Neocloud / AI Cloud Providers

Neocloud — 7 vendors and operators, each broken down by layer L1 to L7: what they sell at that layer, and which layer is their home turf.

CoreWeave

Home turf: L2 / L7

L1 Physical

This layer needs to state honestly what this domain actually owns versus leases. Like some cloud domains on this page, it typically leases facility space and power capacity from a facility partner rather than buying land and building itself — its real investment and differentiation sit in the server hardware configuration inside the racks and the inter-rack…

L2 Data Link

This is this domain's actual reason for existing, and its biggest difference from this page's other cloud domains: a large-scale GPU training job's performance bottleneck is usually not raw compute — it's the speed at which GPUs exchange gradient data with each other.

L3 Network

The core of this layer isn't conventional routing and forwarding (mechanics here don't fundamentally differ from this page's other cloud domains) — it's topology-aware scheduling: placing a training job's GPU nodes as physically close together on the network topology as possible (under the same switch, within the same aggregation layer), reducing the hop cou…

L4 Transport

This layer continues L2's discussion of RDMA's transport behaviour: InfiniBand natively supports credit-based flow control and is inherently loss-resistant; but on a high-speed-Ethernet-plus-RoCEv2 path, the transport layer instead depends on mechanisms like Explicit Congestion Notification (ECN) and Priority-based Flow Control (PFC) to approximate a lossles…

L5 Session

This layer is a standard cloud-platform access session — console and API authentication, SSH access to an assigned GPU instance, interactive Jupyter Notebook access — mechanics not fundamentally different from the account and access-session model on this page's other cloud domains; this domain hasn't built a distinctive differentiator at this layer.

L6 Presentation

This layer is explicitly empty — no product or service. This domain has no standalone encryption/presentation-layer product — baseline encryption capability at the storage and transport level is typically a built-in feature of the underlying storage and network infrastructure, not a separately sold product line of this domain's own.

L7 Application

This is this domain's second star layer: using Kubernetes rather than a traditional HPC scheduler (Slurm) to manage a GPU cluster, a relatively unusual choice in AI infrastructure — traditional high-performance computing and AI research circles are more familiar with job schedulers like Slurm, while Kubernetes is the container-orchestration platform the clou…

Lambda

Home turf: L2 / L5

L1 Physical

This layer needs to state a real difference from this page's CoreWeave domain up front: CoreWeave pivoted from cryptocurrency-mining hardware; this domain instead started as a deep-learning hardware company — founded in 2012, its original business was manufacturing and selling workstations and servers preloaded with deep-learning frameworks, and it only adde…

L2 Data Link

This is one of this domain's star layers: 1-Click Clusters — letting a user self-service spin up a multi-node GPU training cluster interconnected by InfiniBand in a single step, without going through the traditional customized sales-and-manual-deployment process.

L3 Network

This layer is standard cloud networking capability — assigning IPs to instances and clusters, private network segmentation isolating multi-tenant traffic, and managing network configuration through a public API/console — mechanically not fundamentally different from this page's other cloud/Neocloud domains, and this domain hasn't built a differentiated produ…

L4 Transport

This layer is explicitly empty — no product or service. This domain hasn't publicly disclosed a transport-layer tuning product independent of L2's InfiniBand interconnect — reliable-transport mechanics within a cluster are provided as part of the InfiniBand fabric itself, not a separately documented capability, stated honestly as empty rather than splitting…

L5 Session

This is this domain's second star layer, and the clearest fork in product positioning from this page's CoreWeave domain. CoreWeave focuses almost entirely on large enterprise training jobs, with its session layer being ordinary console/API/SSH access; this domain's other product line, extending from its earlier hardware business, is a retail single-GPU renta…

L6 Presentation

This layer is explicitly empty — no product or service. This domain has no standalone encryption/presentation-layer product; baseline encryption at the storage and transport level is a built-in capability of the underlying infrastructure, not a separately sold product line.

L7 Application

This layer is the control plane for this domain's cloud business: the Lambda Cloud console and API, used to manage single-GPU instances, 1-Click Clusters, and billing. Beyond its cloud products, this domain also maintains an on-premises reference-architecture product line for large enterprise customers (a continuation of the hardware business mentioned at L1…

Crusoe

Home turf: L1

L1 Physical

This is this domain's star layer, and its most substantive difference from this page's other Neocloud domains: rather than building cloud software first and adding hardware later, it started from energy engineering — founded in 2018 doing 'Digital Flare Mitigation,' capturing associated natural gas at oil-and-gas well sites that would otherwise be flared off…

L2 Data Link

Like this page's other Neocloud domains, large-scale distributed training requires extremely high bandwidth and extremely low latency within a GPU cluster's interconnect, and this domain provides InfiniBand-based inter-node interconnect for large GPU clusters, paired with GPUDirect RDMA to reduce latency in GPU-to-GPU memory communication.

L3 Network

Standard cloud networking capability — assigning IPs to instances and clusters, private networking isolating multi-tenant traffic — mechanically not fundamentally different from this page's other cloud/Neocloud domains, and this domain hasn't built a differentiated product at this layer.

L4 Transport

This layer is explicitly empty — no product or service. This domain hasn't publicly disclosed a transport-layer tuning product independent of L2's InfiniBand interconnect; reliable-transport mechanics are provided as part of the network fabric itself, not a separately documented capability.

L5 Session

A standard cloud-platform access session — console and API authentication, SSH access to an assigned GPU instance — mechanically not fundamentally different from this page's other cloud/Neocloud domains, and this domain hasn't built a distinctive differentiator at this layer.

L6 Presentation

L7 Application

This layer is Crusoe Cloud's standard cloud-management capability: managed, Kubernetes-based container orchestration, plus a management console/API for GPU instances and clusters. Similar to this page's CoreWeave domain in choosing Kubernetes over a traditional HPC scheduler, but this domain hasn't publicly disclosed a homegrown Slurm-compatibility layer lik…

Nebius

Home turf: L1 / L7

L1 Physical

This is this domain's star layer, and a structurally distinctive difference from this page's other Neocloud domains: it isn't a startup built from zero, but the international business entity spun off from the Russian technology company Yandex, refocused entirely on AI infrastructure after a 2024 corporate restructuring.

L2 Data Link

Like this page's other Neocloud domains, this domain provides high-bandwidth, low-latency inter-node interconnect based on Nvidia InfiniBand for large-scale GPU training clusters, paired with GPUDirect RDMA to reduce latency in GPU-to-GPU memory communication.

L3 Network

L4 Transport

L5 Session

L6 Presentation

L7 Application

This is this domain's second star layer: the Nebius AI Cloud platform, providing GPU instances, managed Kubernetes cluster management, and inference-deployment capability, with a product form broadly aligned with this page's other Neocloud domains' cloud platforms.

Nscale

Home turf: L1

L1 Physical

This layer is the relatively solid part of this domain's currently available public information: a UK-headquartered GPU cloud provider focused on AI training compute infrastructure, with public reporting showing it building data centres in Norway and other Nordic locations.

L2 Data Link

This layer is explicitly empty — no product or service. This domain has no reliable, public cluster-interconnect technical detail (such as a specific network topology or RDMA scheme) for this file to cite — this page's other Neocloud domains' InfiniBand/GPUDirect content at this layer is backed by relatively solid public information, which this domain curren…

L3 Network

This layer is explicitly empty — no product or service. Same reason as L2: no reliable public information supports specific network-layer product detail, stated honestly as empty.

L4 Transport

This layer is explicitly empty — no product or service. Same reason as L2 and L3: no reliable public information supports specific transport-layer product detail, stated honestly as empty.

L5 Session

This layer is explicitly empty — no product or service. This domain's cloud-platform control-plane public detail is limited; this file doesn't speculate about specific account/session mechanics without grounding, stated honestly as empty.

L6 Presentation

This layer is explicitly empty — no product or service. There's no public information about a standalone encryption/presentation-layer product for this domain, stated honestly as empty.

L7 Application

This layer is another relatively solid public fact about this domain: per public reporting, it has reached a GPU compute-supply partnership with Microsoft, providing supplemental capacity for Microsoft's AI compute needs — a partnership model that has become fairly common in recent years between several Neocloud companies (including a few already covered on…

Together AI

Home turf: L7

L1 Physical

This layer needs to state honestly a structural difference from this page's other Neocloud domains: there's no public information showing it owns self-built data centres at a comparable scale — its GPU cluster compute is obtained more through leasing facility space and multi-cloud partnerships, a different path from this page's CoreWeave, Crusoe, and other d…

L2 Data Link

This domain operates/leases GPU clusters for large-scale distributed training and batch inference, with public information showing some clusters use InfiniBand interconnect to support multi-node distributed training — this technical direction is broadly aligned with this page's CoreWeave, Lambda, and other domains, but this domain hasn't publicly disclosed i…

L3 Network

Standard cloud networking capability — assigning IPs to inference and training jobs, providing multi-tenant network isolation — mechanically not fundamentally different from this page's other cloud/Neocloud domains, and this domain hasn't built a differentiated product at this layer.

L4 Transport

This layer is explicitly empty — no product or service. This domain hasn't publicly disclosed a standalone transport-layer tuning product; reliable-transport mechanics are provided as part of the underlying network fabric, not a separately documented capability.

L5 Session

A standard cloud-platform access session — console and API authentication for calling inference endpoints and managing fine-tuning jobs — mechanically not fundamentally different from this page's other cloud/Neocloud domains, and this domain hasn't built a distinctive differentiator at this layer.

L6 Presentation

L7 Application

This is this domain's star layer, and where its real core asset actually sits: Together Inference — an optimized inference service for open-source large models (such as Llama and Mixtral-class models), with a FlashAttention paper co-author on the team contributing to inference-performance optimization (specific team composition per latest official disclosure…

Fluidstack

Home turf: L1

L1 Physical

This layer is the relatively solid part of this domain's currently available public information: a GPU cloud provider founded around 2017, whose early business reportedly included rendering and other GPU-intensive workloads before shifting to a positioning centred on large-scale AI training cluster rental.

L2 Data Link

This layer is explicitly empty — no product or service. This domain has no reliable, public cluster-interconnect technical detail (such as a specific network topology or RDMA scheme) for this file to cite, stated honestly as empty rather than applying a peer company's technical template without grounding.

L3 Network

This layer is explicitly empty — no product or service. Same reason as L2: no reliable public information supports specific network-layer product detail, stated honestly as empty.

L4 Transport

This layer is explicitly empty — no product or service. Same reason as L2 and L3: no reliable public information supports specific transport-layer product detail, stated honestly as empty.

L5 Session

L6 Presentation

This layer is explicitly empty — no product or service. There's no public information about a standalone encryption/presentation-layer product for this domain, stated honestly as empty.

L7 Application

This layer is explicitly empty — no product or service. The specific form and functional scope of this domain's cloud-management console/API have limited public documentation; this file doesn't describe unverified product detail — unlike this page's other Neocloud domains, which have solid public information at L7, this is stated honestly as empty rather tha…