Engineered for the
AI decade.
HyperNext is India's first data center built around 800VDC power architecture and rack-scale AI systems from NVIDIA and AMD, designed for the workloads of 2028, not 2018.
Why 800VDC is the future
of AI data centers.
AC distribution served the data center for forty years. For the densities AI now asks of us, it stopped being enough. We made the shift. Eight hundred volts of direct current, end to end, because it is the only honest answer at this scale. Pushed forward by the Open Compute Project and adopted by NVIDIA, AMD, Google, Microsoft and Meta, 800VDC is the architecture the hyperscale industry is converging on.
Every conversion costs energy.
The fewer the conversions, the more tokens per watt.
The same megawatt produces more tokens. That is the entire point.
Built for AI density
Modern AI racks draw 100 kW and beyond. 800VDC delivers that power natively, no Frankenstein retrofits, no compromise on stability.
Fewer conversions, more efficiency
Traditional architectures convert power up to six times before it reaches a chip. 800VDC eliminates most of those stages, directly improving facility PUE.
Less copper, lighter cabling
Higher voltage means lower current for the same power. Cable sections shrink dramatically, freeing pathways and reducing material cost across the facility.
Aligned to the OCP standard
800VDC is the architecture NVIDIA, AMD, Google, Microsoft and Meta have been pushing through the Open Compute Project. We're building to the standard the industry is converging on.
Future-proof headroom
The next generation of AI accelerators is being designed for DC power delivery. A facility built for 800VDC today is one that can host the racks shipping in 2028, without re-platforming.
Cleaner power quality
Direct current eliminates the harmonic distortion and reactive losses inherent to AC distribution at scale, translating to cleaner, more predictable power at every rack.
Same megawatt in.
More compute out.
Two differences. What a rack is allowed to draw, and what it costs to run once it does.
India's first Vera Rubin
rack-scale system.
It is deploying across the fully contracted Hyderabad campus now, on the same 800 VDC power and direct-to-chip cooling we built for exactly this load, with NVLink fabric across the pod, the CUDA stack above it, and Vera Rubin Ultra NVL576 waiting at the Kakinada AI Factory on day one.

8 × 2 pods · 240 kW per rack · 800 MW AI cluster deploying
Built for every shape
of AI compute.
A workload should not have to compromise to fit infrastructure. Infrastructure should adapt. From inference at the edge of a transaction to simulations that run for months, the stack we built bends to the work.
Inferencing
High-throughput, low-latency inference for production AI, chatbots, recommendation, fraud detection, vision pipelines and agentic workloads.
RAG Systems
Retrieval-augmented generation that anchors LLMs in your own knowledge base, built on infrastructure tuned for vector throughput and dense storage.
Physical AI
Simulation, robotics, autonomous systems, workloads that bridge digital intelligence and the physical world.
Scientific AI
Drug discovery, climate modelling, computational fluid dynamics, long-running HPC workloads that demand sustained performance.
Open by design.
Vera Rubin and AMD Helios, side by side. Genuine choice, not the appearance of it. For inference, for retrieval, for physical AI, for scientific work. Our partnership with AMD brings the Helios rack-scale system to India for the first time, an open platform that has the potential to reshape India's AI infrastructure economics.
Power & Mainframe
as a Service.
For Indian banks, IBM Power and IBM Mainframe arrive directly. As managed services. From us. The intermediaries that historically sat between Indian banks and the platforms they ran on, do not sit here. Lower cost, faster onboarding, accountable SLAs.


IBM Power11 · IBM Z
Sizing for an AI workload?
Tell us about your model, your latency target, and your ramp plan, we'll size, price and schedule a deployment.
A row at a time.
400 MW of Helios at Kakinada. 192 MW at Johor Bahru, built around it from the ground up. This is what one row looks like as it lands.
EPYC Venice
- CPUs per rack
- 18
- Cores per CPU
- up to 256
- Cores per rack
- ~4,600
Instinct MI455X
- GPUs per rack
- 72
- HBM4 per GPU
- 432 GB
- HBM4 per rack
- 31 TB
- FP4 per rack
- 2.9 EF
UALink over Ethernet
- Scale-up domain
- 72 GPUs
- Scale-up bandwidth
- 260 TB/s
- Scale-out per rack
- 43 TB/s
What it lands on
- Floor loading
- 2,800 kg/m²
- Distribution
- 33 kV MVSST
- Cooling
- direct-to-chip
- On-site cooling water
- zero

Direct-to-Chip Hybrid Cooling
Liquid where the heat is, air for the rest, and a compact sidecar exchanger instead of an expensive CDU. Cools 1400W chips, retrofits into servers you already own, up to 37% less cooling power.
Read it ›
Inline Fluid Quality Valve
An inline valve that reads the dry-cooler coolant as it flows past and projects how many days of life it has left, so coolant is changed on condition, not on a calendar. Live on the BMS.
Read it ›The same instinct runs through the rest of the platform: an 800VDC power architecture, a vendor-agnostic BMS, and a sovereign cloud control plane, all built rather than bought. More of that engineering is on the TechnologyPlatform page.
Things we took out.
Things we built stronger.
Most of what makes a hall run cool, clean and dense is decided before any equipment arrives. These are the choices that shape HN1, and the reasons behind them.
MVSST: 33 kV straight to the rack
A medium-voltage solid-state transformer takes 33 kV down to rack voltage in a single stage. The conventional chain of step-down transformer, switchgear, UPS, PDU and rack PSU is not shortened, it is removed: there is no traditional transformer anywhere between the grid and the GPU. Each stage carried its own conversion loss, so taking them out saves power at every point they used to sit, and that saving compounds across the whole system.
No UPS hall, no battery rooms
With conversion collapsed into the MVSST and backup held at the rack, the central UPS hall and its rooms of batteries are never built. That removes the largest fire load in a conventional facility, along with its suppression, containment and insurance overhead.
Why it matters commercially: in a conventional facility, 55% of the grey space is taken by UPS rooms, battery racks, UPS plant and electrical panels. It is built, cooled, insured and maintained, and none of it can be leased. Removing that plant removes the space it stood in, and returns it to the part of the building that earns.
Fewer chillers, no cooling water
Direct-to-chip liquid takes the heat where it is made, so far less of the load ever reaches the air. Fewer chillers are installed, and a sealed glycol loop with dry coolers rejects heat without evaporating anything. There is no on-site cooling water consumption to reduce later, because none was designed in.
2,800 kg/m² floor loading
Rack-scale systems from NVIDIA and AMD are heavy, and the liquid distribution that serves them adds more. The slab is specified at 2,800 kg per square metre, the highest floor loading in India, so a Vera Rubin or AMD Helios rack can stand anywhere on the floor without local strengthening. Structure is the one thing that cannot be retrofitted, so it was decided first.
The greenest material
is the one never mined.
Most of a data centre's environmental cost is decided in procurement, not in operation. Two of the heaviest materials in a conventional facility are copper and lithium. This design needs far less copper, and none of the central lithium battery plant that normally holds the whole load. It is the first in India built that way.
Far less copper in the ground
Copper is sized by current, not by power. Raising the distribution voltage to 800V direct current carries the same megawatts at a fraction of the amperage, so conductors get thinner and shorter. MVSST removes whole runs of cable by taking medium voltage to the rack in one stage instead of stepping it down through switchgear, UPS and PDU. Less copper to buy, less to pull, less weight in the trays and less to recover at end of life.
Why it matters ecologically: copper is one of the most energy- and water-intensive metals to mine and refine, and demand for it is the single biggest materials constraint on global data centre build-out. Copper never specified is copper never extracted.
No lithium UPS in the core
A conventional facility holds its ride-through in halls of lithium-ion cells sized for the whole load, cooled, fire-suppressed, monitored and replaced on a cycle. With conversion collapsed into the MVSST, that central battery plant is never built. Lithium still exists in the building, inside servers and in the rack BBUs that ride out surges, and in ordinary office UPS. What disappears is the core plant, and that is where almost all of the mass sits.
Why it matters ecologically: lithium extraction is water-intensive and frequently sited in water-stressed regions, and the cells carry cobalt and nickel supply chains plus an end-of-life burden the industry has not solved. Not building the central plant avoids that at the scale where it counts, and takes the thermal-runaway fire load out of the building with it.
A first for India
Taken together, medium voltage direct to the rack, no central UPS, no lithium battery plant in the core, no on-site cooling water and none of the UPS, battery and panel space that normally takes 55% of a hall's grey space, is a combination no data centre in India has been built to before. Each decision removes a system rather than making one more efficient, and every system removed takes its materials, its cooling load, its maintenance and its disposal with it.

The sidecar heat-exchanger prototype on the bench. The window shows the microchannel cold plate and the coolant moving through it.
A single chip now pulls
1400 watts.
A top-end accelerator used to draw a few hundred watts. The newest pull up to 1400. Air alone cannot move that much heat out of a dense rack, and the usual fix, full direct-to-chip with a row-scale coolant distribution unit, is expensive, heavy, and assumes you are building the hall from scratch.
Most operators are not building from scratch. They have racks and servers already in the ground, and a refresh cycle that does not wait for a new cooling plant. So we built for that reality instead.
Liquid where it counts.
Air where it still works.
A hybrid. Direct-to-chip liquid cooling takes the heat off the hottest parts, the GPUs and CPUs, and air handles everything else. You cool what needs liquid and no more.
The part we are proudest of is the sidecar thermal heat exchanger. It sits beside the rack, not in a separate row of plant, and rejects the loop's heat in a fraction of the footprint and pump energy a conventional CDU needs. It runs on any dielectric fluid, or on DI water, and it asks for very little fluid top-up across a year. The chemistry is watched live on the same BMS that runs our inline fluid-quality valves, so the coolant is changed on condition, not on a calendar.
Up to 37% less
cooling power.
Against conventional cooling, the system uses up to 37% less cooling power. Less pump energy, less fan energy, less fluid, less plant. On a hall sized in megawatts, that is real money and real carbon off the bill every month.

The hybrid loop running in a rack. Liquid to the hot parts, air for the rest, all on the HyperNext BMS.
It goes into the servers
you already own.
It was designed to drop into hardware that already exists. No new hall, no row of CDUs, no forklift upgrade. The same approach fits an older-generation server as easily as a new one, which is the difference between a pilot and something you can actually roll out at scale.

Retrofit into existing nodes. The cooling is fitted to the servers, not the other way round.
What it is, in short.
Built in-house, running on our own BMS. The point was never a science project. It was a way to cool 1400W chips in racks that already exist, without a coolant plant that costs more than the servers it cools.
The inline quality valve on the bench. Stainless, flanged, two per loop. PQV-B1-01 on the PGW25 dry-cooler loop.
We used to change coolant
on a calendar.
Pick a date, drain the loop, refill, move on. Half the time the fluid was still perfectly good and we binned it. The other half it had quietly gone off weeks before anyone thought to look. Nobody enjoys draining 25% propylene glycol that still has life left in it, and nobody enjoys finding out the inhibitor packed up after the fact.
So we built this instead.
It reads the fluid
as it flows past.
An inline quality valve that sits in the dry-cooler loop and senses the coolant in real time. All live, all on the BMS. No sample bottles, no waiting on a lab, no guessing in between.
It tells you how many
days the fluid has left.
The part we are happiest with. The valve does not just say the fluid looks fine today. It projects how many days it has left, and maps the sampled colour against condition, clear and fresh at one end through to end of life at the other. So coolant changes when the fluid says so, not when a spreadsheet says so.
Building 1, PGW25 loop on the HyperNext BMS. Live inline sensing, a fluid-life projection, and sampled colour mapped to condition.
Two of our building loops are sitting around 68 to 71% life right now, with well over 1,200 days projected. We will run them until they are actually done, not until a date on a maintenance plan tells us to drain them.
The loop is closed.
On-site water is zero.
The loop is closed and dry-cooled, so on-site water use stays at zero. That is not a line for a brochure, it is the reason the chemistry has to be watched this closely. There is no make-up water topping things off and hiding our mistakes for us.
Small box. Big difference.
Built in-house, running on our own BMS. A small box that stops a lot of good fluid going down a drain for no reason, and catches the bad fluid before it costs us a loop.
Run your CUDA models
on our AMD systems.
At NO transfer cost.
You should not be held hostage by the choice you made two years ago. We don't think that is right. So we built a path. CUDA to ROCm, translated, tuned, and validated by us. Your model still runs. Your numbers still hold. The silicon underneath happens to be AMD Helios. We pass NO engineering cost on to you, because the engineering cost should never have been yours to begin with.
Twenty-two workloads.
No two alike.
Some workloads ask for scale. Others ask for stillness. Some need every watt of every accelerator. Others need a careful answer in twelve milliseconds. Twenty-two workload patterns, each with its own discipline. We considered each one. We did not retrofit anything.
Frontier model training
Multi-thousand GPU clusters with NVIDIA Vera Rubin Ultra NVL576 and AMD Helios. Rack-scale liquid cooling. Lossless InfiniBand and 800G Ethernet fabric. Continuous training at >90% MFU.
Agentic AI & inference
Production-grade inference with sub-millisecond latency for agentic workflows. Multi-tenant isolation. KV-cache optimisation, speculative decoding, and dynamic batching tuned for tool-use latency budgets.
Healthcare & diagnostics
Non-invasive diagnostic AI for early disease detection. Genomic and proteomic inference. Federated medical models with patient data never leaving the sovereign boundary.
Drug discovery
Molecular simulation, protein folding, and hybrid AI-physics models for pharmaceutical R&D. HPC-class precision combined with AI-class scale. IP protected inside India.
Financial services AI
Real-time fraud detection, credit decisioning, algorithmic trading and risk analytics. RBI-compliant deployment with sub-millisecond inference latency for transaction streams.
Physical & embodied AI
Robotics training, simulation, and digital-twin workloads for manufacturing, defence and autonomous systems. NVIDIA Omniverse and Isaac platform ready.
Scientific computing
Climate modelling, materials science, fluid dynamics and quantum chemistry. HPC-grade interconnect with AI-grade accelerators. Suitable for AIRAWAT and CSIR-grade research.
Enterprise RAG & copilots
Sovereign retrieval-augmented generation and enterprise copilot platforms. Customer data never leaves the boundary. Pre-built integrations with leading vector databases and LLM frameworks.
Federated learning
Train across distributed customer enclaves without raw data leaving each enclave. Differential privacy guarantees, secure aggregation, and audit-grade lineage of every gradient update.
Multimodal & generative AI
Vision-language models, video generation, speech and audio synthesis. Fine-tuning on customer multimodal corpora with content-safety guardrails enforced inside the platform.
Telco AI & RAN
O-RAN intelligent controller workloads, network optimisation, beam-forming and self-organising networks. DoT-licensable and edge-deployable to telecom POPs.
Industry 4.0 & manufacturing
Computer-vision quality inspection, predictive maintenance, digital twin of factories, supply-chain optimisation. OT-IT bridge with field-bus integration.
Quantum-classical hybrid
Hybrid quantum-classical workflow execution. Variational quantum eigensolvers, QAOA optimisers, and classical co-processing for emerging quantum computing partners.
Government & sovereign AI
Whole-of-government inference platforms, citizen-data analytics, national language models, and inter-departmental data fabric workloads. Sovereign-by-architecture.
Speech & language
Speech recognition, translation, and natural language understanding at scale. Indic-language model training and inference, multilingual dialog systems, conversational AI for service workloads.
Computer vision at scale
Image and video understanding, manufacturing inspection, surveillance analytics, retail computer vision, and medical imaging classifiers. Pipelines that operate continuously, not in batches.
Recommendation systems
Two-tower retrieval, deep ranking, and personalisation engines running at internet scale. Latency-bound serving with continuous incremental training and tight feature-store integration.
Cybersecurity AI
Threat detection, behavioural anomaly models, malware classification, and security copilots. Real-time inference against streaming telemetry from SOC and NOC pipelines.
Code generation
Developer copilots, code review automation, repository-wide understanding, and SRE assistants. Fine-tuning on customer codebases inside the sovereign boundary.
Time series & forecasting
Energy load forecasting, demand prediction, capacity planning, financial time series, and weather-coupled industrial optimisation. Continuous training on rolling windows.
Reinforcement learning
Policy training for control systems, robotics, energy management, and operations research. Long-horizon simulation environments coupled with GPU-accelerated rollouts.
Synthetic data & simulation
Synthetic data generation for training where real data is scarce or restricted. Domain-specific simulation environments for autonomous systems, industrial AI, and clinical research.
29 patents granted.
14 under approval.
Engineering is something we take seriously. The patents speak more carefully than we could.
Cancer detection AI
Non-invasive early detection models, multi-modal imaging analysis, and tissue-classification networks.
Clear aligner manufacturing
AI-driven orthodontic planning, treatment-path prediction, and aligner manufacturing optimisation.
Antimicrobial resistance
Resistance-pattern prediction, treatment-recommendation models, and pathogen-genome AI.
Hybrid molecular drug discovery
Physics-informed neural networks for protein-ligand binding and lead-compound optimisation.
Multi-threading for AI on CPU
Cache-aware threading and instruction-level scheduling for efficient AI inference on CPU-only hardware.
Memory load balancing
Dynamic memory allocation and NUMA-aware scheduling for optimum utilisation under heavy AI workloads.
Software that saves you
millions in licensing.
We built our own. We have reasons. An operating system that does not extract a tax with every socket. A virtualisation layer that does not hold your workload in escrow. A building management system that does not depend on a single vendor remembering you in five years.
HyperONE OS
A hardened, performance-tuned operating system replacing expensive proprietary alternatives without losing a single capability.
- RHEL-compatible binary interface
- Real-time scheduling for AI workloads
- FIPS 140-3 cryptographic modules
- Zero per-socket licensing fees
- Integrated container and KVM runtime
- Hardened by default, audit-ready
VNext Virtualisation
Enterprise-class virtualisation that meets every demand of regulated, high-density environments. No hostage taking. The VNext platform ›
- Live migration with zero packet loss
- vGPU partitioning for NVIDIA & AMD
- NUMA-aware VM placement
- VMware-compatible APIs
- Integrated SDN with EVPN/VXLAN
- Per-tenant resource isolation guarantees
In-house BMS
A vendor-agnostic Building Management System developed independently in-house. Real-time facility telemetry without lock-in to any single OEM, down to the coolant chemistry in the dry-cooler loops, sensed live by our inline fluid-quality valve.
- BACnet, Modbus, SNMP, OPC-UA native
- Real-time facility monitoring
- Integrated with Digital Twin platform
- Predictive maintenance ML models
- Cross-vendor alarm aggregation
- Open API for customer integration

One control plane. x86 and IBM Power in a single inventory, with a real guest shell and no agent to install.

A real guest shell in the browser. No client, no plugin, no agent.
Best-in-class PUE,
even at 45°C.
India's ambient temperatures peak at 40 to 45 degrees Celsius for months at a time. Conventional data center electrical designs degrade significantly above 35°C. Our patented electrical architecture maintains best-in-class PUE through the hottest peak hours, every day of the year.
Single compartmentalised system.
Hyperscale reliability.
There are systems that fail well, and systems that fail badly. We chose to build one that fails well. Compartmentalised by design. Concurrently redundant. Tier IV at every campus, by the Uptime Institute. Not the rating. The thing itself.
True compartmentalisation
Physical and electrical isolation between power and cooling trains. A single fault cannot propagate beyond its compartment. Concurrent maintenance across the entire facility with zero workload impact.
Tier IV across every site
All HyperNext campuses are Uptime Institute Tier IV. Not Tier III. Not "Tier IV-ready". Tier IV certified by the Institute, with constructed-facility certification scheduled at commissioning. No exceptions.
Concurrent redundancy
2N power, 2N cooling, multiple independent network paths, and physically diverse standby generation. Any component can fail or be taken offline for maintenance without affecting any workload.
Hyperscale-grade reliability
Designed and commissioned to the same operational standards as the top three global hyperscalers. Same fault domains, same redundancy assumptions, same continuous-availability principles.
365 days of security logs.
Free.
For three hundred and sixty-five days, we keep every security log. We don't charge extra. We don't hide it in a tier. We don't quietly rotate it at thirty days. A full year of forensic visibility, because we don't think you should have to ask.

