AI Datacenter Networking Supply Chain — Market Map

5-layer view of frontier AI compute → hyperscalers → networking suppliers → optical components → physical / packaging. Click any citation number [#] to jump to its source in the References section at the bottom. Verified May 17, 2026
Legend
Builder — owns the underlying silicon
Wrapper — packages someone else's silicon
Hybrid — builds own + wraps third-party
End-user / infra (no silicon claim)
Scale-up: NVLink UALink Google ICI PCIe / CXL
Scale-out: InfiniBand Omni-Path RoCEv2 / UEC Spectrum-X (NV) AWS SRD Google Falcon MRC iWARP SONiC (NOS) Google Jupiter SDN
Libraries: NCCL RCCL MSCCL
Optical: CPO LPO LRO NPO OCI / Photonic Fabric Apollo OCS Optical Interposer
★ In-house networking
1

Frontier AI Companies (the demand)

Train/serve frontier models. They pick network architectures, sometimes design their own.
Anthropic
Claude (Opus / Sonnet / Haiku 4.x)
Multi-cloud: AWS (Trainium/EFA), GCP (TPU/OCS), and now xAI Colossus 1 (Spectrum-X) under May 2026 capacity deal.
RoCEv2/EFASpectrum-XNCCL
[3][9][14]
OpenAI
GPT family, Stargate
Primary: Azure (NVIDIA-heavy). Stargate first sites use Spectrum-X. Researching custom MRC (Multipath Reliable Connection) transport.
Spectrum-XInfiniBandNCCL
[4][17][22]
Google DeepMind
Gemini, Veo
Trains on Google TPU pods (Ironwood: 9,216 chips / pod, 48 OCS units). Fully proprietary fabric — Jupiter SDN + Apollo OCS.
ICI scale-upApollo OCSJupiter SDN★ In-house
[9][10][11][13]
Meta AI / FAIR
Llama 3 / 4 / 5
24K+ H100 RoCEv2 cluster on Arista 7800 + OCP Minipack2/Wedge400. Llama 3 was the Ethernet-vs-InfiniBand bake-off. Custom MTIA accelerators on Broadcom-built training systems.
RoCEv2/UECNCCL
[5][6][7][8]
xAI
Grok
Colossus 1 (Memphis): 100K → 200K H100s on NVIDIA Spectrum-X Ethernet (SN5600 + BlueField-3 SuperNIC). Largest pure-Spectrum-X production cluster.
Spectrum-XNCCL
[1][2]
Mistral AI
Mistral, Mixtral, Codestral
Trains on Nebius / CoreWeave / Microsoft Azure — InfiniBand-heavy NVIDIA H100/H200 clusters.
InfiniBandNCCL
[19][20]
Moonshot AI (Kimi K2.6)
Kimi chat models
China-based; trains primarily on domestic GPU/NPU clusters (Huawei Ascend + NVIDIA where available). Ethernet-aligned via Huawei/H3C silicon.
RoCEv2
Cohere
Command R+, enterprise LLMs
Trains on Google Cloud TPUs and Oracle OCI. TPU pod uses Google's in-house OCS fabric.
ICI scale-upApollo OCSJupiter SDN
[9][12]
Layer 1 → Layer 2 dependency: Every frontier lab depends on at least one hyperscaler or neocloud for capacity. Notable in-house exception: Google DeepMind → Google Cloud is fully vertically integrated to Layer 2.
2

Hyperscalers & Neoclouds (the buyers)

★ marks providers that design networking silicon, switch OS, or fabric protocol in-house.
Hyperscalers — in-house networking efforts
Google Cloud Platform ★
TPU + Jupiter + Apollo OCS
Deepest in-house stack. Designs TPU compute, Jupiter SDN-Ethernet fabric, and Apollo MEMS-based optical-circuit switches. 5× faster, 40% lower power, 30% lower capex vs. prior gen. Ironwood TPU pods use 48 OCS units to interconnect 9,216 chips.
★ In-house OCS + fabricICIApollo OCSJupiter SDNFalcon
[9][10][11][12][13]
AWS ★
Annapurna Labs / Nitro / EFA / SRD
Designs Nitro DPU, the Elastic Fabric Adapter NIC, and a custom transport protocol — SRD (Scalable Reliable Datagram) — to bypass kernel and replace TCP/RoCE inside its fleet. Also designs Trainium/Inferentia accelerators.
★ In-house NIC + protocolAWS SRDRoCEv2 (legacy)
[14][15]
Microsoft Azure ★
SONiC + MAIA + custom optics
Created and open-sourced SONiC switch OS (runs on multi-vendor ASICs incl. Broadcom + NVIDIA Spectrum). Co-develops hollow-core fiber with Huber+Suhner. Designs MAIA accelerator. Still significant NVIDIA InfiniBand for inference. Microsoft co-developed and runs MRC in production across its largest training clusters (per the OpenAI/Microsoft/AMD/Broadcom/NVIDIA joint paper).
★ In-house OS + HCFSONiC (owner)RoCEv2/UECInfiniBandMRC (production)MSCCL
[16][17][18][59][84][138]
Oracle OCI
Project Jupiter / Gov't AI campus
Largely off-the-shelf NVIDIA stack — Quantum InfiniBand + Spectrum-X. Sovereign-cloud focus. No public custom-silicon program.
InfiniBandSpectrum-X
[22]
Neoclouds — predominantly NVIDIA-fabric resellers
CoreWeave
H100/H200/B200 fleets
Quantum InfiniBand + Spectrum-X Ethernet, mostly Quantum-2 400G. Largest pure-play NVIDIA cloud.
InfiniBandSpectrum-XNCCL
[19][20]
Lambda Labs
1-Click Clusters
Quantum-2 400G InfiniBand, rail-optimized, GPUDirect RDMA.
InfiniBandNCCL
[20]
Together AI
Inference + fine-tuning cloud
NVIDIA InfiniBand on H100/H200; lighter scale than CoreWeave but specialized for open-model serving.
InfiniBandNCCL
[19][20]
Nebius
European GPU cloud
Quantum InfiniBand fabric; AI-first. Q4 2025 revenue +547% YoY, ARR ~$1.2B at year-end.
InfiniBandNCCL
[19][20]
Voltage Park
24K H100 across 6 sites
InfiniBand. Non-profit-backed neocloud.
InfiniBand
[21]
Crusoe
Stranded-energy AI cloud
NVIDIA InfiniBand fabrics, expanding into Spectrum-X for scale-out builds.
InfiniBandSpectrum-X
[19][20]
Layer 2 → Layer 3 dependency: Hyperscalers buy switches/NICs from Layer 3 (NVIDIA, Arista, Cisco, Broadcom, Marvell). The in-house exceptions: GCP designs its own OCS + TPU NIC, AWS designs its own EFA NIC, and Azure writes its own switch OS (SONiC). Meta (technically Layer 1 but acts as a Layer 2 buyer) designs OCP switch hardware but uses Broadcom/NVIDIA ASICs underneath.
3

Networking Equipment Suppliers

Solid-green border = builder (owns silicon). Dashed-orange = wrapper (uses third-party silicon — labeled).
Vertically integrated giants — full stack (switch ASIC + DSP + SiPho + optics)
NVIDIA
switches · DPU · interconnect · optics
Switch ASICs: Spectrum-4 (SN5600 800G), Spectrum-6 (SN6810/SN6800), Quantum-3 InfiniBand (3450-LD 144×800G).
AI interconnect: NVLink / NVSwitch (proprietary scale-up).
DPU: BlueField-3 / Thor-equivalent SuperNIC.
Optics: LinkX transceivers; integrating CPO into Quantum-X & Spectrum-X for Rubin generation. EML lasers sourced from Lumentum/Coherent — see Layer 4.
Spectrum-X InfiniBand NVLink CPO (Rubin+) NCCL
[22][23][24][25]
Broadcom (AVGO)
switch ASIC · DSP · in-house SiPho + VCSEL lasers
Switch ASICs: Tomahawk 6 / Tomahawk Ultra (102.4 Tbps, chiplet) — and the Davisson CPO variant; Jericho4 deep-buffer; Trident enterprise.
DSP / SerDes: deep IP portfolio.
NIC: Thor Ultra 800G.
Optics: in-house silicon photonics + VCSEL lasers; ships 50K+ TH5-Bailly CPO switches. CPO uses TSMC COUPE (Layer 5).
Custom-ASIC partner to Meta, Google, Apple for accelerators.
RoCEv2/UEC CPO (Davisson) SUE/UALink
[26][27][28][73][76]
Marvell (MRVL)
switch · optical DSP · SiPho · CXL switch (4 acquisitions)
Optical DSPs: Alaska A 1.6T (AEC) and Alaska C / Spica (PAM4) — from $10B 2021 Inphi acquisition.
Silicon photonics: Inphi SiPho + Celestial AI Photonic Fabric chiplet (16 Tbps scale-up) — $3.25B (up to $5.5B) Dec 2025.
CXL/PCIe switching: XConn (Apollo 2: hybrid CXL 3.1 + PCIe Gen 6.2, 256 lanes) — $540M closed Feb 2026.
Plasmonics modulators: Polariton Technologies (ETH Zurich spin-out) — announced Apr 2026 for 3.2T+ scaling.
Switch ASICs: Teralynx Ethernet line + custom-ASIC business (AWS Trainium etc.).
RoCEv2/UEC CPO UALink/CXL
[29][30][31][32][123][124][125][126]
Cisco (CSCO)
Silicon One + wraps Broadcom in legacy
Switch ASIC: Silicon One G300 102.4 Tbps (64×1600 GbE, 512×200G SerDes) — directly competes with TH6.
Router ASIC: P200 (51.2T deep-buffer router, Cisco 8223).
Pushing LPO as the lower-risk alternative to CPO. Legacy product lines still use Broadcom Tomahawk.
RoCEv2/UEC LPO
[33][34][35]
Switch / system builders — own systems & software, source switch silicon
Arista Networks (ANET)
wraps → Broadcom
Etherlink AI portfolio: 7060X6 leaf (Tomahawk 5), 7800R4 spine (Jericho3-AI), 7700R4 Distributed Etherlink Switch — scales to 10K XPUs single-tier, 100K+ XPUs two-tier. Big differentiator is EOS software, not silicon.
EOS implements SRv6 on Broadcom Tomahawk 5 for MRC source-routing — co-developed with OpenAI / Microsoft / AMD / Broadcom / NVIDIA.
RoCEv2/UEC MRC SRv6 (EOS on TH5)
Silicon: Broadcom Tomahawk 5/6 + Jericho 3/4 [36][37][38][138]
HPE
Slingshot (own) + Aruba (own + → Broadcom)
HPE Cray Slingshot 400: in-house Rosetta-2 switch ASIC + Cassini-2 NIC ASIC (fall 2025) — only non-NVIDIA full-stack HPC fabric.
Aruba CX: some switches use own ASICs, others use Broadcom + AMD DPU partnership.
RoCEv2
Silicon: HPE Rosetta-2 / Cassini-2 (own) + Broadcom (Aruba) [39][40]
AI NICs / DPUs / Endpoint-RDMA silicon
NVIDIA — BlueField & ConnectX (NICs)
DPU / SuperNIC line
BlueField-3 SuperNIC powers Spectrum-X clusters (xAI Colossus reference). ConnectX-8 supports 800 Gb/s Ethernet + InfiniBand and ships the MRC transport per the joint OpenAI/Microsoft/AMD/Broadcom/NVIDIA paper. Spectrum-4 and Spectrum-5 switches implement SRv6 under both Cumulus and SONiC for MRC source-routing. Thor-equivalent for Quantum-X800. NVIDIA's NIC line is the in-house counterpart to AMD Pollara & Intel E810.
InfiniBandSpectrum-XRoCEv2MRC (ConnectX-8)SONiC + Cumulus
(Cross-reference with NVIDIA card above) [22][23][138]
Intel — Ethernet 800 / E810
RoCEv2 + iWARP RDMA NICs
Ethernet 800 Series (E810-CQDA2 etc.) is one of the only RDMA NICs supporting both RoCEv2 (UDP/IP) and iWARP (TCP/IP). 256K queue pairs; 10–200 GbE range. The conservative-IT alternative to AMD Pollara and NVIDIA BlueField.
RoCEv2iWARP
[94][95]
AMD Pensando — Pollara 400 / Vulcano 800
UEC NICs · MRC transport · DPU
Pollara 400 = industry-first UEC-ready AI NIC at 400 Gbps; co-validated with OpenAI on AMD Instinct MI350/MI355. Vulcano 800 (Gen6 PCIe, 2026) brings full general-availability MRC support — the protocol AMD co-developed with OpenAI and contributed to OCP.
RoCEv2/UECRCCL
[4][96][97][98][99]
Broadcom — Thor Ultra
800G AI Ethernet NIC
Industry-first 800G AI Ethernet NIC; pairs with Tomahawk 6 / Jericho4 to give Broadcom an end-to-end UEC stack. Thor Ultra ships MRC per the joint OpenAI/Microsoft/AMD/Broadcom/NVIDIA paper. Tomahawk 5 implements SRv6 source-routing for MRC inside Arista EOS. Cross-reference with the Broadcom card above.
RoCEv2/UECMRC (Thor Ultra)
(Cross-reference with Broadcom card above) [27][138]
Pure-play scale-up fabric & alternative-network challengers
Upscale AI (private)
SkyHammer™ scale-up ASIC · UALink/UEC/ESUN
Pure-play AI networking startup. SkyHammer ASIC connects up to 1,024 devices per rack via UALink 1.0 (200 Gb/s bidirectional, ~100–150 ns latency). Multi-standard: UALink, UEC, ESUN, SONiC. Raised ~$300M (Series A $200M, Jan 2026), unicorn status. Direct counterweight to NVLink/NVSwitch. Samples Q4 2026, volume 2027.
UALinkUEC/ESUNSONiC
[85][86][87][88]
Cornelis Networks
Omni-Path — direct InfiniBand alternative
Intel spin-off (2020). CN5000 400G (48-port: 19.2 Tbps / 576-port director: 230.4 Tbps) shipping Q3 2025. Claims 2× message rate / 35% lower latency vs InfiniBand NDR; 6× faster collectives vs. RoCE on AI workloads. CN6000 800G unifies Omni-Path + RoCE; CN7000 1.6T integrates UEC.
Omni-Path (IB-class)RoCE/UEC (CN6000+)
[89][90][91]
Microchip Technology (MCHP)
PCIe Gen 6 switches · timing chips
Switchtec Gen 6 — industry-first 3 nm PCIe Gen 6 switch family: up to 160 lanes, 64 GT/s/lane, 20 ports, 10 internal stacks. ~15% power advantage per lane vs. 5 nm node. Competes with Astera Scorpio for the PCIe-fabric design slot.
UALink/CXL/PCIe
[92][93]
Specialist interconnect & signal-processing chip builders
Credo (CRDO)
AEC DSPs · SerDes · SiPho (DustPhotonics)
AECs, ZeroFlap optical DSPs, retimers. April 2026: $750M (+ ~0.92M shares) deal to acquire DustPhotonics — gains 400G / 800G / 1.6T (roadmap 3.2T) SiPho PIC. Targets >$500M optical revenue FY27.
RoCEv2/UEC CPO/NPO (post-Dust)
[41][42]
Astera Labs (ALAB)
PCIe/CXL retimers · fabric switches · photonic chiplets (+aiXscale)
Aries PCIe Gen 6 retimers (only one shipping in volume); Taurus Ethernet smart cables; Leo CXL memory controller; Scorpio P-Series PCIe 6 fabric switch (32–320 lanes, >10% of 2025 revenue); COSMOS management software.
aiXscale Photonics (closed Nov 2025) — precision glass fiber-chip coupling IP for photonic scale-up chiplets.
UALink (PCIe/CXL) CPO (post-aiXscale)
[43][44][45][121][122]
MaxLinear (MXL)
PAM4 DSPs · TIAs · copper retimers
Rushmore 1.6T PAM4 DSP (200G/lane, <25W, IB+Eth); Annapurna 224G scale-up copper retimer (8/16-lane, 1.6T/3.2T, Q2 2026); Washington 200G TIA. Positioned as challenger to Credo, Astera, and Marvell.
RoCEv2 InfiniBand LPO supplier
[46][47]
Semtech (SMTC)
LPO + 800G/1.6T transceivers · DFB lasers (HieFo)
Driver-amplifier and TIA supplier for optical modules. Pioneered Linear Pluggable Optics (LPO) — removes power-hungry DSP from the module. March 2026: acquired HieFo for $34M, adding in-house DFB laser optoelectronics for IMDD links. Positioned to integrate electronic + optoelectronic + photonic chipsets.
LPO
[100][101]
Alphawave Semi (now Qualcomm)
optical DSP (Banias) · SerDes IP
Independent optical DSP + SerDes IP house. Acquired Banias Labs (2022, $240M) for data-center optical DSP silicon. Qualcomm closed $2.4B acquisition Dec 2025 — folds Alphawave's IP into Qualcomm's data-center push alongside Oryon CPU + Hexagon NPU.
RoCEv2/UEC
[105][106]
Lightwave Logic (LWLG)
electro-optic polymer modulator IP
Materials platform: proprietary EO polymers that replace Si/InP inside modulators — sub-2V drive at 100 GBaud PAM4. Partners with Tower (PH18 PDK) and GlobalFoundries (GDSFactory PDK on GF Fotonix, March 2026) for polymer-on-SiPho integration. Aims at 200G/400G per lane validation tape-outs.
Low-power LPO/CPO enabler
Foundries: Tower + GlobalFoundries (Layer 5) [102][103][104]
Next-generation silicon photonics & optical I/O
Ayar Labs (private)
optical I/O chiplet · external CW laser
TeraPHY 8 Tbps UCIe optical chiplet + SuperNova 16-wavelength remote laser. Joint solution with Alchip + TSMC delivers >100 Tbps scale-up bandwidth per accelerator. Replaces SerDes copper inside the package.
CPO/Optical I/OUALink-adjacent
Fab: TSMC (CoWoS, COUPE) [48][49]
Lightmatter (private)
3D photonic interposer · light engine
Passage M1000 active photonic interposer (114 Tbps optical, 4,000 mm², 256 fibers, 34 chiplets) + Guide light engine. Wafer-scale "photonic floor" for AI superchips.
CPO/Photonic interposer
SiPho process: GlobalFoundries Fotonix [50][51][75]
Celestial AI (now Marvell)
Photonic Fabric chiplet · scale-up optics
Photonic Fabric chiplet — 16 Tbps in a single chiplet, 10× scale-out 1.6T ports. Targets disaggregated memory (chip-to-memory optics). Acquired by Marvell for $3.25B in Dec 2025.
CPO/Scale-up
Now within Marvell stack [29][30][52]
POET Technologies (POET)
Optical Interposer · hybrid PIC
POET Optical Interposer: hybrid-integrated 1.6T 2xFR4 transmitter PIC, 800G production order ($5M+). 3.2T joint dev with Quantum Computing Inc. (TFLN modulators).
CPO-enabler
[53][54][55]
Lucidean (private)
CohZero™ "coherent-lite" IP
Bridges IMDD and long-haul coherent: lets data-center modules adopt coherent-class link budgets without changing existing IMDD lasers / DSPs. Targets 3.2 Tbps and 6.4 Tbps transceivers for AI fabrics, including CPO-compatible. $18M seed Dec 2025 led by Entrada + Koch Disruptive Tech.
CPO-compatible
[107]
Neurophos (private)
metasurface optical processor (in-network compute)
Duke / Metacept spin-out. Metasurface modulator acts as a tensor core that performs matrix-vector multiplication while data is still in flight as light — enabling "in-network" all-reduce in CPO modules. Test chip claims 300 TOPS/W. Raised $110M Jan 2026. First production chips mid-2028.
CPO-integrated compute
[108][109]
Polaris Electro-Optics (private)
ferroelectric nematic glass + SiPho
CMOS-compatible electro-optic modulator material — drop-in for SiPho processes. Claims 400 Gbps/lane and beyond in pluggables or CPO form. Closed $10M Series A (2025) following $3M seed. UCSD / CU Boulder-affiliated, marked publicly as a "holy-grail" in-package optical I/O bet.
CPO/in-package I/O
[110][111]
Optical transceiver wrappers — the Chinese duopoly + Swiss fiber
Zhongji Innolight (300308.SZ)
wraps → Broadcom / Marvell DSP + Layer-4 EML
800G & 1.6T (DR8 / 2×FR4 / 4×FR2) OSFP transceivers; ~60% of NVIDIA's incremental 800G shipments (shared with Eoptolink). Module assembly leader, not a chip designer.
RoCEv2/UECSpectrum-XInfiniBand
DSP: Broadcom / Marvell · EML: Lumentum / Coherent [56][58]
Eoptolink (300502.SZ)
wraps → Broadcom / Marvell DSP + Layer-4 EML
800G + 1.6T DR8 / 2FR4 / 4FR2 OSFP-XD transceivers; LPO-compatible product lines. Industry-first multicore-fiber 800G at OFC.
RoCEv2/UECSpectrum-XInfiniBandLPO-ready
DSP: Broadcom / Marvell · EML: Lumentum / Coherent [56][57]
Huber+Suhner (HUBN.SW)
hollow-core fiber + connectors (no silicon)
Hollow-core fiber: ~47% lower latency, 15–20% power reduction. Already deployed in production Microsoft Azure backbone — scaled deployment partnership renewed April 2026.
Fabric-agnostic
[18][59]
Layer 3 → Layer 4 dependencies: Every transceiver (Innolight, Eoptolink, plus NVIDIA LinkX, Coherent, Lumentum module lines) needs Layer 4 EML lasers. Broadcom is partially self-supplied on VCSEL but still buys EMLs externally. Credo and Marvell now have in-package SiPho via Dust / Celestial but still need external CW lasers for CPO. Ayar Labs & Lightmatter need external CW lasers for their architectures.
4

EML / Laser / Modulator / Photodetector Suppliers

The pure-play InP-based optical-component layer. EML is the #1 bottleneck in 2025–26.
Lumentum (LITE)
EML lasers · CW lasers · pumps
~50–60% EML market share; only volume supplier of 200G/lane EMLs needed for 1.6T transceivers. Investing $663M to expand capacity ~80% by 2025. Primary supplier to NVIDIA, Innolight, Eoptolink.
CW lasers for CPO
InP substrates: AXT, Sumitomo, Coherent [60][61][62][66]
Coherent (COHR)
EMLs · transceivers · in-house SiPho
Second-largest EML supplier. Ramping 6-inch InP wafers — potential 4× die yield. Also one of the few vertically integrated transceiver makers with in-house SiPho + EML.
CPO/SiPhoTransceivers
InP substrate self-supply (ex-II-VI) [60][63][66]
Mitsubishi Electric
EML lasers
Established merchant EML supplier into 100/200/400/800G market. One of the few Japanese InP/EML producers besides Sumitomo. Less volume than Lumentum but vital for diversification.
[60][67]
Sumitomo Electric
EMLs · CW lasers · InP wafers (vertical)
53 GBaud EML-integrated lasers for intra-DC links. Vertically integrated: Sumiden Semiconductor Materials owns the InP wafer slicing too. Targeted growth strategy disclosed Nov 2025.
[64][65][66]
Hamamatsu Photonics
photodiodes · photodetectors · imaging
Photodetector (PD) and avalanche photodiode (APD) supplier — paired with EMLs on the receive side of every optical link. Smaller revenue exposure than Lumentum/Coherent but supply-critical.
[67]
AXT (AXTI)
InP wafer substrates
Reported 60–70% share of the InP substrate market. The upstream-most input — every EML & CW laser is fabbed on AXT, Sumitomo, or Coherent InP wafers.
[66]
NTT Electronics
EMLs · DSPs · components
Japanese vertically integrated InP-to-component player; partners with Sumitomo on InP supply chain. Niche but strategically relevant for diversification.
[60][65]
Broadcom (in-house lasers)
VCSEL lasers (in-house)
Notable: Broadcom is one of the few non-pure-play optical companies that fabricates its own VCSEL lasers for short-reach optics. It does not manufacture EMLs in volume — those still come from Lumentum/Coherent.
CPO
(Also appears in Layer 3) [26][27]
Layer 4 → Layer 5 dependency: EML and CW lasers need InP wafer substrates (AXT, Sumitomo, Coherent) and increasingly advanced packaging (TSMC CoWoS / COUPE) to integrate with switch ASICs in CPO form.
5

Physical Layer & Advanced Packaging

Copper (intra-rack) + the foundry / packaging that lets CPO exist at all.
Passive copper / DACs / AECs (intra-rack, scale-up)
Amphenol (APH)
DACs · AECs · backplane interconnect
A top-3 high-speed copper interconnect supplier. Demonstrated Marvell-powered 1.6T AEC ecosystem at OFC 2025.
[31][70]
Molex
AECs · DACs · scale-up cabling
Active electrical cable + connector portfolio for AI rack-to-rack. Together with Amphenol + TE: >45% of AEC market.
[68][70]
TE Connectivity (TEL)
connectors · DACs · backplanes
Marvell ecosystem partner at OFC 2025 for 1.6T AECs. Strong NVIDIA/Broadcom reference designs.
[31][70]
3M
twin-axial cable · interconnect
High-performance twinax used in DACs and AECs; named Marvell partner.
[31]
Luxshare-Tech (002475.SZ)
connectors · cables · transceiver assembly
Major Chinese AEC/DAC/transceiver-assembly partner; ecosystem member for 800G + 1.6T builds.
[31][70]
Optical connectors & CPO hardware (high-density / expanded-beam)
US Conec (private)
MMC connectors · PRIZM™ TMT expanded-beam ferrules
Designs the MMC 16-/32-fiber compact connector and the PRIZM TMT expanded-beam ferrule — passive alignment, debris tolerance, blind-mate; 3× density of MPO. Multi-source framework with Corning, Fujikura, Sumitomo Electric Lightwave. Standard physical interface for CPO/NPO architectures.
CPO/NPO interface
[112][113]
Foxconn Interconnect Technology (FIT)
CPO sockets · PLS I/O cages · 102.4T laser platform
Supplies key CPO hardware to Broadcom: solderless LGA-to-LGA sockets and Remote Pluggable Laser Source (PLS) I/O cages. Partnered with MediaTek on a 51.2T CPO co-development. FIT's 102.4T CPO External Laser Solution has entered ecosystem validation.
CPO physical layer
Designed for Broadcom Tomahawk-6 / Davisson [114][115]
Composable / disaggregated infrastructure (CDI & CXL pooling)
One Stop Systems (OSS)
PCIe 5.0 accelerator chassis · Ponto Reef
Ponto Reef 32-slot PCIe 5.0 expansion system — industry-first 32 PCIe accelerator devices in a single chassis. Sold into CDI deployments for GenAI/ML, HPC, visualization. Initial AI-infrastructure design-win $2M PO, production Q1 2025.
PCIe/CXL CDI
[116][117]
Liqid (private)
composable GPU + CXL 2.0 memory pooling
Leading independent CDI vendor (with GigaIO). EX-5410P 10-slot 600W GPU box supporting H200 / RTX Pro 6000 / Intel Gaudi 3 + up to 100 TB CXL 2.0 disaggregated memory pool. Memory-wall play complementary to Marvell/XConn switching silicon.
CXL pooling
[117][118]
System management & control silicon
Lattice Semiconductor (LSCC)
low-power FPGAs · DC-SCM · PQC root-of-trust
"Companion-chip" strategy: small FPGAs that manage boot, security, control, and I/O adjacent to CPU/GPU/accelerators. 70–130 FPGAs per hyperscaler rack. AI-server revenue +85% YoY in FY 2025. $1.65B AMI acquisition adds firmware orchestration for AI DCs. Certus-NX / MachXO5-NX with DC-SCM + post-quantum crypto.
Rack control plane
[119][120]
Advanced packaging & silicon-photonics foundries (the CPO enablers)
TSMC (TSM)
CoWoS · SoIC · COUPE silicon photonics
COUPE = 6 nm EIC + 65 nm SOI PIC, SoIC-X hybrid bonding, on a CoWoS roadmap. 40% lower power or +170% bandwidth vs. micro-bump. Used by Broadcom Tomahawk-6 Davisson CPO, Ayar Labs, others. Qualified 2025 → CoWoS integration 2026.
CPO foundry
[71][72][73]
Intel Foundry
advanced packaging · hybrid laser SiPho
Long-running SiPho program with hybrid-laser coupling (>90% efficiency claimed). More selective vs. TSMC's aggressive COUPE rollout. Foveros packaging used for some AI accelerators.
CPO foundry (selective)
[74]
GlobalFoundries (GFS)
GF Fotonix silicon-photonics platform
Monolithic CMOS + silicon photonics platform. Used by Lightmatter Passage M1000 for 3D photonic interposer. DMEA Cat-1A Trusted Supplier for defense AI work.
SiPho foundry
[75]
Tower Semiconductor
silicon-photonics foundry
~85% historical SiPho foundry share (pre-TSMC ramp) per defense supply-chain mapping. Still the volume foundry for many of today's SiPho PICs.
SiPho foundry
M

Recent Networking M&A — Innovators absorbed into Layer-3 mega-caps

The companies below were independent in 2024; now they live inside the parent stack shown.
aiXscale Photonics → Astera Labs
closed Nov 2025
German precision-glass fiber-to-chip coupling for photonic chiplets. Gives Astera Labs a credible CPO physical-layer story to bolt onto Scorpio/Aries.
[121][122]
XConn Technologies → Marvell
$540M · closed Feb 10, 2026
Industry-first hybrid CXL 3.1 + PCIe Gen 6.2 switch on a single chip (Apollo 2: 256 lanes, 2,048 Gbps). Pairs with Marvell's CXL memory controllers + Celestial AI photonics for full UALink/CXL stack. Direct shot at Astera Scorpio and Microchip Switchtec Gen 6.
[123][124]
Polariton Technologies → Marvell
announced Apr 22, 2026
Swiss / ETH Zurich spin-out — plasmonics-based silicon photonics. Ultra-high-bandwidth, low-power modulators for 3.2T+ coherent / DCI. Marvell's third 2025–26 optical acquisition after Celestial AI and XConn.
[125][126]
DustPhotonics → Credo
$750M cash + ~0.92M shares · April 2026
Israeli SiPho PIC house — 400G/800G/1.6T PIC with roadmap to 3.2T. Turns Credo into a vertically-integrated SerDes-to-SiPho competitor to Marvell.
[41][42]
Celestial AI → Marvell
$3.25B (up to $5.5B) · Dec 2025
Photonic Fabric chiplet — 16 Tbps single chiplet for scale-up optical I/O and disaggregated memory. Now the core of Marvell's chip-to-memory optics roadmap.
[29][30][52]
HieFo → Semtech
$34M cash · March 2026
InP DFB-laser optoelectronics. Pulls Semtech from a pure analog/LPO play into a vertically-integrated DSP-laser-driver chipset competitor.
[100][101]
Alphawave Semi (Banias Labs) → Qualcomm
$2.4B Qualcomm deal · closed Dec 2025
Optical DSP + SerDes IP. Effectively gives Qualcomm a credible data-center networking-IP arm overnight — a new entrant to watch in Layer 3.
[105][106]
★

Protocol / Architecture Alignment Matrix (v3)

Four functional groups: scale-up · scale-out · communication libraries · optical strategies. Each cell shows which specific items inside the group a vendor supports.
Group 1 — Scale-up interconnect protocols (intra-rack / chip-to-chip)
Vendor NVLink UALink Google ICI PCIe / CXL
NVIDIA✓ owner——· (host PCIe)
AMD—✓ founder—✓
Intel—✓ founder—✓
Google——✓ owner (3D torus)·
Broadcom—✓ (SUE/UALink)—✓
Marvell (+ XConn)—✓—✓ Apollo 2 hybrid
Cisco—✓ (UALink consortium)—·
Astera Labs—✓—✓ Aries / Scorpio
Microchip (Switchtec)———✓ PCIe Gen 6
Upscale AI—✓ flagship (SkyHammer)—·
Ayar Labs—✓ (UCIe optical chiplet)—·
Celestial AI (Marvell)—✓ (scale-up fabric)—·
Liqid———✓ CXL 2.0 pool
One Stop Systems———✓ PCIe 5.0 CDI
Sources: [22][23][29][41][43][45][48][52][76][78][85][92][96][118][123][127][128]
Group 2 — Scale-out transport protocols & control planes (rack-to-rack)
Vendor InfiniBand Omni-Path RoCEv2 / UEC Spectrum-X AWS SRD Google Falcon MRC iWARP SONiC (NOS) Jupiter SDN
NVIDIA✓ Quantum (owner)—· (Spectrum-X is modified RoCEv2)✓ owner——✓ ConnectX-8 NIC + SRv6 on Spectrum-4/5 [138]—✓ Spectrum runs SONiC—
Broadcom——✓ Tomahawk/Jericho———✓ Thor Ultra NIC + SRv6 on Tomahawk 5 [138]—✓ supported—
Marvell——✓ Teralynx + DSPs—————✓ supported—
Cisco——✓ G300/P200—————✓ supported—
Arista——✓ Etherlink (flagship)———✓ SRv6 in EOS on Broadcom TH5 [138]—✓ supported—
HPE——✓ Slingshot—————✓ Aruba supports—
Intel (E810)——✓——✓ Falcon on E2100 IPU—✓——
AMD (Pollara / Vulcano)——✓ UEC-first NIC———✓ Pollara + Vulcano 800 [138]———
Cornelis Networks· (compares against)✓ owner (CN5000)✓ CN6000+ unified———————
Upscale AI——✓ UEC/ESUN—————✓ supported—
Microsoft Azure✓ for some clusters—✓ UEC· (consumer)——✓ production deployment [138]—✓ owner—
AWS——· (legacy)—✓ owner (Nitro/EFA)—————
Google—————✓ owner (HW transport)———✓ owner
xAI Colossus——·✓ flagship——————
OpenAI (training)·—✓ UEC✓ (Stargate)——✓ co-developed———
Meta——✓ flagship (Llama)—————✓ supported—
Innolight / Eoptolink✓—✓✓——————
MaxLinear✓—✓———————
Semtech (+ HieFo)——✓———————
Alphawave (Qualcomm)——✓———————
Credo (+ Dust)——✓———————
CoreWeave / Lambda / Nebius✓ flagship—··——————
Sources: [2][4][9][10][14][15][16][17][19][20][22][23][24][27][29][33][36][39][46][56][76][77][78][89][91][94][95][96][97][98][127][129][130][131][137][138]
Group 3 — Collective communication libraries (software orchestration layer)
Vendor NCCL RCCL MSCCL / MSCCL++
NVIDIA✓ owner——
AMD· (NCCL-compatible API)✓ owner (ROCm)—
Microsoft Azure✓ uses·✓ owner
Anthropic✓ uses—· (when on Azure)
OpenAI✓ uses· (AMD MI355 path)✓ (Azure clusters)
Meta✓ flagship——
xAI✓ flagship——
Google DeepMind· (XLA collectives on TPU; uses NCCL on GPU)——
Mistral / Nebius✓——
Libraries are software-only — they orchestrate collectives (all-reduce, all-gather) over whatever scale-up / scale-out fabric is underneath. Sources: [82][83][84]
Group 4 — Optical connection strategies (form factor & integration approach)
Vendor CPO LPO LRO NPO OCI / Photonic Fabric Apollo OCS Optical Interposer
NVIDIA✓ Quantum-X / Spectrum-X (Rubin+)——————
Broadcom✓ flagship (TH6-Davisson)·—————
Marvell (+ Celestial + Polariton)✓ + plasmonics·——✓ owner via Celestial——
Cisco· (future)✓ priority·————
Credo (+ Dust)✓·✓ Dove 850 (industry-first 800G LRO DSP)✓ explicit roadmap———
Astera Labs (+ aiXscale)✓ (fiber-chip glass coupler)——————
MaxLinear—✓ analog supplier (Rushmore)✓ Optomind / OFC2024 demo————
Ayar Labs✓ TeraPHY + SuperNova——✓ near-port optical I/O· (optical I/O)——
Lightmatter————· (photonic compute floor)—✓ Passage M1000
Celestial AI (Marvell)✓ scale-up———✓ flagship (Photonic Fabric)——
POET Technologies✓ enabler—————✓ POET Optical Interposer
Semtech (+ HieFo)—✓ flagship✓————
Lightwave Logic✓ EO-polymer materials✓ EO-polymer materials·————
Lucidean✓ CohZero-compatible——————
Neurophos✓ + in-network compute——————
Polaris Electro-Optics✓ EO-glass materials——————
Innolight / Eoptolink·✓ Eoptolink·————
Lumentum / Coherent✓ CW lasers for CPO·—————
Foxconn (FIT)✓ 102.4T CPO socket / PLS cage——————
US Conec✓ MMC / PRIZM TMT (CPO/NPO)——✓ same physical interface———
Google—————✓ owner (TPU pods)—
Huber+Suhner (HCF)———————
Sources: [11][24][26][29][30][33][41][42][46][48][49][50][52][53][55][56][57][60][61][102][103][104][107][108][110][112][114][115][125][132][133][134][135][136]
Reading the matrix: ✓ owner = vendor created / owns the standard. ✓ flagship = primary product line aligned to that standard. ✓ = ships a product or has explicit public commitment. · = supports / participates but not the primary bet. — = not in scope.

Spectrum-X note [137]: Spectrum-X is NVIDIA's Ethernet hardware platform, not a wire-level protocol. Its on-the-wire format is a heavily modified RoCEv2 with NVIDIA-proprietary adaptive routing & telemetry — that is why both columns light up for NVIDIA / xAI / OpenAI.
⇣

Cross-Layer Dependency Chain (worked examples)

Three end-to-end paths from frontier model → InP wafer.
Path A — xAI Colossus (Spectrum-X)
L1: xAI / Anthropic (renter) → L1.5: xAI-operated datacenter → L3: NVIDIA Spectrum-X SN5600 switch + BlueField-3 SuperNIC + LinkX optics → L3 wrapper: Innolight / Eoptolink transceiver modules → L4: Lumentum / Coherent 200G EML + Hamamatsu PD → L5: AXT/Sumitomo InP substrate + TSMC packaging.
[1][2][3][22][56][60][66][71]
Path B — Meta Llama (RoCEv2 / Arista)
L1: Meta FAIR → L2: Meta in-house DC with OCP Minipack3N / Wedge400 → L3 wrapper: Arista 7800R4 + 7700R4 DES (EOS software) → L3 builder: Broadcom Jericho4 + Tomahawk 5/6 + Thor Ultra NIC + in-house SiPho → L4: Lumentum/Coherent EMLs + Broadcom VCSEL → L5: TSMC CoWoS + COUPE packaging + Amphenol/Molex AECs intra-rack.
[5][6][7][8][26][27][31][60][68][71][73]
Path C — Google Gemini (TPU + Apollo OCS)
L1: Google DeepMind → L2 in-house: Google Jupiter SDN + Apollo MEMS OCS + custom TPU NIC → L3 selective: Broadcom custom-ASIC partnership for TPU + merchant DSPs/SerDes → L4: EML lasers (Lumentum/Coherent/Mitsubishi/Sumitomo) for inter-cube optical hops → L5: TSMC CoWoS for TPU packaging; intra-cube uses passive DACs.
[9][10][11][12][13][60][66][71]
Path D — Open-stack alt-fabric (UALink / UEC / MRC)
L1: any frontier lab seeking to escape NVLink lock-in (OpenAI on AMD MI355 is the validated reference) → L2: hyperscaler / neocloud running Ultra Ethernet rack with composable infrastructure → L3 scale-up fabric: Upscale AI SkyHammer (UALink) / Astera Scorpio / Microchip Switchtec / Marvell-XConn Apollo 2 → L3 NIC: AMD Pensando Pollara 400 → Vulcano 800 (MRC) → L3 fabric: Broadcom Tomahawk 6 / Cisco G300 / Arista Etherlink → L4: Semtech LPO modules + Lumentum CW lasers → L5: US Conec MMC + FIT CPO sockets + Lattice control FPGAs + Liqid/OSS CDI chassis.
[4][26][33][36][43][45][85][92][96][97][100][112][114][116][118][119][123]
Notes on construction. Layer 3 is the most informationally dense layer and the layer with the most contested classifications. The builder/wrapper/hybrid split is binary by design: a company is a builder if it owns the underlying silicon IP (switch ASIC, DSP, SerDes, SiPho PIC, or laser) it ships; a wrapper if it packages someone else's silicon into a system; hybrid if it does both (Cisco and HPE both qualify because their flagship lines use proprietary ASICs but legacy / non-flagship lines OEM Broadcom). Edge cases: NVIDIA is unambiguously a builder, but its transceiver (LinkX) bodies are built by ODM partners against NVIDIA-designed silicon. Marvell has recently moved from primarily a SerDes/DSP builder to a near-full-stack player after Celestial AI. Broadcom is the only pure builder spanning switch ASIC + DSP + SerDes + SiPho + VCSEL laser + custom-ASIC business.

On in-house networking at hyperscalers (Layer 2): the strict definition used here is owns at least one of (switch ASIC, switch OS, fabric protocol, optical-circuit-switch design, NIC ASIC). By that bar: Google (OCS + Jupiter SDN), AWS (EFA NIC + SRD), Microsoft (SONiC) all qualify. Meta is debatable — it designs OCP hardware and an RDMA transport stack but uses merchant ASICs. Oracle does not qualify. Neoclouds are all merchant-fabric resellers.

Verification. Every supplier-component claim in the map was cross-checked against at least one primary source (company press release, 10-Q, official documentation) and at least one secondary source (trade press) where available.

Last verified: 2026-05-17.

References

141 sources. Click any [#] citation above to jump here. Last verified May 18, 2026.
[1] xAI, "Colossus: The World's Largest AI Supercomputer." — Confirms 100K → 200K NVIDIA Hopper GPUs on Spectrum-X Ethernet https://x.ai/colossus
[2] NVIDIA Newsroom, "NVIDIA Ethernet Networking Accelerates World's Largest AI Supercomputer, Built by xAI." — Spectrum-X 95% throughput vs 60% standard Ethernet https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus
[3] xAI, "New Compute Partnership with Anthropic" (2026). — Anthropic renting Colossus 1 capacity, Spectrum-X exposure https://x.ai/news/anthropic-compute-partnership
[4] OpenAI, "MRC: Multipath Reliable Connection — Supercomputer Networking." — OpenAI custom transport stack signal https://openai.com/index/mrc-supercomputer-networking/
[5] Meta Engineering, "Meta open-sources transport stack to scale AI training to over 100K GPUs." — Meta's RDMA-over-Ethernet stack on Arista https://www.sdxcentral.com/news/meta-open-sources-transport-stack-to-scale-ai-training-to-over-100k-gpus/
[6] Arista Networks Blog, "Meta and Arista Build AI at Scale." — Meta's 24K H100 cluster using 7800 + Minipack2 + Wedge400 RoCE fabric https://blogs.arista.com/blog/meta-and-arista-build-ai-at-scale
[7] Data Center Frontier, "Closer Look: Meta's Custom ASIC for AI Computing (MTIA)." — Meta MTIA v2 inference accelerator https://www.datacenterfrontier.com/servers/article/33005340/closer-look-metas-custom-asic-for-ai-computing
[8] Data Center Dynamics, "Meta places order for its next-gen ASIC-powered AI servers, partners with Broadcom and Quanta." — Meta–Broadcom MTIA training collaboration https://www.datacenterdynamics.com/en/news/meta-places-order-for-its-next-gen-asic-powered-ai-servers-partners-with-broadcom-and-quanta-computer/
[9] Google Cloud Blog, "The evolution of Google's Jupiter data center network." — Jupiter + MEMS OCS, 5x speed, 40% less power https://cloud.google.com/blog/topics/systems/the-evolution-of-googles-jupiter-data-center-network
[10] Google Research, "Jupiter Evolving: Transforming Google's Datacenter Network via Optical Circuit Switches and Software-Defined Networking" (SIGCOMM '22) https://research.google/pubs/jupiter-evolving-transforming-googles-datacenter-network-via-optical-circuit-switches-and-software-defined-networking/
[11] SemiAnalysis, "Google OCS Apollo: The >$3 Billion Game-Changer in Datacenter Networking." — Apollo OCS scale, custom TPU pod interconnect https://newsletter.semianalysis.com/p/google-apollo-the-3-billion-game
[12] Introl Blog, "Google TPU Architecture: 7 Generations Explained" — TPU v4/v5/Ironwood OCS pod topology https://introl.com/blog/google-tpu-architecture-complete-guide-7-generations
[13] HPCwire, "Google Aims to Flip the Script on AI Inference with New Ironwood TPUs" — Ironwood 9,216 chips, 48 OCS units https://www.hpcwire.com/2025/11/10/google-aims-to-flip-the-script-on-ai-inference-with-new-ironwood-tpus/
[14] AWS, "Elastic Fabric Adapter (EFA)." — In-house OS-bypass NIC with SRD protocol on Nitro card https://aws.amazon.com/hpc/efa/
[15] IEEE Xplore, Shalev et al., "A Cloud-Optimized Transport Protocol for Elastic and Scalable HPC" — AWS SRD paper https://ieeexplore.ieee.org/document/9167399/
[16] Microsoft Azure Blog, "SONiC: the network innovation powerhouse behind Azure." — SONiC powers Azure + AI platforms; multi-ASIC https://azure.microsoft.com/en-us/blog/sonic-the-network-innovation-powerhouse-behind-azure/
[17] NVIDIA Technical Blog, "Choosing NVIDIA Spectrum for Microsoft Azure SONiC." — Azure runs SONiC on NVIDIA Spectrum silicon https://developer.nvidia.com/blog/choosing-spectrum-for-microsoft-azure-sonic/
[18] HUBER+SUHNER, "Strengthening collaboration with Microsoft for HCF" (Apr 2026). — Hollow-core fiber deployment in Azure backbone https://www.hubersuhner.com/en/newsroom/company-news/news-ad-hoc-news/hubersuhner-strengthens-collaboration-with-microsoft-to-advance-global-deployment-of-hollow-core-fi
[19] Network World, "Neoclouds roll in, challenge hyperscalers for AI workloads." — CoreWeave / Lambda / Nebius / Crusoe positioning https://www.networkworld.com/article/4011187/neoclouds-roll-in-challenge-hyperscalers-for-ai-workloads.html
[20] SemiAnalysis, "ClusterMAX™ 2.0: The Industry Standard GPU Cloud Rating System." — Neocloud NIC/fabric tracking (InfiniBand vs Spectrum-X) https://newsletter.semianalysis.com/p/clustermax-20-the-industry-standard
[21] DigitalOcean, "7 CoreWeave Alternatives for Cloud GPU Computing in 2025." — Voltage Park scale (24K H100s), neocloud fabric strategies https://www.digitalocean.com/resources/articles/coreweave-alternatives
[22] NVIDIA Newsroom, "NVIDIA Announces New Switches Optimized for Trillion-Parameter GPU Computing and AI Infrastructure." — Spectrum-X800, Quantum-X800, BlueField-3, LinkX https://nvidianews.nvidia.com/news/networking-switches-gpu-computing-ai
[23] NVIDIA Docs, "NVIDIA Quantum, Spectrum, and LinkX Product Lines." — Product catalog including Spectrum-4 SN5600, Quantum-2/3, ConnectX-8, LinkX optics https://docs.nvidia.com/networking/display/400g100gpam4ovpub/nvidia-quantum,-spectrum,-and-linkx-product-lines
[24] The Next Platform, "Nvidia Weaves Silicon Photonics Into InfiniBand And Ethernet" (Mar 2025). — CPO integration into Quantum-X and Spectrum-X https://www.nextplatform.com/2025/03/18/nvidia-weaves-silicon-photonics-into-infiniband-and-ethernet/
[25] Network World, "Nvidia networking roadmap: Ethernet, InfiniBand, co-packaged optics will shape data center of the future." — Quantum 3450-LD (late 2025), SN6810/SN6800 (2026), Vera Rubin CPO roadmap https://www.networkworld.com/article/4050881/nvidia-networking-roadmap-ethernet-infiniband-co-packaged-optics-will-shape-data-center-of-the-future.html
[26] Broadcom Investors, "Broadcom Announces Tomahawk® 6 – Davisson, the Industry's First 102.4-Tbps Ethernet Switch with Co-Packaged Optics." — 3rd-gen CPO, 70% optical interconnect power reduction https://investors.broadcom.com/news-releases/news-release-details/broadcom-announces-tomahawkr-6-davisson-industrys-first-1024
[27] Broadcom Blog, "Broadcom Advances Industry Support for Ethernet for AI Networking at the 2025 OCP Global Summit." — Tomahawk 6, Tomahawk Ultra, Jericho4, Thor Ultra 800G NIC https://www.broadcom.com/blog/2025-ocp-global-summit-broadcom-advances-industry-support-for-ethernet-for-ai-networking
[28] NADDOD Blog, "Broadcom's Tomahawk 6 Delivers 102.4 Tb/sec Ethernet for AI-Scale Fabrics" — Chiplet architecture, 200G SerDes specifics https://www.naddod.com/blog/broadcom-tomahawk-6-102-4-t-ethernet-switch-chip-for-ai-fabrics
[29] Marvell Investor Relations, "Marvell to Acquire Celestial AI, Accelerating Scale-up Connectivity for Next-Generation Data Centers" (Dec 2025). — $3.25B deal, scale-up optical I/O https://investor.marvell.com/news-events/press-releases/detail/1000/marvell-to-acquire-celestial-ai-accelerating-scale-up-connectivity-for-next-generation-data-centers
[30] The Next Platform, "With Celestial AI Buy, Marvell Scales Up The Datacenter And Itself." — Celestial AI 16 Tbps Photonic Fabric chiplet integration https://www.nextplatform.com/2025/12/04/with-celestial-ai-buy-marvell-scales-up-the-datacenter-and-itself/
[31] Marvell, "Marvell Extends Connectivity Leadership with AEC Ecosystem Demonstrations at OFC 2025." — Alaska A 1.6T AEC DSP, partners Amphenol/TE/Molex https://www.marvell.com/company/newsroom/marvell-extends-connectivity-leadership-with-aec-ecosystem-demonstrations-at-ofc-2025.html
[32] Marvell, "Marvell Optical DSPs | Powering the Future of AI Infrastructure." — Inphi-derived optical DSP roadmap https://www.marvell.com/solutions/data-center/optical-dsp.html
[33] Cisco Newsroom, "Cisco Announces New Silicon One G300, Advanced Systems and Optics to Power and Scale AI Data Centers for the Agentic Era" (Feb 2026). — 102.4 Tbps G300 ASIC, LPO-aligned optics https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2026/m02/cisco-announces-new-silicon-one-g300.html
[34] Cisco Data Sheet, "Cisco Silicon One G300." — 64x 1600 GbE, 512x 200G SerDes, scale-up/scale-out roles https://www.cisco.com/c/en/us/solutions/collateral/silicon-one/silicon-one-g300-ds.html
[35] The Register, "Cisco unveils 102.4T Silicon One G300 switch chip." — Independent confirmation of specs and AI positioning https://www.theregister.com/2026/02/10/cisco_challenges_broadcom_nvidia_switch_chips/
[36] Arista Networks Press Release, "Arista Unveils Etherlink AI Networking Platforms." — 7060X6 (Tomahawk 5), 7800R4 (Jericho3-AI), 7700R4 DES https://www.arista.com/en/company/news/press-release/19493-arista-unveils-etherlink-ai-networking-platforms
[37] Arista 7700R4 Data Sheet — Distributed Etherlink Switch specifications https://www.arista.com/assets/data/pdf/Datasheets/7700R4-Distributed-Etherlink-Switch-Datasheet.pdf
[38] Arista Press Release, "Arista Networks Unveils Next Generation Data and AI Centers" (Oct 2025). — Latest AI portfolio expansion https://www.arista.com/en/company/news/press-release/22541-pr-10292025
[39] The Next Platform, "HPE Upgrades Supercomputer Lineup Top To Bottom In 2025." — Slingshot 400 with Rosetta-2 switch + Cassini-2 NIC, in-house silicon https://www.nextplatform.com/2024/11/26/hpe-upgrades-supercomputer-lineup-top-to-bottom-in-2025/
[40] The Next Platform, "How Will Juniper Change HPE's Datacenter Networking Strategy?" — Aruba uses mix of own ASICs and Broadcom; AMD DPU partnership https://www.nextplatform.com/2025/07/02/how-will-juniper-change-hpes-datacenter-networking-strategy/
[41] Credo Investor Relations, "Credo Agrees to Acquire DustPhotonics" (Apr 2026). — $750M cash + ~0.92M shares; vertically integrated SerDes/DSP/SiPho https://investors.credosemi.com/news-events/news/news-details/2026/Credo-Agrees-to-Acquire-DustPhotonics-Accelerating-Expansion-into-Silicon-Photonics-and-Next-Generation-Optical-Connectivity/default.aspx
[42] BusinessWire/Credo, full release with $500M FY2027 optical revenue target https://www.businesswire.com/news/home/20260413933103/en/Credo-Agrees-to-Acquire-DustPhotonics
[43] Astera Labs Product Page, "Aries PCIe/CXL Smart DSP Retimers." — Aries 6 PCIe Gen 6 retimers, AECs https://www.asteralabs.com/products/pcie-cxl-smart-dsp-retimers/
[44] Astera Labs Products Overview — Aries (retimers), Taurus (Ethernet smart cable), Leo (CXL memory), Scorpio (fabric switch), COSMOS (software) https://www.asteralabs.com/products/
[45] Astera Labs Press Release, "Astera Labs Ramps Production of PCIe 6 Connectivity Portfolio." — Scorpio P-Series PCIe 6 fabric switches https://ir.asteralabs.com/news-releases/news-release-details/astera-labs-ramps-production-pcie-6-connectivity-portfolio
[46] MaxLinear Investors, "MaxLinear Unveils Rushmore: Low-Power 1.6T PAM4 DSP for AI/ML and Data Center Networks." — 200G/lane, <25W optical module support, Ethernet + InfiniBand compatible https://investors.maxlinear.com/press-releases/detail/569/maxlinear-unveils-rushmore-low-power-1-6t-pam4-dsp-for
[47] MaxLinear Investors, "MaxLinear Unveils Annapurna 224G Scale-Up Retimer to Extend Copper Connectivity in AI Data Centers." — 224G PAM4 retimer for AECs / backplanes; Q2 2026 availability https://investors.maxlinear.com/press-releases/detail/603/maxlinear-unveils-annapurna-224g-scale-up-retimer-to-extend
[48] Ayar Labs, "Ayar Labs Unveils World's First UCIe Optical Chiplet for AI Scale-Up Architectures" (Mar 2025). — TeraPHY 8 Tbps + SuperNova 16-wavelength laser https://ayarlabs.com/news/ayar-labs-unveils-worlds-first-ucie-optical-chiplet-for-ai-scale-up-architectures/
[49] Semiwiki, "Alchip and Ayar Labs' Co-Packaged Optics Breakthrough at TSMC OIP 2025." — 100+ Tbps per accelerator scale-up demo https://semiwiki.com/semiconductor-services/362133-revolutionizing-ai-infrastructure-alchip-and-ayar-labs-co-packaged-optics-breakthrough-at-tsmc-oip-2025/
[50] Lightmatter, "Lightmatter Unveils Passage M1000 Photonic Superchip, World's Fastest AI Interconnect." — 114 Tbps optical, 4,000 mm² interposer, 256 fibers https://lightmatter.co/press-release/lightmatter-unveils-passage-m1000-photonic-superchip-worlds-fastest-ai-interconnect/
[51] ServeTheHome, "Lightmatter Passage M1000 at Hot Chips 2025." — Architecture detail; runs on GF Fotonix SiPho process https://www.servethehome.com/lightmatter-passage-m1000-at-hot-chips-2025/
[52] Photonics Spectra, "Marvell to Acquire Celestial AI for $3.25B." — Photonic Fabric chiplet, scale-up optical I/O https://www.photonics.com/Articles/Marvell-to-Acquire-Celestial-AI-for-325B/a71734
[53] POET Technologies, "POET Technologies Redefines Optical Integration with Its Hybrid-Integrated 1.6T 2xFR4 Transmitter PIC." — POET Optical Interposer scalable to 3.2T https://www.poet-technologies.com/blog/poet-technologies-redefines-optical-integration-with-its-hybrid-integrated-1-6t-2xfr4-transmitter-pic
[54] POET Technologies, "POET Technologies Receives $5 Million Production Order for 800G Optical Engines." https://www.poet-technologies.com/news/poet-technologies-receives-5-million-production-order-for-800g-optical-engines
[55] POET Technologies / Quantum Computing Inc., "Co-Develop 3.2 Tbps Optical Engines for CPO." — TFLN modulator + POET interposer collaboration https://www.poet-technologies.com/news/poet-technologies-and-quantum-computing-inc-to-co-develop-3-2-tbps-optical-engines-for-cpo-and-next-gen-ai-connectivity
[56] IP-Fiber, "Nvidia Orders Surge: InnoLight and Eoptolink Dominate 60% of 800G SFP Optical Modules Supply." — NVIDIA supplier-share split https://ip-fiber.com/blogs/news/nvidia-orders-surge-innolight-and-eoptolink-dominate-60-of-800g-sfp-optical-modules-supply
[57] Eoptolink, "Eoptolink Unveils Industry-First 800G Optical Transceiver for Multicore Fiber at OFC 2025/2026." — 1.6T OSFP-XD product line https://www.eoptolink.com/news/358-eoptolink-unveils-industry-first-800g-optical-transceiver-for-multicore-fiber-at-ofc-2026
[58] SemiAnalysis, "Nvidia's Optical Ascent: >$1B Revenue; The Missing 800G Ramp." — NVIDIA optical revenue and Innolight/Eoptolink role https://newsletter.semianalysis.com/p/nvidias-optical-ascent-1b-revenue
[60] TrendForce, "AI Data Centers Ignite a Laser Shortage Wave; Nvidia's Strategic Lock-In Reshapes the Global Laser Supply Chain." — 200G EML constraint, multi-supplier capacity https://www.trendforce.com/presscenter/news/20251208-12823.html
[61] Chipstrat (Austin Lyons), "Lumentum and the Laser Bottleneck." — Lumentum 50–60% EML share; only volume supplier of 200G/lane EML https://www.chipstrat.com/p/lumentum-and-the-laser-bottleneck
[62] FinancialContent, "The Light Engine of AI: A Deep Dive into Lumentum Holdings (LITE) and the 1.6T Revolution." — Lumentum $663M capacity expansion https://markets.financialcontent.com/stocks/article/finterra-2026-3-26-the-light-engine-of-ai-a-deep-dive-into-lumentum-holdings-lite-and-the-16t-revolution
[63] Compound Semiconductor Magazine, "A tough 12 months for CS shares." — Coherent 6-inch InP wafer ramp https://compoundsemiconductor.net/article/121810/A_tough_12_months_for_CS_shares
[64] Sumitomo Electric, "53 GBaud Electro-Absorption Modulator Integrated Lasers." — EML technical paper and product line https://sumitomoelectric.com/sites/default/files/2023-04/download_documents/E96-05.pdf
[65] Sumitomo Electric, "Growth strategy for data-center-related business" (Nov 2025 deck). — Vertical InP-to-EML strategy https://sumitomoelectric.com/sites/default/files/2025-11/download_documents/Growth%20strategy%20for%20data%20center-related%20business_2025.pdf
[66] Mordor Intelligence, "Indium Phosphide (InP) Wafer Market — Size, Share & Industry Analysis." — AXT, Sumitomo, Coherent (II-VI) as InP substrate leaders https://www.mordorintelligence.com/industry-reports/indium-phosphide-wafer-market
[67] Journal of Semiconductors, "High-speed electro-absorption modulated laser." — EML technical baseline; Mitsubishi Electric, Hamamatsu cited among manufacturers https://www.jos.ac.cn/en/article/doi/10.1088/1674-4926/25030015
[68] Molex, "Active Electrical Cable Solutions." — AEC product family for AI rack interconnect https://www.molex.com/en-us/products/connectors/high-speed-pluggable-io/active-electrical-cables-aec
[69] Introl, "Cables and Interconnects: DAC, AOC, AEC, and Fiber for 800G AI Data Centers (2025)." — Inventory of copper-vs-optical reach and economics https://introl.com/blog/cables-interconnects-dac-aoc-aec-800g-ai-data-center-2025
[70] OpenPR / IntelMarketResearch, "AI Data Center Active Electrical Cable (AEC) Modules Research." — Molex/Amphenol/TE Connectivity collectively >45% AEC market share, 22% CAGR https://www.intelmarketresearch.com/aec-active-cable-for-data-centers-market-27749
[71] TrendForce, "Silicon Photonics in the Spotlight: TSMC Lifts the Curtain on COUPE at SEMICON Taiwan" (Sep 2025). — COUPE = SoIC-X stacked optical engine on CoWoS roadmap https://www.trendforce.com/news/2025/09/09/news-silicon-photonics-in-the-spotlight-tsmc-lifts-the-curtain-on-coupe-at-semicon-taiwan/
[72] 3D InCites, "IFTLE 642: TSMC Advanced Co-packaged Optics Integrated CoWoS and COUPE." — Technical detail: 6nm EIC + 65nm SOI PIC https://www.3dincites.com/2025/10/iftle-642-tsmc-advanced-co-packaged-optics-integrated-cowos-and-coupe/
[73] DigiTimes, "TSMC integrates CoWoS packaging with Coupe photonics for AI performance boost." — Broadcom Tomahawk 6 CPO uses TSMC COUPE https://www.digitimes.com/news/a20250812VL210/tsmc-cowos-performance-silicon-interposer.html
[74] TechSoda, "IP Insight #3: TSMC Accelerates in Silicon Photonics While Intel Becomes More Selective." — Intel Foundry SiPho positioning and hybrid-laser IP https://techsoda.substack.com/p/ip-insight-3-tsmc-accelerates-in
[75] Lightmatter Hot Chips 2025 paper — GlobalFoundries Fotonix silicon-photonics platform used by Lightmatter Passage https://lightmatter.co/wp-content/uploads/2025/08/57-lightmatter_hotchips_2025_final.pdf
[76] Ultra Ethernet Consortium official site — founding members AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, Microsoft; 100+ members by 2024 https://ultraethernet.org/
[77] HPCwire, "Ultra Ethernet Has Arrived: One Network to Rule Them All?" — UEC 1.0 (Jun 2025), 1.0.1 update https://www.hpcwire.com/2025/09/09/ultra-ethernet-has-arrived-one-network-to-rule-them-all/
[78] Tom's Hardware, "AMD unveils industry's first Ultra Ethernet ready network card for AI and HPC." — Pensando Pollara 400 UEC NIC https://www.tomshardware.com/networking/amd-unveils-industrys-first-ultra-ethernet-ready-network-card-for-ai-and-hpc
[79] SemiAnalysis, "The New AI Networks: Ultra Ethernet UEC, UALink, Broadcom SUE." — Scale-up vs scale-out framing https://newsletter.semianalysis.com/p/the-new-ai-networks-ultra-ethernet-uec-ualink-vs-broadcom-scale-up-ethernet-sue
[80] The Register, "Copackaged optics have officially found their killer app." — CPO momentum at Broadcom + NVIDIA https://www.theregister.com/2025/11/22/cpo_ai_nvidia_broadcom
[81] HPCwire, "Photonic Switches Promise to Keep GPUs Fed, Cool." — CPO power/latency benefits at scale https://www.hpcwire.com/2025/10/01/photonic-switches-promise-to-keep-gpus-fed-cool/
[82] NVIDIA Developer, "NVIDIA Collective Communications Library (NCCL)." https://developer.nvidia.com/nccl
[83] ROCm/rccl on DeepWiki — RCCL drop-in NCCL-compatible AMD library https://deepwiki.com/ROCm/rccl
[84] arXiv, "MSCCL++: Rethinking GPU Communication Abstractions for Cutting-Edge AI Applications." — MSCCL/MSCCL++ background, up to 5.4x speedup https://arxiv.org/html/2504.09014v2
[85] HPCwire, "Upscale AI Eyes Late 2026 for Scale-Up UALink Switch." — SkyHammer ASIC, 1,024 devices/rack via UALink 1.0, ~100–150 ns latency https://www.hpcwire.com/2025/12/02/upscale-ai-eyes-late-2026-for-scale-up-ualink-switch/
[86] Upscale AI, "Upscale AI Unveils SkyHammer™ Architecture." — Multi-standard support (ESUN, UEC, UALink) https://upscaleai.com/upscale-ai-unveils-skyhammer-architecture/
[87] The Next Platform, "Upscale AI Nabs Cash To Forge SkyHammer Scale-Up Fabric Switch." — $200M Series A + ~$100M seed, ~$300M total, unicorn status https://www.nextplatform.com/connect/2026/01/21/upscale-ai-nabs-cash-to-forge-skyhammer-scale-up-fabric-switch/4092095
[88] HPCwire, "Upscale AI Nabs $200M to Build a Scale-Up Interconnect That's Open." https://www.hpcwire.com/2026/01/22/upscale-ai-nabs-200m-to-build-a-scale-up-interconnect-thats-open/
[89] Cornelis Networks, "Cornelis Launches CN5000: Industry Leading AI and HPC Scale-Out Network." — 400G Omni-Path, 19.2 Tbps in 48-port switch, 230.4 Tbps director, 6× faster than RoCE for AI collectives https://www.cornelis.com/stories/cornelis-launches-cn5000-industry-leading-ai-and-hpc-scale-out-network
[90] The Register, "Omni-Path is back to take on InfiniBand?" — Spin-off context, CN5000 vs Quantum-2 NDR https://www.theregister.com/2025/06/09/omnipath_is_back/
[91] Network World, "Cornelis Networks offers alternative to Infiniband or Ethernet for HPC and AI networks." — CN6000 800G (Omni-Path + RoCE unified), CN7000 1.6T (UEC + Omni-Path) https://www.networkworld.com/article/4001125/cornelis-networks-offers-alternative-to-infiniband-or-ethernet-for-hpc-and-ai-networks.html
[92] Microchip Technology Press Release, "Microchip Unveils First 3 nm PCIe Gen 6 Switch to Power Modern AI Infrastructure." — Switchtec Gen 6: up to 160 lanes, 64 GT/s/lane, 15% power advantage vs. 5 nm https://www.microchip.com/en-us/about/news-releases/products/microchip-unveils-first-3-nm-pcie--gen-6-switch-to-power-modern
[94] Intel, "Intel Ethernet Network Adapter E810-CQDA2 — Product Brief." — 100 GbE Ethernet 800 series, RoCEv2 + iWARP, 256K queue pairs https://www.intel.com/content/www/us/en/products/sku/192558/intel-ethernet-network-adapter-e810cqda2/specifications.html
[95] Intel Support, "Which Intel Ethernet Network Adapters Support iWARP and RoCE v2?" — Explicit protocol map for 800 series https://www.intel.com/content/www/us/en/support/articles/000031905/ethernet-products/700-series-controllers-up-to-40gbe.html
[96] AMD, "AMD Pensando Pollara 400 AI Network Interface Card (NIC)." — Industry-first UEC-ready 400 Gbps AI NIC, validated with AMD Instinct MI350/MI355 https://www.amd.com/en/products/network-interface-cards/pensando.html
[97] AMD Blog, "From Innovation to Deployment Ready: AMD Advances AI Networking at Scale with MRC." — Pollara + Vulcano 800 supporting MRC transport https://www.amd.com/en/blogs/2026/amd-advances-ai-networking-at-scale-with-mrc.html
[98] Network World, "AMD rolls out first Ultra Ethernet-compliant NIC." https://www.networkworld.com/article/4011095/amd-rolls-out-first-ultra-ethernet-compliant-nic.html
[99] TechRadar, "Pollara 400GbE AI NIC leads AMD's push into UltraEthernet networking, Vulcano 800G coming for Gen6 clusters in 2026." https://www.techradar.com/pro/amd-debuts-a-400gbe-ai-network-card-with-an-800gbe-pcie-gen6-nic-coming-in-2026-but-will-the-industry-be-ready
[100] Semtech, "Semtech Expands Data Center Portfolio with Acquisition of HieFo Corporation." — $34M cash, DFB-laser optoelectronics, March 2026 close https://www.semtech.com/company/press/semtech-expands-data-center-portfolio-with-acquisition-of-hiefo-corporation
[101] Semiconductor Today, "Semtech expands data-center portfolio by acquiring HieFo for $34m." — IMDD + LPO integration strategy https://www.semiconductor-today.com/news_items/2026/mar/semtech-hiefo-050326.shtml
[102] Lightwave Logic Press Release, "Lightwave Logic Demonstrates World-Class 200 Gbps Heterogeneous Polymer/Silicon Photonic Modulator Results." https://www.prnewswire.com/news-releases/lightwave-logic-demonstrates-world-class-200gbps-heterogeneous-polymersilicon-photonic-modulator-results-302097749.html
[103] Lightwave Logic + Tower Semiconductor partnership — EO polymer integration on Tower PH18 SiPho PDK https://finance.yahoo.com/news/lightwave-logic-tower-semiconductor-announce-203000189.html
[104] StockTitan / Lightwave Logic, "Lightwave Logic modulators added to GF silicon photonics PDK." — GlobalFoundries Fotonix integration (Mar 2026) https://www.stocktitan.net/news/LWLG/lightwave-logic-high-speed-modulator-platform-now-available-in-gds-nsl6r3dsd39v.html
[105] Alphawave Semi Press Release, "Acquisition of Optical DSP Developer Banias Labs." — $240M, 2022 deal https://awavesemi.com/press-release/acquisition-of-optical-dsp-developer-banias-labs/
[106] Qualcomm Press Release, "Qualcomm Completes Acquisition of Alphawave Semi" (Dec 2025). — $2.4B; Alphawave brings SerDes + Banias optical DSP IP into Qualcomm data-center push https://www.qualcomm.com/news/releases/2025/12/qualcomm-completes-acquisition-of-alphawave-semi
[107] BusinessWire, "Lucidean Raises Series Seed Funding to Accelerate Next-Gen Coherent Optical Links for Data Centers" — $18M seed Dec 2025; CohZero coherent-lite IP for 3.2T/6.4T transceivers, compatible with CPO https://www.businesswire.com/news/home/20251223818151/en/Lucidean-Raises-Series-Seed-Funding-to-Accelerate-Next-Gen-Coherent-Optical-Links-for-Data-Centers-Announces-Dr.-James-Raring-as-Chief-Executive-Officer
[108] TechCrunch, "From invisibility cloaks to AI chips: Neurophos raises $110M to build tiny optical processors for inferencing" (Jan 2026). — Metasurface optical processor for in-network compute https://techcrunch.com/2026/01/22/from-invisibility-cloaks-to-ai-chips-neurophos-raises-110m-to-build-tiny-optical-processors-for-inferencing/
[109] EE Times, "Neurophos Photonic AI Test Chip Hits 300 TOPS/W." — Performance benchmark for the metasurface OPU https://www.eetimes.com/neurophos-photonic-ai-test-chip-hits-300-tops-w/
[110] Polaris Electro-Optics company site. — Ferroelectric nematic glass integrated with SiPho; 400 Gbps/lane and beyond; CPO-ready drop-in https://polariseo.com/
[111] Yole / Polaris Electro-Optics, "Integrated photonics startup polaris electro-optics adds $3m in financing" (and subsequent $10M Series A 2025) https://www.yolegroup.com/industry-news/integrated-photonics-startup-polaris-electro-optics-adds-3m-in-financing/
[112] US Conec / PRNewswire (multi-source framework), "US Conec, Corning, Fujikura, and Sumitomo Electric Lightwave Announce Multi-Source Framework for PRIZM TMT Expanded Beam Ferrules and Associated MMC Connectors." — 3× density vs MPO; CPO/NPO-ready https://www.prnewswire.com/news-releases/us-conec-corning-incorporated-fujikura-and-sumitomo-electric-lightwave-announce-multi-source-framework-for-prizm-tmt-expanded-beam-ferrules-and-associated-mmc-connectors-302710483.html
[113] US Conec, "MMC Connector Solutions." — Product overview + CPO/NPO applications https://www.usconec.com/featured-products/mmc-connector
[114] Foxconn Interconnect Technology (FIT), "FIT Advances Broadcom's Co-Packaged Optics (CPO) Initiative with High-Performance Interconnect Solutions." — Solderless LGA sockets + Remote PLS I/O cages for Broadcom CPO https://www.fit-foxconn.com/main/modules/MySpace/BlogInfo.php?xmlid=27625
[115] FIT, "FIT Advances 102.4T CPO External Laser Platform into Validation." — 102.4T CPO External Laser Solution validation https://www.fit-foxconn.com/main/modules/MySpace/index.php?xmlid=28162
[116] StockTitan / One Stop Systems, "One Stop Systems Lands $2M AI Infrastructure Deal for Revolutionary PCIe 5.0 System." — Ponto Reef 32-accelerator PCIe 5.0 system for CDI https://www.stocktitan.net/news/OSS/oss-announces-design-win-with-an-award-winning-ai-infrastructure-0rwi8kixm1nz.html
[117] Introl, "Disaggregated Computing for AI" — Liqid + GigaIO CDI vendor framing; CXL pooling https://introl.com/blog/disaggregated-computing-composable-infrastructure-cxl-ai-2025
[118] Liqid Blog, "Getting the Most Out of Your GPUs — How Composable Disaggregated Infrastructure Can Help." — Liqid EX-5410P with H200 / RTX Pro 6000 / Gaudi 3 support, CXL 2.0 memory pooling up to 100 TB https://www.liqid.com/blog/getting-the-most-out-of-your-gpus-how-composable-disaggregated-infrastructure-can-help
[119] Lattice Semiconductor + AMI acquisition / FY25 server narrative. — AI servers +85% YoY; 70–130 FPGAs per hyperscaler rack; $1.65B AMI deal https://futurumgroup.com/insights/lattice-semiconductor-q4-fy-2025-record-comms-compute-ai-servers-85/
[120] Lattice Semiconductor Press Release, "Lattice Expands Low Power, Small FPGA Portfolio with High I/O Density and Secure Device Options." — Certus-NX / MachXO5-NX; DC-SCM; PQC root-of-trust https://ir.latticesemi.com/news-releases/news-release-details/lattice-expands-low-power-small-fpga-portfolio-high-io-density
[121] Astera Labs, "Astera Labs to Acquire aiXscale Photonics" (Oct 22, 2025). — Glass fiber-chip coupling IP for rack-scale photonic chiplets https://www.asteralabs.com/news/astera-labs-to-acquire-aixscale-photonics/
[122] Astera Labs, "Astera Labs Completes Acquisition of aiXscale Photonics" (Nov 2025) https://www.asteralabs.com/astera-labs-completes-acquisition-of-aixscale-photonics/
[123] Marvell Investor Relations, "Marvell to Acquire XConn Technologies, Expanding Leadership in AI Data Center Connectivity." — $540M; XConn Apollo 2 hybrid CXL 3.1 + PCIe Gen 6.2 switch on one die https://investor.marvell.com/news-events/press-releases/detail/1004/marvell-to-acquire-xconn-technologies-expanding-leadership-in-ai-data-center-connectivity
[124] Marvell Press Release, "Marvell Completes Acquisition of XConn Technologies" (Feb 10, 2026) https://www.marvell.com/company/newsroom/marvell-completes-acquisition-of-xconn-technologies.html
[125] Marvell Press Release, "Marvell Announces Acquisition of Polariton Technologies, Advancing Optical Performance Scaling to 3.2T and Beyond" (Apr 22, 2026). — Swiss plasmonics-based modulator IP for 3.2T+ coherent / DCI https://investor.marvell.com/news-events/press-releases/detail/1020/marvell-announces-acquisition-of-polariton-technologies-advancing-optical-performance-scaling-to-3-2t-and-beyond
[126] Data Center Dynamics, "Marvell acquires silicon photonics device startup Polariton Technologies to advance optical scaling." — ETH Zurich spin-out context https://www.datacenterdynamics.com/en/news/marvell-acquires-silicon-photonics-device-startup-polariton-technologies-to-advance-optical-scaling/
[127] Google Cloud Blog, "Inside the Ironwood TPU codesigned AI stack." — ICI 3D-torus topology, 6 neighbors per chip, up to 1.2 TBps bidirectional per chip; underpinning of TPU scale-up domain https://cloud.google.com/blog/products/compute/inside-the-ironwood-tpu-codesigned-ai-stack
[128] FiberMall, "Unveiling Google's TPU Architecture: OCS Optical Circuit Switching — The Evolution Engine from 4×4×4 Cube to 9216-Chip Ironwood." — ICI + Apollo OCS combined to form TPU pods https://www.fibermall.com/blog/unveiling-google-tpu-architecture.htm
[129] Google Cloud Blog, "Introducing Falcon: a reliable low-latency hardware transport." — Hardware NIC transport at 200 Gbps, ~3 µs one-way, scalable to 400/800 Gbps; sits on UDP/IP; supports IB Verbs RDMA + NVMe ULPs; deployed in Intel E2100 IPU https://cloud.google.com/blog/topics/systems/introducing-falcon-a-reliable-low-latency-hardware-transport
[130] ACM SIGCOMM 2025, "Falcon: A Reliable, Low Latency Hardware Transport." — Peer-reviewed protocol paper https://dl.acm.org/doi/10.1145/3718958.3754353
[131] Converge Digest, "Hot Interconnects 2025: Google Unveils Falcon, Swift, and Firefly for AI Data Centers." — Falcon + Swift (congestion control) + Firefly (clock sync) trio for AI fabrics https://convergedigest.com/at-hoti-2025-google-unveils-falcon-swift-and-firefly-for-ai-data-centers/
[132] Credo Semiconductor Press Release, "Credo Introduces World's First 800G DSP for Linear Receive Optics, Targeting Hyperscale and AI Data Centers" (Nov 2023). — Dove 850 LRO DSP, up to 50% power reduction; foundational LRO product https://credosemi.com/news/credo-introduces-worlds-first-800g-dsp-for-linear-receive-optics-targeting-hyperscale-and-ai-data-centers/
[133] Juniper Networks, "Types of Optics" (800G optics cable guide). — Authoritative definitions for LPO vs LRO (Half-Retimed Linear Optics / HALO) https://www.juniper.net/documentation/us/en/hardware/800g-optics-cables-guide/optics/topics/concept/800g-optic-types-lpolro.html
[134] MaxLinear Investor Relations, "Optomind and MaxLinear Demonstrate 800G LRO Optical Modules and AOCs at OFC2024." — Multiple-vendor LRO ecosystem confirmation https://investors.maxlinear.com/press-releases/detail/533/optomind-and-maxlinear-demonstrate-800g-lro-optical-modules
[136] Lightwave Online, "Credo's DustPhotonics acquisition solidifies its AI interconnect position." — NPO/CPO context for hyperscale AI optical roadmap https://www.lightwaveonline.com/home/article/55370917/credos-dustphotonics-acquisition-solidifies-its-ai-interconnect-position
[137] SDxCentral, "Inside Spectrum-X: NVIDIA's Ethernet networking platform." — Spectrum-X is the platform brand for NVIDIA's modified-RoCEv2 stack on Spectrum-4/Spectrum-6 silicon + BlueField SuperNICs. Not a separate transport protocol https://www.sdxcentral.com/analysis/inside-spectrum-x-nvidias-ethernet-networking-platform/
[138] OpenAI / Microsoft / AMD / Broadcom / NVIDIA joint whitepaper, "Resilient AI Supercomputer Networking using MRC and SRv6." — Authoritative implementation list: MRC on NVIDIA ConnectX-8, AMD Pollara + Vulcano, Broadcom Thor Ultra; SRv6 on NVIDIA Spectrum-4/Spectrum-5 (Cumulus + SONiC) and on Broadcom Tomahawk 5 inside Arista EOS. In production at OpenAI and Microsoft training frontier LLMs for ChatGPT and Codex. Spec released via OCP under an open license. (User-supplied PDF, 2026.)
[139] LightCounting, "AI creates a new wave in demand for optical transceivers and accelerates LPO/CPO adoption" (Jan 2025); "Scale-up networks in AI Clusters is a new market for optical interconnects" (Jul 2025). — AI cluster optics forecast doubles from $5B (2024) to $10B+ (2026); >100 product categories sorted by reach/form factor including LPO/LRO and CPO/NPO; CPO emerges as best option for scale-up by 2028. https://www.lightcounting.com/newsletter/en/january-2025-optics-for-ai-clusters-319
[140] Dell'Oro Group, "Large Scale AI Clusters to Fuel More Growth in Coherent Optical Transceiver Shipments." — Forecasts LPO/CPO ports together >30% of all 800G and 1.6T ports deployed in 2026–2028; AI back-end networks >$20B by 2028. https://www.delloro.com/news/large-scale-ai-clusters-to-fuel-more-growth-in-coherent-optical-transceiver-shipments/
[141] SemiAnalysis, "Co Packaged Optics (CPO) – Scaling with Light for the Next Wave of Interconnect." — NVIDIA copper-first for intra-rack scale-up through 2027–2028; CPO used to extend scale-up *between* Rubin Ultra racks in NVL576; Celestial AI scale-up CPO shipping with Amazon Trainium 4 at ~$1B run rate by end of CY28; "power tax of 1.6T pluggables forces CPO out of R&D and into mandatory deployment." https://newsletter.semianalysis.com/p/co-packaged-optics-cpo-book-scaling
↑ Back to top