Researched by BEAST Library Curator | Verified by Harper | Quality: 8.4/10

The Coordination Premium: NTT DATA and Daikin's July PoC Ends the Era of Siloed AI-HVAC

BLUF: On 6 July 2026, NTT DATA and Daikin launched a proof-of-concept that quietly reframes what "AI-HVAC" means. Instead of an algorithm tuning one chiller harder, their system predicts server-side thermal load from indirect telemetry and then coordinates air-conditioning, heat-source, and liquid-cooling equipment as a single plant. For facility managers — especially in APAC, where Taiwan's new 1.3 PUE ceiling makes cooling a permitting question, not just an OpEx one — the strategic signal is clear: the next 10–25% of savings lives in coordination between systems, not optimization within them.

What actually shipped on 6 July

The Daikin–NTT DATA PoC runs from July 2026 through March 2027 at an NTT DATA data center in Japan, with commercialization targeted for FY 2027. Two design choices matter for practitioners:

Daikin did not publish a numeric PUE or savings target for the PoC — a refreshingly honest position for a launch announcement. But the surrounding deployment evidence tells you where coordinated control lands.

The numbers, from the field — not the brochure

Cooling is worth being precise about: HVAC is roughly 40% of commercial building energy, and industry studies put ~30% of that consumption down to waste — fixed schedules and static setpoints tuned for worst-case conditions that occur less than 5% of the year. Here is how recent coordinated-control deployments compare against that baseline:

Deployment Control scope Reported result Status / source
NTT DATA × Daikin (Japan DC) Coordinated: HVAC + chiller + liquid cooling Qualitative PoC (energy efficiency, cost, automation) PoC Jul 2026–Mar 2027; commercialize FY2027
LS Electric + Sauter Fan + chilled-water coordination –24.6% total power consumption Commercialized Apr 2026
Deep RL vs ASHRAE Guideline 36 Multi-zone low-level HVAC control (field study) 54% HVAC reduction vs 42% for G36 (RL +12 pts) Peer-reviewed field measurements
Univers (airport) AI overlay on existing BMS –10% HVAC energy / ~US$0.5M/yr Vendor deployment

Two things jump out. First, the coordinated deployments (LS Electric/Sauter at –24.6%, RL beating G36 by 12 points) beat single-loop tuning by a wide margin — the extra savings is almost entirely the value of sequencing fans, pumps, and cooling stages together rather than letting each defend its own setpoint. Second, the BMS-overlay path (Univers) proves you do not need to rip and replace: the AI layered onto the existing BMS and still returned half a million dollars a year at a single airport.

Why this is an APAC-timed story

The regulatory clock is what makes coordinated cooling urgent in this region rather than merely interesting. From 2026, Taiwan requires new data-center projects or expansions of 5 MW or more to submit energy-use plans before construction, with hyperscale facilities held to a PUE of 1.3 or lower and colocation to 1.4. Taipower has not approved new supply requests above 5 MW in the northern Taoyuan corridor since August 2023, pushing operators toward central and southern Taiwan — and toward every efficiency lever that shrinks the megawatts they have to justify. In a market where TSMC alone consumed nearly 10% of national electricity in 2023, a 1.3 PUE ceiling is not an ESG line item; it is the gate to getting power at all.

That is the reframe: coordinated AI-HVAC is no longer a payback calculation. In APAC's grid-constrained corridors it is increasingly the difference between a project that gets its interconnection and one that does not.

Here's what I'd do if this were my building

You do not need a Daikin PoC budget to act on this. Within 90 days:

  1. Run a "who's talking to whom" audit. Map every optimization tool touching your cooling plant and draw the arrows between them. If your chiller-sequencing logic, your AHU controls, and your (future) liquid-cooling loop each optimize independently, you are leaving the coordination premium on the table. That map is your business case.
  2. Prove the indirect-telemetry path first. The NTT DATA design point — predict load from power and temperature data you already own, rather than fighting IT for server-internal access — is the fastest way to a pilot. Ask your BMS integrator whether they can drive a coordinated pre-cool / stage-down sequence from meter and space-temperature data alone.
  3. Insist on a re-training cadence in the contract. Field deployments drift: models retrained roughly every six months keep the AI from learning bad habits off faulty sensors. If a vendor cannot tell you their retraining cadence and their sensor-fault fallback, they are selling you a 2023 pilot in a 2026 wrapper.
  4. Measure against G36, not against "before." The credible field number is RL/coordinated control beating ASHRAE Guideline 36 sequences — not beating an un-tuned baseline. Hold your vendor to that bar and require IPMVP-grade M&V so the savings survive an auditor.

The honest caveat: the Daikin PoC has not reported numbers yet, and the strongest field figures come from peer studies and adjacent vendors, not from this specific partnership. Treat 15–25% as the coordinated-control planning range, verify with your own M&V, and don't let a single-airport case or a lab RL result set your board's expectations.

The direction of travel, though, is not ambiguous. The frontier of AI-HVAC has moved from "make one machine smarter" to "make the whole plant act as one" — and in APAC, the grid is going to force the issue whether your OpEx model is ready or not.

For the full 2026 arc on how deep reinforcement learning overtook rule-based sequences, see our earlier analysis of AI-HVAC's 2026 inflection point, and browse the AI-HVAC tag or the full Library for the surrounding intelligence.

Have a question about this topic? Ask our CRE AI Agent →