Inside the rack
Scale-up
GPUs and switches within the rack. Passive DAC where the run is short; ACC and AEC when the rack gets dense or the run gets long.
DAC · ACC · AEC
Sovereign AI data centres
Sovereignty is decided in the cable plant.
A sovereign AI cluster means the compute, the data and the operator sit under Indian jurisdiction — so a regulated Indian business can train and run models without its data leaving the country’s legal reach. That is a procurement and engineering problem long before it is a legal one, and most of it lands on the links between the GPUs.
A hall like this is mostly cable.
The GPUs get the budget and the headlines. What decides whether they reach their benchmark is the metre of copper or glass between them.
- At 400G and abovethe margin between a link that works and one that retrains under load is small
- And it is set byinsertion loss, the reach of the optics actually fitted, the temperature at the top of the rack at 3am
- None of whichappears in the bill of materials
Sovereign is a supply chain, not a sticker.
If every spare optic is a six-week import and every fault needs an engineer in another time zone, the operator is holding risk that was supposed to be designed out.
We do the India side.
Architecture, design-in, integration, qualification and support — on local time, close to the clusters being built.
Illustrative visualisation. Not a Praverse facility — we supply and design in the interconnect, we do not operate data centres.
The test
Four questions an operator should be able to answer.
“Sovereign” is doing a lot of marketing work at the moment. These are the questions that separate the ones who mean it from the ones who have put an Indian flag on a slide.
- 01
Where does the data physically sit
Not where the contract is signed — which racks, in which building, under whose operating licence.
- 02
Who can be compelled to hand it over
The operator’s jurisdiction decides this, and it is not always the same as the customer’s.
- 03
How long is the supply chain
A cluster that cannot be repaired without a six-week import is sovereign on paper and fragile in practice.
- 04
Who answers at 3am
Support in the same time zone is not a nicety at this scale; it is part of the availability figure.
The fabric
Your GPUs are ready. Is the interconnect?
As clusters scale, the hardest decisions move to the links between everything — GPUs, NICs, switches, servers, the whole rack. Get them wrong and you rebuild, burn power, or watch a deployment window disappear. Get them right and everything downstream gets easier.
Between racks
Scale-out
Rack to rack and row to spine. Retimed copper to about 9 m, then AOCs and pluggable optics, with LPO where power matters.
AEC · AOC · SR / DR optics
Across the fabric
Scale-across
Between pods, halls and buildings. Single-mode DR and FR optics to 2 km, with coherent optics for the longest links.
DR / FR optics · coherent
Reach
Match the medium to the distance at every hop. Over-engineering every link wastes money; under-engineering one link means a rebuild.
Power
Every watt moving data is a watt not spent on compute. Copper where reach is short; low-power LPO optics where distance demands glass.
Qualification
Weeks spent validating links are weeks closer to a missed deployment window. Retimed copper with proven link stability shortens bring-up.
The market
The capacity is being built now.
India’s operational data-centre capacity is concentrated in a handful of markets, and our people are in four of the five largest — which matters, because interconnect is a job somebody has to do on site.
- Mumbai & Navi MumbaiPraverse team790 MW
- ChennaiPraverse team305 MW
- BengaluruPraverse team182 MW
- Hyderabad152 MW
- Delhi NCRPraverse team76 MW
Vadodara, Gujarat — headquarters, about 400 km from Mumbai. Not a capacity market, which is rather the point: the engineering sits close enough to reach the clusters, and the people sit in them.
Operational data-centre capacity by market. Source: MeitY (actuals, March 2026); Savills India (July 2025); Colliers (May 2025); Wood Mackenzie via IANS (July 2026). Forecasts vary widely.
Where it bites
Three situations we are built for.
Written as the shape of a real decision rather than as case studies: the question that gets asked, the constraint that turns out to be the real one, and what changes.
These are illustrative scenarios, not customers. We do not publish a client’s deployment without their written permission, and we will not dress a hypothetical up as a reference.
By workload
Don’t design around the cable. Design around the cluster.
Training clusters
Reach, power and density hit their limits at once. The fabric is designed around the GPUs before the rack is built, and today's choices are made with the next generation in view.
- Reach
- Power
- Density
- Interoperability
1.6T today. 3.2T on the roadmap.
Inference
Little room for wasted power or latency. Link distance, data rate, latency, power budget, port density, thermals and the deployment environment are all sized before a part is chosen.
- Power
- Latency
- Reach
- Density
Keep the power on compute.
At national scale
A great design is only the beginning. The interconnect has to qualify, repeat, ship and perform the same way across thousands of units — through architecture, design-in, production and delivery.
- Qualification
- Yield
- Supply
- Responsiveness
Same-day answers, not same-quarter.
800G in practice
Every way a switch port meets a ConnectX-8.
The ConnectX-8 SuperNIC brings 800G to the server on PCIe Gen6, in a single-OSFP card (C8180) and a dual-QSFP112 card (C8240). Against an 800G switch such as the Spectrum SN5600 there are four Ethernet configurations — and a DAC, ACC, AEC, AOC or LPO optic for each, chosen by the distance.
| Link | Result | |
|---|---|---|
| Switch → 1 × CX8 C8180 | 800G OSFP to OSFP-RHS | One 800G link, or 2 × 400G |
| Switch → 2 × CX8 C8180 | 800G OSFP broken out to 2 × 400G | Each leg to its own NIC in 400G mode |
| Switch → 1 × CX8 C8240 | 800G OSFP to 2 × QSFP112 | Both cages on one NIC, full 800G |
| Switch → 2 × CX8 C8240 | 800G OSFP to 2 × QSFP112 | One 400G cage per NIC |
From 224G to 448G per lane
Doubling the lane rate means new signalling, stronger error correction and moving the connection closer to the silicon. Copper keeps up — retimed 1.6T copper already runs 3 m at 18–20 W per end, against about 22 W for a 1.6T optic over 500 m — but where copper ends and optics begins keeps moving. The job is to measure that boundary for each build, not argue where it should be.
What we supply
The layer underneath.
High-speed copper and optical interconnect for AI clusters, with our USA technology partner building the technology and Praverse doing the Indian architecture, design-in, integration, qualification and support.
We do not manufacture in India today. Local testing and qualification come first; assembly follows when volumes justify it. Phases 2 and 3 below are plans, not facilities.
PHASE 1 · NOW
Supply and design-in
Interconnect supplied with our US technology partner, with India-side programming, testing, labelling and qualification, and support on local time.
PHASE 2 · NEXT
Semi-finished assembly in India
Assembly of semi-finished interconnect in India, close to the clusters being built.
PHASE 3 · PLANNED
Indian manufacturing
An Indian manufacturing facility for copper interconnect first, then optical, as volumes justify it.
Building one?
Send us the topology and the distances. You get a written plan within two business days — link budgets, breakout plan, and a parts list that still makes sense at the tenth pod.