400G Optical Modules Suit Datacenters

 

Every time I think the datacenter world might calm down for a second, someone walks in and says something like, "Hey, why aren't we doing 400G everywhere yet?" as if ripping and replacing half the network is as easy as swapping a power strip. But yes - 400G is here, and whether we like it or not, we're dealing with it.

 

 

A Not-So-Linear Upgrade Path

 

People talk about networking speeds as if they scale predictably, like "10G → 40G → 100G → 400G → 800G."
That is absolutely not how it feels when you're actually the one plugging these things in at 3 AM, hoping the link LEDs blink the way they're supposed to.

I still remember a 2019 deployment where the idea of a 400G backbone made half the team stare at the floor because none of us trusted the early QSFP-DD thermals. The modules got hot enough to make you question your life choices. One guy actually kept a thermal camera on his desk for weeks just to watch them.

Anyway-yes, 400G matters. But not because a marketing slide says so. It matters because traffic spikes no longer surprise anybody. AI workloads are chewing through bandwidth in a way that makes traditional East-West traffic patterns look like a warm-up exercise.

 

Formats, or "Which Rectangle Ruins Your Budget Less?"

 

If you've worked with physical hardware long enough, you eventually develop strong opinions about module shapes. I personally never liked CFP8. It's too big. It feels like someone melted a VHS tape into a transceiver. We used a few in a backbone router upgrade, and those things ran warm enough that we actually put off the rest of the line cards until facilities improved airflow in that aisle.

QSFP-DD became the "fine, we'll go with this" option mostly because vendors rallied behind it. OSFP is nice too - I actually prefer the thermals - but unless your switch vendor is on board, you're not using it. Datacenter infrastructure is 40% technology and 60% politics; nobody tells you this until your bill of materials gets rejected for the third time.

 

 

Multimode vs. Single-mode (And the Fiber Graveyards We All Pretend Don't Exist)

 

I've done enough datacenter audits to know one universal truth: fiber plants are never as clean as the documentation claims.
Half the time, the labels don't match reality. Sometimes the patch panels have mysteries from the previous decade.

This absolutely affects 400G.

SR8 over multimode? Great - on paper. In the real world, you blow on an MPO connector the wrong way and suddenly lane 6 looks like it's dying. I once spent an entire afternoon cleaning connectors only to eventually discover the installer had pinched the cable trying to close a rack door. That was fun.

DR4, FR4, LR4 - yes yes, the distances are in the brochures. But price is the real differentiator. One of our clients switched half their planned DR4 links back to AOCs simply because the optics budget exploded past procurement limits. Practicality beats elegance every time.

 

Power, Heat, and Other Things That Burn Your Hands

 

Let me be blunt: 400G runs hot.
Not "warm."
Not "slightly toasty."
Hot enough that the first time someone hands you a module freshly yanked from a running switch, you reconsider your friendships.

Some vendors claim sub-8W optics now. I'll believe it when I can touch one without making a noise.

Cooling becomes a real architectural decision. In one deployment, our hottest aisle climbed nearly 4°C after we populated the last 8 ports on a 1RU switch with LR4s. Facilities blamed networking. Networking blamed facilities. In the end we just moved two PDUs and rotated a rack. The problem disappeared, nobody talked about it again.

 

 

The Technology Jump (PAM4 Isn't Magic, It's Just… Fine)

 

The shift from NRZ to PAM4 gets talked about like it was a grand revolution. It wasn't. It was simply the least painful way to cram more bits down the same pipes.

PAM4 is noisy.
It's fussy.
It demands better DSPs, better PCBs, better everything.

But it works, and once vendors ironed out the equalization issues, the practical problems mostly faded. Mostly. I've still seen DR4 modules that refuse to lock at full distance until the fiber is re-terminated. No idea why. Worked after rework. That's how optics goes sometimes.

 

The Modules You Actually See in the Field

 

If you've spent time in a real datacenter - not a slide deck - you've probably encountered these, in roughly this order of frequency:

400G-SR8

Cheap(ish), predictable. MPO connectors that collect dust like magnets. Good for short runs, good for labs. Bad if you don't have proper cleaning tools.

400G-DR4

I see this the most. Fits the typical "few hundred meters but not kilometers" use case. We once used DR4 to extend a link between two buildings because trenching new conduit got blocked by city permits. Sometimes the simplest answer is just buying more expensive optics.

400G-FR4

Better reach, CWDM lanes, dual LC. I like these but they're pricier. Most enterprises won't touch them unless they absolutely need the 2km.

400G-LR4

These are like fire extinguishers: important, expensive, and you hope you never need them.

Everything else exists somewhere, somehow, but I've never seen anyone excited about PLR4 or ER8 outside of telco-heavy environments.

 

The Money Part Nobody Likes Discussing

 

Procurement departments get nervous when they hear "400G."
Like visibly nervous.

SR8 can be affordable. DR4 even more so if you've got a vendor who doesn't gouge you. But LR optics still cost enough that you hesitate before ordering them in quantity.

AOCs and DACs? Sure - for top-of-rack or if your switches are hugging each other in the same rack. We had one deployment where the entire leaf layer sat so close the longest cable was 4 meters, so DACs saved us thousands. Other times, AOCs are band-aids for poor fiber planning.

 

 

Deployment Reality (Not the Clean Version You Hear at Conferences)

 

I'll keep this short:
400G exposes every weakness in your fiber plant.
And your labeling.
And probably your documentation.

I've seen:

cross-connected panels mislabeled since 2015

MPO cassettes installed backwards

cleaners used incorrectly (yes, this happens)

optics failing because someone bent the cable at the wrong radius

None of this shows up in vendor webinars. But in real deployments, these little imperfections eat entire weekends.

 

And About That 800G Talk…

 

It's coming. Sure. But the idea that 800G will "replace" 400G soon is nonsense. Most enterprises still have 10G islands that refuse to die. Hyperscalers move fast because scale demands it, not because technology wants them to.

400G will stay with us for years. Frankly, the amount of money tied to existing infrastructure all but guarantees it.

 

Final Notes Before I Go Back to My Lab

 

400G isn't elegant. It isn't painless. It isn't the clean, glossy upgrade people imagine when they see new hardware announcements.

But it works. And once you survive your first deployment - including the sweating over fiber, the heat, the occasional mysterious link flaps - you end up appreciating how far the industry has come.

If you're planning for 400G in a new build, good. Just don't forget the boring stuff: cleaning kits, proper airflow, someone who actually knows how to terminate MPO correctly, and at least one person on your team who can stay calm when a link refuses to come up for no apparent reason.

That last one is harder to find than good optics.