Networking Watch: Data Movement Becomes the AI Bottleneck, Qualcomm x Amazon, the Fiber Gold Rush, and 1.6T Goes Native

The CODEW Networking Watch | September 9, 2026

When Data Movement Becomes the AI Bottleneck — Qualcomm x Amazon, the Fiber Gold Rush, and 1.6T Goes Native.

Executive Brief

The AI Infrastructure Race Becomes a Data-Movement Race

The biggest signal this week: networking, optics, and fiber have become first-order constraints, as compute expansion pulls connectivity deeper into the strategic infrastructure stack.

On Sept 8, Qualcomm and Amazon announced a multi-generation custom silicon partnership for AI inference covering up to $60B of Qualcomm AI data-center chips and products, with a warrant giving Amazon the right to buy a ~$4B stake at $161.26/share. Critically, the deal includes high-speed optical connectivity co-development extending to 1.6T and future-gen — compute and interconnect engineered together. It follows a cascade of fiber commitments: Amazon x Corning multibillion-dollar deal (June 8, 1,000 jobs NC), Verizon x Google >$1B dark fiber deal (July 24), Zayo x Corning long-term supply through 2030 for 15,000 route miles (including 8,000 long-haul miles with Nvidia), Nvidia $500M to Corning (May, 3 plants, 10x optical connectivity capacity), and Meta $6B with Corning (Jan 2026). Underneath: Broadcom Tomahawk 6 at 102.4 Tbps is now baseline for 1.6T fabrics, CPO breaks the power wall, and Nvidia Spectrum-XGS adds scale-across to link distributed data centers into giga-scale super-factories.

1. Networking at a Glance

Top Developments

Qualcomm x Amazon (Sept 8): Multi-generation custom AI inference chips for AWS, up to $60B product commitment, $4B warrant at $161.26 vesting with purchases. Joint optical connectivity development to 1.6T. Qualcomm expanding AWS use for chip design. Target: $15B data-center chip revenue by 2029. Amazon custom chip business run-rate >$25B annualized end of June quarter.
Fiber gold rush: Amazon-Corning multibillion deal (June 8, 1,000 NC jobs + training program), Verizon-Google >$1B dark fiber (July 24, CEO Dan Schulman says more multi-billion deals coming), Zayo-Corning long-term fiber through 2030 for 15,000 route miles (8,000 new long-haul with Nvidia), Nvidia $500M to Corning (May, 10x capacity, 50%+ US fiber production increase).
1.6T becomes product: Arista 7060XE7 family (64x 1.6T air/liquid-cooled, 128x 800G) on Broadcom Tomahawk 6 102.4T switch; Broadcom Taurus BCM83640 3nm DSP enabling 1.6T optics + foundation for 3.2T. Multiple deployments >100K accelerators planned.
Scale-across emerges: Nvidia Spectrum-XGS Ethernet connects distributed data centers into a unified supercomputer, 1.6x bandwidth density vs off-the-shelf Ethernet, with ConnectX-8 SuperNICs and distance-adaptive algorithms. Third pillar beyond scale-up/out.
Corning targets $10B inside-the-box opportunity: As AI-driven optics demand moves from outside plant to inside data center (high-density trunk cables with hundreds/thousands of fibers, GlassWorks AI™).

2. AI Network Watch

Inference Is the New Networking Driver

Training grabbed headlines, but this week's deal centers on AI inference — running trained models at scale, now the fastest-growing market and battleground among semiconductor firms. Inference traffic pattern differs from training: less synchronized collective burst, more fan-out, low-latency response, and massive east-west model copy, dataset movement, and retrieval-augmented generation (RAG) lookups between storage, compute, and edge.

Follow the data implications:

Latency budget: Inference SLOs require predictable tail latency. Qualcomm-Amazon explicitly pairs custom inference silicon with optical interconnect (1.6T) — acknowledging that interconnect latency, not just FLOPS, determines tokens/sec at scale.
Topology shift: From single gigawatt campus to distributed campuses. Nvidia Spectrum-XGS description: "connect our data centers into a single, unified supercomputer" — algorithms adapt the network to the distance between facilities. Data movement becomes an inter-DC fabric problem, not just intra-DC.
Power wall: Traditional pluggable copper/electrical interconnects hit a power wall at 102.4T fabrics. CPO integrates optical engines onto the switch package, slashing energy per bit. Broadcom claims 30% less power, 40% less cost per bit vs conventional. Without photonics, 1.6T cannot scale.

3. Data Center Networking

From Rack-Scale to Gigascale: 102.4T Fabrics Become Baseline

Data center Ethernet switch market hit $15.4B in Q1 2026 alone (+39.8% YoY), almost entirely AI build-outs. Three forces converging this week:

Radix doubling: Tomahawk 6 at 102.4 Tbps — double any prior Ethernet switch — enables 64x 1.6T or 128x 800G in a single chassis. Arista Networks CEO: collaboration showcases co-innovation to meet relentless bandwidth/latency demands of next-gen AI.
Liquid + optics: Arista 7060XE7 offers air-cooled and liquid-cooled variants, with low-power optics. Broadcom Sian3 (3nm) delivers 20% power reduction for 1.6T modules. Sian2M with integrated VCSEL drivers targets short-reach cost/power efficiency.
Scale-up/out/across convergence: Traditionally, scale-up (NVLink) = inside server/rack ultra-low latency; scale-out = across thousands of servers (Ethernet vs InfiniBand). Now Spectrum-XGS adds a third pillar: linking data centers across cities/nations/continents into giga-scale AI super-factories. Arista notes working with AMD on Infinity Fabric + Tomahawk 6 for massive-scale low-latency GPU clusters.

4. Ethernet & Interconnect Watch

Ultra Ethernet 1.0 Matures While Vendors Ship

Ultra Ethernet Consortium (UEC) released Specification 1.0 on June 11, 2025 — 560+ pages, comprehensive Ethernet-based stack for AI/HPC: modern RDMA, Ultra Ethernet Transport over UDP 4793, open NIC/switch/optics/cable interfaces. Hosted under the Linux Foundation, founded in 2023 by AMD, Arista, Broadcom, Cisco Systems, HPE, Intel, Meta, Microsoft — now 50+ members including Dell, Samsung, Huawei. Chair Dr. J Metz called it a major milestone to redefine Ethernet for the AI age.

Where we are Sept 2026:

Spec vs product gap: Most 2026 production deployments still run vendor-specific RoCEv2 rather than native UET. Enterprises should treat "UEC-ready" as a roadmap, not a certified implementation.
Competing architectures: Nvidia Spectrum-X provides 1.6x bandwidth density vs off-the-shelf Ethernet with adaptive routing, Direct Data Placement, and congestion control. Nvidia participates in UEC while shipping Spectrum-X — hedging proprietary (NVLink/InfiniBand) and open (UALink, UEC).
Interoperability test: IP Infusion OcNOS AI Fabric tracks UEC 1.0 fabric profile details, positioning open AI fabric vs closed single-vendor stack. Risk: orgs delaying fabric decisions until UEC matures may find themselves 12-18 months behind early adopters.

5. Optical Networking

The 1.6T Surge: From Pluggable to Co-Packaged

As Nvidia and Broadcom race to deploy 102.4T switching fabrics, limitations of traditional pluggable copper/electrical interconnects have become a power wall that only photonics can scale. Breakthrough: integrating optical engines directly onto the processor package — Co-Packaged Optics (CPO).

1.6T modules now commercial: Cisco Systems, Ciena, Naddod offer 1.6T optics; Marvell joining, Broadcom Taurus DSP enables more designers with compact 3nm design. Sian3 sampling to early customers, foundation for 3.2T leap.
CPO economics: Broadcom co-packaged optical switch claims 30% less power, 40% less cost per bit. Optical interface moves from pluggable module into chassis — reduces physical distance between switch silicon and photonics. Nvidia Quantum-X Photonics InfiniBand switches expected later in 2025, Spectrum-X Photonics Ethernet switches in 2026 from infrastructure vendors.
Qualcomm entry: Deal includes optical communications technology from Qualcomm after $2.4B AlphaWave acquisition last year. CEO Tony Pialis is now Qualcomm's data center chip chief. Analyst Bob O'Donnell: "By incorporating both compute and optical interconnect, Qualcomm is highlighting range of semiconductor technologies they can uniquely bring to AI infrastructure."
Hollow-core frontier: Microsoft x Corning x Heraeus manufacturing agreement focused on hollow-core fiber (HCF) offering lower latency (reduced group index) for latency-sensitive AI workloads and inter-DC transport — aligned with ITU-T G.650.

6. Cloud & Enterprise Networking

Hyperscalers Vertically Integrate the Fiber Layer

The structural shift: hyperscalers transition from passive consumers of optical components to active participants in manufacturing scale-up via anchor tenancy and vertical influence. Not just buying bandwidth — underwriting capacity expansion, securing preferential access, influencing fiber specifications and roadmaps.

Deal inventory 2026:

Amazon-Corning (June 8): Multiyear, multibillion — optical fiber, cable, connectivity for expanding US data center footprint. 1,000 new highly skilled jobs in NC, hundreds of construction jobs, workforce training with Catawba Valley Community College. Amazon invested >$20B in NC, 26,000 jobs.
Zayo-Corning (Aug 20): Expanded strategic supply agreement securing a significant portion of projected fiber through the remainder of the decade. Execution behind Zayo plan to add 15,000 route miles through 2030 (8,000 new long-haul in collaboration with Nvidia).
Meta-Corning (Jan 2026): Multiyear up to $6B for data center fabrics — leaf-spine topologies, IEEE 802.3bs/ck (400G/800G), parallel single-mode fiber (PSM) and WDM (ITU-T G.694.x).
Nvidia-Corning (May 2026): $500M commitment for 3 new optical manufacturing facilities in NC and Texas — 10x US-based optical connectivity manufacturing capacity, 50%+ domestic fiber production increase — targeting high fiber count, low-latency links for AI cluster interconnect, aligned with IEEE 802.3 Ethernet and emerging CPO ecosystems.
Lumen Private Connectivity Fabric: Modular network fabric allowing custom design — dedicated access to existing fiber, installation of new fiber on existing/new routes, digital services. Microsoft is deploying for AI-ready fiber, boosting capacity/performance/stability/speed between DCs. Lumen reserves 10% of Corning global fiber capacity each of the next two years — largest cable purchase.

Enterprise implication: AI corridors extending fiber into areas where power exists, and data centers are planned. Tier 2 markets become next frontier as Northern Virginia constrained. Enterprises need programmable, cloud-ready network that reaches every part of business — Lumen extending Multi-Cloud Gateway beyond own fiber footprint to 10M+ US business locations.

7. Network Security

SASE Surge as AI Expands Attack Surface

SASE consolidation accelerated ahead of AI infrastructure build-out. Key moves tracked:

CyberFOX acquires Timus Networks (June 29, 2026): 100% cloud-native SASE with ZTNA added to MSP-focused cybersecurity platform (PAM, password management, DNS filtering). Driver: product consolidation for secure network access.
Zscaler Zenith Live 2026 (Las Vegas): Extended Zero Trust SASE to eliminate firewall-based segmentation, arguing legacy SASE built on firewalls/VPNs is not designed for unmanaged devices and supply chains. Focus on disrupting SD-WAN, NAC, and firewall-based security for east-west and OT networks.
Cradlepoint (Ericsson) acquires Ericom: Build platform for cellular-first customers to integrate SASE, zero trust, cloud security across wireless and 5G/cellular networks — critical as AI inference moves to the edge.

Why it matters for networking:

As data moves from data center to distributed super-factories (Spectrum-XGS), east-west security becomes a north-south security challenge — requiring zero trust networking at optical and IP layers.
AI infrastructure multi-tenancy (CoreWeave, xAI largest AI supercomputer on Spectrum-X) demands performance isolation + security isolation simultaneously.

8. Wireless & Edge

Two Networks for AI: Inside the Brain and Between Brains

Analysis from IEEE ComSoc framing gains traction: AI requires two different networks. The first connects processors inside the brain (scale-up NVLink/UALink, scale-out Ethernet/InfiniBand). The second connects different brains with the outside world — where telcos have a stronger position.

Verizon-Google $1B+ dark fiber (July 24): Verizon will provide dark fiber connectivity for Google data centers, using existing fiber where possible, constructing new routes where necessary, offering dark or lit managed capacity. Google already operates >2M miles of lit fiber, 33 subsea cables, 200+ edge locations, thousands of CDN sites — yet needs Verizon for additional routes. Signal: hyperscalers cannot build every nerve themselves at this pace.
Telco opportunity = when data leaves campus: Models copied between regions, datasets moved storage→compute, private connections to cloud platforms, inference results to factories/vehicles/hospitals/stores. AI therefore elevates long-haul/metro fiber as strategic infrastructure, not commodity bandwidth.
S&P Global Market Intelligence survey (June 17, 2026): 87% of telecom providers in NA and Europe are deploying fiber optics; FTTH seen as end-game wireline, but now a dual-use asset supporting residential + hyperscalers. AI as fiber catalyst intensifies.

9. Silicon & Hardware

Compute + Photonics Convergence

This week collapses compute and networking silicon into a single story:

Qualcomm: After $2.4B AlphaWave purchase (CEO Tony Pialis becomes data-center chip chief), now couples custom AI inference chips with optical communications tech. By incorporating both compute and optical interconnects, Qualcomm highlights a range of semiconductor tech uniquely for AI infrastructure.
Broadcom: Tomahawk 6 (102.4 Tbps) + Sian3 (3nm, 20% power reduction for 1.6T optical modules) + Sian2M (integrated VCSEL drivers + 200G VCSELs for short-reach cost/power). Sampling to early customers. Taurus BCM83640 compact 3nm DSP powering 1.6T optical modules, foundation for 3.2T.
Arista: 15-year partnership with Broadcom, EOS feature-rich + Tomahawk 6 for AI-aware routing based on open Ethernet standards, intelligent packet buffering for microbursts typical of AI communication patterns.
Nvidia: Spectrum-4 Ethernet switch + BlueField-3 DPU + ConnectX-8 SuperNIC + Quantum-X Photonics InfiniBand (later 2025) + Spectrum-X Photonics Ethernet (2026) from vendors including Coherent, Eoptolink, Fabrinet, Innolight. CoreWeave first customer for XGS — "access to giga-scale AI."
Google angle: Marvell Technology struck a similar custom AI chip deal with Alphabet's Google weeks earlier, with the right to buy a stake worth up to $12.2B — a pattern of warrant-based financing tying cloud giants to custom silicon suppliers.

10. M&A & Competitive Landscape

Deal / Move	Parties	Value	Strategic Rationale
Custom AI inference chips + optical interconnect	Qualcomm → Amazon/AWS (multi-gen)	Up to $60B product + $4B warrant @ $161.26	Enable customized silicon at scale for AI data centers + 1.6T optical; Qualcomm diversifies beyond smartphones, Amazon drives $25B+ custom chip run-rate
Dark fiber for AI DCs	Verizon → Google	>$1B (multi-year), more multi-B deals expected	Connect Google data centers for AI workloads; signal dark fiber as competitive differentiator.r
Fiber supply chain lock-up	Zayo + Corning (expanded)	Long-term through 2030	Secure fiber for 15,000 route miles (8,000 new long-haul with Nvidia), AI-ready infrastructure
Optical manufacturing expansion	Amazon + Corning	Multibillion	1,000 NC jobs, fiber for expanding US DC footprint, workforce training program
Optical manufacturing expansion	Nvidia + Corning	$500M	3 new facilities in NC/Texas, 10x optical connectivity capacity, 50%+ fiber production increase for AI clusters
Data center fiber fabrics	Meta + Corning	Up to $6B (Jan 2026)	Large-scale fabrics, high-fiber-density leaf-spine, 400G/800G Ethernet
SASE consolidation	CyberFOX → Timus Networks; Zscaler expands; Cradlepoint → Ericom	Undisclosed/organic	Zero trust for AI multi-tenant factories, east-west OT, 5G/cellular edge

Pattern:

Financing innovation: Warrant-based deals (Qualcomm-Amazon $4B, Marvell-Google $12.2B) tie equity upside to product purchases — intertwined financing in AI boom.
Supply chain as moat: Hyperscalers co-investing in Corning manufacturing (Meta, Amazon, Nvidia, Lumen 10% of global capacity) — fiber becomes strategic inventory, not a commodity.
Three AI networking stacks clarify: Broadcom merchant silicon (Tomahawk 6) de facto for scale-out, Nvidia full-stack (Spectrum-X/XGS + NVLink) for proprietary optimization, Qualcomm emerging as custom inference + optical challenger leveraging AlphaWave.

11. The Networking Shift

The structural shift: data movement now dictates infrastructure economics. For two decades, networking was best-effort connectivity supporting compute. Now, with 100K+ accelerator clusters planned on Tomahawk 6, giga-scale super-factories linking data centers across cities/nations/continents, and inference requiring tokens/sec at fixed SLOs, the network determines how efficiently intelligence can move, synchronize, and scale.

Bandwidth: 1.6T per port (800G→1.6T in 24 months) with 102.4T single-chip radix — but electrical reaches the power wall. CPO + silicon photonics mandatory, not optional.
Latency: Inference tail latency + collective communication sensitivity → lossless Ethernet + job-aware routing + hollow-core fiber for inter-DC (lower group index). Distance-adaptive algorithms in Spectrum-XGS adjust for inter-DC distance.
Topology: From leaf-spine inside a DC to scale across DCs. AI requires two networks: inside the brain (scale-up/out) and between brains + the outside world (dark fiber, high-capacity backbone). Google's >$1B dark fiber deal shows even the most advanced private network (2M+ miles of lit fiber) needs telco for additional routes.
Power: Optics now 20-30% of data center power. Co-packaged optics slashes energy per bit, enabling hundreds of chips to function as a single virtual processor. Power, not silicon, is the binding constraint.
Interoperability: UEC 1.0 gives an open alternative to proprietary, but production still uses RoCEv2. Nvidia hedges: member of UEC yet ships proprietary-enhanced Spectrum-X. Open standard ≠ vendor-neutral outcome.

12. The CODEW Analysis

Follow the Movement of Data

If we follow the movement of data this week, it flows like this: datasets moved from storage to compute → models copied between regions → training synchronized across 100K+ accelerators via 1.6T lossless fabrics → inference results retrieved via RAG and fanned out to factories/vehicles/hospitals → all requiring dark fiber backbones between distributed super-factories.

Qualcomm x Amazon is not a chip deal — it's a data-movement deal. By pairing custom inference silicon with 1.6T optical co-development, they acknowledge that inference economics = compute + interconnect. The warrant structure aligns incentives: Amazon gets equity upside as it buys more chips, Qualcomm gets an anchor customer for its $15B data-center ambition. Similar to the Marvell-Google $12.2B pattern.
Fiber is the new wafer: Just as TSMC capacity constrained AI chips in 2024-25, fiber supply now constrains AI data centers in 2026. Corning's deals (Amazon multibillion, Zayo through 2030, Meta $6B, Nvidia $500M, Lumen 10% global capacity) show hyperscalers treating fiber manufacturing as strategic inventory. Zayo's 15,000 route miles, including 8,000 new long-haul with Nvidia collaboration, indicates AI corridors deliberately built where power exists, not just where demand was.
Three pillars of AI computing are now defined: Scale-up (inside server/rack, NVLink/UALink), scale-out (across thousands of servers, Ethernet/InfiniBand), scale-across (across distributed DCs, Spectrum-XGS). Each has different latency/bandwidth/power trade-offs. Enterprises must design for all three.
Open vs closed tension intensifies: UEC 1.0 (560 pages) provides an open Ethernet-based stack, but Nvidia's Spectrum-XGS provides an immediate product with 1.6x bandwidth density and distance-adaptive algorithms. As one analysis warned: orgs delaying fabric decisions until UEC matures may be 12-18 months behind competitors who adopt accelerated Ethernet now.

The AI infrastructure race is no longer just about who has most GPUs. It's about who can move data between them most efficiently, at lowest power, with lowest tail latency, across an increasingly distributed footprint. Compute determines how much intelligence can be generated; networking determines how efficiently that intelligence can move, synchronize, and scale.

13. Strategic Takeaway

Treat networking as strategic infrastructure, not plumbing — with direct impact on AI economics.

For enterprises & cloud architects: 1) Audit fiber route diversity and dark fiber options now — hyperscalers are locking up Corning capacity through 2030. 2) Design fabrics for 1.6T + CPO readiness: liquid cooling, high-fiber-count trunk cables, modular chassis supporting 800G→1.6T upgrade. 3) Plan for scale-across: if inference must reach factories/vehicles/hospitals, you need both intra-DC lossless fabric and inter-DC high-capacity backbone. 4) Evaluate zero trust SASE for east-west AI traffic — multi-tenant AI factories need performance isolation + security isolation.
For investors & vendors: Watch warrant-based financing as a leading indicator of custom silicon traction (Qualcomm $4B, Marvell $12.2B). Corning is integral to AI grid — fiber supply agreements worth billions becoming a recurring theme. Broadcom Tomahawk 6 retains 60-70% switch silicon share, but Qualcomm entry via AlphaWave + optical changes competitive set. Verizon-Google $1B+ dark fiber signals telco monetization path beyond best-effort connectivity.
Positioning: Networking Watch asks "How is data moving?" — this week's answer: data is moving from centralized training to distributed inference, from electrical to optical, from campus to giga-scale super-factories, and from shared transport to dark fiber with explicit performance objectives. Cloud Watch tracks how cloud platforms evolve to support that; Edge Watch tracks what happens when compute moves closer to generation; Hardware Watch tracks physical systems enabling it; Infrastructure Software Watch tracks software operating it.

14. What to Watch Next Week

Qualcomm-Amazon chip specifics: Inference chip architecture details (AlphaWave integration, power per token), optical connectivity timeline to 1.6T and beyond, AWS design workload expansion impact on development cycles.
More dark fiber announcements: Verizon CEO Dan Schulman said more multi-billion deals expected by year-end — watch AT&T, Lumen, Zayo for similar hyperscaler announcements. Indicator of fiber scarcity pricing.
OCP Global Summit & UEC interop: Progress on UEC 1.0 adoption vs Spectrum-X/MRC OCP specifications — will Broadcom, AMD, Intel, Arista adopt, adapt, or compete? Fragmented response benefits Nvidia.
1.6T optics availability: Broadcom Taurus, Sian3 sampling to early customers — who reaches volume production first? Marvell, Cisco, Ciena roadmaps. CPO vs pluggable TCO analysis for 100K+ accelerator clusters.
Enterprise AIOps: TPG Telecom March 2026 transition to AIOps-driven Service Operations Centre using Cisco + Splunk — early production results, MTTR improvements for AI fabrics.
Fiber Connect / optical manufacturing: Corning's $10B inside-the-box opportunity execution, workforce training programs (Catawba Valley), and ability to meet 15,000 route-mile build-outs without becoming a barrier to execution.

The CODEW Stat

Verizon signed a>$1B dark fiber deal with Google (July 2026) and expects multiple billions more by year-end — while Amazon, Meta, Nvidia, Zayo, and Lumen have collectively committed >$13B to Corning fiber and optical manufacturing (Meta $6B, Amazon multibillion, Nvidia $500M, plus Zayo 15,000 route miles and Lumen 10% of global capacity). Hyperscalers are now underwriting fiber factories: Corning's US optical connectivity manufacturing capacity to increase 10x, fiber production 50%+, with 1,000 new NC jobs — proving fiber is no longer a passive enabler but a competitive differentiator in the data-center supply chain.

The CODEW