27 articles on machine learning — what drives the cost, how it is priced, and where the savings actually are.
GPU nodes cost 10 to 40 times a general-purpose node per hour, and Kubernetes gives you a whole GPU per pod by default. Idle GPU nodes are the most expensive waste in any cluster. Here are the rates and the sharing strategies.
An H100 costs roughly three times an A100 per GPU-hour, but it can be two to four times faster on modern training workloads. The cheaper GPU is the one with the lower cost per unit of work, not the lower hourly rate. Here is how to do that math.
The T4 has been the default cheap inference GPU for years, but the L4 costs about 53 percent more per hour while delivering two to three times the throughput on modern models. For inference fleets, that inversion is worth real money.
A hosted API charges per token with no floor. A self-hosted model charges per GPU-hour whether you use it or not. The crossover sits at a specific monthly token volume, and most teams guess it wrong by an order of magnitude.
Adding pgvector to a Postgres instance you already run looks free. A managed vector service starts at several hundred dollars a month. The honest comparison involves index memory, replica sizing, and how much your team wants to operate.
Fine-tuning cost is a product of four numbers: model size, dataset tokens, GPU rate, and how many times you will redo it. Get those on paper and a run that felt unbounded turns into a figure you can approve.
Spot GPUs cost 50 to 70 percent less and can vanish with two minutes of warning. Whether that trade is good depends on one number: how much work you lose per interruption, which is a function of your checkpoint interval.
A web server scales in seconds. A GPU replica loading a 40GB model takes minutes. That gap forces you to keep warm capacity, and warm capacity is where most model serving budgets quietly go.
A GPU at 25 percent utilization is not 25 percent efficient, it is 75 percent wasted at ten dollars an hour. Measuring real utilization usually reveals the largest single optimization available in an AI budget.
Running twelve models on twelve GPUs is simple and expensive. Packing them onto three, with time-slicing or partitioning, can cut the serving bill by 70 percent when each model only needs a fraction of a device.
Training pulls terabytes from object storage, distributed jobs push gradients across zones, and inference ships embeddings between services. At $0.01 to $0.09 per GB, AI data movement turns into a five-figure line quickly.
Training checkpoints accumulate silently. A team running dozens of experiments can hold hundreds of terabytes of optimizer states they will never load again, at $0.023 per GB-month it adds up to a real recurring bill.
Hourly rates hide enormous differences in delivered compute. Dividing price by teraflops and by memory bandwidth turns a confusing price list into a ranking you can actually reason about.
SageMaker instance rates carry a premium of roughly 15 to 40 percent over the equivalent EC2 instance. Whether that is good value depends on how much of the surrounding platform you would otherwise build yourself.
On Google Cloud you can run models on Vertex AI or on your own GKE cluster with GPU node pools. The hardware rates are similar, so the decision comes down to cluster fees, discount eligibility, and who operates it.
Azure Machine Learning charges you the underlying VM rate with no platform surcharge, which makes cluster configuration the whole cost story. Minimum node counts and idle timeouts decide most of the bill.
The KV cache grows linearly with context length and batch size, and it competes with model weights for the same expensive GPU memory. Long contexts are not a feature you add for free, they are a capacity purchase.
Committing to GPU capacity saves 30 to 60 percent, but GPU generations turn over every 18 to 24 months. Getting the commitment horizon right matters more here than anywhere else in cloud purchasing.
Cost per million tokens is the one number that compares a self-hosted GPU with any hosted alternative. Computing it takes one benchmark run and turns inference budgeting into arithmetic.
GPU instances are the most expensive compute you can rent, from a few dollars an hour for inference cards to tens of dollars an hour for training clusters. Understanding what drives the price, and how to cut it, keeps GPU bills sane. Here is the breakdown.
Cost per 1000 inference requests is the unit metric that tells you whether serving a model is economical. It ties GPU or serverless cost to actual usage, and it is the number to optimize as you scale. Here is how to compute and reduce it.
AI cost divides into training (building the model, often GPU-heavy and periodic) and inference (running it, continuous and scaling with usage). They have different profiles and different levers. Here is how to think about the split.
Azure Machine Learning cost is dominated by compute, GPUs for training, endpoints for inference, and idle clusters. Right-sizing, autoscaling to zero, spot compute, and batch endpoints are the biggest levers. Here is the playbook.
GPU instances cost many times a CPU instance per hour, so idle GPUs, oversized accelerators, and full-price on-demand are expensive mistakes. Spot, right-sizing, and sharing are the big levers. Here is how to keep GPU cost under control.