A Hefes Project · Live at tethysx.com

TethysX

The AI compute aggregator — every GPU you'll ever need, on one contract. Capacity contracted across regions and resold as a single metered service.

Visit tethysx.com

---

1 · What TethysX Is

TethysX is an AI compute aggregator. We contract GPU capacity ahead of demand across multiple regions, keep it warm, and resell it as a single metered service — a drop-in API, one rate card, one invoice.

Buying GPUs one vendor at a time is the expensive way. An aggregator carries the supply risk so you don't: you pick the region when you create the key, and every product carries that tag while the scheduler enforces it end to end.

In one line: every GPU you'll ever need, on one contract.

2 · Why Aggregate

• Aggregate pricing, not list pricing. We commit volume upstream and pass the blended rate down — scale pricing from the first month, without a scale commitment of your own.
• Capacity that doesn't run out. Several independent supply lines per region mean a queue on one never becomes your outage. Committed blocks are reserved, not best-effort.
• One counterparty, one invoice. One master agreement, one rate card, one usage ledger — instead of a procurement cycle per vendor and a spreadsheet to reconcile them.

3 · Products

Three ways to buy. Same region tag, same ledger.

• Inference API — point your existing SDK at api.tethysx.com and keep shipping. Open-weight frontier models on datacenter GPUs, compact models on the cost tier, streaming and failover included. Priced per 1M tokens, by model class and region.
• Training jobs — LoRA/QLoRA, vision models, PyTorch and TensorFlow from curated templates or your own image. Dataset in at /input, weights out from /output, with live logs and hard credit caps. Priced per GPU-second or per task.
• Reserved blocks — a monthly block for a team, faculty, or research group, with a notebook surface on top: persistent workspaces, dataset mounts, idle auto-stop, and your weights downloadable any time. Priced per monthly CU block, by region.

4 · Pricing — the Compute Unit

Everything we sell normalises to the Compute Unit (CU): one RTX 4070-class GPU-hour, delivered at ≥ 90% availability, in the region on your key. A quote, a job, and an invoice all speak the same language.

• Regions price separately — CU-apac and CU-na — and per-token and per-job rates derive from the CU basis of the region you run in. Nothing is marked up twice.
• Every hardware class converts by a published multiple, blending measured throughput with the class's market reference rate.

On-demand rates in apac-1 are benchmarked against the open market and set below it, with committed blocks taking a further discount. Indicative numbers and a live estimator are published on the TethysX rate card; final rates are set in the rate card attached to your agreement and applied identically to everyone.

5 · How It Works

From key to invoice in four steps. The control plane sees metadata, usage, and billing; your payloads take the short path between you and the region you chose.

• 1 · Tag the key. Create a key and choose its region. Sub-tenants inherit it or pick their own.
• 2 · Resolve in milliseconds. We authenticate, check your commitment, and place the job on a node inside that region only.
• 3 · Stream over private transit. Prompts, datasets, and outputs move over licensed private links. Pre-empted jobs checkpoint and resume.
• 4 · Settle in CUs. Usage lands in the ledger with its region column. Monthly statements and exports show exactly where every unit ran.

6 · What We Guarantee at Every Stage

• No silent cross-region fallback. A key tagged for one region is never served by another — not on overflow, not on failover. It's in the code and in the contract.
• Nothing retained in the middle. Only account data, usage, and billing live at the control plane. Prompts and datasets never land there.
• Your models stay yours. Encrypted models and datasets, minimal retention on nodes, and an acceptable-use policy that flows down the whole chain.

7 · The Transparency Ledger

Delivered CUs, measured availability, and service credits issued are published every month, for every region — whether the numbers flatter us or not.

• Metered at the node. Every job writes a signed usage record at completion. The ledger is the sum of those records, not an estimate.
• Availability is measured, not claimed. Synthetic probes run every 60 seconds per region, and a month below 90% triggers credits automatically.
• Published within 5 business days. Customer statements and the public page are generated from the same ledger export, on the same schedule.

8 · Wholesale & Reserved Fleets

Committed capacity for a small number of partners who resell under their own brand, and reserved blocks direct to institutions. Nothing is sold as a GPU-hour you have to babysit.

• ✅ Structure. Master agreement plus a capacity schedule per region.
• ✅ Commitment. Committed CUs, take-or-pay, split by region — stepping up as we add supply.
• ✅ Price. Wholesale per CU in USD, with price protection on the committed block.
• ✅ Rights. First refusal on new capacity, white-label products, and sub-tenant keys with region choice.
• ✅ SLA. Defined on CUs delivered per month per region, with service credits below threshold.
• ✅ Reporting. Usage export with a region column, monthly CU statements, and the public transparency ledger.

You contract with TethysX and nobody else — we carry the supply relationships, the capacity risk, and the regional plumbing. Most partners are live inside a week: a schedule signed, keys issued, and the first committed block metering on day one.

---

Next Steps

Tell us the region, the monthly volume in CUs or tokens, and whether you resell. We come back with a rate card and a capacity schedule.

Visit tethysx.com Talk to Us ← Back to Hefes

TethysX is an active Hefes project. This page summarises the platform and its commercial model at a high level. Rates, regions, and capacity schedules are established per agreement; indicative pricing is published on the TethysX rate card.