800G and 1.6T Optical Modules Explained: Form Factors, Power, and When You Need Them
800G and 1.6T optical modules, explained in plain terms: OSFP and QSFP-DD form factors, DR and FR reach classes, and the power trade-off between DSP and LPO — plus how to tell which generation your AI fabric actually needs.
On this page
The short answer
An 800G optical module is a pluggable transceiver that moves 800 gigabits per second - usually as eight 100-gigabit lanes - between a switch or accelerator port and a fiber; a 1.6T module doubles that to 1.6 terabits using eight 200-gigabit lanes. The module generation you need is set by your switch generation: 51.2T switches take 800G modules, 102.4T switches take 1.6T modules, and the two are not interchangeable.
That one constraint simplifies most buying questions, but it leaves the ones that actually burn money: which form factor, which reach class, and whether the module carries a power-hungry DSP or a leaner linear drive. Get the reach class wrong and you pay for kilometers of capability inside a 30-meter hall; get the power architecture wrong and thousands of modules quietly add a megawatt to the facility. This guide walks the lanes, form factors, reach classes, and power options in plain terms.
It is the third article in our optical interconnect series, following what CPO is and copper versus optical interconnect - the article that explains why these modules exist at all, since everything beyond a few meters of copper lands on them. For market context on the hardware cycle behind these generations, see our GPU Market Watch.
Key facts
- Lane math: 800G = 8 lanes x 100G (the common build) or 4 x 200G (emerging); 1.6T = 8 lanes x 200G. Lane rate, not the headline number, is what new silicon has to deliver. (Industry-standard configurations, as of October 2026.)
- Switch pairing: a 51.2T switch ASIC provides 64 ports of 800G; a 102.4T ASIC provides 64 ports of 1.6T. Modules and switches upgrade together, generation by generation.
- Form factors: OSFP and QSFP-DD are the two volume 800G form factors; OSFP224 is the common 1.6T shape. All are hot-pluggable and field-replaceable.
- Reach classes: DR-class about 500 m, FR-class about 2 km, LR-class about 10 km, over single-mode fiber. AI halls overwhelmingly deploy DR-class.
- Power: a DSP-based 800G module typically draws roughly 15-20 W; linear-drive (LPO) equivalents remove the DSP and several watts per module; 1.6T DSP modules draw more per module but less per gigabit than 800G.
- Count the ends: every optical link uses two modules. A 10,000-link fabric is a 20,000-module purchase, power draw, and sparing pool.
- Status as of October 2026: 800G is the volume shipping generation in AI fabrics; 1.6T is ramping with 102.4T switch platforms.
It is lanes all the way down
Nobody builds a single 800-gigabit laser beam. A module's headline speed is a bundle of parallel lanes: the electrical side talks to the switch in lanes (eight lanes of 100G for mainstream 800G), and the optical side maps those lanes onto fiber - either one wavelength per lane, or several wavelengths sharing a fiber pair, depending on the reach class. When the industry says "200G lanes," it means every element in that chain - switch SerDes, module electronics, modulators and lasers - has doubled its per-lane signaling rate, using PAM4 modulation pushed to roughly 106 gigabaud and beyond.
This is why module generations feel like platform generations. A 1.6T module is not an 800G module with the dial turned up; it needs a 200G-capable switch ASIC on one side and 200G optics on the other, and it will not link with 800G equipment at 1.6T no matter how similar the cage looks. Plan module purchases as part of the switch purchase, because that is what they are.
Form factors: OSFP and QSFP-DD
The form factor is the module's physical shell and connector standard - it decides what fits in the switch's cages. Two standards carry essentially all 800G volume. OSFP is the slightly larger shell, with an integrated heat-sink design that suits high power draw; it is the common choice in AI and InfiniBand fabrics. QSFP-DD is the smaller, double-density evolution of the QSFP family, backward-compatible with older QSFP modules in the same cage, and common in large Ethernet cloud fabrics.
For 1.6T, OSFP224 - an OSFP shell wired for 200G lanes - is the shape most platforms have converged on, with QSFP-DD variants following. The practical guidance is dull but reliable: your switch vendor has already chosen, the fabric should use one form factor consistently, and the genuinely important module decisions are reach class and power architecture, which come next.
Reach classes: DR, FR, and beyond
Reach classes describe how far a module type is specified to carry its signal over single-mode fiber, and they are the easiest place to waste money, because longer-reach modules cost more and the surplus reach is invisible. DR-class modules are specified to about 500 meters - the datacenter class, using one wavelength per lane across eight fiber pairs or an equivalent parallel design. FR-class reaches about 2 kilometers, typically by putting multiple wavelengths on one fiber pair. LR-class reaches about 10 kilometers, for links between buildings or sites.
Inside an AI data hall, switch-to-switch runs are almost always well under 500 meters, so DR-class modules do nearly all the work - including the spine layer. FR appears at building exits and campus links; LR between sites. A useful audit of any cluster BOM: count modules whose reach class exceeds the longest actual run in the building. Every one of them is paying a premium for distance it will never cover. (Separate from single-mode classes, short multimode SR-class modules exist for very short, lower-speed runs; at 800G and above, AI fabrics have largely standardized on single-mode.)
Power: DSP, LPO, and the road to CPO
The module's internal architecture decides its power draw. A conventional module carries a DSP - a digital signal processor that cleans up the electrical signal from the switch, drives the optics, and recovers the incoming signal. The DSP is what makes modules forgiving: it tolerates varied switch boards and longer channels. It is also most of the module's roughly 15-20 W at 800G, and it adds latency.
LPO (linear pluggable optics) removes the DSP and lets the switch SerDes drive a simpler linear optical engine directly. Savings run to several watts per module - across 20,000 module ends, a facility-scale number - with lower latency as a bonus. The trade is discipline: with no DSP to rescue the signal, the switch-to-module channel must be short, clean, and validated as a pair, and interoperability testing matters more. CPO is the endpoint of the same logic - remove the module's electronics entirely by moving conversion onto the switch package - covered in depth in our CPO guide. Read the three as one progression: DSP modules maximize flexibility, LPO trades some flexibility for power, CPO trades the most for the least power per bit.
For 1.6T the same menu applies at higher numbers: DSP modules draw more watts per module than 800G equivalents, but roughly half the watts per gigabit. When comparing quotes, always normalize to watts and dollars per gigabit at the fabric level - per-module figures make the newer generation look worse for delivering twice the bandwidth.
Which generation do you actually need?
- Your switches are 51.2T (64 x 800G): 800G modules throughout. This is the volume, best-priced, multi-vendor position in October 2026.
- Your switches are 102.4T (64 x 1.6T): 1.6T modules, with OSFP224 the default shape; expect pricing and second sources to be younger than 800G's.
- You are speccing a fabric not yet bought: size the bandwidth first, per GPU generation and model shape - our AI cluster bandwidth estimator does this from GPU count and parallelism choices - then let the switch quote set the module generation, and audit reach classes before signing.
- You rent GPU capacity: modules are your provider's problem, but the same generations explain the bandwidth and power claims in their offers; the GPU cost calculator helps compare what those offers mean per hour.
Common mistakes
- Buying reach you do not use. FR or LR modules on 30-meter hall runs are pure waste; measure the longest run, then buy the class that covers it.
- Mixing form factors or vendors link-by-link. Both ends of a link should be the specified pair; "800G" on two datasheets does not guarantee a validated combination, especially with LPO.
- Comparing per-module power instead of per-gigabit. 1.6T modules draw more and deliver double; normalize before concluding anything.
- Forgetting the second end. Budgets that count links instead of module ends understate optics cost, power, and sparing by exactly half.
- Treating modules as an afterthought to the switch order. Modules are commonly a third or more of total fabric cost. They deserve the same negotiation and validation effort as the switches themselves.
Next in this series: how much bandwidth an AI cluster actually needs, and the power bill of the interconnect itself. Previous articles: copper versus optical interconnect and co-packaged optics explained.
Frequently asked questions
What is the difference between 800G and 1.6T optical modules?
Bandwidth per module: 800G carries 800 gigabits per second, 1.6T carries twice that. The practical difference is the lane generation underneath - most 800G modules run eight 100G lanes, while 1.6T modules run eight 200G lanes, which needs newer SerDes, newer lasers and modulators, and draws more power per module while delivering more bandwidth per watt and per switch port.
OSFP or QSFP-DD: which form factor should an AI fabric use?
Both carry 800G and both are widely deployed. OSFP is slightly larger with an integrated heat sink and is common in AI and InfiniBand fabrics; QSFP-DD is smaller, backward-compatible with the huge QSFP installed base, and common in Ethernet cloud fabrics. The honest answer is that the switch you buy decides: module choice follows the platform, and mixing form factors inside one fabric is what to avoid.
What do DR, FR, and LR mean on an optical module?
They are reach classes. DR (datacenter reach) covers about 500 meters on single-mode fiber for intra-building runs, FR (fiber reach) covers about 2 kilometers, and LR (long reach) covers about 10 kilometers. In an AI data hall, most fabric links are far shorter than 500 meters, so DR-class modules do most of the work; buying FR or LR for in-hall runs pays for reach you never use.
What is an LPO module, and why does it use less power?
A linear pluggable optics (LPO) module removes the digital signal processor (DSP) chip and lets the switch's own SerDes drive the optics directly through a simpler linear driver. Removing the DSP removes several watts per module and some latency, at the cost of a tighter electrical channel between switch and module - the switch and module must be validated together, and reach is shorter than DSP-based equivalents.
Do I need 1.6T modules for an AI cluster today?
Only if your switches are 102.4T-generation platforms with 200G SerDes; 1.6T modules physically do not run in 800G switch ports. Most clusters deploying in 2026 on 51.2T switches use 800G modules throughout the fabric. The right sequence is to size the fabric bandwidth first - our bandwidth estimator does this from GPU count and model shape - and let the switch generation, not the headline number, set the module generation.