🚀 1. Why AI Data Center Upgrades in 2025 Are All About Optical Speed

The explosion in AI and machine learning model sizes, the proliferation of "super pod" GPU racks, and the relentless push for lower total cost of ownership are making 400G and 800G optics the new backbone of next-generation AI infrastructure. If your data center is still operating on legacy 100G or 200G links, you are not just falling behind on performance — you are creating structural bottlenecks that compound with every additional GPU, every additional training run, and every incremental scale-out event. The physical layer is no longer an afterthought in AI infrastructure planning; it is the rate-limiting constraint that determines whether your clusters can train at full capacity or spend cycles waiting on network congestion.

Upgrading to 400G and 800G optics is no longer just futureproofing — it is a present-day operational requirement for performance, manageability, and scale at the GPU cluster densities that modern AI workloads demand. This guide covers why the upgrade is happening now, how to select the right transceivers and cabling, what immersion cooling and open networking mean for your infrastructure decisions, and which Vitex products are the right fit for each deployment scenario.

The Three Drivers Making 400G/800G a 2025 Requirement

Bandwidth demand at the rack level is the primary driver. A single AI rack populated with 16 GPUs operating at full training throughput generates 400Gbps or more of east-west traffic between GPUs performing AllReduce gradient synchronization. That traffic volume exceeds the capacity of 100G network links within a single rack, let alone across an entire cluster fabric. Even 200G configurations that served adequately for smaller models are now bottlenecking larger training runs where every millisecond of network latency adds to per-epoch training time at a scale that compounds over weeks of continuous operation.

Physical density constraints are the second driver. Data center space and power are scarce and expensive. Higher-speed optics allow engineers to serve the same number of GPU endpoints with fewer switch ports, fewer cable runs, and less physical infrastructure. An 800G uplink carries eight times the bandwidth of a 100G uplink in the same port count and roughly the same physical cable footprint. That reduction in cable volume has direct implications for airflow, cooling efficiency, and the ability to add additional GPU racks without triggering a full infrastructure redesign.

Multi-vendor flexibility requirements are the third driver. As AI infrastructure has matured, organizations have moved away from single-vendor network fabrics toward open, MSA-compliant architectures where GPU servers, top-of-rack switches, and spine switches can be sourced independently and upgraded on different refresh cycles. That flexibility requires transceivers and cabling that operate correctly across different vendor platforms without proprietary lock-in — a requirement that modern 400G and 800G MSA-compliant optics meet and that legacy proprietary 100G solutions often do not.

📊 2. How to Select 800G OSFP & QSFP-DD Transceivers for High-Density AI Clusters

800G transceivers are the backbone of today's highest-density AI clusters, but selecting the right module requires understanding both the form factor and the reach variant — two independent decisions that together determine whether a transceiver is compatible with your switch hardware, your fiber plant, and your distance requirements. Getting either decision wrong results in transceivers that do not fit, do not link up, or do not reach — all of which are expensive field failures that professional specification work prevents.

OSFP vs QSFP-DD: Choosing the Right Form Factor

OSFP (Octal Small Form-factor Pluggable) modules offer outstanding thermal performance and are optimized for dense, high-speed 800G environments. The OSFP form factor is physically larger than QSFP-DD, which allows for better heat dissipation — a critical advantage in AI racks where GPU servers, top-of-rack switches, and optical transceivers are all operating at high power densities simultaneously. OSFP is the recommended form factor for new 800G AI deployments where the switch hardware supports it, because its thermal design directly addresses the cooling challenges that are otherwise managed with additional airflow infrastructure.

QSFP-DD (Quad Small Form-factor Pluggable Double Density) modules are widely adopted for their backward compatibility. QSFP-DD ports support both 400G and 800G modules, allowing organizations to deploy 400G modules in their existing QSFP-DD infrastructure today and upgrade to 800G as cluster requirements grow — without replacing switch hardware. This makes QSFP-DD the right choice for brownfield environments where existing switch infrastructure uses QSFP-DD cages and where a phased migration from 400G to 800G is the operational plan.

Selecting the Right 800G Reach Variant

Once you have selected the correct form factor, the reach variant determines which fiber type and distance the transceiver supports. SR8 and DR8 modules are designed for short-reach connections — up to 100 meters over OM4 or OM5 multimode fiber — making them the standard choice for intra-rack, within-row, and adjacent rack GPU-to-switch connections where the cable distance is measured in meters rather than hundreds of meters. SR8 uses MPO-16 connectors with parallel fiber lanes, requiring MTP/MPO-16 trunk cables and cassettes throughout the cabling infrastructure.

2FR4 and 2LR4 modules extend the reach to 2 kilometers and beyond over singlemode fiber, enabling connections between clusters housed in different halls, different buildings, or different pods within a large campus AI facility. These modules use duplex LC connectors and operate over OS2 singlemode fiber, integrating naturally with existing singlemode patch panel infrastructure that many organizations already have in place for 100G and 400G singlemode links. For any AI cluster connectivity requirement beyond 150 meters — the ceiling of OM5 multimode — 2FR4 and 2LR4 on singlemode are the correct specification regardless of the form factor choice.

🔵 3. 400G QSFP-DD & OSFP Transceivers — The Workhorse of AI Data Center Networks

400G optics remain the performance sweet spot for the majority of AI data centers transitioning from 100G or 200G infrastructure — particularly for leaf-spine backbone links, aggregation connections, and inter-pod connectivity where 800G is not yet required but 100G is clearly insufficient. The 400G transceiver ecosystem is mature, widely available across multiple vendors, and cost-effective in ways that 800G modules are still approaching. For organizations building or upgrading AI clusters in 2025 where 800G is a 2026 or 2027 consideration, 400G is the right investment today.

Key 400G Module Types and Their Applications

400G QSFP-DD DR4 is the standard parallel-fiber module for short-to-medium reach connections up to 500 meters over OS2 singlemode fiber. DR4 uses MPO-12 connectors with four parallel lanes at 100G each, making it the natural choice for direct in-row connections, top-of-rack to spine connections, and aggregation switch uplinks in leaf-spine topologies where cable runs typically range from tens to a few hundred meters. The DR4 variant's 500-meter reach covers the vast majority of intra-campus AI cluster connectivity scenarios without requiring the longer-reach and higher-cost LR4 variant.

400G QSFP-DD FR4 and LR4 are duplex fiber variants that extend reach to 2 kilometers and 10 kilometers respectively over OS2 singlemode fiber with LC duplex connectors. FR4 is the right choice for connecting different parts of a large AI cluster across a building or campus where singlemode infrastructure already exists. LR4 serves data center interconnect scenarios where AI clusters in separate buildings or facilities need high-bandwidth connectivity without deploying coherent optics. Both FR4 and LR4 use standard WDM technology — four wavelengths per fiber pair — making them compatible with existing singlemode patch panel infrastructure.

400G OSFP DR4 and FR4 provide the same reach capabilities as their QSFP-DD counterparts in the OSFP form factor, adding the thermal management advantages of OSFP for deployments on OSFP-equipped switch platforms. Organizations running OSFP-based spine switches who need 400G for specific links while running 800G OSFP on others will find the 400G OSFP variants essential for maintaining a consistent form factor across their switch infrastructure without mixed cage types.

TAA/NDAA Compliance for Public Sector and Regulated AI Deployments

Many AI data centers — particularly those serving government agencies, defense contractors, healthcare organizations, and research institutions receiving federal funding — require TAA-compliant (Trade Agreements Act) and NDAA-compliant (National Defense Authorization Act) optical transceivers. Non-compliant modules sourced from restricted-country manufacturers are increasingly being rejected at procurement review for any project touching federal funding or government networks, and the downstream consequences of a compliance failure — module replacement, project delay, and potential loss of contract — are significantly more expensive than specifying compliant modules at the outset. Vitex offers multiple TAA and NDAA-compliant 400G and 800G SKUs with complete compliance documentation available for procurement review, enabling public sector and regulated industry organizations to deploy the latest high-speed optical technology without regulatory risk.

🗂️ 4. Structured MTP/MPO Cabling for High-Density AI Networks — Why Getting It Right Matters

Cabling infrastructure is the most frequently underestimated element of an AI data center upgrade — and the most consequential when it goes wrong. A transceiver selection that is technically correct but paired with an incorrectly specified, incorrectly polarized, or poorly labeled cable plant will fail to establish links, generate intermittent errors, or create troubleshooting nightmares that consume engineering time disproportionate to the cost of the cables themselves. Structured, pre-terminated, professionally specified cabling is not a premium option in a high-density AI deployment — it is the baseline that prevents a known class of preventable field failures.

Why Structured Cabling Is Critical for AI Clusters

Reliability is the first benefit. Pre-terminated trunk assemblies tested and certified at the factory eliminate the field variability that causes connector-related insertion loss failures, polarity errors, and intermittent link instability. A GPU cluster training an LLM for weeks at a time cannot afford network link resets caused by a poorly terminated field-polished connector on a cable that a structured pre-term assembly would have delivered as a factory-tested, certified component. The cost of a single training run restart caused by a cable issue exceeds the cost of the entire structured cabling system for a mid-size cluster.

Airflow management is the second benefit. Dense GPU racks operating at high thermal loads are critically sensitive to airflow restriction. Bundles of unmanaged patch cords create airflow blockages that raise GPU operating temperatures by several degrees — a difference that, sustained over weeks of continuous training, causes thermal throttling that reduces training throughput independently of any network specification. Structured trunk-and-cassette cabling routes cables in controlled, organized paths that maintain airflow lanes through the rack and across the row, directly contributing to GPU thermal performance.

Scalability and speed of deployment is the third benefit. Pre-terminated assemblies allow new rack deployments to proceed at the speed of physical installation rather than at the speed of field termination and testing. In hyperscale AI build-outs where dozens of racks are deployed simultaneously, the difference between pre-terminated and field-terminated cabling can be measured in weeks of deployment schedule — weeks during which GPUs are powered on but not delivering training capacity because the network is not yet connected.

MTP/MPO Cabling Solutions for 400G and 800G Deployments

MTP/MPO trunk cables are the backbone of parallel-fiber AI cluster cabling. For SR8 and DR8 800G modules using MPO-16 interfaces, MPO-16 trunk cables with 16-fiber connectivity are required between the switch ports and the patch panel cassette system. For DR4 400G modules using MPO-12 interfaces, MPO-12 trunk cables serve the equivalent function. Vitex supplies both MPO-12 and MPO-16 trunk assemblies in OM4, OM5, and OS2 singlemode fiber, custom-cut to the exact lengths required by your rack layout and row configuration — eliminating the excess cable slack that accumulates when standard-length assemblies are used in a non-standard physical environment.

Custom pre-terminated assemblies are the right choice for any deployment where rack layout, cable routing paths, or connection requirements do not match standard off-the-shelf trunk lengths. Vitex offers custom length, custom labeling, and custom connector configuration options that allow every cable in a deployment to be specified to the exact physical requirements of its routing path — ensuring that cables are neither too short to reach nor too long to manage without creating slack bundles that restrict airflow and complicate maintenance access.

⚡ 5. Active Optical Cables (AOC) vs Direct Attach Cables (DAC) for AI Cluster Connectivity

AOCs and DACs are essential complementary tools to discrete transceivers and structured fiber cabling in AI cluster deployments, providing flexible, cost-effective connectivity for intra-rack and very short-reach inter-rack connections where the overhead of separate transceivers and fiber cables is not justified by the distance or the port density requirements. Understanding when to use each type — and when to use discrete transceivers with structured fiber instead — is a practical skill that directly affects the cost, performance, and maintenance characteristics of an AI cluster network.

AOC and DAC Product Examples for 400G and 800G AI Deployments

For 400G connections, the 400G OSFP to 4×100G QSFP56 AOC provides breakout connectivity from a single 400G OSFP switch port to four independent 100G server ports — a common topology in AI clusters where spine switches run 400G uplinks and GPU servers present 100G network interfaces. The 400G QSFP-DD to QSFP-DD DAC serves direct switch-to-switch or switch-to-server connections within a rack where passive copper DAC's zero-power-consumption and cost advantages are the determining factors. For 800G connections, the 800G OSFP to OSFP DAC provides direct high-speed intra-rack connectivity at the lowest possible cost and power overhead, while the 100G QSFP28 to 4×SFP28 AOC continues to serve legacy server connectivity scenarios where 25G per lane is the server-side interface standard and the switch-side port is a 100G aggregate.

🌊 6. Immersion Cooling and TAA-Compliant Optics — Upgrading for Advanced AI Data Center Environments

As AI data centers push rack power densities beyond 50kW and toward 100kW per rack for the most GPU-dense configurations, conventional air cooling is reaching its physical limits. Immersion cooling — submerging server hardware in thermally conductive but electrically non-conductive dielectric fluid — is becoming a practical operational reality in hyperscale AI facilities, not a research curiosity. This transition has a direct and frequently overlooked consequence for optical networking: standard optical transceivers are not designed for long-term submersion in dielectric fluid, and deploying standard modules in immersion-cooled racks results in module degradation, link failures, and ultimately the replacement of non-immersion-rated hardware at a time and cost that are both entirely avoidable with correct specification at the outset.

Immersion-Ready Optical Modules: What Makes Them Different

Immersion-rated optical transceivers are designed with hermetically sealed enclosures and chemically resistant materials that prevent dielectric fluid ingress over the multi-year operational lifetime of the module. The sealing approach — applied at the connector interface, the module housing seams, and the electrical contacts — protects the optical components, laser, and photodetector from chemical exposure that degrades performance in standard modules within months of submersion. Vitex's immersion-ready 800G OSFP modules are specifically engineered for next-generation immersion cooling deployments, providing the same 800G SR8 optical performance as standard air-cooled variants in a hermetically sealed package rated for continuous dielectric fluid operation.

Organizations evaluating immersion cooling for new AI cluster builds should include immersion-rated transceiver specifications in their infrastructure planning from the earliest design stage — not as an afterthought after selecting standard modules. The cost delta between standard and immersion-rated modules is significantly smaller than the cost of replacing a full rack's worth of failed standard modules after deployment into an immersion environment, and the schedule impact of a mid-deployment transceiver replacement in a live AI cluster is an operational consequence that no production AI facility can afford to manage reactively.

TAA/NDAA Compliance: Protecting Regulated AI Deployments

Beyond immersion cooling, a growing number of AI data center deployments serve regulated industries, government agencies, and federally funded research institutions where product compliance with the Trade Agreements Act and the National Defense Authorization Act is a procurement requirement rather than an optional consideration. TAA compliance restricts the country of origin for government-procured goods; NDAA Section 889 prohibits the use of telecommunications equipment from specific restricted manufacturers regardless of procurement vehicle. Transceivers sourced from non-compliant manufacturers — even inadvertently, through distributors who do not clearly disclose country of origin — can trigger contract review, equipment removal, and replacement at the deploying organization's expense.

Vitex provides full TAA and NDAA compliance documentation for its compliant 400G and 800G module SKUs, enabling procurement teams to verify compliance before purchase rather than managing the consequences of non-compliance after deployment. For any AI data center project serving government, defense, healthcare, or research sectors with federal funding exposure, specifying TAA/NDAA-compliant transceivers from the outset is the risk-management decision that prevents a compliance failure from becoming a program-level incident.

🔓 7. Open Networking and Multi-Vendor Interoperability for AI Data Centers

The era of single-vendor, proprietary AI network fabrics is giving way to open, MSA-compliant architectures where GPU servers, top-of-rack switches, spine switches, and optical transceivers are sourced from the best-available vendor for each component rather than from a single vendor ecosystem. This shift is driven by the operational reality that no single vendor offers the optimal solution across all layers of an AI cluster network — and that locking into a proprietary ecosystem means accepting sub-optimal performance, pricing, and upgrade flexibility in exchange for support simplicity that mature operations teams no longer require.

Why Open Networking Matters for 2025 AI Infrastructure

Open networking means deploying MSA-compliant transceivers that operate correctly in any switch or server platform that supports the relevant interface standard — QSFP-DD, OSFP, or SFP — without vendor-specific firmware locks, compatibility validation fees, or proprietary management interfaces. MSA-compliant 400G and 800G transceivers from vendors like Vitex are tested for interoperability with leading switch platforms from Arista, Cisco, Juniper, NVIDIA/Mellanox, and others, ensuring that the modules function correctly regardless of which vendor's switch they are installed in.

The practical benefit of open networking in AI deployments is the ability to mix hardware from different vendors on different refresh cycles. A cluster's GPU servers may refresh on a two-year cycle, its top-of-rack switches on a three-year cycle, and its spine switches on a five-year cycle. Open, MSA-compliant transceivers allow each layer to be upgraded independently without triggering compatibility reassessment of the entire optical infrastructure — a flexibility that proprietary ecosystems explicitly prevent in order to drive full-stack replacement sales.

Vitex rigorously tests its optics with leading switches, GPU servers, and network interfaces, ensuring that its 400G and 800G modules are truly interoperable across multi-vendor environments. This testing regime extends to firmware compatibility, forward error correction configuration, and module management interface behavior — the details that matter in production deployments but are frequently absent from vendor compatibility matrices that list "compatible" based on physical fitment alone rather than validated end-to-end link operation.

🗺️ 8. 7 Essential Steps for Your 2025 AI Data Center Fiber and Optical Upgrade

Upgrading an AI data center network from legacy 100G or 200G infrastructure to 400G and 800G requires a systematic approach that addresses topology, hardware selection, cabling, compliance, and ongoing operations in a defined sequence. The seven steps below represent the professional planning process that prevents the most common upgrade failures — hardware incompatibilities discovered at installation, cabling errors found during link-up testing, compliance issues surfaced at procurement review, and operational gaps that emerge after the cluster goes live.

🏆 9. Real-World Example: From Bottleneck to Breakthrough

The following example illustrates how the seven-step upgrade process translates into measurable infrastructure outcomes in a production AI research environment. While specific organizational details are generalized, the technical scenario, product specifications, and results reflect the outcomes of a real deployment supported by Vitex engineering.

The Challenge: Legacy Infrastructure Blocking LLM Training Scale

A major AI research center had deployed a 100G Ethernet cluster for initial large language model training workloads. As model sizes grew from billions to hundreds of billions of parameters, the east-west traffic generated by AllReduce gradient synchronization across the GPU cluster began saturating 100G switch uplinks during peak training runs. Training throughput measurements showed that GPUs were spending up to 30% of each training step waiting for gradient synchronization — network-idle time that directly extended training duration and increased the compute cost per training run. The organization needed to quadruple the effective east-west bandwidth of its cluster network without interrupting ongoing training runs on adjacent racks.

The Solution: Targeted 800G and 400G Optical Upgrade

Working with Vitex engineering, the team designed a phased upgrade that replaced legacy 100G switch infrastructure with new spine switches supporting 800G OSFP ports while reusing existing OM4 MTP/MPO cabling infrastructure for intra-rack connections that had been installed to the correct specification in the original build. The upgrade deployed the following components: 800G OSFP SR8 modules in spine switch ports serving GPU racks with existing OM4 MPO-16 cabling infrastructure; OM4 MTP/MPO trunk cables for new rack additions that had not been part of the original build; immersion-ready 800G OSFP transceivers for a subset of racks being converted to immersion cooling as part of a parallel thermal management upgrade; and TAA-compliant 400G and 800G optics for all connections serving infrastructure tied to a federally funded research grant requiring NDAA-compliant procurement.

The Results: Measurable Training Performance Improvement

Following the upgrade, the research center measured a reduction in gradient synchronization wait time from 30% of step duration to under 8% — a result of quadrupling the effective east-west bandwidth available to the AllReduce operations running across the upgraded cluster fabric. Training throughput on representative LLM workloads increased by approximately 3.5× relative to the pre-upgrade baseline, with the remaining gap from the theoretical 4× bandwidth increase attributable to compute-bound operations that were not network-limited. Physical cable volume decreased despite the addition of new racks, because 800G OSFP ports served by SR8 modules replaced multiple 100G connections per GPU endpoint — reducing total cable count while increasing total bandwidth. Compliance documentation for all TAA/NDAA-required modules was provided by Vitex at procurement, eliminating compliance review delays and enabling the grant-funded portion of the upgrade to proceed on schedule.

📦 10. Vitex Product Portfolio Reference for AI Data Center Upgrades

Vitex offers the broadest range of optical transceivers, fiber cables, AOC/DAC assemblies, and structured cabling infrastructure for AI data center, high-performance computing, and telecommunications deployments — with rapid delivery, custom length and labeling options, compliance documentation, and compatibility validation across all leading OEM switching platforms and hyperscale architectures. The following table provides a complete reference for the product categories most relevant to 400G and 800G AI cluster upgrades in 2025.

🛠️ 11. Best Practices and Deployment Tips for 400G/800G AI Optical Infrastructure

The difference between an AI cluster network that performs at design specifications from day one and one that requires weeks of troubleshooting before reaching stable operation is almost always traceable to a small number of specification and installation decisions that experienced teams handle correctly and first-time deployers frequently underestimate. The best practices below represent the accumulated experience of optical infrastructure deployments at scale — applied to 400G and 800G AI cluster environments where the consequences of getting them wrong are measured in GPU-hours of lost training capacity.

Transceiver Specification and Validation

Always perform a pre-deployment compatibility check between every transceiver model and every switch firmware version in your deployment before ordering at volume. Switch firmware compatibility tables are updated frequently, and a module that operated correctly on firmware version N-1 may generate errors or fail to initialize on version N due to management interface changes. Request Vitex's compatibility validation reports for your specific switch hardware and firmware version combination, and do not assume that "QSFP-DD" or "OSFP" form factor compatibility implies link-level interoperability — the form factor is the necessary condition, not the sufficient condition, for a working link.

For SR8 and DR8 modules, verify that your switch platform's forward error correction (FEC) configuration matches the module's requirements before enabling the port. RS-FEC (Reed-Solomon Forward Error Correction) is required for 800G SR8 links and must be enabled on both ends of the link simultaneously — a common source of link establishment failures that is trivially fixed once identified but can consume significant diagnostic time when the engineer does not know to check FEC configuration as a first step.

Cabling Installation and Polarity Management

The single most common cause of post-installation link failures in parallel-fiber MTP/MPO deployments is incorrect polarity — a condition where the transmit fiber on one end of a trunk cable connects to the transmit fiber on the far end rather than to the receive fiber as required for a working duplex link. Polarity errors are invisible to visual inspection, pass connector cleanliness checks, and can only be detected with a polarity tester or by observing that the link does not come up when the transceiver is installed and powered. Establish a polarity management plan — selecting Method A, B, or C consistently throughout the deployment — before installing a single cable, and verify polarity of every trunk before connecting transceivers.

Clean connectors before every connection. Contamination is the leading cause of optical insertion loss that pushes links beyond their power budget margins — a condition that produces intermittent link errors rather than clean failures, making the contamination source difficult to localize without systematic connector cleaning and inspection. Use a one-click cleaner on every MPO connector before every connection, and inspect with an MPO-compatible fiber scope before installing transceivers on newly connected trunk cables.

Operational Monitoring and Documentation

Establish optical power monitoring baselines for every link in the cluster during initial commissioning — recording transmitted power, received power, and insertion loss for every port. These baselines serve as the reference against which future degradation is measured, enabling early detection of connector contamination, transceiver aging, or physical cable damage before they cause link failures. Most enterprise switch platforms expose optical power data via SNMP or streaming telemetry; integrating this data into your cluster monitoring platform from day one is the operational investment that transforms reactive fiber troubleshooting into proactive maintenance.

Maintain a living cable plant documentation record that maps every transceiver, trunk cable, and patch cord to its physical location, connection endpoints, and specification. In a cluster with hundreds or thousands of links, this documentation is the difference between a 15-minute maintenance intervention and a 4-hour troubleshooting session when a single link fails. Vitex's custom labeling options for trunk cables and patch cords integrate directly into physical documentation systems — each cable arrives labeled with the information you specify, eliminating the field labeling step that teams under deployment schedule pressure most commonly skip.

🔮 12. FAQs and Getting Expert Support for Your 2025 AI Data Center Upgrade

Frequently Asked Questions

The Bottom Line: Smart Optical Infrastructure Is the Foundation of AI Performance

The AI data center of 2025 runs on smart fiber and the right optical transceivers. Choosing the wrong module form factor, the wrong reach variant, the wrong fiber type, or the wrong cabling approach creates infrastructure constraints that compound with every scale-out event — and the cost of resolving those constraints reactively is always higher than the cost of specifying correctly at the outset. Use the seven-step upgrade plan in Section 8 as your planning framework. Match 800G OSFP or QSFP-DD to your switch hardware. Pair SR8 with OM4/OM5 MTP/MPO-16 and DR4 with OS2 MPO-12. Specify immersion-rated modules for immersion-cooled racks. Require TAA/NDAA compliance documentation before procurement for any regulated deployment. Prioritize MSA-compliant, multi-vendor-tested transceivers throughout — and engage Vitex engineering support before finalizing your bill of materials, not after discovering a field incompatibility during installation.