Our pricing page claims up to 70% savings vs AWS Trainium. This page shows exactly how that number was derived — methodology, data sources, assumptions, and where it doesn't hold.
Two distinct comparisons back our pricing claims. They're measured differently and apply to different use cases.
Raw cost per GPU/Trainium-chip hour for training and fine-tuning workloads.
End-to-end cost per page for document processing tasks (summarization, extraction, classification).
On-demand instance pricing as of April 2025 (us-east-1). ThetaZero routes jobs to Theta EdgeCloud nodes running Trainium hardware; pricing is based on EdgeCloud's published per-hour rates.
| Instance / Config | Chips | AWS On-Demand/hr | ThetaZero/hr | Savings |
|---|---|---|---|---|
| trn1.2xlarge equivalent | 2× | $1.3438 | ~$0.40 est | ~70% |
| trn1.32xlarge equivalent | 32× | $21.50 | ~$6.45 est | ~70% |
| trn2.48xlarge equivalent | 48× (Trainium2) | $24.78 | ~$7.43 est | ~70% |
Data sources: AWS EC2 pricing page (aws.amazon.com/ec2/pricing/on-demand); Theta EdgeCloud published rate schedule (thetatoken.org/edgecloud). "est" = ThetaZero rate derived from EdgeCloud pricing applied to Trainium capacity — not a contractually fixed rate and may vary. AWS reserved instance pricing (1yr / 3yr) will be lower; see Caveats section.
Per-page cost for common document AI tasks. ThetaZero's PAYG rate is $0.004/page and covers the full pipeline: parsing, inference, structured output. AWS equivalent stacks Textract (parsing) + Bedrock (inference).
| Task | Input Size | AWS Equivalent/page | ThetaZero PAYG/page | Savings |
|---|---|---|---|---|
| Document summarization | 1 page / ~350 tokens | ~$0.018 | $0.004 | ~78% |
| Field extraction | 1 page / ~350 tokens | ~$0.020 | $0.004 | ~80% |
| Email classification | ~200 tokens | ~$0.012 | $0.004 | ~67% |
| Q&A over document | 1 page + question | ~$0.022 | $0.004 | ~82% |
AWS equivalent breakdown: AWS Textract Forms & Queries: $0.015/page (analysis) + Bedrock Claude 3 Haiku: ~$0.00025/1K input + $0.00125/1K output tokens, estimated $0.003–0.007/page for typical outputs. Combined estimate: $0.018–0.022/page. ThetaZero runs inference on Trainium hardware through EdgeCloud routing, absorbing both parsing and inference in the flat per-page rate.
The pricing calculator on our pricing page compares ThetaZero against self-managed AWS infrastructure. This overhead applies even before you run a single workload — it's the cost of running the plumbing.
ThetaZero has no infrastructure overhead. You pay per use; no standing resources, no idle charges. This overhead is excluded from the per-hour compute comparison above — it applies on top of any Trainium instance costs.
How these numbers were derived.
AWS on-demand instance prices pulled from aws.amazon.com/ec2/pricing/on-demand (us-east-1 region). Theta EdgeCloud rates sourced from published rate schedule at thetatoken.org. AWS Textract pricing from aws.amazon.com/textract/pricing. AWS Bedrock per-token rates from aws.amazon.com/bedrock/pricing.
We tested four representative tasks: document summarization, structured field extraction, email classification, and document Q&A. Each task processed 50 real business documents (mix of invoices, contracts, and email threads) ranging from 200–800 tokens of input. Runs were conducted between January–March 2025. Average token counts per task type were measured and used as inputs to the pricing formula.
For each document, AWS cost = Textract analysis cost (per-page rate) + Bedrock inference cost (input tokens × $0.00025/1K + output tokens × $0.00125/1K for Haiku). Output token count estimated from actual ThetaZero responses averaged across test runs. Infrastructure overhead (NAT, ECS, ALB) amortized over 10,000 pages/month as a per-page surcharge.
Flat PAYG rate of $0.004/page applied uniformly. No per-token billing, no infrastructure overhead. The rate is fixed regardless of document complexity within the tested range (200–800 input tokens).
Trainium instance hourly rates taken directly from published pricing. ThetaZero EdgeCloud rate for Trainium capacity is the EdgeCloud published per-node-hour rate for equivalent hardware, multiplied by the number of Trainium chips. No actual runtime benchmarks performed — this is a pricing comparison only, not a performance comparison.
We re-pull all published pricing sources monthly (first week of each month) and re-run the inference test suite quarterly. AWS pricing changes infrequently; EdgeCloud rates may update with network conditions. The "last updated" date in the page header reflects the most recent pricing pull.
The scenarios where our claims don't hold, or where the savings are smaller.
1-year reserved Trainium instances typically cost 40–50% less than on-demand. If you commit to reserved capacity, the effective AWS rate for trn1.2xlarge drops to roughly $0.67–0.80/hr — still more expensive than ThetaZero's ~$0.40/hr, but a ~50% gap rather than 70%. 3-year reservations may approach or match EdgeCloud spot pricing at high utilization levels. If you already have reserved capacity, the savings are lower.
AWS Bedrock has volume discount tiers starting at 1M+ tokens/month. At very high throughput (>10M pages/month), a direct Bedrock negotiated rate may undercut ThetaZero's PAYG rate. Contact us for volume pricing at those scales.
The $89/mo overhead estimate assumes a fresh AWS setup with no existing VPC or networking infrastructure. If you already run services on AWS and can share an existing VPC, NAT gateway, and load balancer, your marginal infrastructure overhead is lower — closer to $5–10/mo additional for the Trainium-specific resources.
EdgeCloud routing introduces variable latency depending on node proximity and load. We measured cost only. In tests, median latency for inference tasks was 1.2–2.8 seconds per page — comparable to managed Bedrock endpoints. But latency benchmarks are not included in the tables above.
The $0.004/page PAYG rate is current as of the date shown. We will update this page within 30 days of any pricing change. Locked-in plan subscribers retain their rate for the duration of their subscription term.
EdgeCloud pricing for Trainium capacity is derived from published node rates and may vary with network conditions and chip availability. The ~$0.40/hr and similar figures are representative estimates. Get a firm quote by contacting enterprise support.