Enterprises chasing on-premises AI in 2026 face a genuinely confusing shopping trip. AWS, Microsoft, and Google each sell a version of “the public cloud, but in your own data center,” and each one prices, licenses, and hardware-gates that promise in a completely different way. AWS Outposts locks buyers into a fixed three-year hardware contract. Microsoft rebranded Azure Stack HCI into Azure Local and switched to a pay-as-you-go per-core subscription. Google Distributed Cloud (GDC) ties its Edge tier to a mandatory 36- or 60-month vCPU commitment but is the only one of the three that publicly documents on-premises NVIDIA H100 and Blackwell GPU support today.
That last point matters more than it sounds. Every enterprise architecture conversation in September 2026 eventually turns into “where do we run the model,” and for regulated industries the answer is often “not in a public region.” This comparison walks through the hardware, the exact published pricing, the contract terms, the GPU story, and the real deployments driving adoption of AWS Outposts, Azure Local (formerly Azure Stack HCI), and Google Distributed Cloud so technical buyers can shortlist the right platform instead of guessing from marketing pages.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Are AWS Outposts, Azure Local, and Google Distributed Cloud?
All three products fall under what analysts call “distributed cloud”: public-cloud infrastructure, APIs, and control planes physically extended into a customer’s own data center, colocation facility, retail branch, factory floor, or even a disconnected government site. The pitch is consistency: the same consoles, APIs, and services teams already use in the public region, just running on hardware that never leaves a building the customer controls.
AWS Outposts ships as a fully managed 42U rack or as compact 1U/2U servers. Amazon designs, delivers, installs, patches, and eventually removes the hardware; customers have no say in the physical specs beyond picking a capacity bundle at order time. Azure Local, which Microsoft formally renamed from Azure Stack HCI effective November 19, 2024, flips that model: customers or their integrators buy validated hyperconverged hardware from OEM partners like Dell, HPE, and Lenovo, and Microsoft sells the software subscription that turns it into an Azure-connected cluster. Google Distributed Cloud sits in between, offering both a “connected” tier built on Google-certified partner hardware and a fully air-gapped appliance for classified and sovereign workloads that never touches the public internet.
The three products also target subtly different buyers. Outposts leans on AWS’s enormous regional service catalog and appeals to teams that are already deep into EC2, S3, and RDS and simply need a local extension. Azure Local leans on Windows Server licensing relationships and IT departments that already run Hyper-V or VMware and want a lighter migration path. GDC leans hardest into two extremes: telco/retail edge locations needing Kubernetes-native infrastructure, and defense/intelligence agencies needing an appliance that can run without any connection to the internet at all. Teams already weighing the managed-Kubernetes layer on top of these platforms should also see how EKS, AKS, and GKE compare on pricing and control-plane cost, since all three distributed-cloud products eventually connect back to their parent’s managed Kubernetes offering.
Full Specs Comparison Table
| Spec | AWS Outposts | Azure Local (Azure Stack HCI) | Google Distributed Cloud |
|---|---|---|---|
| Hardware model | AWS-owned, delivered and maintained by Amazon | Customer-owned OEM hardware (Dell, HPE, Lenovo) | Google-certified partner hardware or GDC appliance |
| Form factors | 42U rack (1st/2nd gen), 1U/2U servers | 2- to 16-node hyperconverged clusters | Connected/Edge nodes, Hosted appliance, air-gapped appliance |
| Pricing model | Bundled hardware + capacity, fixed term | ~$10 per physical core, per month (subscription) | $35 per vCPU, per month, 96 vCPU minimum |
| Minimum commitment | 3 years (rack); 1 year (1U server) | None fixed; 60-day free trial, cancel any time | 36 or 60 months mandatory (connected tier) |
| Confirmed on-prem GPU | G4dn (NVIDIA T4-class); newer GPUs “coming soon” | OEM-dependent; no named NVIDIA model documented | NVIDIA H100 94GB (up to 4 per node); Blackwell platform confirmed |
| Kubernetes support | EKS on Outposts (local clusters) | AKS on Azure Local (AKS-HCI) | GKE-based, Kubernetes-native by design |
| Air-gapped option | No public air-gapped SKU | No public air-gapped SKU | Yes — GDC air-gapped appliance, DoD IL6 authorized |
| Uplink networking | 1/10/40/100 Gbps (gen-dependent) | OEM-dependent, typically 10-100 Gbps | Site-dependent, quote-based for hardware |
| Object storage on-prem | S3 on Outposts, ~$0.176/GB-month | Storage Spaces Direct (local, included in cluster) | GDC local object storage, capacity-based billing |
| Best documented use case | Regional AWS extension, 5G telco cores | Branch-office virtualization, VMware migration | Classified/sovereign AI, tactical edge, retail edge |
| Billing granularity | Reserved capacity over term, charged from install | Daily, based on all registered physical cores | Monthly per vCPU, capacity-based for appliances |
| Rebrand/status in 2026 | No rename; 2nd-gen racks shipped in 2025 | Renamed from Azure Stack HCI, Aug 2026 release v12.2608 | No rename; expanded IL6 and Blackwell support in 2026 |
AWS Outposts: Hardware, Pricing, and the 3-Year Lock-In
AWS Outposts comes in two physical shapes. The rack version is a full 42U cabinet, and second-generation racks (which began rolling out through 2025) split into a compute rack drawing 10-30 kVA and a network rack drawing 8.89 kVA, both supporting 10/40/100 Gbps uplinks and newer Nitro-based EC2 instance families. The smaller footprint comes from Outposts servers: 1U and 2U appliances built for a retail branch, a factory floor, or any site too small to justify a full rack.
The pricing model is where Outposts feels most different from the other two. AWS sells racks on a fixed 3-year term, with All Upfront, Partial Upfront, or No Upfront payment options, and the price covers delivery, installation, maintenance, patching, and eventual rack removal. Customers are not billed hourly for usage; they are billed for reserved hardware capacity over the full term, and the meter starts 24 hours after installation whether or not the capacity gets used. Real-world quotes for a mixed M5/C5/R5 rack land in the $150,000 to $500,000 range over three years, and one widely cited “OR-XLARGE” configuration in us-east-1 runs $180,000-$350,000 per year on a no-upfront 3-year term — $540,000 to $1,050,000 total. Outposts servers scale down sharply: a 1U retail unit runs roughly $400-$900 a month on a 1-year lease, and a 2U edge server runs $1,000-$1,500 a month on a 3-year commitment.
Graviton-based instances are listed as “coming soon” for racks as of the current AWS Outposts FAQ, and GPU support on racks is currently anchored to the older G4dn family (Turing-generation NVIDIA T4 GPUs). AWS has not published documentation confirming H100, H200, or Blackwell GPUs running on Outposts hardware, even though those chips are widely available in AWS’s public regions through P5 and other instance families. For teams that specifically need on-prem H100-class training or inference, that gap is the single biggest reason to look elsewhere in this comparison.
Azure Local: The Rebrand From Azure Stack HCI and Its Per-Core Model
Microsoft spent 2025 quietly retiring the “Azure Stack HCI” name. Microsoft Learn now states plainly that “Azure Stack HCI is now part of Azure Local” and that clusters previously called Azure Stack HCI clusters are renamed Azure Local instances, with individual servers renamed Azure Local machines. The name change rolled out across pricing and product pages starting November 19, 2024, and by the August 2026 release (version 12.2608.1003.9), Microsoft’s own release notes still flag the transition as “formerly Azure Stack HCI” for anyone searching under the old name. Industry commentary frames the rebrand as more than cosmetic — Azure Local is positioned as “cloud infrastructure for distributed locations” rather than a narrow hyperconverged virtualization product.
Unlike Outposts and GDC, Azure Local does not bundle hardware into the price. Customers buy validated servers from an OEM (Dell, HPE, Lenovo, and others sell certified Azure Local nodes) and then pay Microsoft a software subscription. The official Azure Stack HCI pricing page confirms the rate: “purchase an Azure subscription and activate only the cores you need for your hyperconverged infrastructure deployment (US$10 per physical core, per month).” Billing runs daily based on every physical core in the cluster while it stays registered, and there’s a 60-day free trial before charges kick in. Customers with Windows Server Datacenter licenses under active Software Assurance can apply Azure Hybrid Benefit to knock that host fee down to $0 per physical core per month, which is a meaningful advantage for enterprises that already own Windows Server Datacenter licensing at scale.
Because Azure Local is a software-only subscription layered on customer-chosen hardware, GPU support depends entirely on which OEM configuration a buyer picks. Microsoft’s documentation doesn’t publish a specific list of validated NVIDIA GPU models for Azure Local the way Google does for GDC. In practice, buyers who want GPU nodes on Azure Local are sourcing them through their OEM’s validated hardware catalog and running Azure Kubernetes Service on Azure Local (AKS-HCI) or GPU-attached VMs, but there’s no public confirmation of H100, H200, or Blackwell support specifically tied to the Azure Local brand as of this writing. Teams weighing whether to route AI agent workloads through this hardware instead should also compare AWS Bedrock AgentCore, Azure AI Foundry, and Vertex AI Agent Builder, since the agent-hosting layer and the hybrid-hardware layer are usually procured as separate decisions.
Google Distributed Cloud: The Only One With Confirmed On-Prem H100s
Google Distributed Cloud splits into three distinct products, and the differences matter a lot for anyone comparison-shopping. GDC connected (also called GDC Edge) runs on Google-certified hardware with a live connection back to Google Cloud, and it’s priced transparently: $35 per vCPU per month, with a 96-vCPU minimum per site and a mandatory 36- or 60-month term commitment. GDC Hosted and the GDC air-gapped appliance move to fully quote-based pricing, since those tiers are built for government and regulated customers whose hardware and support terms get negotiated individually.
What sets GDC apart in 2026 is documentation. Google publishes exact GPU-equipped machine types for GDC Hosted, including the a3-highgpu-1g-gdc (28 vCPUs, 240GB memory, 1x NVIDIA H100 94GB), a3-highgpu-2g-gdc (56 vCPUs, 480GB, 2x H100), and a3-highgpu-4g-gdc (112 vCPUs, 960GB, 4x H100), all with Multi-Instance GPU (MIG) profiles enabled. GDC Sandbox adds an AI Optimized SKU bundling four NVIDIA H100 80GB HBM3 GPUs. Google’s Google Cloud NVIDIA hardware page goes further, stating that the NVIDIA Blackwell platform on Google Distributed Cloud enables secure, on-premises deployment of advanced agentic AI, including Gemini models — language neither AWS nor Microsoft currently publishes about their own hybrid platforms.
GDC’s air-gapped appliance is also the only product of the three with a named U.S. government security authorization: it achieved DoD Impact Level 6 (IL6) certification, the tier required for classified workloads up to Secret. That single fact explains most of GDC’s public-sector traction and separates it structurally from Outposts and Azure Local, neither of which markets an equivalent air-gapped, IL6-grade appliance today.
Pricing Comparison: Contract Terms Side by Side
| Platform | Published rate | Minimum commitment | Estimated 3-year cost (comparable mid-size deployment) |
|---|---|---|---|
| AWS Outposts (rack) | Bundled hardware + capacity, fixed term | 3 years, no early exit | $150,000 – $1,050,000 (mid-range rack, upfront option varies) |
| AWS Outposts (1U/2U server) | $400 – $1,500/month | 1 year (1U) / 3 years (2U) | $14,400 – $54,000 |
| Azure Local | $10/physical core/month (or $0 with Hybrid Benefit) | None — 60-day trial, cancel any time | Hardware cost (separate) + $0 – ~$23,000 software fee for a 64-core cluster |
| Google Distributed Cloud (connected/Edge) | $35/vCPU/month, 96 vCPU minimum | 36 or 60 months mandatory | $120,960 (36-month) – $201,600 (60-month) for the minimum footprint |
| Google Distributed Cloud (air-gapped/Hosted) | Custom quote | Multi-year, negotiated | Varies — government/enterprise quote only |
The pattern that jumps out: Azure Local is the only one of the three with zero mandatory lock-in on the software side, because Microsoft is only selling a subscription layered onto hardware the customer already owns or leases separately. AWS Outposts and Google Distributed Cloud’s connected tier both demand multi-year commitments, but Outposts front-loads a much larger fixed cost because AWS is bundling and maintaining the physical hardware inside that price. GDC’s connected tier looks cheaper at the entry point ($35/vCPU) but the 96-vCPU floor and 36-to-60-month term mean a small pilot still obligates a buyer to five- or six-figure spend before a single workload ships.
Benchmark and Capability Data From Multiple Sources
Because hybrid/on-prem cloud appliances aren’t benchmarked the same way as public-cloud compute, “benchmarks” here mean documented capability and cost data pulled from vendor pricing pages, product documentation, and trade press rather than synthetic performance scores:
- CRN market-share tracking reported that AWS Outposts mindshare fell to 8.3% in August 2026, down from 15.8% a year earlier, while Azure Stack mindshare dropped to 12.0% from 24.6% — both signals of buyers shifting evaluation attention toward newer or differently branded distributed-cloud options.
- AWS’s own documentation confirms second-generation Outposts racks add accelerated networking instances designed for ultra-low latency and high-throughput workloads, and CRN reported in 2026 that AWS is even extending Outposts to orbital data centers through a partnership with startup Starcloud, running AI models on hardware in space.
- Google’s GPU documentation is the most granular of the three vendors, publishing exact vCPU, memory, and GPU-count specs down to the MIG profile level for its a3-highgpu-gdc machine family — a level of transparency AWS and Microsoft don’t match for their own hybrid GPU offerings.
- RFP.wiki’s independent evaluation of GDC Edge pricing rated it 3.6 out of 5 on cost, citing the $35/vCPU rate, 96-vCPU floor, and mandatory 36-/60-month terms as the main friction points for smaller buyers.
- Microsoft’s release notes across the February, April, and August 2026 Azure Local updates repeatedly reference the ongoing Azure Stack HCI-to-Azure Local rename, indicating the transition is still actively surfacing in customer support tickets more than a year after the initial announcement.
Real-World Deployment Examples
Outpost VFX and AWS Outposts for AI training. AWS has documented how visual-effects studio Outpost VFX uses AWS infrastructure, including edge capacity, to accelerate AI model training for rendering pipelines, avoiding the latency of shuttling massive video assets to a distant region.
AWS Outposts in orbit. The Starcloud partnership reported by CRN in 2026 pushes Outposts hardware into a genuinely extreme edge case: running AI inference workloads on satellites, using the same Outposts management model AWS sells for terrestrial retail and telco sites.
U.S. Air Force tactical edge exercise on GDC air-gapped. GDIT’s press materials describe a field exercise where Google’s air-gapped appliance, described internally as “cloud-in-a-box,” ran seven mission-critical workloads for Air Force teams at the tactical edge, including language translation, optical character recognition, and retrieval-augmented generation for AI-assisted analysis, entirely disconnected from the public internet.
Singapore government agencies on GDC air-gapped. GovTech, HTX, and CSIT in Singapore are deploying agentic AI built on Google Gemini models through GDC air-gapped, allowing classified and regulated public-sector workloads to use frontier AI models while keeping all data and inference on sovereign infrastructure.
U.S. intelligence community on GDC Hosted. Google’s public-sector documentation describes GDC Hosted running Top Secret and Secret workloads for defense and intelligence agencies, giving those agencies ML and analytics capability while keeping the entire environment under their own operational control.
Which Platform Fits Which Use Case
Already deep in the AWS ecosystem, need a local extension: AWS Outposts is the obvious pick if a team already runs EC2, RDS, and S3 in a public AWS region and just needs the same APIs available on-prem for latency or residency reasons. The tradeoff is a fixed 3-year contract and hardware AWS fully controls.
Running Windows Server/Hyper-V today, want a lighter migration: Azure Local is the natural landing spot for shops already carrying Windows Server Datacenter licensing with Software Assurance, since Azure Hybrid Benefit can push the software fee to zero and the lack of a fixed hardware contract keeps the commitment flexible.
On-prem AI inference with confirmed GPU hardware: Google Distributed Cloud is currently the only one of the three publishing named H100 and Blackwell support for on-prem deployment, making it the default choice for teams whose core requirement is running LLM inference or training locally rather than in a shared region.
Classified or air-gapped government workloads: GDC’s air-gapped appliance with DoD IL6 authorization has no direct equivalent from AWS or Microsoft in this comparison, making it the practical default for defense and intelligence customers that need a fully disconnected environment.
Retail or branch-office edge with small footprint: AWS Outposts servers (1U/2U) undercut a full Outposts rack and Azure Local’s minimum cluster sizing for very small sites like individual stores, at $400-$1,500 a month rather than a six-figure rack commitment.
Telco 5G core and low-latency networking: AWS markets Outposts racks heavily for telco central-office deployments needing consistent AWS networking primitives, and second-generation racks specifically add accelerated networking instances built for that workload pattern.
Multi-site retail or manufacturing needing Kubernetes-native infra: GDC’s Kubernetes-first architecture, built on the same control-plane lineage as GKE, makes it a strong fit for organizations standardizing container orchestration across dozens or hundreds of small sites rather than running full VM-based infrastructure at each one.
Networking, Data Transfer, and Connectivity Requirements
None of these platforms works as a truly standalone box you can plug in and forget. AWS Outposts racks need a dedicated link back to a parent AWS Region, either through AWS Direct Connect or a VPN, and second-generation racks bump the available uplink speeds to 10, 40, or 100 Gbps depending on the compute-rack or network-rack configuration ordered. That link matters because Outposts still depends on the parent Region for control-plane operations like billing, monitoring, and certain management APIs, even though data-plane traffic for EC2 and EBS stays local. Losing that connection for an extended period can affect a site’s ability to launch new instances or apply certain configuration changes, even if already-running workloads keep serving traffic.
Azure Local takes a similar but not identical approach. Because it’s fundamentally a hyperconverged cluster running customer-owned hardware, local workloads can, in principle, keep running through a connectivity outage, but the daily core-based billing model depends on the cluster staying registered with Azure Arc. Extended disconnection risks billing and support complications rather than an outright workload stop, which is a meaningfully different failure mode than the Outposts dependency on its parent Region. Azure Local clusters typically run on 10-100 Gbps networking depending on the OEM hardware chosen, since Microsoft doesn’t dictate the physical network layer the way AWS does with its own racks.
Google Distributed Cloud is the outlier by design. The connected/Edge tier behaves similarly to Outposts, needing an active link back to Google Cloud for full functionality, while the air-gapped appliance is explicitly built to operate with zero external connectivity at all — that’s the entire point of the DoD IL6 authorization. GDC’s air-gapped nodes sync and update through manually managed processes rather than a live internet link, which is also why Google prices that tier entirely through custom quotes instead of a public per-vCPU rate: connectivity, update cadence, and support logistics vary enormously between a classified government facility and a commercial retail chain.
Security, Compliance, and Data Sovereignty Across the Three Platforms
Data residency requirements are the single biggest driver pushing enterprises toward any of these three platforms in 2026, rather than simply running everything in a public region. Financial services firms facing country-specific data-localization rules, healthcare organizations bound by regional privacy law, and manufacturers protecting proprietary process data on the factory floor all have a version of the same problem: some workloads legally or contractually cannot leave a specific building or jurisdiction, even while the rest of the estate happily runs in AWS, Azure, or Google Cloud regions.
AWS Outposts and Azure Local both market general data-residency and edge-locality benefits — the pitch is “your data never leaves this rack” for Outposts, and “your data never leaves this cluster” for Azure Local — but neither currently publishes a named compliance authorization as specific as GDC’s DoD IL6 certification for its air-gapped appliance. Governance teams standardizing multi-account policy across regional and hybrid footprints at once often pair this decision with a broader look at AWS Control Tower, Azure Landing Zone, and GCP’s equivalent guardrail tooling, since landing-zone policy needs to extend to whichever distributed-cloud platform gets chosen here. That doesn’t mean Outposts or Azure Local are unsuitable for regulated workloads; both platforms inherit relevant SOC, ISO, and FedRAMP-adjacent certifications from their parent clouds’ broader compliance programs. It does mean that for the narrowest, highest-classification use cases — a Secret-level defense workload, for example — Google is currently the only vendor of the three with a named, public authorization matching that exact requirement.
Encryption and identity management follow each platform’s parent cloud model rather than introducing anything unique to the hybrid deployment. Outposts racks use the same IAM, KMS, and encryption-at-rest primitives as EC2 and S3 in a public Region. Azure Local integrates with Azure Arc for identity and policy enforcement, meaning a security team managing Entra ID and Azure Policy for cloud resources can extend largely the same controls to on-prem Azure Local clusters. GDC connected and Hosted tiers similarly extend Google Cloud IAM and encryption defaults, while the air-gapped appliance runs its own local identity and key-management stack precisely because it has no live connection back to the parent Google Cloud control plane to rely on.
Migration Guide: Moving Between Distributed Cloud Platforms
Migrating an existing on-prem cloud footprint from one vendor to another is a multi-month project, not a weekend cutover. Here’s the realistic sequence teams follow when moving off Azure Stack HCI/Azure Local, AWS Outposts, or a legacy on-prem stack toward one of these three platforms.
- Inventory workload dependencies. Catalog every API call, IAM policy, and storage class each workload currently uses. Outposts workloads tied tightly to specific EC2 instance families or S3 on Outposts storage classes need explicit remapping if moving to Azure Local or GDC.
- Confirm hardware lead times before committing. AWS Outposts racks and GDC connected hardware both ship on vendor timelines that have run 8-16 weeks in recent procurement cycles; Azure Local buyers sourcing OEM hardware separately should validate delivery windows with Dell, HPE, or Lenovo directly.
- Right-size the new commitment before signing. Because GDC’s connected tier has a hard 96-vCPU floor and a 36-/60-month minimum term, and Outposts racks lock in three years of capacity, undersizing or oversizing at this step is expensive to unwind later. Azure Local’s core-based subscription is the only one of the three that allows shrinking spend without penalty. Running the projected spend through a dedicated FinOps tool before signing — the way teams already do when comparing CloudZero, Vantage, and Kubecost for public-cloud cost tracking — helps catch an undersized commitment before it becomes a multi-year mistake.
- Stand up the new environment in parallel, not in place. Deploy the target platform (Outposts rack, Azure Local cluster, or GDC node) alongside the existing environment rather than replacing it directly, so workloads can be validated against the new control plane before cutover.
- Migrate data first, compute second. Move object storage (S3 on Outposts, Azure Local’s Storage Spaces Direct, or GDC’s local object storage) and validate integrity before touching compute, since storage migrations are typically the longest-running and hardest-to-rollback step.
- Re-point IAM and networking. Rebuild identity federation and VPN/Direct Connect-equivalent links (AWS Direct Connect, Azure ExpressRoute, or Google’s Dedicated Interconnect) for the new on-prem environment before cutting production traffic over.
- Migrate GPU/AI workloads last. If the migration target is specifically to gain on-prem H100 or Blackwell access via GDC, sequence AI/ML workloads as the final wave, after core infrastructure and storage are already validated on the new platform.
- Run both environments in parallel for at least one full billing cycle. Given the fixed-term contracts on Outposts and GDC connected, keep the legacy environment live through one full cycle to catch workload dependencies that don’t surface until end-of-month batch jobs or reporting runs.
- Decommission and reclaim hardware. AWS handles rack removal for Outposts as part of the original contract; Azure Local customers repurpose or resell OEM hardware directly; GDC connected hardware returns follow the original procurement agreement.
Pros and Cons of Each Platform
AWS Outposts
Pros: Deepest catalog of regional AWS services available locally; fully managed hardware with AWS handling delivery, patching, and removal; strong telco and 5G core positioning with second-generation networking; extends as far as orbital deployments through the Starcloud partnership.
Cons: Fixed 3-year hardware contract with no early exit; no customer choice over physical hardware; GPU support currently capped at older G4dn instances with no confirmed H100/Blackwell path; highest up-front cost of the three for a comparable mid-size deployment.
Azure Local (Azure Stack HCI)
Pros: No fixed multi-year software commitment; cancel by deregistering the cluster; can reach $0/core with Azure Hybrid Benefit for existing Windows Server Datacenter customers; customer chooses and owns the hardware, avoiding vendor lock on physical infrastructure; natural fit for Hyper-V/VMware migration paths.
Cons: No named GPU model support published for AI workloads, unlike GDC’s explicit H100/Blackwell documentation; hardware procurement, delivery, and support fall on the customer and their OEM rather than the cloud vendor; the recent rebrand from Azure Stack HCI still causes confusion in search and documentation as of 2026.
Google Distributed Cloud
Pros: Only platform of the three with publicly documented on-prem H100 and Blackwell GPU support; only platform with a DoD IL6-authorized air-gapped appliance; Kubernetes-native architecture suits multi-site container deployments; transparent per-vCPU list pricing for the connected tier.
Cons: Mandatory 36- or 60-month term on the connected tier with no shorter option; 96-vCPU minimum per site makes small pilots expensive to right-size; air-gapped and Hosted tiers are entirely quote-based, hiding real-world pricing until deep into a sales cycle; smaller overall regional service catalog compared to AWS’s Outposts ecosystem.
The Verdict: Which One Should You Choose?
There’s no single winner here because the three products are solving different procurement problems, but the data points to a clear decision tree. If the requirement is on-premises GPU inference or training with vendor-confirmed hardware, Google Distributed Cloud is the only platform among the three that currently documents H100 and Blackwell support down to the machine-type level — that alone should end the debate for AI-focused buyers, even with the 36-to-60-month commitment attached. If the requirement is minimizing financial lock-in and a team already holds Windows Server Datacenter licensing, Azure Local’s $0-to-$10-per-core model with no fixed term is the most forgiving commercial structure of the three, though buyers give up the vendor-managed hardware convenience of Outposts and GDC. If the requirement is the deepest possible catalog of regional cloud services running locally, and the buyer can absorb a fixed 3-year, six-figure hardware contract, AWS Outposts remains the most complete single-vendor extension of a public cloud into a private facility.
The broader signal from CRN’s 2026 mindshare data — Outposts down to 8.3% from 15.8%, Azure Stack down to 12.0% from 24.6% — suggests all three vendors are fighting for a shrinking pool of buyers willing to commit to any distributed-cloud hardware contract at all, as more workloads either stay fully in public regions or move toward lighter edge-Kubernetes patterns that don’t require a branded appliance. For 2026 buyers, the practical move is to map the GPU and contract-term requirements first, then pick the vendor whose commercial terms match the actual budget cycle, rather than defaulting to whichever public cloud already runs the rest of the estate. For a broader view of how these hybrid products fit into the rest of the cloud market, see our 2026 cloud computing coverage hub.
Support, Patching, and Lifecycle Management
Ongoing operations are where the three products diverge most sharply from a day-two staffing perspective. AWS Outposts is the most hands-off of the three: the rack price already includes delivery, installation, infrastructure maintenance, software patches, and eventual removal, so a customer’s operations team manages workloads the same way they’d manage EC2 and S3 in a Region, without owning hardware lifecycle tasks like firmware updates or component replacement. That convenience is baked into the higher fixed price.
Azure Local pushes more lifecycle responsibility onto the customer and their OEM. Firmware updates, hardware failures, and physical maintenance route through whichever partner sold the validated nodes (Dell, HPE, or Lenovo, among others), while Microsoft handles the software side through Azure Local’s update channel, evidenced by the regular monthly and quarterly release notes such as the August 2026 build. Teams choosing Azure Local should budget for an internal or outsourced hardware-support contract on top of the Microsoft software subscription, since the two are billed and supported separately.
Google splits lifecycle support by tier. GDC connected/Edge nodes run on Google-certified partner hardware with support arrangements similar to Azure Local’s OEM model, while GDC Hosted and the air-gapped appliance ship as more turnkey systems with support terms negotiated as part of the custom quote, closer to the Outposts model of vendor-managed lifecycle. Because air-gapped appliances have no live connection back to Google Cloud, patch and update delivery for that tier depends on scheduled manual processes rather than an always-on update channel, which is a meaningful operational planning difference for teams used to cloud-native patch cadences.
Frequently Asked Questions
Is Azure Stack HCI the same thing as Azure Local?
Yes. Microsoft renamed Azure Stack HCI to Azure Local, with the transition rolling out across product and pricing pages starting November 19, 2024. Azure Stack HCI clusters are now called Azure Local instances, and Azure Stack HCI servers are now Azure Local machines. The underlying product and per-core pricing model did not change, only the branding.
Can AWS Outposts run NVIDIA H100 GPUs on-premises?
AWS’s current public documentation does not confirm H100, H200, or Blackwell GPU availability on Outposts hardware. Outposts racks currently support GPU-optimized instances built on the older G4dn (NVIDIA T4) family, with AWS stating that support for newer GPU-based instances is coming soon for second-generation racks.
How much does a Google Distributed Cloud connected deployment cost at minimum?
GDC connected/Edge pricing starts at $35 per vCPU per month with a 96-vCPU minimum per site, which works out to at least $3,360 per month before any hardware or support add-ons. Customers must also commit to either a 36-month or 60-month term, meaning the smallest possible connected deployment obligates roughly $120,960 to $201,600 in total commitment.
Does Azure Local require a long-term contract?
No. Azure Local bills daily based on physical cores registered to the cluster, with a 60-day free trial and no fixed multi-year commitment. Customers can reduce spend by deregistering cores or the entire cluster at any time, which is structurally different from the fixed 3-year Outposts contract or the mandatory 36-/60-month GDC connected term.
Which platform is authorized for classified U.S. government workloads?
Google Distributed Cloud’s air-gapped appliance holds DoD Impact Level 6 (IL6) authorization, the level required for classified workloads up to Secret. Neither AWS Outposts nor Azure Local currently markets an equivalent IL6-authorized, fully air-gapped appliance in this comparison.
Can I use Azure Hybrid Benefit to reduce Azure Local costs?
Yes. Customers with Windows Server Datacenter licenses under active Software Assurance can apply Azure Hybrid Benefit to waive the Azure Local host service fee entirely, dropping the per-core software cost from roughly $10 per physical core per month to $0.
What is the difference between GDC connected, GDC Hosted, and GDC air-gapped?
GDC connected (Edge) runs on Google-certified hardware with a live link back to Google Cloud and uses transparent per-vCPU pricing. GDC Hosted and the GDC air-gapped appliance are built for government and highly regulated customers, use quote-based pricing, and the air-gapped tier can run entirely disconnected from the public internet, which is why it holds DoD IL6 authorization.
Why did AWS Outposts and Azure Stack mindshare drop in 2026?
CRN’s 2026 tracking showed AWS Outposts mindshare falling to 8.3% (from 15.8% a year earlier) and Azure Stack mindshare falling to 12.0% (from 24.6%). The decline suggests buyers are increasingly weighing lighter, Kubernetes-native edge options or staying in public regions altogether rather than committing to branded, multi-year distributed-cloud hardware contracts.
Related Coverage
- AWS Cost Explorer vs Azure vs Google Billing: $0.01 Gap [2026]
- AWS Lambda vs Azure Functions vs GCP: 4x Timeout Gap [2026]
- AWS Systems Manager Azure Setup: 12 Steps, 90 Min [2026]
- AWS Bedrock AgentCore vs Azure AI Foundry vs Vertex AI: 15 Regions, 1 Deadline [2026]
- AWS Lambda MicroVMs Setup: 12 Steps, 4x Peak Scaling [2026]
