Why "Direct" Beats "Swap"

When people get cloud new-user trial credits, the first instinct is often to burn them in the cloud marketplace on some AI SaaS subscription. That works, but it adds a layer: whether the marketplace listing accepts credits, at what ratio, and whether there is a minimum spend are all variables.

If the goal is simply "turn credits into callable AI tokens," the shorter path is to spend the credits directly on the cloud's own model service:

  • AWS new-user credits → Amazon Bedrock
  • Google Cloud new-user credits → Vertex AI (including Gemini models)
  • Microsoft Azure new-user credits → Azure AI Foundry (Azure OpenAI and the model catalog)

None of these require procurement approval. You only need to enable the service, verify your credit type covers it, and request model quota. Here is the platform-by-platform breakdown.

Steps

Step 1: Confirm your credits can actually be spent here

This is where most people trip. Trial credits are not universal cash. Three common restrictions:

  1. Service allowlist: Some promotional credits only cover compute, storage, and other baseline services; managed AI/ML services may be excluded.
  2. Expiry: AWS promotional credits usually carry an explicit expiration date (often anywhere from tens of days to a year, depending on the promotion); GCP new-user credits typically run 90 days; Azure new-user credits typically run 30 days and do not auto-renew. Always trust what your billing console shows.
  3. Account type: GCP's free-trial credit and its always-free tier are two different things. The former is spent once and gone; the latter refreshes on a monthly quota. Do not conflate them.

Action: Open each platform's billing console, list every credit grant, and read its scope and expiration. Spend the soonest-expiring ones on AI inference first.

Step 2: AWS — route credits into Bedrock

  1. Search for Bedrock in the AWS console. Request model access first — some Bedrock models require an explicit access request, typically approved within minutes to a day.
  2. Choose a region. Bedrock model availability varies by region; us-east-1 and us-west-2 tend to have the widest coverage. Picking the wrong region produces a "model not found" error, not a credit problem.
  3. Send a test request in the Bedrock Playground to confirm billing works.
  4. Go back to Billing and confirm the call was offset by credits. If it shows as cash spend, your credits do not cover Bedrock — switch accounts or platforms.
  5. Once confirmed, create an IAM user, generate access keys, and wire up the Converse / InvokeModel API.

Key limits: Bedrock's default concurrency and per-minute request quotas are low. Submit a Service Quotas increase before production use. Bedrock prices input and output tokens separately, so long-context workloads burn through credits noticeably faster.

Step 3: GCP — route credits into Vertex AI

  1. Enable the Vertex AI API in the Google Cloud console. This is mandatory; calls fail with a permission error otherwise.
  2. Confirm your billing account is linked and the trial credit is still valid.
  3. Run a quick test in Vertex AI Studio with a Gemini model and watch the response and billing.
  4. Connect via gcloud CLI or a service account JSON: grant the service account the Vertex AI User role in IAM, then specify project ID and region in code.
  5. On regions, us-central1 usually has the broadest model coverage; some newer models launch in specific regions first.

Key limits: GCP new-user credits zero out at expiry and do not roll over. Also, some Vertex AI models have separate quota systems you must request per model on the quotas page.

Step 4: Azure — route credits into Azure AI Foundry

  1. Create an Azure AI Foundry (formerly Azure AI Studio) resource in the portal, choosing a supported region.
  2. Deploy a model inside that resource. Azure OpenAI models require an access request first; approval time varies.
  3. Once deployed, test in the Playground and capture the endpoint and key.
  4. Check Cost Management to confirm credit coverage. Azure new-user credits have a clear expiry, and some promotional amounts only cover specific service tiers.

Key limits: Azure model deployments have TPM (tokens per minute) quotas that often default low. Raise them in the deployment config or request an increase on the quotas page. Deployments can also incur idle cost — delete them when done.

Step 5: Funnel everything through one entry point

The three routes have different API shapes. Put a compatibility layer in front (for example, a self-hosted OpenAI-compatible gateway) that wraps Bedrock, Vertex AI, and Azure behind a single endpoint. The payoff: when one vendor's credits run out, you switch the backend without touching business code.

Step 6: Set budget alerts so credits don't run away

  • AWS: set a threshold alert in Budgets.
  • GCP: create a budget in Billing and bind a Pub/Sub notification.
  • Azure: create a budget and action group in Cost Management.

Without alerts, once credits are exhausted the bill flips straight to cash charges. This is the single most common new-user trap.

Caveats

  • Do not run long-lived always-on services on trial credits. Inference is pay-per-call and stops when you stop; that beats renting a GPU instance.
  • Do not farm credits by registering new accounts with the same card. Every platform has risk controls, and triggering them can restrict your main account too.
  • Do not copy credit numbers from someone else's screenshot. Amounts, expiry, and coverage shift with promotion, region, and account type. Trust your own console.
  • Region selection outranks model selection. Confirm which models exist in your target region before deciding which model to use.
  • Validate the billing path with small traffic first, then scale once you have confirmed credits are actually being deducted.

When this fits

  • Solo developers or small teams that need a predictable block of free inference for prototyping.
  • Teams already running infrastructure on one cloud, who want to steer new-user credits into that same cloud's model service and avoid multi-account overhead.
  • Anyone comparing models A/B who wants credits to cover evaluation-phase calls.
  • Not a fit for: production systems needing long-term, stable, high-volume inference — credits are a time-limited subsidy, not a cost strategy.

Bottom line

To convert cloud new-user trial credits into AI tokens, the shortest path is not a detour through the marketplace — it is enabling Bedrock, Vertex AI, or Azure AI Foundry in the same console. What decides success is not "which vendor gives more," but three small things: whether the credit covers the service, whether your chosen region has the model you want, and whether your quota has been raised. Get those three right, in order, and your credits will not quietly expire on the billing page.