| Myth | Reality |
|---|---|
| Bare-metal and neocloud providers are only for tiny startups that can’t afford AWS. | CoreWeave runs production inference for major AI labs and Mistral; OVHcloud hosts regulated European enterprise workloads at multi-petabyte scale. |
| GPU neoclouds always have better GPU availability than hyperscalers. | Availability varies by generation and region; H100/H200 capacity has been tight everywhere, and neoclouds can sell out of a specific SKU just as fast as AWS does. |
| Bare metal means no automation — you’re back to racking servers by hand. | Modern bare-metal platforms provision via API, Terraform, and Ansible in minutes, with IPMI/BMC-level remote control replacing physical data-center visits. |
| Switching to a neocloud means giving up managed databases, IAM, and compliance tooling entirely. | Most teams run a hybrid model: neocloud or bare metal for compute-heavy workloads, hyperscaler for managed data services, identity, and edge distribution. |
| Price lists are apples-to-apples across providers. | On-demand list price ignores egress, support tiers, minimum commitments, and interconnect fees — the effective cost per useful GPU-hour can differ by 40% or more from the sticker price. |
Why This Comparison Matters in Late 2026
Two forces are reshaping the compute market at once. First, GPU scarcity for AI training and inference pushed enterprises to look past the big three hyperscalers toward specialist “neoclouds” that exist almost entirely to sell GPU-hours. Second, the bare-metal market itself just went through a shakeup: Equinix announced it would sunset its Equinix Metal bare-metal-as-a-service platform, with the service discontinued for new sales and existing customers pushed toward migration by June 30, 2026, console access continuing only through the end of the year for account owners. That single event sent a wave of workloads toward OVHcloud, Latitude.sh, and other bare-metal operators, and it’s a useful case study in why relying on a single non-hyperscaler vendor carries its own risk.
Meanwhile, AWS, Azure, and Google Cloud remain the default choice for most enterprise IT simply because nobody gets fired for picking them — a pattern this site has covered in how cloud computing is reshaping IT operations. But “default” and “optimal” have never been the same thing, and for GPU-heavy or infrastructure-heavy workloads the gap between the two has widened, a shift already visible in where cloud computing is headed over the next few years.
What “Bare Metal” and “Neocloud” Actually Mean
Bare metal refers to a dedicated physical server rented from a provider with no hypervisor between you and the hardware. You get the full CPU, full RAM, full local NVMe, and full network interface — not a virtualized slice of it. Neoclouds are a newer category: cloud providers built specifically around GPU rental for AI training and inference, frequently backed by NVIDIA itself as an investor or reference-design partner, and frequently running on some mix of bare-metal and lightly virtualized GPU nodes.
The two categories overlap in practice. CoreWeave, Lambda, and Crusoe are neoclouds; OVHcloud and Latitude.sh are bare-metal generalists that also sell GPU servers. A hyperscaler like AWS sells virtualized instances by default but also offers bare-metal instance types (e.g., its `.metal` EC2 family) for the minority of workloads that need them.
The Provider Landscape
Bare-Metal Generalists: OVHcloud and Latitude.sh
OVHcloud operates one of the largest bare-metal footprints in the world, with more than 30 data centers spread across multiple continents. Its pitch is flat-rate pricing, free bandwidth on many plans, built-in DDoS protection, and hourly or monthly billing with API-driven provisioning compatible with OpenStack, Terraform, and Ansible. It has positioned itself directly as the landing spot for workloads leaving Equinix Metal, publishing migration guidance aimed at exactly that audience.
Latitude.sh takes a similar API-first approach: pre-defined hardware profiles, provisioning entirely through API or CLI, a Terraform provider, bandwidth pooling across an account, and both hourly and monthly pricing with no hidden surcharges. Both companies compete less on being cheap and more on being predictable — a flat invoice instead of a line-item bill full of surprise networking and support charges.
GPU-Focused Neoclouds: CoreWeave, Lambda, Crusoe
CoreWeave is the largest specialist neocloud by scale, was NVIDIA’s first Elite cloud services provider, and claims tens of thousands of GPUs across its fleet with a catalog spanning ten-plus GPU model families including current-generation B200 and B300 parts. Its pricing sits at a premium versus smaller neoclouds — reported at roughly $6.31/hour for a single H200, meaningfully above Crusoe and Lambda list rates for comparable hardware.
Lambda is GPU-only by design, serves a large base of research teams, and ships instances pre-loaded with PyTorch, TensorFlow, CUDA drivers, and a notebook environment — optimized for researchers who want to start training within minutes of provisioning rather than build a stack from scratch.
Crusoe has carved out a position on price, with AMD MI300X pricing reported from around $1.71/hour, among the lowest rates from a named provider of that GPU class, partly because Crusoe’s origin story is built around co-locating compute with stranded or otherwise wasted energy sources.
| Provider | Type | Reference price point | Billing granularity |
|---|---|---|---|
| AWS / Azure / GCP | Hyperscaler, virtualized GPU instances | Premium list price; heavy discounts via committed-use contracts | Per-second/per-hour, complex SKU matrix |
| CoreWeave | Neocloud, GPU-as-a-Service | ~$6.31/hr single H200 (on-demand) | Hourly, on-demand and spot tiers |
| Lambda | Neocloud, GPU-only | Below CoreWeave list rate for comparable H100/H200 SKUs | Hourly, reserved options |
| Crusoe Cloud | Neocloud, GPU + energy-co-located | ~$1.71/hr AMD MI300X | Hourly, reserved and on-demand |
| OVHcloud | Bare metal generalist | Flat-rate, often bandwidth-inclusive | Hourly or monthly |
| Latitude.sh | Bare metal generalist | Predefined hardware profiles | Hourly or monthly, pooled bandwidth |
Pricing: Sticker Price Is Not the Whole Story
List price comparisons flatter neoclouds and bare-metal providers because the hyperscalers’ most punishing costs live outside the compute line item. Standard internet egress runs roughly $0.09/GB on AWS, $0.087/GB on Azure, and $0.12/GB on GCP, with only a small monthly free allowance. Add NAT gateway processing (around $0.045/GB on AWS), cross-AZ transfer, load balancer data processing fees, and per-hour charges for VPC endpoints and public IPv4 addresses, and networking alone can function as a hidden tax approaching high single digits to double-digit percentage points of a total cloud bill. Bare-metal and neocloud providers typically fold bandwidth into the base price or charge a simple flat per-GB rate with a much larger free tier, which is one of the more underrated reasons total cost of ownership diverges from the quoted GPU-hour rate.
The counterpoint: hyperscaler list prices are a starting point, not an ending point. Enterprises spending over $5 million a month on cloud (a group Flexera’s 2026 State of the Cloud research puts at 76% of large enterprises) routinely negotiate committed-use discounts of 30-60% off list. A neocloud’s flat $6/hour on-demand rate can lose to a heavily discounted hyperscaler reserved instance once volume commitments are on the table — pricing comparisons only mean something when both sides reflect what you’d actually pay, not the public rate card. This mirrors a broader shift already visible in cloud spending and adoption data, where hybrid, multi-provider setups have become the norm rather than the exception.
Figure 1: Effective Cost Stack by Provider Category
A stacked bar comparison across four cost components — base compute, networking/egress, support and managed services, and idle/overprovisioning waste — for hyperscaler on-demand, hyperscaler committed-use, neocloud on-demand, and bare-metal flat-rate. Hyperscaler on-demand shows the tallest “networking” and “idle waste” segments; neocloud and bare-metal show taller “compute” segments but near-zero networking segments and a visible “support gap” segment representing work your own team must absorb instead of the vendor.
Provisioning Speed: Minutes vs. Days vs. Weeks
Hyperscaler virtual instances provision in seconds to a couple of minutes for standard SKUs, but GPU-heavy instance types are frequently capacity-constrained, meaning the API call succeeds instantly while the actual resource request queues or fails silently behind availability limits — the request completing fast doesn’t mean the GPU shows up fast.
Neoclouds built for GPU workloads generally provision dedicated GPU nodes in minutes once capacity exists, because that is the entire business — CoreWeave and Lambda both emphasize rapid cluster stand-up as a core differentiator versus hyperscalers, since a research team waiting days for an H100 node is a churned customer. Traditional bare-metal generalists (OVHcloud, Latitude.sh) sit in between: API-driven provisioning against pre-built inventory typically completes in minutes to under an hour, but a fully custom hardware configuration not held in ready stock can take days to weeks, the same constraint that existed under the old physical-colocation model, just compressed.
- Step 1Define the workload’s real constraint — is it GPU-bound, network-bound (distributed training needing InfiniBand/RDMA), storage-bound, or genuinely elastic and bursty?
- Step 2Price the top 2-3 candidate providers using realistic utilization, not peak — include egress, support tier, and any committed-use discount you could actually negotiate on the hyperscaler side.
- Step 3Check capacity reality, not the brochure — ask each vendor for current lead time on the exact GPU SKU and region you need, in writing.
- Step 4Map what managed services you’d lose — IAM, managed databases, WAF, CDN — and price the cost of self-hosting or hybrid-linking those back to a hyperscaler.
- Step 5Pilot on a non-critical workload for 30-60 days before moving anything customer-facing or revenue-critical.
- Step 6Decide with an exit plan already written — know how you’d migrate off this provider before you migrate onto it.
GPU Availability: Where the Scarcity Actually Sits
GPU scarcity has not been evenly distributed. Frontier-generation parts (current Blackwell-class B200/B300 SKUs) have been supply-constrained across every provider, hyperscaler and neocloud alike, largely gated by NVIDIA allocation rather than any single vendor’s operational choices. Where neoclouds have pulled ahead is in previous-generation and specialty inventory — H100, H200, and AMD MI300X capacity — because their entire fleet strategy is built around holding and reselling exactly that inventory at scale, rather than spreading capital across dozens of instance families the way a hyperscaler must. Crusoe’s price advantage on MI300X is a direct artifact of this: capacity dedicated to a narrower set of chips means fewer idle SKUs eating margin, which shows up as lower rates for the workloads that fit.
Practically, this means checking real-time regional availability matters more than checking a provider’s overall reputation for “having GPUs.” A team that needs a cluster of H100s for a six-week fine-tuning run should request a capacity-backed quote, not just an on-demand price, from at least two providers before committing.
Figure 2: GPU Provisioning Timeline by Provider Category
A horizontal timeline comparing time-to-first-training-run across four paths: hyperscaler on-demand (API succeeds in seconds, but capacity queue can add hours to days for scarce SKUs), hyperscaler reserved capacity (multi-week advance booking, then instant availability on the reserved date), neocloud on-demand (minutes to an hour when capacity exists), and bare-metal custom build (days to weeks for non-stocked configurations). The neocloud bar is shortest for pre-stocked inventory; the bare-metal custom-build bar is longest.
Managed Tooling: What You Give Up, and When That’s Fine
The honest case against bare metal and neoclouds is the managed-services gap. Hyperscalers bundle IAM, managed relational and NoSQL databases, global CDN, WAF, serverless compute, and dozens of compliance certifications that took years to accumulate. A neocloud or bare-metal provider gives you raw compute and, at most, basic networking, storage, and orchestration hooks — everything above that, you build or buy separately, a cost this site’s FinOps coverage of autonomous workload costs treats as a first-class line item, not an afterthought. Smaller teams without dedicated platform engineering — the profile covered in this site’s rundown of cloud services suited to small businesses — are usually better served staying fully managed rather than adopting bare metal at all.
The trade-off is worth accepting when three conditions hold simultaneously: the workload is genuinely compute-bound (training, batch inference, rendering, simulation) rather than needing a rich managed-services ecosystem; your team already has the platform engineering capacity to run its own Kubernetes, storage, and observability layer, a pattern explored in this site’s look at hybrid compute strategies for inference scaling; and the cost or capacity advantage is large enough to justify the operational lift — generally when you’re spending enough on GPU-hours that a 30%+ savings meaningfully moves your budget.
Common Mistakes
- Comparing on-demand neocloud rates against non-discounted hyperscaler list prices, ignoring that most serious hyperscaler customers negotiate 30-60% off list.
- Signing a long-term bare-metal or neocloud contract without a written exit and data-portability plan — the Equinix Metal shutdown proved single-vendor bare-metal bets can evaporate with a year’s notice.
- Underestimating the platform engineering headcount needed to replace IAM, secrets management, and observability that a hyperscaler provided for free.
- Treating “bare metal” as a synonym for “cheap” without pricing in support-tier costs, which can be substantial for 24/7 production SLAs.
- Chasing the lowest advertised GPU-hour rate without verifying actual regional capacity and interconnect quality (InfiniBand vs. standard Ethernet) for distributed training jobs.
In Practice
Teams that get this right typically run a deliberately split architecture: customer-facing APIs, managed databases, and compliance-sensitive data stay on a hyperscaler where certifications and managed IAM already exist, while GPU-heavy training and batch inference move to a neocloud or bare-metal provider connected back via a dedicated interconnect or VPN rather than public internet egress. This keeps the egress bill low, keeps the compliance boundary clean, and lets the compute-heavy piece be re-priced or re-homed independently if a provider’s pricing or capacity changes — exactly the kind of workload-specific placement that shows up in mature multi-cloud architecture practice. The pilot-then-scale pattern (30-60 days on a non-critical workload before moving production jobs) shows up repeatedly in successful migrations and rarely in failed ones.
Glossary
- Bare metal
- A dedicated physical server rented without a hypervisor layer, giving the tenant direct, exclusive access to the underlying hardware.
- Neocloud
- A cloud provider built specifically around renting GPU capacity for AI training and inference, as opposed to general-purpose hyperscaler infrastructure.
- Hyperscaler
- A large-scale general-purpose cloud provider — AWS, Microsoft Azure, and Google Cloud — offering broad managed-service catalogs across compute, storage, networking, and AI.
- GPU-as-a-Service (GPUaaS)
- A consumption model where customers rent GPU compute by the hour or minute rather than purchasing and operating physical GPU hardware.
- Egress fee
- A per-gigabyte charge for data leaving a cloud provider’s network, typically the largest hidden cost difference between hyperscalers and bare-metal/neocloud providers.
- Committed-use discount
- A negotiated price reduction in exchange for a customer’s contractual commitment to a minimum spend or usage level over a fixed term.
- InfiniBand
- A high-throughput, low-latency network interconnect used to link GPUs across nodes for distributed training, common in neocloud clusters built for large model training.
- Spot instance
- Discounted, interruptible compute capacity sold from a provider’s unused inventory, offered by both hyperscalers and neoclouds at a fraction of on-demand price.
Key Takeaways
- Bare-metal and neocloud providers generally beat hyperscalers on price-per-GPU-hour and provisioning speed for dedicated hardware, but the comparison only holds when you price in egress and support costs on both sides.
- CoreWeave, Lambda, and Crusoe compete primarily on GPU access, pre-built ML tooling, and capacity for specific chip generations rather than breadth of managed services.
- OVHcloud and Latitude.sh compete on predictable flat pricing and API-driven provisioning as bare-metal generalists, and both absorbed demand after the Equinix Metal shutdown.
- Hyperscaler committed-use discounts of 30-60% off list price change the math significantly — never compare a neocloud’s on-demand rate to a hyperscaler’s undiscounted list price.
- The Equinix Metal sunset (completed June 30, 2026) is a live reminder that single-vendor bare-metal bets carry vendor-continuity risk that hyperscalers are less exposed to.
- The right trade-off pattern for most teams is hybrid: managed services and compliance-sensitive data on a hyperscaler, GPU-heavy compute on a neocloud or bare-metal provider, connected via dedicated interconnect rather than public internet.
- Always pilot on a non-critical workload for 30-60 days and have a written exit plan before committing production traffic to any single non-hyperscaler provider.
FAQs
Is CoreWeave cheaper than AWS for GPU workloads?
Often yes on list price for on-demand GPU-hours, but AWS customers with committed-use agreements can close or exceed that gap. CoreWeave’s advantage is strongest for teams that need on-demand access without long-term negotiation, or need current-generation GPU SKUs that hyperscalers have allocated conservatively.
What happened to Equinix Metal?
Equinix announced it would stop selling its Equinix Metal bare-metal platform and sunset the service by June 30, 2026, citing that it represented a small fraction of overall company revenue. Existing customers were directed toward migration partners including OVHcloud and Latitude.sh, with console access continuing through the end of 2026 for account owners.
Can I run compliance-regulated workloads on a neocloud?
Some neoclouds hold relevant certifications, but coverage is far narrower than hyperscalers’ decade-plus compliance catalogs. Most regulated enterprises keep compliance-sensitive data and identity management on a hyperscaler and route only GPU-heavy, non-regulated compute to a neocloud, connected through a private interconnect.
Do bare-metal providers support Kubernetes and Terraform?
Yes. OVHcloud and Latitude.sh both provide Terraform providers, CLI tooling, and API-first provisioning, and both integrate with standard Kubernetes distributions. The gap versus hyperscalers is depth of managed add-ons (managed control planes, integrated logging, autoscaling policies), not basic infrastructure-as-code support.
How much do egress fees really add to a multi-provider setup?
Standard internet egress runs about $0.09/GB on AWS and $0.087/GB on Azure, with GCP around $0.12/GB, and hidden networking charges like NAT gateway processing can push total networking costs toward a meaningful double-digit share of a cloud bill for data-heavy, multi-cloud workloads.
Are neocloud GPUs actually available, or is it all waitlists?
Availability varies sharply by GPU generation. Frontier-class chips are tight everywhere due to allocation constraints upstream from any single vendor, but neoclouds generally hold deeper inventory of previous-generation and specialty chips (H100, H200, MI300X) than hyperscalers, since that inventory is their core business.
What’s the biggest hidden cost of switching away from a hyperscaler?
Rebuilding the managed-services layer — identity and access management, managed databases, observability, and compliance tooling — that a hyperscaler bundles by default. Teams that skip pricing this in are usually the ones who regret the move within a year.
Should a startup ever choose a neocloud over AWS for its first deployment?
If the startup’s core product is GPU-compute-bound (training, inference, rendering) and the team already has infrastructure engineering capacity, yes — the price and provisioning-speed advantages compound quickly. If the product is a typical web/SaaS stack needing broad managed services, a hyperscaler’s free-tier and managed catalog usually wins early on.
References
- Data Center Dynamics — “Equinix to kill off Metal by June 2026”
- SDxCentral — “Equinix officially retires bare metal offering”
- Flexera — 2026 State of the Cloud Report
- GMI Cloud — “CoreWeave, Lambda, Nebius, and GMI: H200 GPU Provider Pricing 2026”
- AWS — Amazon EC2 On-Demand Pricing
- OVHcloud — “Plan Your Exit from Equinix Bare Metal with OVHcloud”
- Latitude.sh — “Equinix Metal sunset: How to reduce the migration burden”
