tank liquid level, coolant conductivity — all of it lives here. The CDU's internals are compact and sensors are typically embedded directly in the piping. Some media claim leading vendors' CDU flow control reaches ±0.5 L/min (not verified at announcement level). Anyone from industrial process control will recognize it instantly: a miniature chemical-plant control loop.
Primary side (outdoor section). Plate heat exchangers, dry coolers, cooling towers, pumps — the most traditional industrial thermal-control field there is. Armored platinum RTD and PT1000 processes apply directly, and certification barriers are far lower than inside the rack.
Leak detection (the second-largest sensing demand). Ethylene and propylene glycol solutions are conductive; a leak that wicks into a circuit board burns the equipment outright. Leak detection therefore moved from optional to an insurance-and-compliance topic. Engineering practice already shows distributed fiber-optic leak detection (localization at the ±0.5 m class) and "pressure-drop interlock power-off" interlock designs. Temperature sensing and leak sensing are now appearing in the same procurement package.
Shift one: 45 °C warm water plus propylene glycol rewrites corrosion-resistant packaging. The Rubin generation made warm-water loops standard, and long-term immersion in propylene glycol solution demands corrosion-resistant packaging for the sensing elements — "10,000-mate corrosion-resistant propylene-glycol compatible" is already spec language at the quick-disconnect layer, and the sensor layer gets no exemption. For temperature-sensor factories with armored, packaging and corrosion-resistance craft, this is a direct capability mapping.
Shift two: from "temperature measurement" to "millisecond-level thermal monitoring." In the air-cooled era temperature was a slow variable; in liquid cooling the heat capacity is small and runaway is fast, so response time and deployment density (point sensing vs distributed fiber) are being re-weighted. Hybrid architectures (cold plate with retained air cooling) mean both air-path sensing arrays and liquid-path sensing coexist in one rack — the total count stacks.
Shift three: leak detection and temperature sensing fuse. Detection ropes, fiber, point leak sensors and temperature sensors are now co-procured through the same channel. For sensor vendors this may be a bigger crack than the temperature sensor itself — though the technology-route fight (discrete vs integrated) has not been settled.
International majors already run "AI data-center thermometry" as a standalone SKU: TDK published a technical document on NTC temperature sensors for AI data-center thermal management in July 2025, covering liquid loops, ambient monitoring and in-rack measurement with ordering codes attached; Sensata's site explicitly targets liquid-level and flow sensors at cooling and thermal-management applications.
On the domestic (China) side: the public materials of the top-four thermal-control vendors (Envicool, Shenling, Highlan, Tongfei) contain no disclosure of sensor procurement details; the filings of domestic listed sensor factories show no design-win record for data-center liquid cooling. Anrun (Apex Dragons) wrote "temperature sensors for battery cooling plates" in its 2024 annual report — that is the automotive path; the data-center equivalent has not appeared.
There are two readings of this blank: either the thermal vendors don't disclose (procurement is opaque by nature), or domestic sensor makers genuinely haven't entered. Either way it means the same thing for second-tier vendors: the design-win competition here has not started its involution yet.
Worth noting as well, the cross-over player: Tuopu Group entered the ODM chain directly as an automotive Tier 1 and took 1.5 billion RMB of orders across data-center, storage and robotics liquid cooling in 2025. The automotive temperature-sensor market was ~5.62 billion RMB in 2025 with slowing growth — automotive-grade capacity hunting a second curve and data-center demand are landing in the same window.
Path judgment: liquid-cooling procurement is tiered and dual-driven — entry is decided by certification and specifications (MGX ecosystem certification, platform Tier-1 designation, OCP standardization), and price only matters after volume. The sensor itself is mostly bundled and sold with the cold plate, CDU or quick disconnect; pitching sensor prices standalone has no persuasive power. The correct posture is to bundle into a component solution and enter alongside a thermal vendor or connector maker.
While writing this we lined up every liquid-cooling market size we could find — and they differ absurdly: a broker deep-dive says 94.2 billion RMB in 2026 (full-chain scope); CCID says China cold-plate liquid cooling is 21.62 billion RMB in 2026 (infrastructure thermal-control scope); overseas houses give a global figure of only $0.5–0.6B class. The 94.2 and the 21.6 differ by more than 4×, because three scopes are being mixed: thermal-control equipment / servers-with-liquid-cooling-as-a-whole / total data-center heat dissipation.
For judgment purposes, anchor on the thermal-control-equipment scope (CCID's 21.6–43.7 billion RMB range) — that is the market actually reachable by thermal vendors and their sensor suppliers. Before citing any market-size claim, ask about the scope first. Our comparison table maintains this continuously.
The blank described above cuts both ways: buyers evaluating Chinese thermal and sensor suppliers need the factory verified before prepayment. See Supplier Verification Services — remote desktop verification, live video audits, and a bilingual GO / CAUTION / NO-GO scorecard, conducted by a 15-year industry veteran and CCAA-certified ISO 9001 auditor.
About the author: 15 years in the temperature-sensor industry; CCAA-certified ISO 9001 quality-management-system auditor.
Scope-flagged figures (94.2B / 21.62B / $0.5–0.6B) are quoted from the named public reports with their base years; some vendor details are marked pending announcement-level verification in the source material. Related in this series: Fiber-Optic Sensing Walks Out of the Oilfield · Thermometry in Dilution Refrigerators.