PTUs require a monthly or annual reservation commitment. PTUs not used are billed anyway. For bursty traffic, the PAYG model wins because the PTU capacity sits idle during low-traffic periods. The breakeven depends on workload predictability: workloads above 60-70% average utilisation benefit from PTU; workloads below that are cheaper on PAYG. Azure also supports Data Zone deployments and Global deployments at slightly higher rates (Azure pricing docs, September 2026).
Bedrock's add-on stack: the hidden cost
Bedrock inference is rarely the surprise line item — the add-on services stack that production deployments need often runs 20-35% of the inference cost.
Knowledge Bases (Bedrock's RAG service) runs on Amazon OpenSearch Serverless, which starts at $345 per month before a single query. Bedrock Guardrails cost $0.15 per 1,000 text units processed. Bedrock Flows (agent orchestration) bill $0.035 per 1,000 node transitions. Cross-region inference does not add a separate routing charge but has data-residency implications (technologymatch.com, June 2026).
On a production bill, these adjacent charges often run 20-35% of inference cost. Agentic workflows that chain multiple model calls (reasoning + tool selection + tool results + synthesis) can burn 5-10x the visible prompt-and-response tokens once intermediate reasoning is counted. Bedrock's value at that scale is governance, not discount — the per-token rates match Anthropic's direct pricing exactly, so the platform fee is what you pay for IAM integration and unified billing (technologymatch.com, June 2026).
Vertex AI's batch and free tier
Vertex AI's batch prediction 50% discount makes it the cost-competitive option for non-real-time inference workloads — embeddings generation, document summarisation, nightly report generation, and offline data processing.
Vertex AI Batch Mode cuts token rates by 50% across Gemini 2.5 Pro ($0.625/$5 per million tokens) and Gemini 2.5 Flash ($0.15/$1.25 per million tokens). Batch returns results within 24 hours and is the cheapest Gemini route. Vertex AI also has a free tier ($300/90 days for new Google Cloud accounts) that Azure does not match, making development significantly cheaper on Vertex (digiusher.com, March 2026).
Vertex AI's hidden costs come from data pipelines, BigQuery storage ($6.25 per TB/month after the first 1 TB free), Dataflow jobs ($0.018-$0.035 per vCPU hour), and Agent Engine (billed per second of agent runtime). Hybrid pricing complexity adds another layer: Vertex's pricing is separated into model pricing, generative AI pricing, and quota/throughput behaviour, requiring external normalisation for accurate cost-per-workload calculation. Cross-region egress adds 10-20% (usage.ai, August 2026).
What enterprise buyers should do next
Three actions for organisations choosing a hyperscaler LLM platform in 2026.
- Run on PAYG for 30-60 days first. All three platforms look inexpensive at the model layer. Use PAYG pricing across your production workload before committing to PTU or Provisioned Throughput. Measure P95 throughput against reserved capacity before committing.
- Model the add-on stack, not just the token rate. Bedrock Guardrails and Flows, Azure Monitor and AI Search, Vertex BigQuery and Dataflow — these adjacent services often run 15-40% of inference cost. Include them in your TCO model before signing.
- Apply FOCUS normalisation across the three billing models. Azure bills on tokens and PTU capacity; Bedrock on tokens and request dimensions; Vertex on tokens plus compute plus data processing. Map all three to the FOCUS standard before drawing a price conclusion.
What to watch next
Three near-term datapoints. First, Azure OpenAI's introduction of native budget enforcement — Microsoft currently recommends alerts plus automation for stronger cost control, rather than hard limits. Second, AWS Bedrock's expansion of Custom Model Import beyond $0.0785/min/CMU pricing — the current per-CMU pricing is not directly comparable to token rates on other platforms. Third, Vertex AI's continued investment in adapter-based fine-tuning — the most flexible fine-tuning platform of the three, but the per-token training cost is not always published upfront.