800G / 1.6T

"800G" and "1.6T" are the data rates of optical transceivers - the laser-based plugs that move 800 gigabits or 1.6 terabits per second of network traffic between switches and AI servers. They are core plumbing for AI data centers, and demand far outstrips supply of the lasers inside them.

What "800G / 1.6T" actually means

800G and 1.6T refer to the **speed of optical transceivers** - small pluggable modules that convert electrical signals into pulses of laser light and back again so data can travel over fiber. "800G" means 800 gigabits per second; "1.6T" means 1.6 terabits per second (twice as fast). They are the current and next-generation rungs on a ladder that ran through 100G, 200G, and 400G before them. When people talk about "800G / 1.6T stocks," they mean the companies that design and build these modules and the laser and chip components inside them.

How a transceiver works

A transceiver plugs into a network switch or server like a USB stick into a slot. Inside sit tiny lasers, modulators, photodetectors, and a signal-processing chip. To reach 800G, the module typically runs **eight electrical lanes at 100 gigabits each**; a 1.6T module doubles the per-lane speed to **200G using PAM4 signaling over 224G SerDes**. The laser of choice for these speeds is an **EML (electro-absorption modulated laser)**, a III-V compound-semiconductor part that is both the performance bottleneck and the supply bottleneck. Modules ship in form factors with names like QSFP-DD800, OSFP, and OSFP-XD.

Why it matters for AI

Training and serving large AI models means lashing tens of thousands of GPUs together so they behave like one machine. Those GPUs must exchange enormous amounts of data with very low latency, which means the network - not just the chips - sets the ceiling on cluster performance. Each generation of GPU systems (for example NVIDIA's Quantum-X800 networking platform) demands faster optics, so 800G became the default for new AI build-outs around 2025-2026 and 1.6T is ramping right behind it. TrendForce pegged the AI optical-transceiver market at roughly **$16.5 billion in 2025, growing toward $26 billion in 2026**.

Where it sits in the supply chain

The value chain has three layers. At the **component** layer are the EML lasers and the DSP/retimer chips: lasers come from a handful of makers including Lumentum, Coherent, Broadcom, and Mitsubishi; DSPs come mainly from Broadcom and Marvell. At the **module** layer, firms like Applied Optoelectronics and many Chinese manufacturers assemble those parts into finished 800G/1.6T transceivers. At the **system** layer, switch vendors and hyperscalers (Microsoft, Amazon, Google, Meta) and NVIDIA buy the modules by the million. Notably, the components - not the modules - capture most of the profit, because they are the scarcest link.

Who the key players are

Among publicly traded names often discussed here: **Lumentum (LITE)** is a leading EML laser supplier and, by several accounts, the main volume source of 200G-per-lane lasers needed for 1.6T. **Applied Optoelectronics (AAOI)** makes 800G and 1.6T modules and lasers, and began volume 800G shipments to a major hyperscaler in early 2026. **Credo Technology (CRDO)** sits slightly adjacent: it makes the DSP/retimer chips and **Active Electrical Cables (AECs)** - copper interconnects rated to 800G and 1.6T that handle the very short links inside AI racks where fiber is overkill. Coherent, Broadcom, and Marvell are other heavyweight component suppliers.

What's changing now

Two themes dominate 2025-2026. First, **supply is the story**: McKinsey-style analyses cited across the industry project 800G output falling well short of demand through 2027, and 1.6T shortfalls persisting for years - which is why NVIDIA reportedly committed roughly $4 billion combined to Lumentum and Coherent in 2026 to lock up laser capacity. Second, **architecture is shifting**: to cut the 30W+ power a DSP-based 1.6T module burns, the industry is exploring **Linear Pluggable Optics (LPO)**, which drops the DSP, and **Co-Packaged Optics (CPO)**, which moves the optics onto the switch chip package itself. Both could reshape who wins as speeds climb past 1.6T.

Frequently asked

Is 1.6T just two 800G modules?

Roughly, in capacity terms - a 1.6T link carries twice the data of an 800G link. But it is not simply two old modules taped together: 1.6T generally doubles the speed of each electrical lane to 200G (via 224G SerDes) rather than just adding more lanes, which requires faster lasers and signal-processing chips.

Why is there a shortage of these transceivers?

The bottleneck is the EML laser, a specialized III-V semiconductor part that is hard to manufacture at 200G-per-lane speeds and yields. Only a few suppliers can build them at volume, and AI demand has outrun that capacity, with shortfalls projected to last into the late 2020s for 1.6T.

Are 800G and 1.6T the same as copper cables?

No. Transceivers use lasers and fiber for links longer than a meter or two. For the very shortest connections inside a rack, copper-based Active Electrical Cables (AECs) - a Credo specialty - can carry 800G or 1.6T at lower power and cost, so the two technologies are complementary.

What is replacing pluggable transceivers?

Nothing yet, but two approaches aim to cut power: Linear Pluggable Optics (LPO) removes the power-hungry DSP from the module, and Co-Packaged Optics (CPO) integrates the optics directly onto the switch chip. Pluggable 800G/1.6T modules remain mainstream for now because they are interoperable and field-serviceable.

Which companies benefit most from 800G/1.6T?

Component makers tend to capture the most value because they are the scarce link - laser suppliers like Lumentum and Coherent, and DSP suppliers like Broadcom and Marvell. Module makers like Applied Optoelectronics and chip/cable specialists like Credo also participate directly.

Related companies