Author: Marcin Bała, MSc Eng., Chief Technology Officer
Published: June 2026 | Updated: May 2026 | Reading time: 10 minutes
TL;DR — key takeaways
Integrator recommendation in 60 seconds
Below 50,000 GPU ports — always 800G. No economic or technical justification for 1.6T in 2026.
5-year TCO: 800G DSP ~€89,400, 1.6T ~€305,000 per 100 ports. Difference: over €215,000.
Three mistakes that cost money: no cooling plan, LPO without ASIC verification, ignoring lead time.
Order 3 months early. Market is tight — especially LPO and 1.6T.
Your Decision in 90 Seconds
One question puts everything in order: how many GPU ports do you have or plan to have over the next 24 months?
Below 50,000 GPU ports — always 800G
800G OSFP Gen 2 DSP or LPO — no economic or technical justification for 1.6T in 2026
Enterprise DC, colo, first AI workloads in Poland and CEE
10,000–50,000 ports, new buildout
800G LPO where switch is LPO-ready, DSP where it is not
1.6T pilot on 2–5% of ports as proof-of-concept for future migration
Above 50,000 GPU ports — AI factory
1.6T conversation is justified — but only after verifying switch ASICs, cooling plan, and EML availability
Checklist — 10 Questions Before Signing the Order
01
Which generation of 800G?
Require Gen 2 — 4×200G optical, 8×112G electrical. Gen 1 (SR8/multimode) being phased out. Gen 3 entering the market, not yet available in all models.
Critical
02
Which switch ASIC and does it support CMIS 5.x?
CMIS 5.x required for telemetry. Without it = no visibility in DCIM. Confirmed in Poland: Arista 7050X4, Cisco NCS 5500, Juniper QFX10000, Nokia 7750 SR.
03
SM or MM?
DR4 = 500 m SMF. SR8 = 50–100 m MMF. FR4 = 2 km SMF. Match to existing cable infrastructure. Changing fiber type is a separate project — verify physically.
04
Full DSP or LPO?
LPO saves 25–30% energy and 15–25% on module price. Requires LPO-ready ASIC: Broadcom Sian2, Marvell Teralynx 10, Nvidia Spectrum-4. Check compatibility matrix before ordering.
Check compatibility matrix
05
What is the vendor’s lead time?
800G Gen 2 from stock: 2–4 weeks. Specialist LPO: 4–10 weeks. 1.6T and non-Nvidia EML: 8–20 weeks. With a hard deadline: order 3 months early.
Critical
06
Does the vendor guarantee interoperability in writing?
Request written confirmation of compatibility with the specific switch model and firmware version. A good integrator verifies this before delivery.
07
What is the SLA for module replacement on failure?
A production DC requires next-business-day replacement SLA or better. Check whether the vendor holds spare modules in Poland or the CEE region.
08
Is the rack cooling ready for 16–17 W per port?
800G DSP at 32 ports per switch generates 512 W from optics alone. For 1.6T (30 W per port): check rack thermals before ordering.
09
Are you planning migration to 1.6T within 3 years?
If yes: choose switches with 200G/lane ASIC (Broadcom Tomahawk 6, Marvell Teralynx 10) and plan cooling for 30 W per port now. 800G modules now — 1.6T migration 2027–2028.
10
Has anyone audited the infrastructure before selecting modules?
Wrong module selection is the most common cause of problems at launch. The audit covers: switch ASIC verification, rack thermal assessment, firmware verification.
Critical
Three Mistakes That Most Often Cost Money in DC Deployments in Poland and CEE
01
Buying 1.6T modules without a cooling plan
Modules arrived. Existing racks could not handle 28 W per port at 32-port density. Project stalled for 6 weeks to redesign the cooling layout. Cost of delay: a multiple of the module price.
Verify thermals before ordering — not after launch.
Most common mistake
02
Ordering LPO without verifying the switch ASIC
The client ordered 800G LPO modules because they were $200 cheaper per unit. The switch turned out to be too old to support LPO. Modules went back to the vendor — 4-week delay and the full cost of re-ordering.
Checking the compatibility matrix takes 10 minutes and saves weeks.
Costly mistake
03
Ignoring lead time in the project schedule
"We’ll order when we’re ready" — the specific model available in 14 weeks, DC launch planned in 8. Schedule had to be rebuilt. The market is tight — especially for LPO configurations and 1.6T.
Always order with a 3-month buffer.
Scheduling mistake
Key 800G Module Vendors Available Through the Integrator (Q1 2026)
Innolight
Modules: 800G OSFP DR4/FR4
Price: $700–850
Lead time: 2–6 weeks
Certification: Arista and Cisco
Mass production
Eoptolink
Modules: 800G OSFP DR4 and LPO
Price: $720–880
Lead time: 2–6 weeks
Advantage: LPO for selected models
Good price/quality ratio
AOI (Applied Optoelectronics)
Modules: 800G OSFP
Price: $750–900
Lead time: 2–6 weeks
Ref.: NA hyperscaler, Dec 2025
Proven quality
Coherent Corp.
Modules: 800G OSFP and 800ZR+
Price: $820–950
Lead time: 2–6 weeks
Certification: Cisco and Nokia
Coherent DCI
Contact Us!
FAQ — 800G or 1.6T?
An 800G OSFP DR4 module costs $700–900 in Q1 2026, draws 15–17 W per port, and works with existing switches. A 1.6T OSFP DR8 module costs $2,600–3,100 — three times more — draws 20–30 W per port, and requires new switches with 200G/lane ASIC and cooling infrastructure rated for 30 W per port. Lead time for 800G is 2–8 weeks; for 1.6T it is 8–20 weeks.
Yes. The IEEE 802.3df standard for 800G was ratified in 2024. At OFC 2026, 40 companies simultaneously confirmed full interoperability of modules from different vendors. TrendForce estimates over 63 million 800G-and-above optical modules will be shipped in 2026. It is a mature, safe choice with guaranteed compatibility.
The IEEE 802.3dj standard for 1.6T is still in draft as of Q1 2026 — no guarantee of interoperability between modules from different vendors. Beyond that, the price is three times higher than 800G, power draw nearly twice as high, lead time longer, and new switches with 200G/lane ASIC are required. This is technology for hyperscalers with GPU clusters above 50,000 cards — not a typical enterprise-class data centre.
1.6T makes sense if you are building a GPU cluster above 50,000 cards and port density matters more than price. Before buying, verify three things: (1) Your switch ASIC supports 200G/lane electrical — Broadcom Tomahawk 6 or Marvell Teralynx 10. (2) You have cooling infrastructure ready for 30 W per port. (3) The vendor will confirm compliance with the final IEEE 802.3dj standard once approved. Without meeting these conditions you risk compatibility issues and unplanned costs.
Data centres in Warsaw, Prague, Bratislava, and Budapest are receiving first AI projects, analytical models, and image processing systems. Each of these workloads increases the demand for internal bandwidth. 10G no longer cuts it for AI workloads, 100G starts feeling tight at the first GPU cluster. 800G solves this problem now, with certain costs and full interoperability.
No, if you plan the infrastructure correctly. The key step: choose switches and cooling compatible with 1.6T at the 800G deployment stage. Then the 2027–2028 migration will simply mean swapping modules in existing ports — evolution, not revolution. Organisations that buy switches without thinking about 1.6T will have to replace the entire infrastructure when moving to the higher standard.
Typical lead time for 800G modules from stock is 2–8 weeks. By comparison, 1.6T modules take 8–20 weeks. The 800G market is mature and well-stocked — many suppliers hold modules in inventory. When planning a project with a hard launch deadline, order at least 2 months in advance to allow for any logistical delays.
5G is not an evolution of 4G, but a fundamentally different level of demand on the transport network. Every 5G base station generates traffic that would have been unthinkable with LTE, and it has to send it somewhere. That “somewhere” is the backhaul, and it is precisely the backhaul that has become the biggest challenge for operators in Poland and the CEE region. Coherent optical modules change the economics of that challenge — and not in theory, but in production networks already running today.
You already know what CWDM, DWDM, and O-Band are, and you understand the differences. Now you're facing a specific purchasing decision, and you have that nagging feeling that you want to get it right—because choosing the wrong architecture today will mean costly replacements in a few years. This article is here to help you make that decision with confidence. No jargon, just the facts.
