Direct attach copper
DAC
Passive, in-rack. No added latency.
800G to 4 m · 1.6T to 1 m
Copper inside the rack.
Passive and active copper where the distance is short and the budget is power.
Optics across the hall.
Fibre from the row to the spine, where copper runs out of reach.
400G → 800G → 1.6T.
800G is mainstream for AI clusters today. 1.6T is ramping.
Photonics & AI data centres
The layer every AI cluster runs on.
An AI cluster is only as fast as the links between its GPUs. We are India’s partner for the copper and optical interconnect that carries that traffic — supplying and designing it in today, and building the capability to manufacture it here.
Reach
Copper is cheaper. Optics goes further.
Typical maximum reach by media type. Indicative figures — actual reach depends on speed, cable construction and the link budget you design to.
- DACPassive copper0 m
- ACCLinear redriver0 m
- AECRetimer / DSP0 m
- AOCActive optical0 m
- Single-modeAcross the hall0 m
Log scale · copper fades at its reach · optics keeps going
Inside an 800G module
What is actually in there.
- HousingThe metal shell. Grounds the module and holds everything in line.
- Heat sinkPulls heat out. At 800G a module runs hot enough that this sets the limit.
- PCBCarries the high-speed lanes from the electrical edge to the optics.
- DSP / retimerCleans up the signal and re-times it, which is what buys the extra reach.
- Optical engineLasers and detectors. Turns electrical lanes into light and back.
- FibreCarries the light out of the module and across the hall.
Illustrative. Not a specific vendor’s product.
The right answer at every reach
Copper where it wins. Optics where it counts.
Scale-up inside the rack, scale-out between racks, scale-across the fabric. Every link gets the medium its distance and power budget call for — and nothing it does not.
< 5 W
800G multimode LPO, per end
Linear pluggable optics, no DSP. In mass production.
9 m
800G on copper
Retimed AEC. Up to 4 m at 1.6T.
1.6T
2×DR4 / DR8 LPO
Silicon photonics, 500 m single-mode.
1M+
Units shipped in 2024
Automated lines, 150 units an hour at over 98% yield.
| Built on | 800G | 1.6T | Where it goes | |
|---|---|---|---|---|
| DAC | Passive copper | 4 m | 1 m | In-rack, GPU to switch. Lowest power and cost per link, no added latency. |
| ACC | Linear redriver | 5 m | 3 m | Adjacent racks and xPU-to-xPU, keeping copper's cost and power. |
| AEC | Retimer / DSP | 9 m | 4 m | Row-level copper. The reach that keeps a rebuild out of the plan. |
| AOC | VCSEL optics, LPO or retimed | SR 100 m | — | Rack-to-rack and across the row, where copper should stop. |
| Pluggable optics | Multimode and single-mode | SR 100 m · DR 500 m · FR 2 km | DR8 500 m · FR8 2 km | Leaf to spine and across the hall. Reconfigurable. |
| Coherent | Coherent optics | Coming | Coming | Longest reach, between halls and buildings. |
How far, how fast? Find your starting point.
Start from
ACC Active copper cable
A linear redriver extends copper past DAC reach while keeping copper's cost and power.
Also reaches 5 m: AEC · AOC (LPO) · SR8 LPO optics
Get a written link planIndicative, from published maximum reach. Your link budget, cable build and thermals decide the final answer.
Portfolio
Six families, one supply chain.
Supplied through our USA technology partner.
Active copper cable
ACC
Linear redriver, for xPU-to-xPU links.
800G to 5 m · 1.6T to 3 m
Active electrical cable
AEC
Retimer / DSP. Copper at row reach.
AEC: up to 9 m at 800G (partner specification)
Active optical cable
AOC
Optical reach inside the hall, LPO or retimed.
200G – 800G · to 100 m
Multi-mode and single-mode
Optical transceivers
LPO and full-retimed, SR to FR — 67 models.
400G · 800G · 1.6T
Structured fibre
Fibre cable assemblies
Trunks, breakouts, shuffle boxes and fibre array units.
Rolling out
Form factors
- SFP56
- QSFP56
- QSFP112
- QSFP-DD
- OSFP
- OSFP-RHS
- OSFP-XD
Ethernet and InfiniBand.
Speeds
- 400GInstalled base
- 800GMainstream for AI clusters
- 1.6TRamping
- 3.2TNext
Indicative direction, not dated commitments.
Linear pluggable optics
Take the DSP out, and the watts go with it.
Why linear optics
An LPO module drives the laser straight from the switch SerDes, with no DSP or CDR in the module. That takes out the part that burns the most power and adds the most latency, and lowers the cost per port.
What it draws
Under 5 W per end at 800G multimode and about 3 W at 400G. Across a few thousand ports, the watts a DSP would have burned go back to the GPUs.
When to use retimed instead
Full-retimed (DSP) optics stay the safer choice where the host SerDes is marginal or the link budget is tight. The catalogue carries both, so the choice is made per link rather than per vendor.
| Models | 67 pluggable transceivers · 16 AOCs · 27 each of DAC, ACC, AEC |
|---|---|
| Data rates | 400G · 800G · 2×400G · 1.6T |
| Lane rates | 50G, 100G and 200G per lane (PAM4) |
| Architectures | Full-retimed (DSP) and linear pluggable (LPO) |
| Reach classes | SR4 / SR8 to 100 m · DR4 / DR8 to 500 m · FR4 / FR8 to 2 km |
| Form factors | OSFP (flat-top, RHS, closed-top) · QSFP-DD · QSFP112 · QSFP56 · breakouts |
| Cooling | Air, liquid and immersion, single- and two-phase coolant |
| Management | CMIS 4.0 and 5.x · Class 1 laser |
Interoperability
Proven on the platforms your fabric runs on.
The best specification does not matter if the ecosystem cannot keep up. Every design is qualified on current test systems and for interoperability on production switches, by third parties as well as in the lab.
Switches
Broadcom Tomahawk 6 platforms, NVIDIA Spectrum Ethernet and Quantum InfiniBand switches.
NICs and DPUs
NVIDIA ConnectX-8 800G SuperNIC, ConnectX-9 and BlueField-4, in Ethernet and InfiniBand.
Multi-vendor plugfests
Passed CMIS 5.3 compliance and interop at the 43rd InfiniBand and RoCE Plugfest (IBTA / UNH-IOL), at 800G and 1.6T.
Test equipment
Links proven on Keysight and VIAVI test systems — 800G copper at 9 m holding pre-FEC BER better than 1E-8.
Standards
- IEEE 802.3dj
- IEEE 802.3ck
- IEEE 802.3cd
- OSFP MSA
- QSFP-DD MSA
- CMIS 5.x
- InfiniBand NDR · HDR · XDR
- Ethernet 400G / 800G
- OCP
- RoHS · REACH
Loopback testers
31 loopback modules from 25G to 1.6T, SFP56 through OSFP-XD — passive at 0 W, or active up to 60 W to load a port thermally the way a real module does. The fastest way to qualify a switch before the optics arrive.
Design to volume
Proven designs, made to work in your environment, at scale.
Design is part of the work, but it is not the point. The point is everything that turns a design into a qualified product running in your cluster — and the capacity to deliver it in volume.
- 01
Choose
Start from a proven design on leading silicon, matched to the reach, power and timeline of your cluster — not a catalogue to sort through.
- 02
Guarantee
The design is stood behind. When a build genuinely needs something custom, the engineers design and qualify it.
- 03
Prove
Make it work in your environment before deployment: signal integrity, link margin, thermals and interop, on your switches.
- 04
Scale
Once qualified, it is manufactured in volume — from ten units to the millions — on your timeline.
Automated
AI-driven production and test lines. 150 units an hour at over 98% yield on the copper lines.
Elastic
From ten units to the millions the largest deployments need, with capacity stood up in weeks rather than quarters.
Geodiverse
Production replicated across North America and Asia, so no single region is the one your supply depends on — and country-of-origin requirements can be met without re-engineering.
Engineering behind the link
- Signal integrity
- Power integrity
- Thermal integrity
- Optical design
- Optoelectronics packaging
- High-speed twinax
- Advanced firmware
- Testing at 800G · 1.6T · 3.2T
India
The capacity is being built now.
Gigawatts of installed capacity
MeitY (2020 and 2025 actuals, Mar 2026); Savills India (Jul 2025); Colliers (May 2025); Wood Mackenzie via IANS (Jul 2026). Forecasts differ widely — treat 2030 as a range.
- Mumbai790 MWteam
- Chennai305 MWteam
- Bengaluru182 MWteam
- Hyderabad152 MW
- Delhi NCR76 MWteam
- VadodaraPraverse HQ
Operational capacity. Source: MeitY reply in Lok Sabha (Mar 2026). Schematic layout — positions are indicative, not a survey map.
Interconnect is a local business: someone has to be on site when a cluster is qualified, and on the phone in the same time zone when it is not. Four of those five markets are cities where we have team members.
Our model
Building India’s AI interconnect capability — in three phases.
Our USA technology partner builds the interconnect. Praverse does the Indian architecture, design-in, integration, qualification and support, for sovereign AI data centres.
Each phase is unlocked by customer demand, so capacity arrives when India’s AI build-out needs it — not before.
Our team works from the cities that hold most of India’s data-centre capacity, close to where deployments happen. Vadodara to Mumbai is about 400 km.
Data stays under Indian law
Built around the DPDP Act 2023.
Resilient domestic supply
Design-in and qualification happen in India.
Engineers and support in India
On the same clock as you.
Roadmap
Where this goes.
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.
Phases 2 and 3 are plans, not facilities. We will not show you a factory we do not have, and we will not quote a location, a capacity or a date until one is signed.
Designing a cluster?
Send us the topology and the distances. You get a written plan within two business days.