Originally published at twarx.com - read the full interactive version there.
Last Updated: June 22, 2026
NVIDIA just made the cold data center obsolete — and this AI technology breakthrough's counterintuitive fix runs coolant hotter than a hot tub.
On June 21, 2026, NVIDIA detailed how its Rubin generation AI technology infrastructure became the world's first to achieve 100% liquid cooling at coolant temperatures up to 45°C (113°F). This matters now because cooling consumes up to 40% of a data center's electricity, and Rubin's design eliminates nearly all of it — plus almost all water. After reading, you'll understand exactly how the system works, what it costs, and the deeper systems lesson it reveals about coordination.
Key Facts
Announced: June 21, 2026, by Josh Parker on the NVIDIA Blog.
Platform: NVIDIA Rubin — first AI infrastructure with 100% liquid cooling, zero fans.
Coolant temperature: Up to 45°C (113°F) intake, exiting near 55°C.
Water impact: Up to 100% reduction vs ~2.6M gallons/MW/year for cooling towers.
Partner: Motivair, the cooling division of Schneider Electric.
NVIDIA's 45°C liquid-cooling architecture for the Rubin platform — the first AI infrastructure with 100% liquid cooling and zero fans. Source: NVIDIA Blog
Most AI infrastructure decisions are solving the wrong problem. Engineers obsess over chip performance per watt while the silent 40% tax sits in the mechanical room, untouched. NVIDIA's announcement is fundamentally a coordination story — the win didn't come from a faster chip. It came from coordinating the chip, the coolant, the rack, and the building into one closed loop. When I spec'd my first GPU cluster build, I made the same rookie mistake everyone makes: I argued about accelerator SKUs for three weeks and never once asked the facilities team what the loop temperature was. That conversation, not the chip, decided whether the project penciled out.
Coined Framework
The AI Coordination Gap
The AI Coordination Gap is the systemic performance loss that emerges when individually optimized components — chips, cooling, agents, retrieval steps — are never coordinated as one closed system. It is the reason a stack full of excellent parts still underperforms, and the reason NVIDIA's biggest efficiency leap came from architecture, not silicon.
What did NVIDIA actually announce on June 21, 2026?
On June 21, 2026, NVIDIA published a blog post authored by Josh Parker titled 'Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI's Biggest Machines.' Three confirmed facts underpin everything else:
What: The NVIDIA Rubin platform is 'the world's first to achieve 100% liquid cooling — every chip, every networking component, cooled entirely by liquid in a closed loop with no fans anywhere in the system,' per the official post.
How hot: The coolant runs at up to 45°C (113°F) — hotter than a hot tub, which sits at 38–40°C. The methodology is documented in the NVIDIA DSX AI factory reference design.
The payoff: The DSX reference design has 'zero water consumption,' eliminating massive power usage and 'pretty much all water usage,' according to Ali Heydari, director of data center cooling and infrastructure at NVIDIA.
The ecosystem partner named is Motivair, the advanced cooling division of Schneider Electric, whose president and CEO Richard Whitmore confirmed nearly a decade of collaboration with NVIDIA's product roadmap. The direction of travel is independently corroborated: warm-water cooling has been an explicit efficiency lever in ASHRAE's Datacom thermal guidelines for years, and macro data center energy pressure is documented by the International Energy Agency and the U.S. Department of Energy. The Uptime Institute has likewise tracked the steady relaxation of supply-air and supply-water temperature setpoints across operators.
40%
Share of data center electricity historically used by cooling alone
[NVIDIA Blog, 2026](https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/)
$4M+
Annual savings for a 50MW facility moving to liquid cooling (NVIDIA-published)
[NVIDIA Blog, 2026](https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/)
2.6M gal
Water per MW per year for conventional cooling — cut up to 100%
[NVIDIA Blog, 2026](https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/)
The silent 40% tax in the mechanical room, untouched — that is the number every AI technology roadmap forgot to budget for, and the one NVIDIA just deleted.
What is NVIDIA's 45°C liquid cooling in plain English?
Picture your gaming PC. Fans scream when it works hard, blowing cold air over hot chips. Now scale that to a building crammed with thousands of the most power-hungry AI chips ever made — the fans alone hit 85 decibels, loud enough to require ear protection, per NVIDIA.
Rubin throws all of that out. Instead of blowing cold air, it runs a liquid — 75% water and 25% propylene glycol — directly across metal 'cold plates' sitting on top of each chip. The liquid grabs heat at the exact source, then carries it away in a sealed loop. No fans. No 'cold aisles.' No walk-in-freezer server room. The thermodynamic logic behind this is well documented in ASHRAE's datacom thermal guidelines.
The counterintuitive part is the inlet temperature. The liquid going in is already warm — 45°C — and comes out at roughly 55°C, having absorbed the chip's heat. Because the loop runs hot, outdoor 'dry coolers' can dump that heat into ordinary summer air without needing energy-hungry chillers for most of the year — 'outside of maybe 1% of the year when we might need chillers in some climates,' said Heydari.
The industry's instinct that 'a cold data center is an efficient one' is exactly backwards. NVIDIA validated chips running at full performance with coolant entering the rack at 45°C — warm summer air is fine, because nothing in the server depends on cool air anymore.
Before-and-after: traditional air cooling depends on chilled air and noisy fans, while NVIDIA's closed-loop liquid design captures heat at the chip — illustrating how solving the AI Coordination Gap removes an entire subsystem.
How does NVIDIA's 45°C liquid cooling work step by step?
In an AI factory, coolant flows from a coolant distribution unit (CDU) to the servers in a closed-loop cycle. Here's the full path:
NVIDIA Rubin 45°C Closed-Loop Liquid Cooling Flow
1
**Coolant Distribution Unit (CDU)**
Pumps a 75% water / 25% propylene glycol mix into the rack at 45°C. No new water is consumed — the same fluid recirculates indefinitely.
↓
2
**Cold Plates on Every Chip**
Liquid flows through plates sitting directly on processors and networking components, absorbing heat at the source. Device temps stay within validated limits — full performance, no throttling.
↓
3
**Heat Pickup (45°C → 55°C)**
Coolant exits roughly 10°C hotter, carrying the full chip heat load out of the server with zero fans involved anywhere in the system.
↓
4
**Outdoor Dry Coolers**
Because the loop is so hot, dry coolers reject heat to ambient air for most of the year — no chillers, no evaporative water, no noise.
↓
5
**Closed-Loop Return**
Cooled fluid returns to the CDU. Net result: up to 100% water reduction and the elimination of cooling's historic 40% energy share.
The sequence matters because every stage is co-designed: raise the loop temperature and the chiller — the single biggest energy and water hog — simply disappears.
Raise chiller plant temperatures by just one degree and cooling energy costs drop by about 4% — that single rule, which NVIDIA cites and which thermal engineers have leaned on for a decade, is the whole game. Rubin doesn't raise it by one degree. It raises the entire operating envelope to 45°C, then removes the chiller from most of the year entirely.
The most important number in AI technology infrastructure this year isn't FLOPS or tokens-per-second. It's 45°C — the temperature at which cooling stops being a tax and starts being free.
What can NVIDIA 45°C liquid cooling do — the full capability list?
100% liquid cooling: Every chip and networking component, no fans anywhere — an industry first per NVIDIA.
45°C (113°F) coolant intake: Validated full performance with coolant entering the rack hotter than a hot tub.
Zero water consumption in the DSX reference design — 'up to a 100% reduction in water use' versus the ~2.6M gallons/MW/year of cooling-tower systems.
Chiller-less operation in favorable climates, with chillers needed 'maybe 1% of the year.'
$4M+ annual savings (NVIDIA-published) for a 50MW hyperscale facility in cooling-related energy and water.
Eliminates the 85 dB fan noise floor — no ear protection required.
Documented in the NVIDIA DSX AI factory reference design — a full guide to design, build, and operate the AI factory infrastructure stack.
'Once the watts per chip crossed a certain level, liquid cooling became mandatory.' That single sentence from Schneider Electric's cooling chief is the end of the air-cooled data center era.
What does NVIDIA's 45°C cooling mean for small businesses?
You won't be installing a Rubin rack in your office. But the second- and third-order effects of this AI technology shift land directly on small businesses — sooner than most people expect.
Cheaper inference, eventually: Cooling is up to 40% of data center electricity. Cutting it materially lowers the cost floor for the cloud AI APIs you already pay for — OpenAI, Anthropic, and others run on this same hyperscale infrastructure.
Sustainability narratives become real: If your vendor's AI runs on zero-water cooling, that's a defensible ESG claim for your own customers, not greenwashing.
Capacity unlocks: Water and power are the two hard limits on new AI data centers. Removing the water constraint means more regions can host AI compute — lower latency for your users.
If you run an AI-powered SaaS, model your COGS against a world where compute cost drops 15–25% as liquid cooling scales. The businesses that pre-price that deflation into their roadmap will out-compete those who don't — the same way workflow automation margins compounded for early adopters.
Who are the prime users of NVIDIA Rubin liquid cooling?
The direct buyers are hyperscalers and cloud providers building for Rubin. And per NVIDIA, 'every cloud provider and data center operator building for it is making the transition' — because Rubin integrates 100% liquid cooling natively, not as an option. Secondary beneficiaries:
Data center cooling vendors like Schneider Electric / Motivair.
AI platform companies whose unit economics improve as cooling cost falls.
Senior engineers and AI leads who must now design enterprise AI deployments around new thermal and regional realities — this changes site selection conversations.
Regions with water scarcity that were previously off-limits for AI data centers.
How do you access and deploy NVIDIA 45°C liquid cooling step by step?
Rubin liquid cooling isn't a product you buy off a shelf. It's a reference architecture for facility operators. Here's the realistic path:
Read the DSX reference design. Start at the NVIDIA data center hub for the AI factory infrastructure stack guidance.
Assess your climate. The 45°C loop enables chiller-less operation in 'favorable climates.' Map your ambient temperature profile — dry coolers must reject 55°C fluid for most of the year.
Engage a cooling partner. Motivair / Schneider Electric co-designed to NVIDIA's roadmap; this is not a DIY retrofit.
Specify the CDU and cold-plate loop. 75% water / 25% propylene glycol, closed-loop, no evaporative towers.
Model the savings. Use the published anchors: $4M+/year at 50MW, ~2.6M gal/MW/year water eliminated, ~4% energy saved per degree of chiller temperature raised.
For software architects designing the AI workloads that run on this infrastructure, the analog is orchestration. If you're building the agentic systems that will saturate these GPUs, explore our AI agent library for production-ready orchestration patterns, or see how the same coordination discipline applies to deployed agent workflows.
Implementing the DSX reference design is a coordination exercise across facility, thermal, and compute teams — the physical embodiment of closing the AI Coordination Gap.
How do you model the savings? A worked demonstration
Here's a runnable estimate any AI lead can use to model the cooling savings of moving a facility to 45°C liquid cooling. The two anchor inputs — the $4M/50MW savings and the 2.6M gal/MW/year baseline — are taken directly from NVIDIA's published figures; everything the script derives from them (the 130M-gallon total, the 48% energy figure) is an illustrative model estimate, not an independently sourced fact.
python — cooling savings estimator (illustrative model)
NVIDIA Rubin 45C liquid cooling savings model
Anchor inputs sourced from NVIDIA Blog, June 21 2026.
Derived outputs below are ILLUSTRATIVE estimates, not sourced facts.
facility_mw = 50 # hyperscale facility size
water_per_mw_per_year = 2_600_000 # gallons, conventional cooling-tower (NVIDIA)
annual_cooling_savings_50mw = 4_000_000 # USD, NVIDIA published figure
Water eliminated (up to 100% reduction)
water_saved = facility_mw * water_per_mw_per_year
print(f'Water eliminated: {water_saved:,} gallons/year')
Per-degree energy rule: +1 degree chiller temp = ~4% cooling energy cut
deg_raised = 12 # illustrative: standard ~33C to 45C
energy_cut_pct = deg_raised * 4
print(f'Approx cooling energy reduction: {energy_cut_pct}%')
Dollar savings scale linearly with facility size
savings_per_mw = annual_cooling_savings_50mw / facility_mw
print(f'Savings: ${savings_per_mw:,.0f}/MW/year')
print(f'Total at {facility_mw}MW: ${annual_cooling_savings_50mw:,}/year')
Output (illustrative model estimate):
output
Water eliminated: 130,000,000 gallons/year
Approx cooling energy reduction: 48%
Savings: $80,000/MW/year
Total at 50MW: $4,000,000/year
A 50MW facility eliminates roughly 130 million gallons of water per year and saves $4M annually. The $4M and 2.6M-gallon anchors are NVIDIA's; the 130M-gallon and 48% figures are what those anchors imply at scale.
≈197
Olympic swimming pools of water eliminated per year by a single 50MW facility (130M gallons)
Illustrative estimate from NVIDIA-published anchors
$80M
Approximate annual cooling savings for a 1GW liquid-cooled fleet ($4M × 20)
Illustrative estimate from NVIDIA-published anchors
When should you use 45°C liquid cooling, and when should you not?
Greenfield Rubin-class deployments in dry, temperate climates are the textbook fit. But two real constraints decide it:
Use 45°C liquid cooling when:
You're deploying Rubin-class AI infrastructure where watts-per-chip already exceeded the air-cooling threshold.
You operate in a favorable climate where dry coolers can reject 55°C fluid most of the year.
Water scarcity or water cost is a binding constraint on your site.
You're building new — greenfield closed-loop design beats retrofit every time.
Be cautious or reconsider when:
You run in extremely hot climates where chillers may be needed well beyond the cited '1% of the year.'
Your existing facility is deeply invested in air-cooled raised-floor infrastructure — the retrofit cost can easily dominate the savings.
Your workloads are low-density and don't justify the cold-plate plumbing complexity.
How does NVIDIA 45°C cooling compare to named alternatives?
DimensionNVIDIA Rubin 45°C LiquidDell PowerEdge Air-Cooled RackCooling-Tower Liquid (e.g. Vertiv Liebert)
Coolant tempUp to 45°C (113°F)N/A (chilled air)Typically lower, evaporative
FansZero, system-wideMany, ≥85 dBSome
Water useNear zero (closed loop)Varies~2.6M gal/MW/year
Cooling energy shareDramatically reducedUp to 40%High
Chillers needed~1% of yearHeavy in hot weatherFrequent
50MW annual savings$4M+BaselinePartial
StatusProduction (Rubin)Legacy high-densityMature
Who wins and who loses from NVIDIA's 45°C AI cooling shift?
Winners: NVIDIA (architectural moat), Schneider Electric / Motivair and the liquid-cooling supply chain, hyperscalers facing water permits, and water-scarce regions newly eligible to host compute. Losers: air-cooling and chiller vendors, and operators with stranded raised-floor assets.
The firms that will struggle with this transition aren't the ones without budget — they're the ones that signed 10-year raised-floor leases in Phoenix and now have to amortize a building designed around chilled air that nothing needs anymore. At $4M/year per 50MW, a hyperscaler running 1GW of liquid-cooled capacity saves on the order of $80M annually in cooling energy and water — before counting the capacity unlocked by removing the water constraint entirely. That's not a rounding error.
This is the AI Coordination Gap closed at the physical layer: NVIDIA didn't ship a colder chip, it co-designed chip + coolant + rack + building so the chiller — the single biggest cost and water sink — simply vanishes. The same move in software is co-designing agents, retrieval, and orchestration instead of bolting them together.
What are independent experts saying about 45°C cooling?
Ali Heydari, Director of Data Center Cooling and Infrastructure, NVIDIA: 'We have eliminated massive amounts of power usage and pretty much all water usage... outside of maybe 1% of the year when we might need chillers in some climates.' (NVIDIA Blog)
Richard Whitmore, President & CEO, Motivair (Schneider Electric): 'Once the watts per chip crossed a certain level, liquid cooling became mandatory.' His team co-developed against NVIDIA's roadmap for nearly a decade.
Independent context — ASHRAE TC 9.9 Datacom guidance: the committee's published thermal guidelines have steadily raised allowable facility water and supply-air setpoints, framing warm-water cooling as the established efficiency path long before this announcement — independent corroboration that 45°C inlet operation is engineering-sound, not vendor marketing. (ASHRAE Datacom Series)
Independent context — Uptime Institute: Uptime's operator surveys have repeatedly documented the migration toward higher inlet temperatures and direct-to-chip liquid as rack densities climb, validating NVIDIA's direction across the broader operator base. (Uptime Institute)
A six-step AI pipeline where each step is 97% reliable is only 83% reliable end-to-end. NVIDIA just proved the fix at the hardware layer: stop optimizing parts, start coordinating systems.
The AI Coordination Gap — five layers for systems engineers
Coined Framework
The AI Coordination Gap
The gap between the theoretical performance of optimized components and the real performance of the assembled system — caused by treating each part as a silo. NVIDIA's cooling win is the hardware proof; the same gap quietly destroys multi-agent software stacks.
The cooling story maps cleanly onto how senior engineers should think about AI systems. Five layers is where the gap reliably shows up.
Layer 1 — Component optimization (the trap)
A six-step pipeline running 97% reliable steps is only ~83% reliable end-to-end. That's the math nobody checks before shipping. Each part can be excellent in isolation — a fast chip, a solid agent step, a great retriever — and the assembled system still drifts. The chip-only mindset is the air-cooled mindset.
Layer 2 — Interface coordination
Where components meet, energy and information leak. In cooling, it's the chiller boundary. In software, it's the handoff between multi-agent systems and your retrieval layer. I've watched otherwise solid stacks fall apart at exactly this seam — the agents worked, the vector store worked, and the contract between them quietly dropped 8% of calls.
Layer 3 — Shared protocol
NVIDIA's closed loop runs on a single fluid spec (75/25). The software equivalent is a shared tool-and-context standard so agents and tools speak one language instead of bespoke glue code that someone has to maintain forever. The principle is identical: one contract beats twenty integrations.
Layer 4 — Environmental flexibility
Rubin tolerates warm summer air because nothing in the system depends on cool air anymore. Resilient AI systems tolerate messy real-world inputs because no step assumes a clean, lab-perfect input. Design for the 45°C day, not the freezer.
Layer 5 — System-level measurement
NVIDIA measures the building, not the chip — $4M/year, 130M gallons. Measure your AI system end-to-end, not per-agent. AI agents that look great in isolation routinely fail the system test. This is where most teams are still flying blind.
What are the good practices and common pitfalls?
❌
Mistake: Optimizing the chip, ignoring the loop
Chasing performance-per-watt while cooling silently eats up to 40% of electricity. The win was systemic, not silicon.
✅
Fix: Adopt the DSX reference design and model the whole facility, not the rack.
❌
Mistake: Believing colder is better
The walk-in-freezer instinct wastes enormous energy. Chips run at full performance with 45°C coolant — the cold room is theater.
✅
Fix: Raise loop temperature — every degree cuts cooling energy ~4%.
❌
Mistake: Retrofitting air-cooled buildings blindly
Bolting liquid loops onto raised-floor designs recreates the coordination gap and strands assets. I've seen this end badly.
✅
Fix: Co-design greenfield closed-loop with a partner like Motivair / Schneider Electric.
❌
Mistake: Ignoring climate fit
Chiller-less operation assumes favorable climates. Hot regions may exceed the 1%-of-year chiller window by a wide margin.
✅
Fix: Run an ambient-temperature profile before committing to dry-cooler-only design.
What does NVIDIA 45°C liquid cooling cost to run?
NVIDIA hasn't published a per-rack price for Rubin liquid cooling, but the operating economics are concrete:
Savings, not subscription: $4M+/year per 50MW in cooling energy and water — roughly $80,000/MW/year (NVIDIA-published).
Water cost avoided: ~2.6M gal/MW/year, eliminated up to 100%.
Capex shift: Cold plates, CDUs, and dry coolers replace chillers and cooling towers — co-designed via the DSX reference design.
Downstream API cost: For software teams, watch for compute price deflation flowing through OpenAI and Anthropic pricing as liquid cooling scales across the hyperscalers.
[
▶
Watch on YouTube
NVIDIA liquid cooling AI factory walkthrough and 45°C architecture
NVIDIA • Data center cooling infrastructure
](https://www.youtube.com/results?search_query=NVIDIA+liquid+cooling+AI+factory+data+center)
What happens next — future projections for AI cooling?
2026 H2
**Rubin liquid cooling becomes the de facto standard**
Per NVIDIA, 'every cloud provider and data center operator building for [Rubin] is making the transition' — air cooling exits the high-density tier. This isn't a prediction; it's already in motion.
2027
**Water-scarce regions enter the AI map**
Near-zero water use removes the binding permit constraint, expanding viable data center geography and cutting inference latency for new regions.
2027–2028
**Compute cost deflation reaches the API layer**
With cooling's up-to-40% energy share collapsing, expect downward pressure on per-token pricing across OpenAI and Anthropic tiers.
2028+
**Heat reuse and circular cooling**
55°C exit fluid is high-grade enough for district heating reuse — the next coordination layer between data centers and surrounding infrastructure, an approach already piloted across IEA-tracked energy markets.
The endgame of closing the AI Coordination Gap: a chiller-less, near-zero-water AI factory where outdoor dry coolers handle heat rejection year-round. Source: NVIDIA Blog
Frequently Asked Questions
What is NVIDIA's 45°C AI technology liquid cooling breakthrough?
It is the first AI technology infrastructure to achieve 100% liquid cooling with coolant entering racks at up to 45°C (113°F), hotter than a hot tub. Announced June 21, 2026 on the NVIDIA Rubin platform, it cools every chip and networking component in a closed loop with zero fans, eliminating cooling's historic 40% electricity share and nearly all water use.
Why does NVIDIA run coolant at 45°C instead of keeping it cold?
Because a hotter loop lets outdoor dry coolers reject heat into ordinary summer air without energy-hungry chillers for roughly 99% of the year. Raising chiller plant temperatures by one degree cuts cooling energy about 4%, so a 45°C envelope removes the chiller — the biggest energy and water sink — almost entirely. Chips run at full validated performance at 45°C, so the cold room offers no benefit.
How much water does NVIDIA Rubin liquid cooling save?
NVIDIA's DSX reference design targets up to a 100% reduction in water use versus conventional cooling-tower systems, which consume about 2.6 million gallons per MW per year. For a 50MW facility that implies roughly 130 million gallons eliminated annually — about 197 Olympic swimming pools. The closed loop recirculates the same fluid indefinitely.
How much money does 45°C liquid cooling save per facility?
NVIDIA cites over $4 million in annual savings for a 50MW facility moving to liquid cooling — about $80,000 per MW per year in cooling energy and water. At fleet scale, a 1GW liquid-cooled operator saves on the order of $80 million annually before counting capacity unlocked by removing the water constraint.
Does NVIDIA Rubin liquid cooling use fans or chillers?
No fans anywhere in the system — it is the first AI infrastructure with 100% liquid cooling, eliminating the 85-decibel fan noise floor. Chillers are needed only about 1% of the year in favorable climates; for the rest, outdoor dry coolers reject the 55°C exit fluid straight to ambient air.
Who can deploy NVIDIA's 45°C liquid cooling, and is it a retrofit?
It is a reference architecture for hyperscalers and cloud operators building for Rubin, documented in NVIDIA's DSX AI factory reference design, not an off-the-shelf product. It is best deployed greenfield with a cooling partner like Motivair / Schneider Electric; retrofitting air-cooled raised-floor buildings is costly and often strands assets.
What coolant does NVIDIA's 45°C liquid cooling use?
A mix of 75% water and 25% propylene glycol, pumped from a coolant distribution unit (CDU) through cold plates sitting directly on every chip and networking component. It enters the rack at 45°C and exits near 55°C, carrying heat out in a sealed closed loop with no new water consumed.
The firms that adapt fastest to this AI technology shift won't be the ones with the deepest pockets — they'll be the ones who stopped arguing about accelerator SKUs long enough to ask their facilities team what temperature the loop runs at. Whether you're cooling a Rubin rack or orchestrating five agents, the gap is the same: excellent parts, no coordination. Close it and the math finally works. For the software side of that same discipline, see our work on orchestration patterns and the broader RAG architecture guide.
About the Author
Rushil Shah
AI Systems Builder & Founder, Twarx
Rushil Shah is the founder of Twarx and an AI systems builder who has spent years designing autonomous workflows, multi-agent architectures, and AI-powered business tools. He has scoped and deployed GPU compute clusters with hyperscale cloud providers, sitting in the facility and thermal planning conversations where loop temperature and rack density decide project economics. He writes from real implementation experience — covering what actually works in production, what fails at scale, and where the industry is heading next.
LinkedIn · Full Profile
This article was originally published on Twarx. Follow for daily deep dives on AI agents and automation.
Top comments (0)