This photo taken on March 3, 2026 shows the Samsung SDI battery factory of South Korean industrial giant Samsung.
Attila KISBENEDEK/AFP via Getty Images
Samsung Electronics signed an agreement to build a 240 billion won — approximately $170 million at current exchange rates, versus the $158 million cited in Samsung’s own official conversion — HVAC production line in Gwangju, South Korea today, the largest manufacturing commitment at that campus in 37 years, planting its FläktGroup cooling subsidiary in the one country where the demand for coolant distribution units is being created fastest. The agreement, signed at Gwangju City Hall with FläktGroup Korea and the Jeonnam-Gwangju Special Metropolitan City government, turns a formerly appliance-focused campus into Samsung’s global anchor factory for the piece of hardware that makes AI data centers thermally viable: the Coolant Distribution Unit.
What CDUs Actually Do — and Why AI Made Them Mandatory
A Coolant Distribution Unit is not a conventional air conditioner. It does not cool a data hall. It cools individual chips.
In a direct-to-chip liquid cooling architecture, a CDU sits between the facility’s main chilled water supply and the IT hardware’s own cooling loop — the tubes and cold plates that run directly to each GPU and CPU package. This separation is not incidental: facility water contains dissolved minerals, biological organisms, and chemicals that would corrode the precision copper cold plates mounted on billion-dollar processors. The CDU provides thermal isolation, filtering and conditioning a separate water-glycol mixture that circulates through the IT loop, while a heat exchanger transfers the chip heat outward to the facility’s chilled water without the two fluids ever mixing. Vertiv’s technical documentation describes how CDUs create an isolated secondary cooling loop that prevents mineral and biological contamination of precision cold plates.
At the heart of the loop, a cold plate — a flat metal block with micro-channels machined through it — bolts directly onto the GPU or CPU package. Coolant flows through those micro-channels at a precise flow rate and temperature setpoint managed by the CDU, absorbing heat through direct conduction at the chip surface, where it can be removed most efficiently. The warmed coolant then routes through the CDU’s heat exchanger, releases heat to the facility loop, chills back down, and recirculates. The CDU also monitors pressure, detects leaks, and manages flow distribution across every rack row it serves.
The reason this architecture has become mandatory is physics. NVIDIA’s Blackwell B200 GPU draws approximately 1,200 watts in liquid-cooled configurations — roughly 71% more than the H100 SXM’s 700 watts. Packed into the GB200 NVL72 chassis (72 GPUs and 36 Grace CPUs in a single rack), that translates to 120–132 kilowatts per rack at full load. Air cooling works adequately below about 30–50 kW per rack; the GB200 NVL72 requires roughly three to four times that. The Rubin generation, targeted for 2027, is projected to reach 250–900 kW per rack. At those densities, the volume of chilled air required to dissipate the heat would be physically impractical — liquid carries approximately 3,500 times more heat per unit volume than air at equivalent flow rates. Liquid cooling is not optional for GB200 NVL72 racks; NVIDIA designed the B200 to require it by default, not as an option.
Cheolgi Kim, Executive Vice President and Head of Digital Appliances at Samsung Electronics, placed the Gwangju commitment squarely in this context: “Samsung’s acquisition of FläktGroup last year puts us in a unique position to expand the HVAC business and transform it into a new growth engine, to become a global top-tier company by 2030.”
A 37-Year Campus Record — and What the Scale Signals
The Gwangju campus has manufactured Samsung consumer appliances — Bespoke refrigerators, washers, dryers, and air conditioners — since Samsung established it in 1989. The new HVAC production line, spanning 21,800 square meters (approximately 234,700 square feet, or roughly the area of three football fields) at Campus 3 in Buk-gu, is the first major production facility investment at the site in those 37 years.
That is not a routine capital expenditure. That is a board-level decision to reposition an existing campus around a new industrial priority — one that Samsung is signaling it expects to grow faster than its legacy appliance business. At 240 billion won, the Gwangju commitment is roughly 24 times larger than the €5.7 million (approximately $6.6 million) Samsung and FläktGroup invested in the Pune, India facility that opened on August 12. India’s Pune plant, which produces up to 6,500 HVAC units annually, was the first production line under the FläktGroup acquisition. Gwangju is something different in kind: a manufacturing hub investment at Samsung’s home base.
The new factory will become FläktGroup’s 15th global manufacturing facility, joining an existing network of 14 production sites across the U.S., Europe, and India that serves customers in 65 countries. FläktGroup CEO David Dorney stated Samsung’s geographic calculus plainly in the official Gwangju announcement: “By combining Samsung’s AI expertise with FläktGroup’s HVAC product capabilities, we will be positioned to become an HVAC partner of choice in Korea and across Asia.”
Full operations are scheduled to begin in early 2028, a timeline the company says aligns with the wave of new AI campuses completing construction across South Korea.
Building CDUs in the Country That Needs Them Most
South Korea is not a routine data center market. It is arguably the most compressed AI infrastructure buildout in the world.
The national government has committed to 8.4 gigawatts of AI capacity by 2029, backed by roughly 550 trillion won in public and private investment. The AIDC Alliance — a 12-company public-private coalition launched on July 29, 2026 — is targeting 18.4 GW by 2035, backed by 200 trillion won in state financing. Samsung Electronics itself committed to a group-wide investment of 2,450 trillion won across all businesses from 2026 to 2040, with a semiconductor cluster in Gwangju as its anchor.
Jeollanam-do Province — where Gwangju sits — is running the fastest capacity growth in the country, at a 23.05% compound annual growth rate on new data center additions. The overall South Korean data center construction market was valued at approximately $7 billion in 2026 and is projected to reach $14.6 billion by 2031.
A domestic CDU manufacturer, operational by 2028, means Korean data center operators building AI campuses right now can plan cooling procurement with shorter lead times and without the logistics costs of importing from European or American competitors. That localization advantage matters in a market that is building fast enough that cooling delivery schedules can become a genuine bottleneck.
Samsung has explicitly stated that the Gwangju line is a “strategic springboard for a full-scale expansion into the Korean market” — with HVAC supply targeted across multiple industrial sectors, AI data centers foremost among them.
The Stack No Competitor Can Replicate
The investment’s deepest significance is not the factory itself. It is what the factory completes.
Samsung’s FläktGroup acquisition brought it CDU manufacturing capability. Its semiconductor division produces HBM4 — the high-bandwidth memory that populates every AI accelerator. Its DRAM division supplies the standard memory that AI servers require. Its building management platforms, SmartThings Pro and b.IoT, provide the software layer for data center energy efficiency and climate control integration.
No competitor in the CDU or liquid cooling space — not Vertiv, which led the data center liquid cooling market in 2025 with over 11% market share; not Schneider Electric, which acquired CDU manufacturer Motivair in early 2025; not LiquidStack, CoolIT, or Alfa Laval — also manufactures the AI accelerator memory their cooling systems are designed to protect. Vertiv can sell a CDU. Samsung can sell the CDU, the HBM inside the GPU the CDU cools, the server chassis, and the software that manages the building the server lives in. That is a vertical integration position in AI data center infrastructure that no HVAC or cooling specialist can replicate.
The CDU market itself is growing fast enough that Samsung does not need to displace Vertiv or Schneider to build a large business: Grand View Research pegged the global CDU market at $1.6 billion in 2025, projected to reach $7.1 billion by 2033 at a 21.4% compound annual growth rate. The Asia-Pacific region led CDU market share at 37.9% in 2025 — meaning the market Samsung is building domestic manufacturing for is already the world’s largest by geography.
The question Samsung still has to answer is whether manufacturing scale and platform integration translate into winning enterprise contracts in a market where Vertiv and Schneider Electric have decade-long relationships with data center operators globally. FläktGroup’s existing hyperscale client base in Europe and Asia gives it a foot in those conversations. The Gwangju factory gives it the domestic delivery network to back them up.
Who Makes the Industry’s Cooling Hardware — and Why Samsung Is Entering
The global data center liquid cooling market was valued at approximately $4.8–6.7 billion in 2025, depending on scope of coverage, and is forecast to reach $18–29 billion by 2031–2035 at a compound annual growth rate of 18–31%. The range in projections reflects how quickly actual deployment is exceeding earlier forecasts.
Within that market, Vertiv holds the largest single vendor share at approximately 11% of the broader liquid cooling market (its CoolChip CDU family, launched at SC24 in November 2024, spans 70–2,300 kW capacity). Schneider Electric acquired CDU specialist Motivair in early 2025, securing cold plate technology and CDU manufacturing capability, and holds a reference design partnership with NVIDIA specifically for GB200 NVL72 systems. Specialized competitors include LiquidStack, which launched its 10 MW GigaModular CDU in June 2025, and CoolIT Systems (direct-to-chip cold plate designs).
FläktGroup’s HVAC manufacturing heritage — more than a century of engineering expertise in air handling, ventilation, and precision cooling for pharmaceutical, hospital, marine, and data center customers — is what Samsung paid €1.5 billion (approximately $1.74 billion at current exchange rates) to acquire in November 2025. The Gwangju factory begins translating that heritage into domestic Korean production, scheduled for early 2028.
How to Think About the 2028 Delivery Date
The timing is deliberate. Large-scale AI campuses in South Korea are completing planning and permitting cycles that began in 2025–2026, meaning construction and equipment installation timelines converge in 2028–2029. A CDU manufacturer operational in early 2028 can bid on cooling infrastructure for campuses that are already under development — not ones still on drawing boards.
The broader window Samsung is targeting is real and measurable. South Korea’s government-designated data center capacity target implies completing more than half of its planned 8.4 GW by early in the next decade. That volume of AI-grade facilities — each requiring direct liquid cooling for GPU rack densities that air cooling cannot serve — creates a CDU procurement pipeline that will need filling for years. Domestically manufactured units reduce delivery risk for operators who cannot afford cooling delays to stop GPU deployments.
The Gwangju production line, once operational, will also serve Asian markets beyond Korea — consistent with FläktGroup CEO Dorney’s stated “Korea and across Asia” positioning. A Korea-based hub reduces shipping distances and times for AI data center projects across Japan, Taiwan, Singapore, and Southeast Asia.
Frequently Asked QuestionsWhat is a Coolant Distribution Unit, and why does every new AI data center need one?
A Coolant Distribution Unit is the thermal isolation and flow management system that separates a building’s chilled water supply from the cooling loop running directly to GPU and CPU chip packages. It pumps a water-glycol mixture through cold plates bolted onto chips, absorbs heat at the source, and routes the warmed coolant through a heat exchanger — keeping facility water and IT equipment’s cooling fluid separate to prevent mineral and biological contamination of precision hardware. CDUs became mandatory for AI data centers because modern GPU racks draw 120 kilowatts or more per rack — three to four times what air cooling can handle — and the 2027-generation hardware is projected to reach 250–900 kW. At those densities, liquid cooling is not an upgrade; it is the only architecture that physically works.
Why does it matter that Samsung is building the factory in South Korea specifically?
South Korea is running the most compressed AI data center buildout in the world, with a government target of 8.4 gigawatts of AI data center capacity and an AIDC Alliance of 12 companies targeting 18.4 GW by 2035. AI campuses completing construction in 2028–2029 will need CDUs domestically sourced to avoid supply chain delays. Samsung’s Gwangju factory, operational in early 2028, positions it as the local supplier at the exact moment the Korean market’s procurement pipeline peaks. Domestic manufacturing also means shorter lead times, lower logistics costs, and local content advantages in government or state-adjacent procurement.
How does this give Samsung an advantage that Vertiv or Schneider Electric cannot easily replicate?
Vertiv and Schneider Electric are the dominant global CDU vendors, but neither manufactures AI accelerator memory. Samsung does — its HBM4 memory populates virtually every high-end GPU on the market. The Gwangju factory extends Samsung’s position into the cooling hardware that protects the chips its memory division ships, alongside the servers those chips run in, and the building management software (SmartThings Pro, b.IoT) that controls the facilities they occupy. No HVAC or cooling competitor can replicate that vertical integration without separately acquiring semiconductor memory manufacturing — a barrier that is practically prohibitive. The moat is not CDU expertise alone; it is the only company that sells AI memory, AI servers, and AI cooling infrastructure all made domestically in Korea.
When will the Gwangju factory be producing CDUs, and how does that timeline compare to competitors?
Full operations at Gwangju are scheduled for early 2028, making it Samsung’s second FläktGroup production line in Asia after the Pune, India facility that opened August 12, 2026. Existing competitors — Vertiv, Schneider Electric, LiquidStack — already have CDU manufacturing capacity operational globally, meaning Samsung enters this market as a new production-scale entrant. Samsung’s advantage is not first-mover status in manufacturing; it is home-market position in the world’s fastest-growing AI data center geography, combined with the vertical integration advantage described above. Whether that translates into contract wins against entrenched competitors with decade-long data center relationships is the operational test that follows the groundbreaking.