Elice Plans Warm-Water AI Data Center and Launches Helpy Code

Elice has reached google news with two firsts: a proposed warm-water AI data center and the public debut of its Helpy Code agent. The Korean company is connecting 2,560 planned Nvidia B300 GPUs with software designed to work inside private enterprise environments.

That combination matters more than either announcement alone. Elice is presenting itself as the operator of an integrated Korean AI stack, from cooling pipes and cloud orchestration to company knowledge and generated code.

The strategy also creates a demanding test. Building efficient GPU infrastructure is difficult, while earning developer trust in a coding agent presents a different set of technical and commercial problems. Elice must prove it can do both.

The data center has not yet established its promised operating efficiency through public production measurements. Helpy Code also lacks independently published benchmarks, adoption data, or detailed model disclosures. The story is therefore about an ambitious integration plan, not a completed victory.

What the Elice AI Data Center Actually Changes

Elice is moving from operating smaller modular facilities toward a government-backed cluster designed around high-density Blackwell Ultra hardware.

On August 4, Elice said it had been selected for a South Korean government program intended to expand national AI computing capacity. The Ministry of Science and ICT and the National IT Industry Promotion Agency are leading that program.

Under the plan, the company will build and operate four portable modular data center units. Each unit will contain 640 Nvidia B300 GPUs, producing a combined total of 2,560 GPUs.

Portable modular data centers package computing, electrical, cooling, and security systems inside prefabricated modules. Operators can install capacity in stages instead of constructing every element of a conventional facility on-site.

Elice says its modular method can shorten deployment and make future expansion more flexible. However, the August announcement did not disclose the final site, construction schedule, electrical capacity, or expected production date.

The planned cooling system is the defining feature. According to the company’s cooling announcement, the facility will circulate coolant at 40 degrees Celsius or higher while using outside air to reject heat.

Warm-water cooling sounds counterintuitive because the liquid can feel hot to a person. It remains much cooler than operating processors, however, so it can still absorb heat from server components.

Higher coolant temperatures reduce the need to chill water far below outdoor conditions. Under suitable weather and load conditions, that difference can reduce compressor use and lower cooling electricity consumption.

The design will use direct liquid cooling, or DLC. In a DLC system, coolant passes through cold plates attached near high-heat components instead of relying mainly on room air.

Elice says it has also developed a separated-piping design intended to contain leaks. That detail matters because liquid near processors, power systems, and networking equipment creates maintenance risks that air cooling avoids.

The company targets a power usage effectiveness, or PUE, of approximately 1.1. PUE divides total facility electricity by electricity consumed by IT equipment. A value of 1.1 means one additional unit supports cooling and other overhead for every ten units used by computing equipment.

A target is not the same as an audited annual result. PUE changes with utilization, outdoor temperature, facility boundaries, redundancy choices, and measurement practices.

The Elice AI data center must therefore be judged after its B300 systems run under sustained customer workloads. Commissioning data will reveal more than the design target alone.

This project expands a modular approach that Elice already markets to Korean organizations. Its existing facilities have supported cloud computing, education, and research workloads, but the proposed B300 deployment operates at another density level.

Nvidia’s reference architecture shows why the change is necessary. A GB300 NVL72 rack can require up to 142 kilowatts and uses rack-scale liquid cooling.

That density transforms cooling from a building service into part of the computing architecture. Flow rates, coolant temperatures, pumps, heat exchangers, leak detection, and software monitoring can all affect usable GPU capacity.

The project is therefore not simply another order for accelerators. Elice is redesigning its physical infrastructure around the thermal requirements of a newer computing generation.

Why Warm Water Has Returned to Google News

Warm-water cooling is not a new invention, but rising rack density has turned it from an efficiency experiment into a deployment requirement.

The “first” in Elice’s claim applies to a specific South Korean AI data center category. It should not be interpreted as the first use of warm-water computing anywhere.

Researchers and supercomputer operators have tested similar designs for years. IBM’s Aquasar prototype circulated coolant at temperatures reaching 60 degrees Celsius and reused captured heat for building heating.

The resulting Aquasar research reported that hot water removed processor heat without conventional chilling. It also demonstrated an 80 percent heat-recovery efficiency in that installation.

IBM later used warm inlet water to cool CPUs and GPUs in the Summit and Sierra supercomputers. Those systems combined direct water cooling for high-density components with air cooling for lower-density hardware.

These precedents do not validate Elice’s specific design. They establish that warm-water cooling has a credible engineering history and is not merely a marketing label.

The renewed attention on google news reflects a change in scale. AI servers now concentrate more electrical power inside each rack, producing heat that becomes difficult to move efficiently through air alone.

Air cooling first transfers heat from chips into a server’s internal airflow. Facility fans then move that heated air toward cooling equipment, adding several thermal and mechanical steps.

Direct liquid cooling shortens the path. Water also transports heat more effectively than air, allowing operators to remove concentrated loads with less fan power.

The US Department of Energy’s cooling guidance explains that liquid systems move server heat into a closed loop and then through a coolant distribution unit. Pumps usually require less energy than the fans needed for comparable heat movement.

Warm water adds another potential efficiency gain. A facility that accepts higher coolant temperatures has more opportunities to discharge heat through outdoor air without running energy-intensive chillers.

A joint IBM and Department of Energy test illustrates the potential. Its long-duration study operated a chillerless warm-water facility for more than 60 days.

Researchers reported average cooling energy equal to 3.5 percent of IT energy during that study. The result belonged to a controlled test facility, so it cannot be transferred directly to Elice’s planned cluster.

South Korea’s climate also complicates the comparison. Hot, humid summer conditions can reduce the number of hours when outside air alone handles the full cooling load.

Facility design must accommodate those difficult periods without throttling GPUs. It must also manage water quality, corrosion, pump failures, coolant pressure, and changing server configurations.

Water consumption presents another distinction. Warm-water operation can reduce chiller electricity, but it does not automatically make a facility water-free.

The final water footprint depends on the heat-rejection equipment. Dry coolers, cooling towers, hybrid systems, and local climate conditions create different tradeoffs between electricity and water.

Elice’s announcement specifies warm-water and outside-air cooling. It does not provide an expected water usage effectiveness figure, a heat-reuse plan, or a full description of its heat-rejection system.

Those omissions do not invalidate the project. They mean the environmental label remains broader than the publicly available operating evidence.

A PUE of 1.1 would represent limited facility overhead if Elice sustains it over a full year. Yet PUE says nothing about the carbon intensity of electricity or the amount of water consumed.

The strongest version of Elice’s argument is therefore narrow. Warm-water cooling can reduce energy spent moving heat, especially when outdoor conditions support compressor-free operation.

The weaker version treats one design choice as proof of an environmentally superior facility. That conclusion requires annual energy, water, utilization, and emissions data that Elice has not published.

Helpy Code Connects Private Knowledge to the Computing Stack

Elice is pairing its infrastructure expansion with a coding agent designed to keep sensitive code and organizational context inside controlled environments.

On August 18, Elice unveiled Helpy Code as part of its Elice AX enterprise software portfolio. The company presented the product alongside Helpy Context, Helpy Vision, and Helpy Chat.

Helpy Code is an agent rather than a basic autocomplete feature. A coding agent can inspect repositories, plan changes, generate files, and perform multi-step development work under user supervision.

The available announcement does not detail every supported repository, programming language, model, or development environment. It instead emphasizes deployment architecture and access to internal knowledge.

Elice says Helpy Code can run on-premises, meaning an organization operates it inside infrastructure that the organization controls. This design targets companies and public institutions that cannot freely send source code to external services.

The agent incorporates two other Elice components. Helpy Vision converts documents, diagrams, and tables into structured information that software can process.

Helpy Context organizes documents, people, projects, and permissions into knowledge layers. Elice describes it as an internal ontology, which means a structured map of organizational concepts and their relationships.

Together, those systems are intended to give Helpy Code more than repository access. The agent could retrieve architecture policies, project ownership, internal standards, and technical documents while preparing a code change.

That context can matter as much as model capability. A technically valid patch may still violate a company’s deployment rules, naming conventions, security policy, or undocumented service dependencies.

Elice says role-based access controls determine which knowledge the agent can retrieve. The company also says information can remain within the customer’s environment.

These are company claims, not independent security findings. Elice has not published a Helpy Code threat model, external penetration test, or detailed account of how generated actions are isolated.

The company also claims the agent operates at less than one-fifth the cost of frontier models. That comparison lacks enough public methodology to evaluate.

It remains unclear which models, workloads, token volumes, infrastructure expenses, or labor costs are included. The figure should therefore be treated as Elice’s positioning rather than an established economic result.

The important product idea does not depend on that number. Elice is trying to sell a vertically connected system in which local infrastructure supports local models, enterprise knowledge, and an action-taking agent.

Its AX product suite links Helpy Code with document analysis and knowledge structuring. That architecture targets customers whose main objection to coding agents involves control rather than raw capability.

A regulated company may need to restrict which model receives source code. It may also need logs showing which files the agent read and which commands it executed.

Public institutions can face additional sovereignty requirements. They may need domestic hosting, controlled network access, and clear boundaries around sensitive datasets.

Elice’s cloud and modular data center businesses give it an unusual position in that sales conversation. The company can offer physical capacity, cloud orchestration, and application software through one relationship.

That integration resembles a searchable knowledge base at a larger organizational scale. Code agents become more useful when they can retrieve trusted technical context rather than guess from isolated files.

The relationship also runs in the opposite direction. Helpy Code and other agents can create demand for Elice’s GPU cloud, storage, networking, and data center capacity.

This feedback loop is the real strategic link between the announcements. The facility supplies compute for agents, while the agents give customers a reason to consume the facility’s compute.

Elice Is Challenging Specialized Vendors With One Integrated Stack

The primary contest is integrated control against specialized services that customers can combine from several vendors.

Elice is not entering an empty coding-agent market. Developers already have access to products backed by major model providers, cloud platforms, and software development companies.

Those services benefit from broad distribution, frequent model updates, and large developer communities. Some also integrate tightly with popular repositories, editors, issue trackers, and deployment systems.

Elice cannot assume that owning infrastructure produces a better coding experience. Developers will judge Helpy Code through patch quality, latency, tool reliability, review burden, and compatibility with existing workflows.

The company’s advantage is different. Elice can argue that a customer does not need to assemble a model gateway, private cloud, document parser, knowledge layer, and coding interface independently.

Its full-stack strategy also includes ECI, an infrastructure service for provisioning GPU and NPU resources. ECI supports virtual machines, storage, networking, monitoring, and automated infrastructure management.

This creates a coherent sales story for Korean enterprises. Elice can deploy a modular facility, orchestrate its accelerators, structure private information, and place an agent on top.

A specialized approach gives customers more choice. An organization can select separate vendors for infrastructure, foundation models, retrieval, security, and developer tools.

That flexibility can reduce dependence on one supplier. It also allows teams to replace a weak component without migrating the entire system.

Integration shifts that tradeoff. One vendor can simplify accountability and data governance, but customers become more exposed to its technical roadmap and operating stability.

Elice must prove that its integration creates measurable benefits. A list of connected products will not be enough if each product trails a specialized alternative.

Helpy Code explained as part of this stack has three plausible advantages. It can keep selected data in-region, retrieve governed internal knowledge, and run on infrastructure configured for the customer.

Each advantage also introduces questions. In-region processing depends on the actual model route, not only the location of the user interface.

Knowledge retrieval depends on document quality, permissions, and freshness. On-premises operation depends on administrators who can maintain models, indexes, agent runtimes, and security controls.

The infrastructure side carries similar tradeoffs. Portable modules can reduce construction work, but modular buildings still need power connections, networking, permits, maintenance access, and heat rejection.

A three-month module assembly claim does not guarantee three-month service availability. Grid interconnection and accelerator delivery can determine the real schedule.

Elice also depends heavily on Nvidia’s hardware roadmap. Its planned cluster uses B300 GPUs, while its future designs anticipate even denser Vera Rubin systems.

That alignment provides access to a widely adopted computing platform. It also exposes Elice to product transitions, supply constraints, and rapidly changing rack requirements.

The company has discussed Ethernet-based clustering as an alternative to InfiniBand-centered designs. Ethernet can offer broader supplier choice, but large training clusters require careful congestion control and operational tuning.

A customer buying the integrated stack is therefore buying Elice’s engineering judgment across several layers. A failure in networking, cooling, retrieval, or agent safety can weaken the entire proposition.

This is why Elice’s strategy pressures more than other Korean cloud providers. It also challenges software vendors that expect enterprises to consume coding agents through foreign public clouds.

The outcome will not be decided by who announces the longest product list. It will depend on whether integrated control improves security, cost, and developer output without creating operational friction.

What the Google News Headline Does Not Prove

The headline combines two promising announcements, but neither has reached the point where public evidence supports the broadest interpretation.

The warm-water facility remains a construction and commissioning project. Elice has published a GPU count, cooling approach, modular layout, and target PUE, but not production telemetry.

Readers should also treat “South Korea’s first” as a company-framed category claim. The phrase depends on how an AI data center and warm-water deployment are defined.

A prior Korean research or high-performance computing installation might use related cooling techniques without matching Elice’s commercial modular configuration. Public reporting has not established a comprehensive national inventory.

The word “build” carries another ambiguity. Elice says it will build and operate the facility through a government-backed program, but its announcement lacks a completion date.

GPU allocation does not equal available computing capacity. Servers must pass acceptance testing, networks must sustain distributed workloads, and cooling must perform across changing utilization.

PUE is equally easy to overread. A 1.1 target suggests efficient supporting infrastructure, not efficient AI models or high customer utilization.

An idle facility can waste capital even with excellent cooling. A busy facility can still carry a large carbon footprint if it draws electricity from emissions-intensive generation.

Helpy Code faces a different verification gap. Elice has not released a benchmark comparable across coding agents, nor has it published task-completion rates from enterprise deployments.

Software engineering benchmarks also have limits. They can measure whether an agent solves selected repository issues, but they may not capture security, maintainability, or human review time.

The more important evidence will come from production teams. Useful measures include accepted patches, reverted changes, review duration, incident rates, and developer time saved.

Security claims need similar specificity. On-premises deployment can reduce data transfer to external services, but it does not remove every risk.

An agent may access files beyond a task’s intended scope. It may execute unsafe commands, expose secrets through logs, or follow malicious instructions embedded in documents.

Role-based retrieval controls help only when permissions are accurate. Organizations often carry inconsistent file ownership and overly broad access from older collaboration systems.

Helpy Context must preserve those boundaries while keeping information current. Otherwise, Helpy Code could act on obsolete guidance or reveal material to the wrong developer.

Model provenance is another unresolved issue. Elice says the system can support secure enterprise environments, but its public announcement does not specify the default model behind Helpy Code.

Customers need to know whether prompts reach third-party endpoints, where inference logs are stored, and whether their data affects later model training. They also need controls for changing model providers.

These questions are not reasons to dismiss the product. They are the practical requirements that separate a controlled demonstration from production adoption.

The timing of the announcements deserves care as well. Elice dated the data center announcement August 4 and the Elice AX announcement August 18.

The August 18 release says the company would showcase its products at AI Summit Seoul & Expo from June 19 through June 21. That date sequence creates uncertainty about whether the English release was posted later than the underlying event.

The public record still establishes the product names and Elice’s stated architecture. It does not support treating every announcement date as the actual first demonstration date without qualification.

Google news readers therefore received a compressed narrative. The underlying story consists of separate infrastructure and software initiatives that Elice is deliberately presenting as one strategy.

That strategy is credible enough to follow. Its largest claims remain open to measurement.

Three Signals Will Show Whether Elice Can Deliver

Commissioning results, Helpy Code deployments, and customer workload data will determine whether Elice has built a business or only a compelling diagram.

The first signal is a confirmed operating milestone for the 2,560-GPU cluster. Elice should identify when all four modules enter production and publish measurements under sustained load.

Annualized PUE would be more informative than a short commissioning result. Water consumption, coolant temperatures, equipment availability, and weather conditions would add necessary context.

A sustained result near 1.1 would strengthen Elice’s cooling argument. Major delays, extensive chiller dependence, or lower usable GPU capacity would weaken it.

The second signal is evidence from Helpy Code deployments. The company does not need to expose customer source code, but it can publish anonymized operational measures.

Useful evidence would include completed tasks, patch acceptance, review time, failure rates, and security interventions. Independent customer accounts would carry more weight than demonstration scripts.

Elice should also explain Helpy Code’s model architecture and data path. Buyers need clear answers about model selection, local inference, external endpoints, permissions, and audit logging.

Production adoption inside a regulated enterprise would support the integrated-control thesis. Limited trials without continued usage would suggest that private deployment alone does not overcome workflow friction.

The third signal is utilization across Elice’s stack. The company must show that customers use its GPUs for sustained training, inference, or agent workloads.

High utilization would connect the data center investment to real demand. It would also test whether ECI can schedule large clusters while maintaining network, storage, and cooling performance.

Low utilization would expose the central risk in infrastructure expansion. Accelerators age quickly, while electrical and cooling systems require long-term capital commitments.

Customers should also watch whether Elice wins workloads beyond government-supported projects. Public programs can create valuable national capacity, but commercial demand tests pricing and service quality more directly.

Competitor responses matter within this signal. Korean telecommunications groups, cloud providers, and system integrators can combine domestic infrastructure with third-party coding agents.

Global vendors can also strengthen private deployment and regional processing options. Those moves would reduce the uniqueness of Elice’s integrated pitch.

Elice does not need to beat every specialized provider. It needs to make the combined system easier to govern and operate for customers with strict data boundaries.

For developers, the immediate question is practical: does Helpy Code understand an organization well enough to reduce review work without gaining excessive access?

For infrastructure buyers, the question is measurable: can warm-water modular cooling keep B300 clusters available and efficient through a full operating year?

For knowledge teams, the challenge is foundational. Agents cannot use private context reliably when documents, permissions, and ownership remain disorganized.

The latest google news attention gives Elice visibility, but visibility does not settle those questions. The next meaningful update should include operating evidence, not another collection of product labels.

Watch the first production modules, the first independently described Helpy Code deployment, and the first sustained utilization results. Together, those signals will show whether Elice’s integrated Korean AI stack works beyond the exhibition floor.