Blog

1.6T optical modules have moved a long way in three years: from conference demonstrations, to qualified production parts, to catalogue items you can request a quote for. That progress is real. What is not yet true is the claim that 1.6T has become an ordinary purchase for an ordinary enterprise data centre.

Those two statements get blurred together constantly, and the confusion is expensive. Buying a 1.6T module for a switch that cannot drive 200G electrical lanes, or ordering a DR8 part when the cabling plant is built for duplex LC, wastes both budget and a deployment window. This guide separates what is available, what is standardised, and what an enterprise can actually deploy today.

Are 1.6T Optical Transceivers Really Mass-Market in 2026?

"Mass market" is used loosely in optics coverage. It helps to break commercial maturity into five distinct stages, because 1.6T sits in a different stage depending on who is buying:

1. Product announcement and demonstration - a working module shown at OFC or ECOC.
2. Sampling and customer qualification - parts in the hands of switch vendors and hyperscalers for interoperability testing.
3. Volume production - repeatable manufacturing, published datasheets, defined part numbers.
4. Open commercial availability - anyone can request a quote and receive stock or a committed lead time.
5. Broad enterprise adoption - 1.6T becomes a default line item in mid-sized data centre refreshes.

As of mid-2026, 1.6T is comfortably at stages 3 and 4. Multiple vendors publish full datasheets, and third-party channels list 1.6T DR8 and FR8 parts for order. Stage 5 has not arrived. Demand remains heavily concentrated in AI training clusters and cloud back-end fabrics, where 1.6T is paired with new switch silicon rather than retrofitted into existing rooms.

The practical reading for an enterprise buyer: you can buy 1.6T today. Whether you should depends almost entirely on the switch platform and fibre plant you already own - which is exactly what the rest of this article covers. For a wider view of where the module market sits this year, our summary of what was real and what was still roadmap at OFC 2026 is a useful companion read.

What 1.6T Transceivers Can You Actually Buy?

Rather than generalising, it is more useful to look at concrete, published parts. The three below are representative of three genuinely different design approaches:

Three observations matter more than the individual specifications.

First, these are not interchangeable. OSFP and OSFP-XD are mechanically keyed so that a module cannot be inserted into the wrong cage type, as the OSFP MSA makes explicit. A 1.6T module is only useful in a port designed for it.

Second, the NVIDIA part is a twin-port design - one cage presenting two independent 800G optical ports. It is engineered for specific switch platforms and InfiniBand XDR fabrics, not as a generic 1.6T Ethernet module that any enterprise switch will recognise.

Third, on price. Claims that mass-produced 1.6T modules have dropped below a four-figure US dollar figure are, as far as published vendor and channel pricing shows, not supportable in general. Publicly listed prices for generic 1.6T DR8 and FR8 parts have remained in the four-figure range, and the parts that matter most for AI fabrics are often sold through qualified channels without public list prices at all. Treat any single price point as meaningless unless it comes with a part number, a date, a region and a quantity.

1.6T DR8, FR8 and 2×FR4: Why the Naming Actually Matters

A common error in coverage of this generation is writing "1.6T FR4". No such product designation exists. FR4 describes a four-lane wavelength-multiplexed interface; at 1.6T the real options look like this:

• DR8 - eight parallel single-mode lanes at 200G PAM4, typically around 500 m, terminated on parallel connectors such as dual MPO-12 or a single MPO-16.
• FR8 - eight lanes engineered for roughly 2 km reach, again on parallel fibre.
• 2×FR4 - two independent 800G FR4 engines in one module, each using wavelength multiplexing over a duplex LC pair, so the module presents two duplex LC interfaces and reaches about 2 km.
• 2×DR4 (twin-port) - two 800G parallel engines in one cage, each on its own MPO connector, used to break out to two 800G endpoints.

The consequence is entirely practical. DR8 and 2×FR4 both deliver 1.6 Tb/s, but one needs eight fibres on MPO and the other needs four fibres on two LC duplex pairs. They are not drop-in substitutes for each other, and the difference determines your patch panel, your polarity scheme and your spares inventory. If you are still mapping how these interfaces relate to earlier generations, our primer on optical module types covers the underlying concepts.

The practical rule when buying: never accept "1.6T module" on a quotation. Require the exact interface designation, the connector type and the operating wavelength in writing.

800G vs 1.6T: What Actually Changes

The power row deserves emphasis, because it is the figure most often misreported. InnoLight's own announcement puts its second-generation 800G modules below 14 W and its 1.6T OSFP-XD DR8+ demonstration module below 23 W - two different products, two different numbers. NVIDIA, arguing the case for co-packaged optics, describes pluggable interfaces at this class as drawing often 30 W per interface. Anyone planning rack power and faceplate cooling should budget in the 20–30 W band per 1.6T module, not the 800G figure.

Note also what the table does not say. Doubling port bandwidth halves port count only under a strict set of conditions: identical total bandwidth, a genuine one-for-one replacement, no breakout configurations, no twin-port modules feeding two endpoints, and no change in redundancy design. Break any of those and the 50% figure stops holding. Our guide to 800G OSFP deployment covers the baseline most enterprises are actually upgrading from.

Can an Enterprise Data Centre Deploy 1.6T Today?

Buying the module is the easy part. Work through these seven checks before committing budget.

1. Switch platform and SerDes generation

A 1.6T OSFP224 module needs a host that drives eight 200G electrical lanes. An existing 800G switch built on 100G SerDes cannot do this, regardless of whether the module physically fits. This single point disqualifies most in-place upgrades.

2. Cage type

OSFP and OSFP-XD are keyed to prevent cross-insertion. Confirm which cage your switch presents before ordering anything.

3. CMIS version and firmware

Current 1.6T parts are typically specified against CMIS 5.x. Host firmware must recognise the module's management interface, or the port will not come up cleanly even when the optics are correct.

4. Protocol: Ethernet or InfiniBand

Some of the highest-volume 1.6T parts target InfiniBand XDR fabrics and specific switch families. If your network is standard Ethernet, those parts are not a shortcut to 1.6T - they are a different product line.

5. Thermal design

At 20–30 W per module, faceplate density becomes a cooling problem before it becomes a bandwidth problem. Check whether your platform expects finned-top or flat-top modules, and whether the deployment is air-cooled or liquid-cooled.

6. Fibre plant and connectors

Parallel interfaces need eight fibres and MPO terminations with correct polarity; 2×FR4 needs duplex LC pairs. End-face contamination that a 400G link tolerated can produce errors at 200G per lane. This is the least glamorous item on the list and the most common cause of a failed turn-up. If you are extending or rebuilding trunk cabling for this, our MPO trunk guide for AI data centres and the MPO/MTP product range cover the connector and polarity choices involved. All of these interfaces run on single-mode fibre, so if your existing plant is multimode, see why single-mode is the required medium at these rates.

7. Standard maturity and vendor qualification

The IEEE P802.3dj task force covering 200G, 400G, 800G and 1.6 Tb/s Ethernet has been working through ballot, with completion targeted for September 2026 in the task force's adopted timeline. Shipping parts reference draft compliance. That is normal for this stage of a generation, but it means interoperability rests on vendor qualification and testing rather than on a finished published standard.

When Does a 1.6T Upgrade Make Financial Sense?

Rather than quoting a savings percentage that cannot travel between deployments, build the number yourself. The comparison that matters is delivered cost per Gb/s over the asset life, and it needs six inputs:

• Switch cost per usable port, at the port speed you will actually run
• Module cost per port, both ends, plus a spares ratio
• Structured cabling: trunks, patch panels, adapters, and any re-termination
• Module power × ports × hours × your electricity rate
• Cooling overhead applied to that power figure
• Rack space and, if relevant, the cost of avoiding an additional switch tier

1.6T tends to win when you are buying a new fabric anyway, when faceplate ports or rack space are the binding constraint, and when the workload genuinely saturates the links. It tends to lose when you are retrofitting a switch platform that would need replacing first, when links run well below capacity, or when the fibre plant would need rebuilding to support a parallel interface. Planning guidance for the surrounding infrastructure is collected in our data centre connectivity solutions overview.

Where 800G - or Even 400G - Is Still the Right Answer

Presenting 800G as a universal bottleneck is misleading. It remains the better choice in several common situations:

• Inference and mixed workloads, where east-west traffic rarely approaches link capacity.
• Storage networks, where latency and consistency usually matter more than raw port bandwidth.
• Rooms with fixed power and cooling budgets, where a 30 W faceplate is simply not viable.
• Environments needing supply stability, since 1.6T supply is tight and heavily allocated to large AI builds.
• Short-reach in-rack links, where direct-attach copper still competes on cost and power - a trade-off explored in our comparison of copper and fibre in AI data centres.

Frequently Asked Questions

Is 1.6T twice as fast as 800G?

Per port, yes: 1.6 Tb/s versus 800 Gb/s. Application performance is a different question, and depends on whether the workload was link-limited in the first place.

How much does a 1.6T transceiver cost?

There is no single answer worth quoting. Pricing varies by interface type, vendor qualification, brand compatibility and order quantity, and much of the volume moves through channels without public list prices. Ask for a dated quotation against a specific part number.

What is the power consumption of a 1.6T OSFP module?

Plan for roughly 20–30 W. InnoLight has stated below 23 W for its 1.6T OSFP-XD DR8+ demonstration module, while NVIDIA characterises pluggable interfaces at this class as often drawing about 30 W. Figures near 14 W refer to second-generation 800G modules, not 1.6T.

What is the difference between 1.6T DR8 and FR8?

Both use eight lanes over single-mode fibre. DR8 targets data-centre reach in the region of 500 m; FR8 is engineered for roughly 2 km. Confirm the exact figure on the datasheet, since implementations vary.

Can a 1.6T OSFP module work in an 800G switch?

Generally no. The host must supply 200G-class electrical lanes and a compatible cage and management interface. Physical fit is not the same as electrical compatibility.

Is 1.6T Ethernet standardised?

Not yet finalised. IEEE P802.3dj covers 1.6 Tb/s Ethernet and has been progressing through ballot with completion targeted for September 2026. Products currently ship against draft compliance plus vendor interoperability testing.