The Kit

The hyperscalers

AWS — The largest cloud, and the broadest menu of AI compute: NVIDIA GPU instances, its own Trainium/Inferentia silicon at a discount for workloads that port, and managed layers (SageMaker, Bedrock) up the abstraction stack. Strengths: capacity breadth, every adjacent service you'll need, mature committed-pricing machinery (savings plans, capacity blocks for short-term GPU reservations). Honest caveats: frontier-GPU availability has historically demanded commitments or capacity blocks rather than on-demand hope; egress pricing punishes data gravity; the custom-silicon discount is real but earned through porting work. Reach for it when your stack already lives there, or when you need AI compute adjacent to everything else.

Microsoft Azure — The hyperscaler that bet the company on AI infrastructure, building some of the largest AI campuses on earth, anchored by its OpenAI partnership. Strengths: frontier-scale training infrastructure (InfiniBand-backed clusters are standard, a legacy of its HPC lineage), the deepest integration with the OpenAI model family, enterprise distribution. Honest caveats: the hottest capacity has tended to be spoken for by anchor tenants; the enterprise-first console and quota experience can be heavy for small teams. Reach for it for large committed AI capacity, or when the Microsoft enterprise stack is already your world.

Google Cloud (GCP) — The most vertically integrated option: Google designs the TPU, the fabric, the data center, and the models. TPU pods offer frontier-scale training on non-NVIDIA silicon with genuinely different economics — if your stack speaks JAX (or increasingly PyTorch/XLA). Strengths: TPU price-performance at scale, first-rate internal networking, strong data/analytics adjacency. Honest caveats: TPU lock-in is architectural, not contractual — porting off is work; NVIDIA capacity exists but TPU is the house specialty; quota negotiation is its own project. Reach for it for TPU-scale training or data-pipeline-heavy AI work.

Oracle Cloud (OCI) — The insurgent fourth, which won frontier training deals by offering what labs actually wanted: huge dedicated GPU superclusters, RDMA networking, aggressive pricing, less abstraction in the way. Strengths: bare-metal GPU clusters at very large scale, price competitiveness, speed of standing up dedicated capacity. Honest caveats: a narrower surrounding-services ecosystem than the big three; the business is concentrated in a small number of enormous AI contracts, which is a structural fact about the vendor, not a defect of the product. Reach for it for dedicated training capacity at committed scale.

The neoclouds

CoreWeave — The definitional neocloud: GPUs as the entire business, grown from a crypto- mining pivot into one of the largest pure-play AI clouds, with early access to each NVIDIA generation and Slurm-and-Kubernetes-native cluster delivery. Strengths: dense, training-grade clusters (InfiniBand standard) available faster and often cheaper than hyperscaler equivalents; an org that speaks ML infrastructure natively. Honest caveats: the balance sheet is the story — GPU-collateralized debt and heavy customer concentration are structural risks the trade press tracks for good reason (and this site's Bubble Watch does too); it is a compute vendor, not a full cloud. Reach for it for serious GPU capacity without a hyperscaler commitment.

Lambda — The ML-researcher's GPU cloud, built from a GPU-workstation vendor into on-demand and reserved clusters. Strengths: simple pricing, low friction from sign-up to a running instance, one-click multi-node clusters, a long-standing focus on the training niche. Honest caveats: capacity depth is a fraction of a hyperscaler's — popular instance types sell out; the services layer around raw compute is thin by design. Reach for it for research workloads, fine-tuning, and teams that want a GPU this afternoon without a procurement process.

Crusoe — The energy-first neocloud: its founding trick was powering compute with stranded energy (flared gas, curtailed renewables), and it has scaled into building large AI campuses where power is, not where the internet is. Strengths: the power problem is its core competency — an advantage that compounds as energy becomes the binding constraint; competitively priced clusters. Honest caveats: smaller and newer than CoreWeave in cloud terms; the energy-siting model means capacity arrives in large discrete chunks tied to specific sites. Reach for it when the power story matters to you — or watch it regardless, because its model is a preview of where the whole layer is going.

The abstraction layer

Modal — Serverless compute for AI workloads: you write a Python function, decorate it, and Modal runs it on a GPU in the cloud — cold-started in seconds, billed by the second, autoscaled to zero. Strengths: the best developer experience in GPU compute; genuinely eliminates the cluster-management layer for inference, batch jobs, and fine-tuning-sized training; per-second billing makes idle cost vanish. Honest caveats: unit prices above raw rented metal — you are paying for the utilization problem to become Modal's problem, which is often a good trade at small-to-medium scale but inverts at sustained high volume; frontier-scale multi-node training is not the use case. Reach for it when engineering time is your scarce resource and your workload is spiky.

SkyPilot — Open-source broker for the multi-cloud shortage era: describe a job (GPUs needed, data, command) and SkyPilot finds capacity across your cloud accounts, launches it wherever it's cheapest or available, and manages spot-instance failover with automatic recovery. Strengths: turns "which cloud has H100s today" from a morning ritual into a scheduler's job; excellent spot-recovery machinery; no vendor to pay. Honest caveats: it brokers the clouds you already have accounts and quota with — it does not conjure capacity; you still own the underlying bills and quotas. Reach for it if you run workloads across two or more providers, or live on spot instances.

Ray / Anyscale — Ray is the open-source distributed-compute framework that lets ordinary Python fan out across a cluster: tasks, actors, and libraries for the ML lifecycle (Ray Train, Ray Serve, Ray Data, RLlib). It has become default plumbing for large-scale data preprocessing, RL pipelines, and model serving. Anyscale is the managed platform from Ray's creators — hosted clusters, an optimized runtime, production hardening. Strengths: the standard answer to "make this Python run on 500 machines"; serving and preprocessing at scale without writing systems code. Honest caveats: Ray is its own distributed system to operate — cluster tuning and debugging are real work (that operational burden is precisely what Anyscale sells relief from); for pure synchronized training, Slurm-style scheduling often remains the simpler tool. Reach for Ray when your bottleneck is scaling Python, and Anyscale when you'd rather not run Ray yourself.

The schedulers

Kubernetes — The cloud-native orchestrator that won: declarative infrastructure, autoscaling, self-healing services — the default substrate for inference serving and the platform every cloud speaks. For AI it brings device plugins for GPUs, and a growing retrofit of batch capabilities (gang scheduling, queueing via projects like Kueue and Volcano) to cover training's shape. Strengths: the ecosystem — every tool integrates with it; unmatched for long-running, autoscaled serving. Honest caveats: it was not designed for gang-scheduled batch training, and the retrofits, while real, still trail Slurm's decades of batch-scheduling maturity; its complexity budget is famous, and a small team can drown in it. Reach for it for inference and AI platforms; think twice before making it your training scheduler out of fashion.

Slurm — The HPC batch scheduler, decades old, unglamorous, and still what a very large share of serious training runs sit on. Whole-node allocations, gang scheduling, fair-share queues, topology-aware placement — the exact shape of a synchronized training job, solved long before AI arrived. Strengths: does the training-cluster job with brutal directness; battle-tested at national-lab scale; simple mental model (queues and jobs). Honest caveats: it is a scheduler, not a platform — no autoscaling services, no declarative app management, and a user experience from another era; cloud integration is workable but bolted on. Reach for it for dedicated training clusters; nearly every neocloud will hand you a cluster with it preinstalled, which tells you what the market actually uses.