Goldman Sachs Raises Optical Transceiver Market Forecast to $148.5 Billion by 2028

Goldman Sachs’ latest optical-transceiver outlook sharply increases the expected value of the market: approximately $67.7 billion in 2026, $131.4 billion in 2027, and $148.5 billion in 2028. Those figures are 33%, 81%, and 115% above its earlier forecasts, respectively. 2

The core thesis is not simply that more AI servers will ship. It is that AI infrastructure is moving toward larger, more networking-intensive rack-scale systems, with faster optical links and more optical content attached to each accelerator or custom ASIC.

The forecast revision at a glance

Goldman raised its global optical-transceiver shipment forecasts by 21% for 2026, 31% for 2027, and 31% for 2028. The updated estimates imply total shipments of roughly 548 million, 691 million, and 729 million units, respectively. 319

The upgrade is concentrated in high-speed products:

• 800G-and-above shipments were raised by 31%, 39%, and 36% for 2026–28. 18
• 1.6T-and-above shipments were raised by 29%, 61%, and 50%, reflecting higher expected rack-scale and ASIC-server volumes, more modules per GPU, and faster specification upgrades. 3
• Goldman now expects 800G-and-above module shipments of 78.22 million in 2026, 144 million in 2027, and 171 million in 2028. 2
• Revenue from that high-speed segment is forecast at $45.2 billion, $107.5 billion, and $129.5 billion across those years—an implied 2026–28 CAGR of about 69%. 2

Why rack-scale AI changes optical demand

The important shift is from networking primarily within a rack toward more elaborate scale-up and scale-out designs spanning larger groups of accelerators and racks. Goldman’s earlier optical-networking research estimated a combined scale-up and scale-out networking opportunity of roughly $154 billion for a Rubin Ultra NVL576 configuration, compared with about $15 billion around GB300 NVL72. 11

That is an analytical scenario rather than a confirmed spending outcome, but it illustrates the mechanism behind the new transceiver forecast: larger AI systems need substantially more high-bandwidth interconnect.

Goldman’s revised assumptions for Nvidia Rubin Ultra are especially notable. It raised its modeled optical content from zero to two 1.6T modules per GPU, and from three to five 3.2T modules per GPU. 27

The firm also added Google TPU v9 servers to its model. Its assumptions call for 12 1.6T modules per chip from the second half of 2027 through the first half of 2028, followed by six 3.2T modules per chip in the second half of 2028. 27

Custom-ASIC deployments are another part of the rationale. Goldman cited higher projected rack-level and ASIC-server shipments as a reason for lifting high-speed optical shipment estimates. 3 The exact adoption path, architecture, and module content remain key variables in the forecast.

The speed transition: 800G to 1.6T, then 3.2T

800G remains a major data-center product, but Goldman’s forecast points to a rapid migration toward 1.6T and then 3.2T links.

Module speed	2026 forecast shipments	2027 forecast shipments	2028 forecast shipments
800G	45.39 million	49.48 million	48.66 million
1.6T	32.83 million	71.36 million	54.85 million
3.2T	—	23.04 million	67.73 million

Source: Goldman Sachs forecasts as reported by Futunn/NetEase. 1927

The projected decline in 1.6T shipments in 2028 does not necessarily indicate a weaker market. In this model, it reflects the beginning of a transition to 3.2T. By 2028, 3.2T’s projected 67.73 million units would represent roughly 40% of Goldman’s projected 800G-and-above volume. 227

Why silicon photonics is central to the thesis

Silicon photonics is expected to take a growing share of high-speed modules, particularly at 1.6T and above. Goldman’s reported assumptions put silicon-photonics penetration at about 60% for 800G, 80% for 1.6T, and 80% for 3.2T during the forecast period. Within the 800G-and-above market, silicon photonics is projected to account for 68% of shipments in 2026, 73% in 2027, and 74% in 2028. 27

One reason is laser cost. Goldman’s cited comparison for a 1.6T module estimates that a silicon-photonics design uses four 70 mW continuous-wave lasers costing about $15–$20 in total. A comparable EML-based design uses eight 200G EMLs, with laser cost of about $160. 127

This is a comparison of the laser portion of the bill of materials—not a complete module cost comparison. It nevertheless explains why silicon photonics may be attractive as port speeds and lane counts increase.

Where co-packaged optics fits

Co-packaged optics, or CPO, is part of Goldman’s longer-term AI networking opportunity rather than a simple near-term replacement for pluggable transceivers. In its Rubin Ultra scenario, Goldman estimated that CPO-related components could account for $91 billion, or 59%, of the modeled $154 billion networking opportunity. 11

That scenario should not be read as a forecast that all high-speed pluggable modules disappear. Goldman’s separate transceiver forecast still anticipates major volume growth in 800G, 1.6T, and 3.2T modules through 2028. 227 The technology mix will depend on system architecture, qualification progress, manufacturability, power requirements, and hyperscaler deployment choices.

Zhongji Innolight: Goldman’s ratings and earnings view

Goldman’s industry upgrade was accompanied by a strongly positive view of Zhongji Innolight:

• A-shares: Buy rating reiterated, with a 12-month target price of RMB 2,645. 4
• H-shares: Buy coverage initiated, with a 12-month target price of HK$3,267. 4
• Revenue outlook: Goldman projects a 78% revenue CAGR from 2026 to 2028, primarily tied to rising silicon-photonics module shipments and expansion beyond conventional scale-out networking. 13
• Earnings expectations: Goldman’s 2026 and 2027 net-profit forecasts are 25% and 42% above market consensus, respectively. 13

What would validate—or challenge—the forecast

The $148.5 billion 2028 projection is an analyst model, not confirmed market demand. Its outcome depends on whether hyperscalers deploy rack-scale systems on the expected schedule, whether custom-ASIC programs ramp as assumed, and how quickly 1.6T and 3.2T modules qualify and move into volume production.

The central takeaway is clear: Goldman’s revision is a bet that AI networking is becoming much more optics-intensive. If that occurs, the market’s growth will be led less by conventional lower-speed modules and more by high-speed 800G, 1.6T, and 3.2T products—alongside a rising role for silicon photonics. 327