Training Infrastructure on the Clouds

Lesson 3 of 3 in In Production: Should You Ever Pre-train?.

If your organization does train — from scratch or continued — you will almost certainly do it on rented infrastructure, and all three hyperscalers sell it in managed form. At the level that stays true across hardware generations: AWS offers managed training through SageMaker — jobs that provision clusters, wire up distribution, watch node health, and resume from checkpoints — on GPU fleets and on purpose-built training accelerators AWS designs itself. Azure runs distributed training through Azure Machine Learning on GPU clusters built around high-bandwidth RDMA-class interconnect, with the identity, network isolation, and compliance surface of the rest of Azure attached. Google Cloud offers TPU training — accelerators and interconnect co-designed as pods you provision by topology — alongside GPU fleets, with the deepest assumptions about the software stack (TPUs reward compiler-first frameworks that emit XLA).

The trade-offs are the same three on every cloud. Managed versus assembled: managed training buys you the orchestration layer from last lesson — provisioning, health checks, restart-from-checkpoint — at the cost of fitting your job into the platform’s shapes; assembling your own cluster from raw instances returns the control and the pager. Capacity: accelerators at cluster scale are a reservation negotiation, not an API call — large blocks are booked ahead, and how far ahead, in what shapes, on what terms, is often the actual constraint on when your run can start. Data gravity: a trillion-token corpus and a steady stream of checkpoints live in object storage next to the cluster; moving either across clouds mid-project is painful enough that the storage decision quietly becomes the cloud decision.

The training-infrastructure conversation, as questions. Rows are the mechanisms this domain taught; cells are what to ask each provider. Deliberately no part numbers — accelerator generations, quotas, and prices are dated facts, and the LLMs on the Cloud domain (L-09) carries those tables.
DimensionAsk AWSAsk AzureAsk Google Cloud

Capacity & reservations

What cluster sizes can SageMaker training actually schedule for me, and how far ahead must a large block be reserved — on GPUs versus your own training accelerators?

What GPU cluster shapes can Azure ML schedule at my scale, and what do cluster-scale reservations look like in my region?

What TPU slice topologies (or GPU clusters) can I reserve, on what lead time, and in which regions?

Interconnect

How are my nodes placed relative to each other, and what inter-node bandwidth does the fabric sustain at my job size?

Which cluster shapes come with RDMA-class interconnect end to end, and is it uniform across the whole allocation?

What does the pod topology I am buying mean for tensor- versus data-parallel traffic in my planned layout?

Failure & resume

What does managed training detect and replace automatically, and what is the measured resume time from my checkpoint size?

What health monitoring and restart behavior does Azure ML provide for multi-node jobs, and where does my restart logic take over?

How do slices behave on hardware fault — what is replaced, what restarts, and what does my framework need to handle?

Data & checkpoints

What throughput can S3 sustain to my cluster for shard streaming, and how long does a full checkpoint take to write and restore?

Same questions against Blob Storage — sustained read for the data pipeline, burst write for checkpoints at my cadence.

Same questions against Cloud Storage — and does the TPU data path change how I should shard the corpus?

Software stack

Which distribution strategies does the managed layer configure for me, and which do I bring (my own launcher, my own parallelism)?

What does Azure ML assume about my training framework, and how much of the job definition is mine to own?

How much of my stack must be XLA-friendly to earn the TPU’s economics — and what is the porting cost from my current codebase?

In production

What “managed training” actually manages, cloud by cloud — and the line where the platform’s responsibility ends and yours begins.

AWS

SageMaker training jobs take a container and a cluster spec and handle provisioning, inter-node wiring, health monitoring, node replacement, and restart-from-checkpoint; purpose-built training accelerators exist as an AWS-designed alternative to GPUs, trading ecosystem breadth for a price-performance argument you should evaluate on your own workload. What stays yours on any path: the corpus and its mixture, the parallelism layout, the hyperparameters, the loss curve, and the judgment call when it spikes.

Azure

Azure Machine Learning runs multi-node distributed jobs on GPU clusters with RDMA-class interconnect, providing scheduling, health monitoring, and integration with Azure’s identity and network-isolation machinery — often the deciding surface for regulated buyers whose sovereignty requirements stop at infrastructure rather than reaching the weights. The training science remains entirely yours; the platform keeps nodes alive, not loss curves healthy.

Google Cloud

Cloud TPU training provisions pod slices where accelerator and interconnect arrive as one co-designed unit — you choose a topology rather than assembling a network — and managed layers handle scheduling and fault recovery. The distinctive trade-off is software: TPU economics reward XLA-compiled stacks, so porting cost is part of the price. Across all three clouds the boundary is identical: platforms manage machines; convergence is never a managed service.

And that is the pre-training domain, closed. You can read a data pipeline, an objective, a scaling-law plot (the Scaling Law Plotter at /tools/scaling-law-plotter will stay useful), a FLOPs budget, a parallelism layout, a loss-spike postmortem — and now a build-versus-adapt decision — the way this domain has insisted throughout: mechanisms first, numbers cited or not stated at all.

But notice what all of it produces: a Base model — a formidable text-completer with no instructions followed, no refusals, no conversation. Every path through this module ends at the same door, whether you pre-trained from scratch, continued on an open-weights base, or wisely did neither: the model that people actually use is made in the next stage. That is the Post-training & Alignment domain — SFT, RLHF, DPO, and reasoning training; how a predictor becomes an assistant.

Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.