Beyond the Hyperscaler: How Cisco, HPE, and Arista Are Building the Enterprise AI Network

The CODEW Special Report  | August 5, 2026

Enterprise AI Ethernet Fabrics 2026: Cisco vs. HPE vs. Arista

The Enterprise AI On-Ramp: Ethernet Fabrics Beyond the Hyperscaler 

Executive Summary

For most of the last decade, enterprise networking was a solved problem — a slow-moving budget line refreshed every five to seven years. Private AI infrastructure has broken that cycle. As organizations move generative and agentic AI workloads out of pilot mode and into production, the network connecting GPUs, storage, and inference endpoints has become one of the most consequential — and least understood — line items in the entire AI stack.

This report examines how the enterprise networking incumbents are responding. Cisco is rebuilding its data center portfolio around Silicon One custom silicon and a unified Nexus One operating model. HPE has spent the past year digesting its $14 billion acquisition of Juniper Networks, merging Juniper's AI-native Mist platform with Aruba's campus franchise into a single self-driving network strategy. Arista Networks has extended its Etherlink portfolio to 1.6 terabit speeds and, by its own account, now counts over 100 customers running AI fabrics — a business that barely existed three years ago.

Underpinning all three is a bigger structural story: the Ultra Ethernet Consortium shipped its 1.0 specification in June 2025, giving the industry a credible open alternative to proprietary interconnects for the first time. And yet the vendor best positioned to benefit from Ethernet's rise in 2026 has been Nvidia itself, whose Spectrum-X switching line pushed its share of the data center Ethernet switch market from under 4% to roughly 21.5% in about a year, per IDC data — a reminder that "open standard" and "vendor-neutral outcome" are not the same thing.

For enterprise IT leaders, the questions this report addresses are practical: why does AI traffic break assumptions baked into twenty years of Ethernet design; what does each of the major platform vendors actually offer once the marketing language is stripped away; and how should a buyer structure a decision that will shape their infrastructure roadmap for the rest of the decade.

Key Findings

• 🚩The data center Ethernet switch market hit $15.4 billion in Q1 2026 alone, up 39.8% year-over-year — a growth rate almost entirely attributable to AI cluster build-outs rather than traditional enterprise refresh cycles.
• 🚩HPE's completed acquisition of Juniper Networks (closed July 2025) has collapsed what the assignment brief treated as two separate vendors into one; Juniper's Mist AI-ops platform and Aruba's Central platform are now being merged under a single "self-driving network" strategy.
• 🚩Nvidia's entry into Ethernet switching via Spectrum-X is the single biggest disruption to the competitive landscape in 2026, taking an estimated 21.5% share of data center Ethernet switch revenue from a near-standing start.
• 🚩The Ultra Ethernet Consortium's 1.0 specification (June 2025) gives Ethernet a standards-based answer to InfiniBand's latency advantage for the first time, but adoption is uneven — most current deployments still run vendor-specific RoCEv2 implementations rather than native UEC transport.
• 🚩Enterprise motivations for private AI infrastructure increasingly center on data sovereignty and cost predictability, not just performance — multiple 2026 surveys show enterprises pulling AI inference workloads back from public cloud, with private cloud now hosting production inference at roughly half of surveyed organizations.
• 🚩Power and cooling, not switch silicon, are emerging as the binding constraint on enterprise AI network design — liquid cooling has moved from a hyperscaler-only concern to a mainstream requirement in vendor roadmaps across Cisco, Arista, and HPE.

Why Enterprise AI Changes Networking

The case for private AI infrastructure

Enterprises are not building private AI infrastructure primarily because it is cheaper than the cloud, though cost predictability is a factor. Recent survey data points to three converging pressures. First, data sovereignty and data residency requirements have become boardroom-level concerns rather than compliance footnotes — one 2026 industry survey found roughly three-quarters of respondents reporting that shadow-AI incidents or residency requirements had directly affected their AI initiatives. Second, cloud cost overruns are common: a meaningful share of organizations report that actual cloud AI spending has exceeded initial projections, making the fixed-cost profile of owned infrastructure more attractive on a multi-year view. Third, latency-sensitive use cases — real-time inference embedded in operational workflows — are pushing compute physically closer to the data and the user.

The result is not a wholesale return to on-premises computing but a layered hybrid model: public cloud for elastic, non-sensitive workloads; private data centers or colocation for fine-tuning and sensitive inference; and edge deployments for latency-critical processing. Separate 2026 research from Broadcom's Private Cloud Outlook found that the share of enterprises naming public cloud as their primary environment for production AI inference fell year-over-year, while more than half of enterprises now run or plan to run production inference on private cloud.

How AI workloads break traditional network assumptions

Enterprise data center networks were designed around north-south traffic patterns — client requests flowing in, responses flowing out — with switches optimized for average-case throughput and best-effort delivery. AI training and large-scale inference invert that model. GPU clusters generate overwhelmingly east-west traffic, synchronizing gradients and parameters across thousands of accelerators in tight, bursty patterns where a single dropped or delayed packet can stall an entire training job, not just one flow.

That changes the engineering priorities in three ways. Networks need to be effectively lossless, since AI collective-communication libraries are far more sensitive to packet loss than typical enterprise applications. They need much higher radix and bandwidth per port — the industry has moved through 400G to 800G and is now shipping 1.6T platforms in barely two years, a pace of speed transition without recent precedent in enterprise networking. And they need topologies (multi-planar leaf-spine, single-tier and two-tier fat-tree designs) that minimize the number of hops between accelerators, because every additional tier adds latency that idles expensive GPU time.

The Rise of AI Ethernet Fabrics

The speed curve: 400G to 800G and 1.6T

800G ports already account for a large and rapidly growing share of data center Ethernet switch revenue — over a third of segment revenue in early 2026, by IDC's count — while vendors including Arista and Cisco have begun shipping 1.6T-capable platforms built around next-generation merchant and custom silicon (Broadcom's Tomahawk 6 and Cisco's Silicon One G300 both clear the 100 Tbps single-chip threshold). The practical effect for enterprise buyers is that hardware refresh cycles for AI-adjacent infrastructure are compressing from five-plus years to roughly two, forcing procurement teams to think in terms of modular, upgradeable fabrics rather than fixed-configuration purchases.

Ultra Ethernet and the open-standards push

The Ultra Ethernet Consortium — founded in 2023 by AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, and Microsoft, and hosted under the Linux Foundation — released its 1.0 specification on June 11, 2025. The spec, more than 560 pages, defines a full Ethernet-based communication stack purpose-built for AI and HPC workloads: a modern RDMA approach, a new Ultra Ethernet Transport protocol for congestion control, and open interfaces spanning NICs, switches, optics, and cables. The explicit goal is interoperability across vendors, in contrast to proprietary scale-up fabrics such as Nvidia's NVLink.

UEC-compliant hardware began reaching the market in the second half of 2025, and by mid-2026 most of the major switch vendors — Cisco, Arista, and HPE among them — describe their platforms as UEC-ready or forward-compatible, even where production deployments still lean on established RoCEv2-based lossless Ethernet rather than the full UEC transport stack. Enterprises evaluating vendors should treat "UEC support" claims carefully in 2026: consortium membership and roadmap compatibility are not the same as a shipping, certified UEC 1.0 implementation.

Ethernet vs. proprietary approaches: where InfiniBand still wins

The Ethernet-versus-InfiniBand debate is often framed as binary, but the more accurate 2026 picture splits by network function. For scale-up networking — the ultra-low-latency fabric connecting accelerators within a single server or rack — InfiniBand and Nvidia's proprietary NVLink retain a real technical edge, and Nvidia's own scale-up numbers still run alongside a fast-growing Ethernet-based Spectrum-X line rather than being displaced by it. For scale-out networking — connecting thousands of servers across a training cluster — Ethernet's cost advantage (switch ports typically running 20-40% cheaper than equivalent InfiniBand hardware) and the depth of the existing Ethernet operations talent pool are winning most new enterprise deployments. That segmentation, more than a clean Ethernet "win," is the operative reality enterprise buyers are navigating.

Vendor Landscape

Editorial note: This assignment's original brief treated Cisco, HPE, Juniper Networks, and Arista as four independent vendors. That is no longer accurate. HPE completed its $14 billion acquisition of Juniper Networks on July 2, 2025, and by mid-2026 the two companies' AI-ops platforms — Juniper Mist and Aruba Central — were being actively merged under a unified HPE Networking strategy. We have restructured this section accordingly: HPE, Aruba, and Juniper appear as a single combined vendor profile, reflecting the market as it actually stands today.

Cisco

AI-ready portfolio. Cisco's data center strategy now centers on Nexus One, a unified operating model announced through 2026 that spans silicon, systems, optics, and software across a single management plane (Nexus Dashboard). Rather than locking customers into one chip architecture, Nexus One supports Cisco's own Silicon One and Cloud Scale ASICs alongside Nvidia's Spectrum-X silicon inside the same N9000 switch family — a notable concession that reflects how central Nvidia has become even to competitors' hardware roadmaps.

Silicon One strategy. At Cisco Live EMEA in February 2026, Cisco introduced Silicon One G300, a 102.4 Tbps switching chip aimed at massive AI cluster build-outs, paired with new liquid-cooled N9000 and 8000-series systems and linear pluggable optics designed to cut per-module power draw by roughly half. Cisco frames this as a full-stack play — silicon, systems, and optics engineered together — intended to compete directly with Broadcom-based merchant-silicon platforms on power efficiency per port.

Nexus platform evolution. Nexus Hyperfabric, Cisco's cloud-managed fabric offering, has been extended with "job-aware" telemetry that correlates network performance directly with GPU workload behavior — a direct response to the operational reality that AI infrastructure teams need to distinguish network-caused slowdowns from compute-caused ones.

AI operations and automation. Cisco Cloud Control, introduced in 2026, is positioned as a single AgenticOps platform across Cisco's networking portfolio, with an "AI Canvas" feature that lets both engineers and AI agents interact with live telemetry through a conversational interface to speed up troubleshooting.

Market position. Cisco remains the incumbent with the largest installed base of any enterprise networking vendor, but independent market-share tracking puts it behind both Arista and Nvidia in high-speed data center Ethernet switching as of early 2026 — a meaningful reversal for a company that dominated this category for two decades.

Hewlett Packard Enterprise (HPE) — with Aruba and Juniper

The combined entity. HPE closed its acquisition of Juniper Networks in July 2025, doubling the size of its networking business and combining HPE Aruba Networking's campus and branch strength with Juniper's data center switching and AI-native Mist operations platform. As a condition of DOJ approval, HPE agreed to license Juniper's Mist AI-Ops source code to a third party through an auction process — a structural check on HPE's combined market power in enterprise wireless networking worth watching through 2026 and beyond.

Integration progress. At HPE Discover events in December 2025 and June 2026, HPE detailed a "build once, deploy twice" integration approach: shared AI-ops capabilities rolling out across both Aruba Central and Juniper Mist rather than forcing an immediate platform migration. Juniper's Mist Large Experience Model — which analyzes application performance data from tools like Zoom and Teams — is being extended into Aruba Central, while Aruba's Agentic Mesh anomaly-detection technology is being brought into Mist. New Wi-Fi 7 access points designed to run on either platform, and Juniper QFX data center switches supporting the open UALink scale-up interconnect standard (a direct alternative to Nvidia's NVLink), shipped as part of the June 2026 announcements.

Aruba networking and GreenLake. Aruba remains HPE's campus and branch franchise, now being extended into data center operations through integration with the Mist platform. GreenLake, HPE's as-a-service consumption model, has been extended with an "Intelligence Mesh" for governing AI agents and a marketplace for direct ISV transactions — positioning GreenLake less as a financing mechanism and more as the operational control plane for HPE's combined AI infrastructure stack.

Enterprise private AI strategy. HPE's pitch to enterprise buyers is full-stack integration: compute, storage, and now a unified Aruba-plus-Juniper networking layer, sold and financed through GreenLake, with shared AI-driven insights across domains. The strategic logic mirrors Cisco's — bundling networking with broader infrastructure — but HPE is doing it from a networking base that was, until mid-2025, two separate and overlapping vendors now being actively merged under a single partner program (effective November 1, 2026).

Arista Networks

EOS architecture. Arista's differentiation has long rested on a single-OS architecture — Extensible Operating System (EOS) — running across its entire hardware line from campus edge to AI spine, in contrast to vendors that operate multiple, less consistent operating system variants across product families. For AI workloads specifically, EOS implements RoCEv2, DCQCN congestion control, and priority flow control alongside Arista's own load-balancing technology.

CloudVision. CloudVision is Arista's network-wide management and telemetry layer, feeding into a broader Network Data Lake (NetDL) that ingests telemetry from Arista hardware, third-party systems, server NICs, and AI job schedulers — explicitly built to correlate network behavior with AI training-job performance, the same operational problem Cisco is addressing through Nexus One's job-aware telemetry.

AI Ethernet fabrics. Arista's Etherlink portfolio spans 400G, 800G, and — as of a June 2026 launch — 1.6T switching, built on Broadcom's Tomahawk 6 silicon and delivering up to 100 Tbps of system bandwidth with roughly 60% lower interconnect power consumption via linear pluggable optics. Arista also ships the 7700R4 Distributed Etherlink Switch, a single-hop architecture designed to scale to more than 30,000 connected accelerators while preserving lossless, deterministic forwarding — an alternative to conventional multi-tier spine-leaf designs for very large clusters. By Arista's own Q2 2026 disclosure, its Etherlink AI fabric portfolio has been adopted by more than 100 cumulative customers, up from roughly four or five discussed publicly in 2024.

Open networking ecosystem. As a founding member of the Ultra Ethernet Consortium, Arista has positioned itself as the leading advocate for open, standards-based Ethernet over proprietary scale-up fabrics — a strategic bet that the company's leadership has described publicly as backing Ethernet to become the eventual default for AI networking broadly, not just scale-out traffic.

Financial trajectory. Arista reported Q2 2026 revenue of $3.036 billion, up 37.7% year-over-year, and was named a Leader in Gartner's 2026 Magic Quadrant for Enterprise Wired and Wireless LAN — evidence that its single-EOS architecture is gaining traction in enterprise campus deployments, not just hyperscale and AI-titan data centers.

Competitive Comparison

Two figures should anchor how enterprise buyers read the competitive landscape in 2026. First, per IDC's Q1 2026 tracking, the overall data center Ethernet switch market reached $15.4 billion in a single quarter, up 39.8% year-over-year. Second, and more consequential for incumbent vendors: Nvidia's own data center Ethernet switch share rose from under 4% roughly two years earlier to approximately 21.5% in Q1 2026 — making Nvidia, by revenue, the largest single vendor in a category it barely competed in until recently, ahead of Arista's roughly 19% share and well ahead of Cisco.

Dimension	Cisco	HPE (Aruba + Juniper)	Arista Networks
Core AI silicon	Silicon One G300 (102.4 Tbps), plus support for Nvidia Spectrum-X silicon in the same chassis	Broadcom Tomahawk 6 (first OEM deal, per HPE's Dec. 2025 announcement); Juniper QFX supports open UALink	Broadcom Tomahawk 6-based Etherlink, up to 1.6T (7060XE7 Series)
Unified fabric platform	Nexus One (single management plane across ACI, Hyperfabric, SONiC)	Merging Aruba Central + Juniper Mist under a unified AIOps layer ("build once, deploy twice")	Single EOS operating system across campus-to-AI-spine; CloudVision management
AI-native ops / automation	Cisco Cloud Control with AI Canvas (agentic, human-in-the-loop troubleshooting)	Marvis self-driving network (Mist), extended into Aruba Central; GreenLake Intelligence Mesh for AI agent governance	CloudVision + Network Data Lake (NetDL) correlating telemetry with AI job performance
Ultra Ethernet Consortium role	Member	HPE is a founding member; Juniper (pre-acquisition) was not	Founding member; publicly the most vocal Ethernet-over-InfiniBand advocate
Est. data center Ethernet switch share (Q1 2026)	~14%	~7% (Juniper legacy figure; combined HPE share not yet separately disclosed)	~19%
Recent growth signal	Low-single-digit switching growth; broader company revenue growth driven by security and software	Networking business roughly doubled in scale post-acquisition; integration still in progress	Q2 2026 revenue of $3.036B, up 37.7% YoY; full-year 2026 revenue guided above $10B
Enterprise strength	Largest installed base; deep integration with security and collaboration portfolio	Strongest combined campus + data center + WAN breadth of any single vendor; GreenLake consumption model	Gartner Magic Quadrant Leader (2026) for enterprise wired/wireless LAN; single-OS operational simplicity
Key open question for buyers	Can Silicon One and Nexus One regain share lost to Arista and Nvidia in high-speed switching?	Will Aruba/Juniper integration disrupt existing deployments, and what happens once Mist source code is licensed to a competitor?	Can Arista's pure-play focus keep pace once Cisco and HPE bundle networking into broader compute/storage deals?

Market-share figures are blended estimates drawn from IDC's Worldwide Quarterly Ethernet Switch Tracker and vendor earnings disclosures as reported by third-party analysis in mid-2026; individual vendors do not uniformly break out AI-specific networking revenue, so figures should be read as directional rather than precise.

Enterprise Buying Considerations

Enterprise IT leaders evaluating AI networking platforms in 2026 are effectively making a decade-long infrastructure bet under conditions where the underlying technology is still moving quickly. A few practical framing points:

• ✒Match architecture to workload, not vendor narrative. Training clusters, inference serving, and storage networking have distinct traffic patterns and latency tolerances. A single-vendor "one fabric for everything" pitch should be tested against the specific east-west traffic profile of the workloads actually planned, not the vendor's largest hyperscale reference deployment.
• ✒Separate genuine open-standards support from roadmap marketing. UEC Consortium membership is now close to universal among major vendors; a shipping, certified UEC 1.0 implementation is not. Buyers should ask vendors directly which parts of the UEC stack are in production versus roadmap.
• ✒Price the operational model, not just the hardware. GreenLake-style consumption pricing, CloudVision/Nexus Dashboard/Mist software licensing, and support contracts materially change total cost of ownership relative to sticker price on switches — especially as AI-specific management features increasingly sit behind separate subscriptions.
• ✒Factor in integration risk explicitly for HPE. The Aruba-Juniper merger is real and material progress has shipped, but enterprises standardizing on HPE Networking in 2026 are buying into an integration still in motion, with a DOJ-mandated Mist source-code license to a future competitor as an added variable.
• ✒Treat power and cooling as a networking decision, not just a facilities one. Liquid-cooled switch options from Cisco, Arista, and HPE are no longer hyperscaler-exclusive; enterprises planning AI clusters above roughly a few hundred accelerators should model facility power and cooling requirements alongside the network bill of materials, not after it.

Future Outlook

Dell'Oro Group projects that cumulative spending on data center switches deployed in AI back-end networks will approach $1 trillion over the 2026-2030 period, with scale-up networking — the tightly coupled, ultra-low-latency fabric inside a rack — projected to account for more than half of that spending by 2030. That shift matters for enterprise buyers because scale-up is precisely the segment where proprietary interconnects (Nvidia's NVLink and increasingly open alternatives like UALink) compete most directly with Ethernet, meaning the vendor architecture decisions enterprises make in the next 12-18 months will shape which ecosystem they are locked into for scale-up expansion later in the decade.

Three trends look durable heading into 2027: continued compression of the switch refresh cycle as port speeds climb past 1.6T toward 3.2T; broader adoption of co-packaged and linear pluggable optics as power efficiency becomes a hard constraint rather than a preference; and a steady shift of AI-ops software from optional add-on to the primary battleground, as Cisco's AI Canvas, HPE's Marvis, and Arista's NetDL each compete to become the control plane enterprises actually manage their AI infrastructure through.

The CODEW Analysis

The following section reflects CODEW's editorial assessment, not vendor-supplied claims.

The most underreported story in this category is not Ethernet's contest with InfiniBand — it is that the vendor best positioned to win enterprise AI networking dollars in 2026 is a compute company. Nvidia's rise to roughly a fifth of data center Ethernet switch revenue in about two years is a scale of share shift this market has not seen from any single vendor in the past two decades, and it happened by bundling switching with the GPUs enterprises already had to buy. That should reframe how IT leaders read "open, vendor-neutral Ethernet" pitches from Cisco, HPE, and Arista: the standards fight is real and worth backing, but it is currently a check on Nvidia's dominance, not evidence that the check is working yet.

We also think enterprises are underweighting integration risk in the HPE-Juniper combination relative to how confidently it is being marketed. "Build once, deploy twice" is a sensible near-term integration strategy, but it also means enterprises buying HPE Networking today are functionally choosing a vendor mid-merger, with a competitor eventually gaining licensed access to Mist's source code as a condition of regulatory approval. That is not a reason to avoid HPE — Aruba and Juniper are each strong on their own technical merits — but it is a reason to ask harder questions about platform roadmap timing than a stable, single-company vendor would require.

Finally, we would push back gently on the framing, common in vendor materials, that AI networking is primarily a silicon and speed story. Every technology executive we track in this space is converging on the same operational bottleneck: most enterprises do not have network engineers who understand both classic Ethernet operations and AI cluster scheduling well enough to run these fabrics without vendor hand-holding. The AIOps layer — Cisco's AI Canvas, HPE's Marvis, Arista's NetDL — is becoming the actual product being sold, with the switch hardware underneath it increasingly commoditized. Buyers evaluating "AI-ready" networking should weigh the operations software and the skills required to run it at least as heavily as the switch specs.

Risks worth tracking

• ✒Deployment complexity: Multi-tier AI fabrics with lossless-Ethernet tuning, RDMA configuration, and job-aware telemetry require specialized expertise most enterprise network teams do not yet have in-house.
• ✒Skills shortages: Demand for engineers fluent in both classic Ethernet operations and AI cluster networking is outpacing supply, pushing enterprises toward vendor-managed or consumption-based models (GreenLake, Cisco-managed Hyperfabric) partly to close the skills gap rather than purely for cost reasons.
• ✒Power and cooling constraints: Liquid cooling is now a baseline requirement for high-density AI switch deployments, and facility power availability is increasingly the actual limiting factor on how quickly enterprises can scale private AI clusters, not switch procurement lead times.
• ✒Vendor lock-in: Despite the open-standards push, most current lossless-Ethernet AI deployments still rely on vendor-specific tuning and management software, meaning switching costs remain real even on "open" Ethernet fabrics.
• ✒ROI uncertainty: With AI back-end network spending projected to approach $1 trillion industry-wide over five years, enterprises building private AI infrastructure on multi-year vendor commitments are making a bet on AI workload growth that, for many organizations outside the hyperscalers, remains genuinely unproven at the budgeted scale.

Conclusion

Enterprise AI networking in 2026 is not a settled market — it is a fast-moving contest between three re-architected incumbents, a newly aggressive compute vendor turned networking company, and an open-standards effort still proving itself in production. Cisco is betting on full-stack silicon-to-software integration through Nexus One. HPE is betting that a combined Aruba-Juniper portfolio, sold through GreenLake, can out-integrate two separate best-of-breed vendors. Arista is betting that operational simplicity and open Ethernet standards win out over proprietary bundling, even as Nvidia's own Ethernet push tests that thesis directly.

For enterprise buyers, the decision that matters most in the next 12 months is not which vendor has the fastest switch — port speeds will keep climbing regardless of which logo is on the chassis. It is which vendor's AI-ops software and operational model the organization's own team can actually run at scale, and how much lock-in the organization is willing to accept in exchange for that operational simplicity. That trade-off, more than any spec sheet, is what will determine which of these platforms enterprises are still running five years from now.

Source Attribution

1. Cisco — Networking for the Agentic Era: Cisco Unveils New Innovations in Scale and Simplicity
2. Cisco Newsroom — Cisco Announces New Silicon One G300
3. HPE Newsroom — HPE Closes Acquisition of Juniper Networks
4. HPE Newsroom — HPE Expands Self-Driving Networks Across Edge, Campus, Data Center, and AI Factories
5. Arista Networks — Arista Introduces Next-Generation 1.6 Terabit Portfolio for AI Fabrics
6. BusinessWire — Arista Networks Reports Second Quarter 2026 Financial Results
7. Ultra Ethernet Consortium — UEC Launches Specification 1.0
8. IDC — Nvidia Becomes #1 in Datacenter Ethernet Switching as 1Q26 Market Surges to $15.4 Billion
9. Dell'Oro Group — AI Back-end Switch Sales to Approach $1 Trillion Over the Next Five Years
10. NTT DATA — 2026 Global AI Report: A Playbook for Private and Sovereign AI

Reporting for this Special Report combines public company disclosures, vendor technical documentation, and third-party industry research current as of August 5, 2026. Market-share and revenue figures are analyst estimates unless otherwise attributed to a company's own earnings disclosures, and should be read as directional. This report reflects available public information and CODEW's editorial analysis; forward-looking statements are clearly marked as such and are not guarantees of future vendor performance.

[full_width]