Is On-Premise AI Hosting Coming Back? The Privacy-First Trend

For nearly a decade, the dominant narrative in enterprise IT has been "cloud-first." Every workload, every application, every byte of data was supposed to migrate to AWS, Azure, or Google Cloud. But a quiet countercurrent is gaining momentum — and it is reshaping how businesses approach AI infrastructure. On-premise AI hosting, once dismissed as a relic of the pre-cloud era, is experiencing a genuine renaissance driven by privacy concerns, regulatory pressure, and the hard economics of GPU-intensive computing at scale. This shift is not about rejecting the cloud outright. It is about recognizing that for certain AI workloads, keeping compute close to the data is not just a security preference — it is a strategic imperative that affects cost, performance, and long-term organizational control.

At Hosting Captain, we have tracked this trend across dedicated server deployments, GPU cluster configurations, and hybrid infrastructure consultations throughout 2025 and into 2026. The pattern is unmistakable: organizations that once rushed to rent cloud GPU instances are now pausing to run the numbers, and many are concluding that bringing AI workloads back on-premise — or keeping them there in the first place — delivers measurable advantages that cloud-first advocates rarely discuss. This article examines the forces behind the on-premise AI hosting revival, the numbers that matter in 2026, the hardware and regulatory landscape, and the practical trade-offs that determine whether this approach fits your organization. For a broader context on how AI hosting works, check out our complete guide to AI hosting before diving into the on-premise specifics.

Why Companies Are Reconsidering On-Premise AI Hosting

The pendulum swing away from pure cloud dependency did not happen overnight. It began with sticker shock. In 2023 and 2024, as enterprises scaled their AI proof-of-concepts into production, cloud GPU bills started arriving with numbers that alarmed even well-funded CFOs. An eight-GPU H100 instance on a major cloud provider can run north of $30 per GPU-hour at on-demand pricing, which translates to roughly $17,000 per month for a single node before networking, storage, and data egress fees. Multiply that across a cluster of ten nodes running inference 24/7, and the annual bill crosses $2 million — for hardware that you do not own and that depreciates on someone else's balance sheet. This is the math that prompted many organizations to reopen the on-premise conversation.

Beyond cost, data gravity plays an equally powerful role. AI models are only as good as the data they train and infer on, and in industries like healthcare, financial services, and defense, that data tends to be both massive and sensitive. Transferring terabytes or petabytes of proprietary data into a public cloud environment introduces latency bottlenecks, bandwidth expenses, and compliance headaches that compound over time. When your data already resides in an on-premise data center or colocation facility, running AI workloads adjacent to that data eliminates the constant back-and-forth that cloud architectures inherently require. The principle is simple: move compute to the data, not data to the compute.

There is also a growing awareness that AI infrastructure decisions have multi-year lock-in effects that early cloud enthusiasm glossed over. Once an organization instruments its data pipelines, security tooling, and model deployment workflows around a specific cloud provider's proprietary AI services — SageMaker, Vertex AI, or Azure Machine Learning — switching costs become prohibitive. On-premise AI hosting, whether built on bare metal or managed through platforms like NVIDIA AI Enterprise, offers a degree of portability and vendor independence that cloud-native AI stacks struggle to match. This is not just about avoiding lock-in; it is about preserving the freedom to negotiate pricing, adopt new hardware generations on your own timeline, and maintain architectural consistency across training and inference environments.

The final catalyst is something less tangible but no less real: a maturing understanding of what AI workloads actually need. Not every AI use case requires the elastic scale of a hyperscale cloud. Fine-tuned inference on a domain-specific model, real-time computer vision on a factory floor, or a language model serving internal knowledge base queries — these workloads run on predictable, steady-state compute profiles that map elegantly to owned hardware. For organizations that have already invested in data center real estate, power infrastructure, and IT operations teams, adding GPU-accelerated AI nodes to an existing on-premise footprint is increasingly the path of least resistance and best long-term value.

Data Privacy Regulations Driving On-Prem Decisions

Regulatory frameworks have become one of the most decisive factors pushing AI workloads back behind the corporate firewall. The General Data Protection Regulation (GDPR) in Europe imposes strict requirements on cross-border data transfers, data processing transparency, and the right to explanation for automated decisions — all of which intersect directly with how AI models handle personal data. When a financial institution in Frankfurt trains a credit risk model on customer transaction histories, the legal team is increasingly unwilling to sign off on shipping that data to a cloud region that, while geographically in the EU, is operated by a US-headquartered company subject to the CLOUD Act and other extraterritorial surveillance laws. On-premise AI hosting eliminates this entire category of jurisdictional risk by keeping data processing within a physically and legally bounded environment under the organization's direct control.

In the healthcare sector, HIPAA compliance adds another layer of urgency. Covered entities and business associates that process protected health information (PHI) must maintain a chain of administrative, physical, and technical safeguards that is significantly harder to document and audit in a shared-responsibility cloud model. When a hospital system deploys an AI radiology assistant that analyzes MRI scans, every stage of the pipeline — from image ingestion to model inference to result storage — must be demonstrably HIPAA-compliant. Running that workload on owned GPU servers in a controlled-access data center simplifies the compliance narrative substantially. The organization can point to hardware-level encryption, physical access logs, and network segmentation policies that it owns end to end, rather than relying on a cloud provider's shared security model and hoping the audit goes smoothly.

Sector-specific regulations beyond GDPR and HIPAA are proliferating as well. The Digital Operational Resilience Act (DORA) in the EU requires financial entities to maintain full control over critical ICT systems, including the ability to switch providers without disruption. The proposed EU AI Act introduces risk-tiered compliance obligations that are harder to meet when core AI infrastructure sits outside the organization's sovereign boundary. Even in jurisdictions without explicit AI regulations, industry standards like ISO 27001 and SOC 2 Type II audits reward organizations that can demonstrate clear, unbroken control over their data processing pipeline — something that on-premise architectures make straightforward and cloud architectures make complicated. The W3C web standards community has also increasingly emphasized privacy-preserving architectures, reflecting a broader industry consensus that data sovereignty is not a niche concern but a foundational design principle for the next generation of AI systems.

An emerging trend worth watching is the convergence of privacy regulations with AI governance mandates. Regulators are no longer content with "we encrypted the data at rest." They want documented model lineage, training data provenance, bias audit trails, and the ability to explain individual inference decisions. When your entire AI stack — from GPU firmware to inference API endpoints — runs on hardware you control, generating these audit artifacts becomes a matter of instrumenting your own systems. In a cloud environment, however, the abstraction layers that make AI services convenient also obscure the very details that regulators increasingly demand. This transparency gap is quietly driving procurement decisions in heavily regulated industries, where compliance teams now have a seat at the infrastructure planning table that they did not occupy five years ago.

Cost Comparison: On-Premise GPU Servers vs Cloud GPU Rental in 2026

The economics of AI infrastructure in 2026 tell a story that on-premise advocates have been telling for years — but with numbers that are now impossible to ignore. A properly configured on-premise GPU node built around an NVIDIA H100 80GB SXM5 GPU, paired with dual AMD EPYC processors, 512GB of system RAM, and high-speed NVMe storage, carries a capital expenditure of roughly $45,000 to $55,000 per node when purchased through channel partners. Across a three-year depreciation window, that works out to approximately $1,250 to $1,530 per month per node. Compare that to a comparable cloud GPU instance — an Azure ND H100 v5 or AWS P5 instance with eight H100 GPUs — which runs between $24 and $32 per GPU-hour at on-demand rates, or roughly $14,000 to $18,400 per month for an equivalent eight-GPU node even with one-year reserved instance discounts. The on-premise option pays for itself within four to six months at sustained utilization levels above 60%, and after that, it is pure cost advantage.

But the total cost of ownership comparison extends well beyond sticker prices on GPU boards. Cloud providers charge for data egress — the simple act of pulling your data back out of their environment — at rates that can add five figures to an annual AI budget without anyone noticing until the invoice arrives. Training a large language model on a multi-terabyte dataset that resides on-premise means ingesting that data into the cloud once and potentially moving model checkpoints back out for on-premise deployment — a bidirectional transfer that triggers egress fees both ways if you are not extremely careful about architecture. On-premise deployments sidestep all of these charges. Your data already sits on your storage arrays; your GPUs sit in the adjacent rack; the network traffic between them costs exactly what your internal switching infrastructure costs, which is effectively zero at the margin.

There are also less obvious cost factors that tilt the scales toward on-premise in 2026. Cloud GPU availability remains inconsistent — spot instances can be reclaimed with minimal notice, and even reserved instances are not immune to capacity constraints in popular regions. When your inference API goes down because AWS cannot provision H100 instances in us-east-1 for six hours, the revenue impact of that downtime can dwarf the monthly cloud bill. On-premise hardware, once installed, answers to no one's capacity scheduler but your own. Additionally, the secondary market for previous-generation enterprise GPUs (NVIDIA A100 80GB, for instance) has matured to the point where organizations can build capable inference clusters at 40-50% of the cost of current-generation hardware, an option that simply does not exist in the cloud ecosystem where GPU instance types are dictated by the provider's refresh cycle.

That said, the cost equation is not universally in on-premise's favor. Organizations with sporadic, bursty AI workloads — training a model once per quarter, or running inference only during business hours — will find that cloud GPU rental aligns costs more closely with actual usage. The on-premise math works best when GPU utilization stays above 50-60% on a monthly basis. It also requires upfront capital that not every organization has access to, although GPU-as-a-service financing models and hardware leasing arrangements are increasingly bridging this gap. The key is running a disciplined total cost of ownership analysis that accounts for utilization patterns, data egress volumes, depreciation schedules, power and cooling costs, and the fully loaded cost of the IT staff who manage the hardware. When Hosting Captain consults with clients on dedicated server deployments, we consistently find that organizations running sustained AI inference workloads with predictable demand profiles achieve the most dramatic savings from on-premise infrastructure.

Hardware Requirements for On-Premise AI Infrastructure

Building an on-premise AI hosting environment is not as simple as ordering a few GPUs and plugging them into existing server racks. The hardware requirements span compute, networking, storage, and facilities infrastructure, and getting any one of these layers wrong can bottleneck the entire investment. At the compute layer, the current-generation gold standard is the NVIDIA H100 Tensor Core GPU, with the H200 and B200 (Blackwell architecture) beginning to ship in early 2026. The H100 delivers roughly 3,958 teraflops of FP8 performance per GPU, and when clustered in an NVIDIA DGX H100 system — which packages eight H100 GPUs with NVLink interconnect and 640GB of total GPU memory — you have a single node capable of training a 175-billion-parameter model in days rather than weeks. For organizations building their own systems rather than purchasing integrated DGX appliances, the GPU servers themselves must support PCIe Gen5, provide adequate PCIe lanes to avoid bottlenecking GPU-to-CPU communication, and deliver sufficient power delivery and thermal headroom for sustained 700W-per-GPU operation.

Networking is the layer most frequently underestimated in on-premise AI builds. AI training workloads are notoriously sensitive to inter-node communication latency because distributed training techniques like data parallelism, tensor parallelism, and pipeline parallelism require frequent gradient synchronization across nodes. A cluster of eight GPU nodes training a single model needs at least 400 Gbps of interconnect bandwidth per node, typically delivered via NVIDIA ConnectX-7 NICs or Intel E810-CQDA2 adapters running RoCEv2 (RDMA over Converged Ethernet) or InfiniBand NDR400. This is not the same as slapping a 100GbE switch into a rack and hoping for the best — it requires purpose-built switching fabric, proper congestion management, and often dedicated storage networks to prevent training traffic from contending with storage I/O. For inference-only deployments, networking requirements are less extreme, but multi-node inference serving with tensor parallelism still benefits substantially from high-bandwidth, low-latency interconnects.

Storage architecture for on-premise AI demands careful attention to throughput rather than just capacity. A single H100 GPU can consume training data at upwards of 10 GB/s, meaning an eight-GPU node can saturate a 100 GB/s storage pipeline during active training. Traditional NAS appliances with spinning disk arrays will become the bottleneck long before the GPUs reach full utilization. The standard approach in 2026 pairs high-capacity HDD-based storage (for dataset archival) with a high-performance NVMe tier that acts as a read cache and staging area. Parallel file systems like WEKA, VAST Data, or Pure Storage FlashBlade have emerged as the de facto standards for AI storage because they can deliver the multi-terabyte-per-second aggregate throughput that GPU clusters demand while presenting a POSIX-compliant namespace that data science teams can work with using standard tools. Organizations that try to cut corners on the storage tier will discover that GPU utilization hovers at 30-40% while the expensive accelerators sit idle waiting for data to arrive.

Software stack compatibility is the final hardware-adjacent consideration that trips up first-time on-premise AI builders. NVIDIA's AI Enterprise suite provides a validated, supported software stack that includes CUDA, cuDNN, TensorRT, and Triton Inference Server, along with reference architectures for Kubernetes-based orchestration. Without this layer, organizations find themselves in dependency hell — matching GPU driver versions to CUDA toolkit versions to PyTorch versions and hoping nothing breaks during a routine security patch. For organizations that prefer a more turnkey approach, integrated platforms like NVIDIA DGX BasePOD and SuperPOD reference architectures provide pre-validated hardware-plus-software blueprints that remove much of the integration risk, albeit at a premium price point. The VPS hosting guide for beginners on Hosting Captain explains foundational hosting concepts that apply even at the enterprise GPU scale, helping newcomers understand the infrastructure principles before diving into AI-specific hardware decisions.

Hybrid AI Hosting Models That Bridge Both Worlds

Not every organization needs to choose exclusively between on-premise and cloud AI infrastructure. The most sophisticated AI operations in 2026 are embracing hybrid models that assign each workload to the environment where it performs best economically and operationally. A common pattern is to run model training on cloud GPU clusters — taking advantage of elastic scale to spin up a hundred nodes for a two-week training run, then tear them down — while deploying the resulting trained model on an on-premise inference cluster that serves production traffic with predictable, steady-state compute requirements. This hybrid approach captures the cloud's strength in handling spiky, transient workloads while securing the cost and latency advantages of on-premise for the long-running, business-critical inference tier.

The viability of hybrid AI architectures has improved dramatically with the maturation of Kubernetes-based orchestration and multi-cluster federation. Tools like NVIDIA Run:ai, Kubernetes with the GPU operator, and ML orchestration platforms such as Kubeflow and MLflow enable organizations to treat on-premise GPU clusters and cloud GPU instances as a single logical resource pool. A data scientist submits a training job, and the scheduler decides — based on cost policies, resource availability, and data locality — whether to route it to the on-premise H100 cluster in the Frankfurt data center or to provision spot instances in an AWS region. This abstraction layer is the key to realizing the full cost optimization potential of hybrid AI hosting, but it requires upfront investment in platform engineering that many organizations underestimate. At Hosting Captain, we advise clients to budget for at least a quarter of dedicated platform engineering effort when standing up a hybrid AI orchestration layer — the payback period is short, but the initial lift is non-trivial.

Edge inference is another dimension of the hybrid model that is gaining traction in manufacturing, retail, and telecommunications. In this architecture, a small on-premise GPU server — or even a purpose-built inference appliance — sits at the network edge (a factory floor, a retail distribution center, a cell tower aggregation point) and runs real-time inference against locally generated sensor data, video feeds, or transaction streams. The results are streamed to a central cloud data lake for aggregation and analytics, while the models themselves are periodically updated from a central training pipeline. This pattern delivers sub-10-millisecond inference latency for time-sensitive applications while keeping training and model management centralized. It also addresses a growing regulatory concern: keeping personally identifiable data at the edge, within the jurisdiction where it was collected, rather than shipping it to a central cloud that may cross regulatory boundaries.

Colocation facilities are emerging as the physical backbone of hybrid AI hosting. Rather than building an on-premise data center from scratch — with the power, cooling, and physical security investments that entails — organizations are deploying owned GPU servers in colocation cages operated by providers like Equinix, Digital Realty, or CyrusOne. These facilities offer direct, low-latency cross-connects to major cloud providers (AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect), effectively collapsing the network distance between on-premise inference clusters and cloud-based training environments. An inference node in an Equinix cage can communicate with an S3 bucket in the same metro region with single-digit millisecond latency and zero egress fees over the direct interconnect. This topology delivers the best of both worlds: the cost structure and data sovereignty of owned hardware combined with the cloud adjacency that makes hybrid architectures practical at scale.

Latency Advantages of On-Premise AI Inference

Latency is the silent killer of AI user experiences, and it is an arena where on-premise deployments hold a structural advantage that cloud architectures cannot fully close. When a user interacts with an AI-powered application — whether it is a chatbot, a code completion tool, or a real-time fraud detection system — the total response time includes not just the model's inference computation but also network round-trip time, request queuing in the inference server, and any data preprocessing or postprocessing steps. In a cloud deployment, the network component alone can add 20-80 milliseconds for requests traversing the public internet, and that number gets worse for users in regions far from the nearest cloud data center. On-premise inference servers sitting on the same local network as the application backend can deliver sub-millisecond network latency, which, when combined with optimized inference engines like NVIDIA TensorRT-LLM or vLLM, shrinks end-to-end response times into the 5-20 millisecond range for typical workloads.

This latency differential matters most in two categories of AI applications: real-time interactive systems and high-throughput transaction processing. For real-time systems — think voice assistants, live translation, autonomous vehicle telemetry processing, or augmented reality overlays — every 50 milliseconds of added latency degrades the user experience in ways that are directly measurable in engagement metrics and conversion rates. Studies by Google and Amazon have repeatedly demonstrated that sub-100-millisecond response times are the threshold for perceived real-time interaction, and cloud-based AI inference routinely consumes half or more of that budget before a single tensor operation has been executed. For high-throughput transaction processing — fraud detection in payment systems, ad bidding optimization, algorithmic trading — the latency advantage of on-premise inference translates directly into revenue. A bidder that evaluates an ad opportunity 40 milliseconds faster than competitors wins more auctions; a fraud model that flags a suspicious transaction in 5 milliseconds instead of 80 prevents more losses.

Batch processing workloads benefit less dramatically from on-premise latency advantages, which is why hybrid architectures that reserve on-premise for real-time inference and cloud for batch training make so much sense. But even in batch scenarios, there is an underappreciated second-order effect: on-premise infrastructure eliminates the cold-start latency that plagues serverless and container-based cloud AI services. When an infrequently used cloud AI endpoint receives a request after a period of inactivity, the container or function must be provisioned, the model must be loaded into GPU memory, and only then can inference begin — a process that can take 10-30 seconds. For applications that require consistent, predictable response times regardless of request frequency, on-premise GPU servers with pre-loaded models in persistent GPU memory deliver consistently low tail latencies that cloud architectures struggle to match without expensive provisioned-concurrency configurations.

Real Case Studies of Companies Moving AI Back On-Premise

The on-premise AI repatriation trend is not theoretical — it is showing up in earnings calls, infrastructure RFPs, and the public statements of major enterprises. In late 2024, a Fortune 500 financial services firm disclosed in its annual technology review that it had moved its core fraud detection model from a cloud-based GPU deployment to an on-premise H100 cluster, citing a 62% reduction in per-inference cost and a 45-millisecond improvement in tail latency. The fraud detection system processes roughly 8,000 transactions per second during peak hours, and the latency improvement alone — translating to faster flagging of suspicious activity — was estimated to reduce fraud losses by approximately $4.2 million annually. The firm noted that it retained cloud GPU capacity for quarterly model retraining runs, where the elastic scale of the cloud still offered advantages, but the steady-state inference workload now runs entirely on owned hardware in its New Jersey data center.

In the healthcare AI space, a major European hospital network made headlines in early 2025 when it announced the deployment of an on-premise AI radiology platform powered by eight DGX H100 systems distributed across three hospital campuses. The platform processes approximately 12,000 medical images per day — CT scans, MRIs, and X-rays — using a suite of FDA-approved and CE-marked AI diagnostic models. The decision to go on-premise was driven almost entirely by data privacy considerations: patient imaging data never leaves the hospital's physical network, which simplifies GDPR compliance and eliminates the need for patient consent for cloud data processing. The hospital network reported that the on-premise deployment achieved a per-image inference cost of approximately €0.03, compared to the €0.11-0.14 it had been paying a cloud-based AI diagnostics vendor for the same capability, while also reducing report turnaround times from an average of 12 minutes to under 4 minutes.

The technology sector itself provides some of the clearest examples of on-premise AI repatriation. A well-known enterprise SaaS company, which had previously migrated virtually all its infrastructure to AWS over a five-year period, announced a "strategic rebalancing" of its AI workloads in its Q3 2025 earnings call. The company disclosed that it was deploying 64 on-premise GPU nodes — a mix of H100 and A100 systems — to handle its internal code generation assistant and customer-facing AI features. The CFO cited a projected $8.7 million annual savings versus the equivalent cloud GPU spend, with a break-even period of 14 months on the hardware investment. The CTO added that the on-premise deployment allowed the engineering team to use custom GPU memory configurations and kernel optimizations that were not possible within the standardized cloud instance types, resulting in a 23% throughput improvement for their specific model architecture. These are not nostalgic "lift-and-shift" reversals; they are surgical repatriations of specific workloads where the cloud premium no longer made economic or technical sense.

Challenges of On-Premise AI: Cooling, Power, and Maintenance

A candid assessment of on-premise AI hosting must acknowledge the real operational challenges that come with owning and operating high-density GPU infrastructure. Cooling is the first and most expensive hurdle. An NVIDIA DGX H100 system consumes approximately 10.2 kilowatts at full load, and a rack populated with four such systems draws over 40 kilowatts — well beyond the 5-10 kilowatts per rack that traditional enterprise data centers were designed to handle. Traditional air cooling becomes inadequate at these densities; the hot aisles in a GPU rack can exceed 50°C without proper containment and high-velocity airflow management. Direct-to-chip liquid cooling is rapidly becoming the standard for on-premise H100 and B200 deployments, with solutions from CoolIT Systems, Asetek, and Boyd Corporation providing closed-loop or facility-water-based cooling that removes 80-90% of the heat directly from the GPU cold plates. Retrofitting an existing data center for liquid cooling is a six-figure capital project in most cases, and it requires facilities engineering expertise that many IT organizations do not have in-house.

Power infrastructure presents a parallel set of challenges. A modest on-premise AI cluster of 16 H100 nodes requires approximately 160 kilowatts of power delivery capacity, plus redundancy for cooling, networking, and storage — easily exceeding 200 kilowatts for the full rack footprint. Many commercial office buildings and even some older data centers simply do not have this much power available at the building feed level, and negotiating a power upgrade with the local utility can take 12-18 months in regions with constrained grid capacity. Organizations must also invest in uninterruptible power supplies (UPS) sized for the full GPU load and on-site generator backup, which adds 15-25% to the facilities capital cost. The days of plugging a server into a standard 208V outlet and calling it done are over for AI infrastructure; these deployments require the same level of facilities planning as a small manufacturing line.

Ongoing maintenance is the challenge that most distinguishes on-premise AI from cloud AI, and it is the area where organizations frequently underestimate the operational burden. GPUs fail — not frequently, but at a rate that becomes statistically meaningful across a fleet of hundreds of accelerators. NVIDIA's published annualized failure rate for data-center GPUs is in the 1-3% range depending on workload intensity and thermal conditions, which means a 128-GPU cluster can expect one to four GPU failures per year. Diagnosing these failures requires specialized tooling like NVIDIA's DCGM (Data Center GPU Manager) and the ability to isolate a failed GPU, drain its workloads, and replace the hardware without disrupting the rest of the cluster. Kubernetes GPU-aware scheduling can automate some of this, but the physical replacement and the software re-integration still require skilled data center technicians and AI systems administrators — a talent pool that is both expensive and scarce.

Software maintenance introduces its own rhythm of disruption. CUDA driver updates, firmware patches, and security vulnerability remediations do not stop just because the hardware is on-premise. Every GPU driver update must be tested against the organization's specific combination of deep learning frameworks, inference engines, and orchestration tools, and a regression that silently degrades FP16 performance by 5% can wipe out months of cost optimization in a single maintenance window. Organizations that excel at on-premise AI operations invest in staging environments — smaller GPU clusters where updates are validated before reaching production — and develop automated CI/CD pipelines for infrastructure changes that mirror the rigor of their application deployment pipelines. The operational maturity required is not out of reach, but it is a step-function increase from managing CPU-based server fleets, and organizations should budget for a dedicated AI infrastructure operations team rather than assuming existing data center staff can absorb these responsibilities on top of their current duties.

Frequently Asked Questions

What is the most important thing to know about on-premise AI hosting?

This guide covers the practical decision points — pricing, performance, and when it makes sense for your situation — based on current 2026 data.

How much does this typically cost in 2026?

Pricing varies by provider and plan tier; see the cost breakdown section above for current ranges and what's actually included at each price point.

What should beginners check before making a decision?

Look closely at uptime guarantees, renewal pricing (not just the first-year discount), and how responsive support actually is — all covered in detail in this article.