The Demand Side

Customer archetypes

1. The academic researcher

Grad student or postdoc training models for a paper. Has a small grant or department budget. Needs cheap GPU access for experiments — sometimes overnight runs, sometimes weeks of training.

• Price-sensitivity: High. The grant has a fixed dollar amount.
• Reliability tolerance: Moderate. A crashed instance costs a day but isn't existential.
• Volume: Bursty. Heavy use around paper deadlines, light otherwise.

2. The indie AI builder

Solo developer or small team building an AI product. Could be a model startup, a tools company, or a hobbyist shipping an open-source project. Uses Vast for training, fine-tuning, or self-hosted inference.

• Price-sensitivity: High. Burning founder savings.
• Reliability tolerance: Moderate. Will retry interrupted jobs.
• Volume: Steady once the workload stabilizes.

3. The hobbyist / enthusiast

ML / AI enthusiast running personal projects. Could be training Stable Diffusion LoRAs, running image generation, experimenting with LLMs. Doesn't have local GPUs or has outgrown their local rig.

• Price-sensitivity: Very high. This is a hobby.
• Reliability tolerance: High. Will absorb instance churn.
• Volume: Bursty.

4. The small-to-mid AI company

Series-A through Series-B AI startup. Has venture funding but is still cost-conscious. Uses Vast for non-production workloads — research, ablations, smaller-scale training — and a more enterprise-y cloud for production.

• Price-sensitivity: Moderate. Cost matters but isn't the only factor.
• Reliability tolerance: Moderate. Will design around variability.
• Volume: Real money. Tens of thousands of dollars per month at this segment.

5. The ML / GPU consultant or service provider

Person or small firm whose business is doing ML work for other people. Uses Vast as the back-end for client work. Buys GPU-hours at marketplace prices; sells time-and-materials or fixed-fee project work.

• Price-sensitivity: Indirectly high — their margin depends on it.
• Reliability tolerance: Variable; depends on the client work.
• Volume: Moderate.

6. The occasional enterprise

Some enterprise users do appear on Vast, typically for non-production research workloads or for one-off bursts that exceed their internal capacity. Rarely the production deployment.

• Price-sensitivity: Low (they have budget).
• Reliability tolerance: Often surprised by variability; sometimes a poor fit.
• Volume: Episodic.

Workload types

What people actually run on Vast:

Training and fine-tuning

• Small-to-mid scale language model fine-tuning (LoRA, QLoRA, full fine-tunes for sub-70B models).
• Image-generation model training (Stable Diffusion fine-tunes, LoRAs).
• Computer-vision model training (object detection, segmentation, classification).
• Research prototyping (model architecture experiments, new optimization techniques).
• Limited large-scale pre-training — Vast's multi-node infrastructure is too weak for the biggest runs.

Inference

• Self-hosted LLM endpoints for projects too small to pay API rates.
• Image and video generation services.
• Indie-product inference back-ends.
• Batch inference jobs.

Other

• Rendering (3D rendering, video processing) — niche but real.
• Scientific computing (molecular dynamics, climate models, etc.) — academic users.
• Crypto mining — historically big, much smaller today, mostly limited to coins that still favor GPU.

Why customers choose Vast

• Price. Often 60-80% cheaper than hyperscaler list prices for the same nominal GPU. The single biggest factor.
• Self-service. Sign up, add a credit card or crypto, launch an instance in minutes. No sales call.
• Variety. Many GPU models available. Older cards (RTX 3090, A6000) are cheap on Vast and not available on most clouds.
• No commitments. Pay-as-you-go; spin up and shut down any time.
• API + CLI flexibility. Scriptable end-to-end. Power users can build workflows that span dozens of instances.

Friction points

The honest list of what makes Vast hard to use:

• Variable quality. Some instances are great; some are flaky. Choosing well requires reading reliability scores and DLPerf benchmarks.
• Linux / Docker required. No Windows GPU instances; no managed Jupyter (though Docker images can include it); little hand-holding.
• Networking limits. Most providers don't offer cluster-grade interconnect. Single-node only for many users.
• Storage churn. Local disk only on the instance. Persistent storage needs to be managed by the user (S3, etc.).
• No SLA. If your instance dies mid-job, you eat the loss. There's no support team that will compensate you for downtime.
• Trust assumptions. The host could theoretically access your VM. Don't run regulated workloads.

The friction is the price of the cheap price. For users who can absorb the friction, Vast is great. For users who can't, the cheap price doesn't make up for the variability.

Where Vast doesn't fit

The customer segments that effectively can't use Vast:

• Regulated industries (healthcare, finance, government) with PII or PHI data. The trust model doesn't support it.
• Large-scale distributed training requiring 100+ GPUs with high-bandwidth interconnect. The fleet structure doesn't exist on Vast.
• Production inference with five-nines uptime requirements. The interruptibility profile is wrong.
• Customers who need account managers and SLAs. Vast doesn't sell that way.
• Customers who need SOC 2 / HIPAA / FedRAMP attestation from the cloud provider. Vast doesn't provide it across the marketplace.

These segments are CoreWeave, Lambda, Crusoe, and Together territory, not Vast. That's by design.

Geographic spread

Vast's demand is global. Notably more international than enterprise neoclouds because:

• Crypto-payment acceptance reduces friction for international users.
• Self-service onboarding means no need for region-specific sales coverage.
• The indie / research / hobbyist user base is geographically distributed.

The supply is also more international than competitors. Providers in Eastern Europe, Asia, and Latin America list on Vast at meaningful scale. Some of the cheapest GPU-hours on the platform come from regions with cheap electricity.

Takeaway

Vast's customer base is structurally different from enterprise neoclouds. It's bigger by user count but smaller by enterprise revenue. The price sensitivity, the technical sophistication, and the tolerance for variability of Vast's users are all higher than at CoreWeave / Lambda / Crusoe. That shapes everything else: the product surface, the support model, the strategic ceiling.

The next chapter does pricing and bidding, the mechanical heart of how the market works.