AI Networking and Optics: NVLink, Ethernet, CPO — 2026-09-19
The AI networking landscape shifted this week as Apple reportedly explores Nvidia's NVLink for its M8 Ultra servers, signaling a potential end to proprietary silos. Meanwhile, Chinese optical module giants like Innolight and Eoptolink reported accelerating shipments of 1.6T transceivers, driven by hyperscaler demand for 2027. Broadcom reaffirmed massive growth projections, forecasting $115 billion in AI chip revenue by FY2027, while startups like Cornelis and Delos Data pitch open alternatives to NVLink.
AI Networking and Optics: NVLink, Ethernet, CPO — 2026-09-19
Top developments
Apple Reportedly Teams Up with Rival Nvidia on NVLink
In a surprising move for historically bitter rivals, reports indicate that Apple is exploring the use of Nvidia's NVLink interconnect technology for its upcoming custom M8 Ultra AI servers, targeted for a 2029 data center push. This potential integration suggests that even major tech incumbents are finding it difficult to bypass Nvidia's dominant scale-up interconnect ecosystem when building high-bandwidth AI infrastructure. If confirmed, this would mark a significant shift in Apple's hardware strategy, leveraging NVLink's high-speed GPU-to-GPU communication capabilities rather than relying solely on proprietary or standard Ethernet solutions.
1.6T Optical Modules Enter Mass Production Phase
Leading Chinese optical transceiver manufacturers are reporting a decisive shift toward 1.6T modules, moving beyond the 800G era that dominated 2025. According to recent investor relations disclosures from Innolight (Zhongji Xuchuang) and Eoptolink (Xinyisheng), 1.6T pluggable modules saw quarter-over-quarter shipment increases in Q2 2026, with expectations for accelerated ramp-up in the second half of the year. These components are becoming the universal standard for AI data centers, serving large cloud service providers, equipment manufacturers, and new cloud entities building trillion-parameter models. The demand is so acute that industry analysts estimate global requirements for 1.6T modules will reach 25–30 million units in 2026, creating supply bottlenecks in 200G EML lasers and 3nm DSPs.
Startups Challenge NVLink Dominance with Open Alternatives
Intel spin-off Cornelis and newcomer Delos Data are actively pitching open networking alternatives designed to scale AI clusters beyond the single rack, challenging Nvidia's proprietary NVLink monopoly. These startups aim to provide cost-effective, scalable interconnects that avoid the vendor lock-in associated with Nvidia's proprietary fabrics. Their emergence highlights a growing market segment seeking to democratize scale-up networking, particularly as the Ultra Ethernet Consortium (UEC) spec 1.0.3 gains traction for its support of AI-specific transport protocols and congestion control mechanisms like NSCC/RCCC.
Broadcom Projects $115B AI Revenue by FY2027
Broadcom updated its financial outlook, now projecting approximately $115 billion in AI chip revenue for the fiscal year ending October 2027, a significant increase from prior forecasts. This guidance underscores the insatiable appetite of hyperscalers for custom AI accelerators (XPUs) and associated networking silicon, including Tomahawk Ethernet switches and Jericho routers. The forecast reflects Broadcom's success in capturing share in both compute and networking layers, positioning it as a primary counterweight to Nvidia’s integrated platform strategy.
Local view
Chinese financial media and local stakeholders are closely monitoring the "light module" (optical transceiver) sector as a bellwether for AI infrastructure health. Outlets like Sina Finance and Phoenix Tech highlight the divergence between first-tier suppliers (Innolight, Eoptolink) who are securing long-term orders for 1.6T products, and second-tier vendors struggling with material constraints. Huxiu notes that while first-tier players have secured capacity for 200G EML chips, the massive demand gap is forcing hyperscalers to look at alternative suppliers, potentially benefiting smaller firms if they can overcome yield challenges. Additionally, the CPO/NPO sector saw stock volatility with companies like T&S Communications and Cambridge Technology rising on news of liquid-cooling integrated optical engines entering verification at top cloud vendors.
Context & numbers
- 1.6T Demand Forecast: Industry estimates suggest global demand for 1.6T optical modules will hit 25–30 million units in 2026.
- Broadcom AI Guidance: Projected AI revenue of ~$115 billion for FY2027.
- Arista Networks: Recently raised its 2026 revenue outlook for the third time, citing strong AI data center demand and nearly tripling multi-year supplier purchase commitments.
- Nvidia Market Share: Nvidia's Ethernet switch market share has surged from under 4% to 21.5% in one year, challenging traditional Ethernet leaders.
On the radar
- Ultra Ethernet Spec 1.0.3: Released in July 2026, this specification is gaining attention for its "AI Full" profile, which includes packet spraying and advanced congestion control, though it still faces competition from InfiniBand and NVLink Fusion.
- Liquid Cooling Integration: A growing trend in CPO (Co-Packaged Optics) is the integration of liquid cooling cages and VC vapor chambers directly into optical modules, with the market projected to grow from $1B in 2026 to $6.3B by 2030.
- FS.com New Portfolio: FS has launched a new 1.6T scale-out optics portfolio featuring Broadcom 3nm DSPs and coherent modules capable of 500km transmission, targeting broader AI fabric deployments beyond short-reach data center links.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
