In 2026, 800G is the deployment target for serious AI data centers. 1.6T is already entering early commercial rollout for hyperscalers who need to stay ahead of that curve.
This article covers what's driving the shift, where 800G and 1.6T stand today, what the form factor and thermal realities look like on the floor, and how to source 800G optics without paying OEM prices that have no business being that high.
Table of Contents
When GPU clusters were small, 100G per port was workable. That changed fast. A single H100 NVLink domain can generate over 3.2 Tbps of east-west traffic. Scale that across a rack and 400G per spine link starts looking tight. Scale it across a pod and 400G becomes the minimum viable interconnect, not the upgrade target.
Three workload types are driving this:
LLM training. Distributed training across thousands of GPUs requires all-reduce operations that saturate every link simultaneously. Latency and bandwidth both matter. A single 400G bottleneck in the fabric can stall gradient synchronization across an entire training run.
Inference at scale. High-throughput inference clusters serving millions of daily requests need fat east-west pipes for model sharding and KV-cache transfers. The traffic pattern is less predictable than training, which makes oversubscription riskier.
Storage and checkpoint I/O. LLM checkpointing at multi-terabyte scale requires high-bandwidth paths to distributed storage. These aren't GPU-to-GPU flows, but they compete for the same fabric bandwidth.
The result: 400G is the baseline spec for any AI data center fabric being designed in 2026. 800G is the target for new builds. 1.6T is on the roadmap for anything expected to serve AI workloads past 2027.
800G optical modules are past the early-adopter phase. Volume production is real, pricing has come down from stratospheric launch numbers, and switch silicon from Broadcom (Tomahawk 5) and Marvell (Teralynx 10) now supports 800G port densities that make large-scale deployment practical.
Projected 800G transceiver shipments in 2026 exceed 40 million units globally, driven almost entirely by AI infrastructure buildouts at hyperscalers and large colocation facilities. That number reflects both new deployments and the replacement of 400G ports in clusters that hit bandwidth limits faster than expected.
The dominant form factors for 800G are OSFP (Octal Small Form Factor Pluggable) and QSFP-DD (Quad Small Form Factor Pluggable Double Density). Both support 8 electrical lanes at 100G PAM4 each, delivering 800G aggregate. OSFP has a slightly larger mechanical footprint, which gives it better thermal headroom for higher-power coherent and long-reach variants. QSFP-DD fits into existing QSFP-DD port cages, making it easier to drop into infrastructure originally designed for 400G.
QSFP112 is also entering the picture for some 800G applications, using 4 lanes at 200G each. It's less common in AI data center deployments today but relevant for certain high-density switch designs.
HYTOPTODEVICE stocks 800G modules in both OSFP and QSFP-DD form factors, including the Arista-compatible 800G QSFP-DD DR8 — covering the most common AI fabric interconnect scenario.
1.6T is not vaporware. It's in early commercial deployment at hyperscalers as of mid-2026, primarily using OSFP and co-packaged optics (CPO) configurations. The electrical interface moves to 200G per lane, which requires new switch ASICs and updated SerDes implementations.
For most enterprise and mid-market data centers, 1.6T is a 2027–2028 procurement question. But if you're designing infrastructure today that will need to scale, your physical layer choices matter now. OSFP cages supporting 800G today can be designed to accept 1.6T OSFP modules when they reach volume pricing. Locking into a form factor that dead-ends at 400G is a mistake you'll pay for in two years.
The practical implication: if your 2026 build uses OSFP for 800G, you're on the right upgrade path. If you're still deciding whether to skip 400G and go straight to 800G, the answer for AI-facing infrastructure is almost always yes.
800G modules consume significantly more power than their 400G predecessors. Typical 800G QSFP-DD SR8 modules run at 10 to 14W per port. OSFP variants for longer-reach applications can reach 20W or higher. That's not a minor footnote — it's a rack power and cooling design constraint.
A 32-port 800G switch at 14W per port adds 448W of transceiver-only thermal load per switch. Multiply that across a spine layer and you're dealing with kilowatts of additional heat your cooling infrastructure needs to handle.
Key thermal considerations for 800G deployments:
None of this is a reason to avoid 800G. These are engineering constraints to plan around, and they're manageable with proper infrastructure design.
| Attribute | OSFP | QSFP-DD |
|---|---|---|
| Electrical lanes | 8 x 100G PAM4 | 8 x 100G PAM4 |
| Mechanical size | Larger (better thermal) | QSFP-compatible footprint |
| Max power (typical) | Up to 20W+ | Up to 15W |
| Cage compatibility | OSFP-specific | Backward-compatible with QSFP-DD 400G cages |
| 1.6T upgrade path | Yes (OSFP 1.6T in development) | Limited |
| Common AI use case | Spine, long-reach, coherent | Leaf-spine, short-reach SR8, DR8 |
For AI GPU cluster interconnects at short reach — SR8, DR8 — QSFP-DD 800G is the practical choice in 2026 because it fits existing port infrastructure. For new spine builds or anything requiring a longer reach, OSFP is the better long-term investment.
OEM 800G modules from Cisco, Arista, and Juniper are priced at 500 to 1,500 dollars or more per unit depending on reach and vendor. For a 128-port 800G spine switch, that's 64,000 to 192,000 dollars in transceiver costs alone — before you touch the switch hardware itself.
Compatible third-party 800G modules deliver the same electrical and optical specifications at 70 to 90 percent lower cost. At scale, that gap isn't marginal. It's the difference between a 200,000-dollar transceiver line item and a 20,000 to 60,000-dollar one.
The compatibility hesitation that once followed third-party modules at 100G carries less weight at 800G for a specific reason: at this price point and this scale, hyperscalers and large colocation operators have already validated compatible modules extensively. The ecosystem is mature enough that compatibility testing is standard practice, not a risk mitigation exercise.
HYTOPTODEVICE publishes compatibility test videos and product datasheets on-site for 800G modules, including the Arista-compatible 800G QSFP-DD DR8 and additional OSFP variants. OEM and ODM options are available for operators who need white-label or custom-programmed modules for specific platform requirements. Review the full 800G catalog and request a quote at hytoptodevice.com.
If you're a network engineer or infrastructure manager evaluating 800G for an AI data center build in 2026, the decision framework is straightforward:
What is the difference between 800G OSFP and 800G QSFP-DD?
Both use 8 electrical lanes at 100G PAM4 for 800G aggregate bandwidth. OSFP has a larger mechanical footprint with better thermal headroom, making it suited for higher-power and longer-reach applications. QSFP-DD fits into existing QSFP-DD port cages, which simplifies retrofitting into infrastructure originally designed for 400G.
Is 1.6T available for standard data center deployments in 2026?
1.6T is in early commercial rollout at hyperscalers as of mid-2026 but isn't yet a standard procurement option for enterprise or mid-market data centers. Most non-hyperscale operators will evaluate 1.6T in 2027 or 2028. Designing your 2026 infrastructure around OSFP form factors positions you for that transition.
How much power does an 800G transceiver consume?
Typical 800G QSFP-DD SR8 modules consume 10 to 14W per port. OSFP modules for longer-reach applications can reach 20W or higher. That's a significant increase over 400G and requires careful power supply and cooling planning at scale.
Can I use compatible third-party 800G modules with Arista and Cisco switches?
Yes, with proper validation. Third-party compatible 800G modules are widely deployed in production environments. HYTOPTODEVICE publishes compatibility test videos and datasheets for its 800G modules, including the Arista-compatible 800G QSFP-DD DR8, to support pre-purchase validation.
Why are AI data centers moving to 800G instead of scaling 400G further?
LLM training and large-scale inference generate east-west traffic that saturates 400G links in dense GPU clusters. 800G doubles per-port bandwidth without doubling port count, improving both fabric efficiency and power-per-bit ratios at scale.
What is QSFP112 and how does it relate to 800G?
QSFP112 uses 4 electrical lanes at 200G each to deliver 800G aggregate. It's less common than OSFP and QSFP-DD in current AI data center deployments but appears in certain high-density switch designs. The 4-lane architecture also aligns with the electrical interface used for 1.6T modules.
Does HYTOPTODEVICE offer OEM or white-label 800G modules?
Yes. HYTOPTODEVICE provides OEM and ODM services for custom-programmed and white-label 800G modules, supporting operators who need branded or application-specific optics. Contact the team via hytoptodevice.com for OEM inquiry details.
800G is not a future consideration for AI data centers — it's the current deployment standard, and 1.6T is already defining the next design cycle. The engineering constraints are real but manageable. The cost savings from compatible modules at this price tier are substantial. Plan the upgrade now, get the form factor right, and source optics your budget can actually support.
