Fleet-hours in Seoul · sales open 1 December 2026
Your own AI fleet in Korea, by the hour.
Rent a whole SparkPipe fleet and, for those hours, it serves only you. It runs Kimi K3 and other open models over a terabyte, too large for most GPU servers, with million-token context and speculation, and Korea's own large models: LG's K-EXAONE 2.0, SKT's A.X K2 and Upstage's Solar Open 2. Your data stays in Korea, there is no queue of other customers' requests, and the price is per hour, not per token.
For businesses only. The terms are published for review and come into force on 1 December 2026; nothing can be bought or paid before then.
Korea
Data stays here
Built for Korean companies.
Data stays in Korea
The fleets and our servers are in Seoul, and your requests go to no foreign AI service. Cached request state is deleted within 24 hours of its last use.
A Korean company
You contract with HYOR AI Tech Ltd., a Korean company, under Korean law. Both contracts are in Korean and English, and the Korean text prevails.
Pay in won
Korean businesses pay by bank transfer against our electronic tax invoice (세금계산서), with 10% VAT. Customers abroad pay in USDC or by USD wire.
Korean models
LG's K-EXAONE 2.0 and SKT's A.X K2, both under Apache 2.0, and Upstage's Solar Open 2 serve on Fleet 1, each the whole model at its publisher's precision, beside Kimi K3, GLM 5.3 and the other open models. They are hard to get elsewhere: in October 2026 neither Artificial Analysis nor Hugging Face lists a public API provider for any of the three.
Dedicated, not shared
During your hours the whole fleet serves only your requests. Nobody else's traffic waits in front of yours, so answers start as soon as the fleet is free.
By the hour
You pay for the fleet's time, not for tokens: long prompts, million-token contexts and agents that re-read their history cost nothing extra.
Compare
Hardware by the hour
What an hour of comparable hardware costs.
A fleet-hour rents a machine, so it compares with renting hardware, not with per-token APIs.
| Option | What you get | Per hour |
|---|---|---|
| SparkPipe fleet-hour | 16 DGX Sparks run as one machine, Kimi K3 loaded with a million-token context, speculation, operated for you | $20 |
| 16 rented DGX Sparks | The same boxes, without the engine that runs one model across all sixteen | about $10–12 |
| One 8× B300 GPU server | 2.3 TB of GPU memory: holds Kimi K3 at full precision and serves many requests at once | about $64–90 |
| One 8× H200 GPU server | 1.1 TB of GPU memory: too small for Kimi K3 at full precision | about $38 |
Public rental listings checked in October 2026: hosted DGX Sparks at $0.60–0.75 per hour (Spheron), B300 at $8.00–11.19 per GPU-hour (Latitude.sh, DigitalOcean), H200 at $4.80 per GPU-hour (Spheron). Prices vary by provider and change often.
Speed
Measured and estimated
Fast where agents need it.
Kimi K3 rewriting whole files with speculation, measured; 15 tok/s without it. Search-and-replace edits run at 28 tok/s.
to the first token of a follow-up turn whose conversation is cached, measured on Kimi K3 up to 137,700 tokens.
ESTIMATED: a 128K-token Kimi K3 prompt from cold with four prefill accelerator boxes, 18 s with eight.
ESTIMATED for Kimi K3 on Fleet 2, eight M5 Ultra Mac Studios: 23 tok/s for one request without speculation, and about 35 tok/s with it on agent coding turns. SKT's A.X K2 is estimated at 33 and about 49 tok/s. Estimates are projections from our measurements and published hardware figures, published only where we expect to meet them; the Fleet List shows how each is made, and each is replaced by a measurement once the hardware runs.
Roadmap
Next
What comes next.
How it works
Four steps
Account, hours, reservation, use.
Open an account
Give your company's details and connect an Ethereum wallet: its address identifies the account. Accept the Hourly Terms by signing their text with the wallet. Every business is screened against sanctions lists.
Buy hours
Pay in advance: Korean businesses in won, by bank transfer against our electronic tax invoice; customers abroad in USDC, through an invoice from our payment processor. Korean VAT of 10% is added. Your hours are issued once the payment has cleared.
Reserve
Choose a fleet, a model and the hours on the fleet's calendar, and sign; on hourly terms you reserve up to 7 days ahead. A confirmed reservation is final and is taken from your balance as the hour passes.
Use
During your hours the whole fleet serves only your requests, through the SparkPipe API, which follows the OpenAI chat format. You issue the access tokens, so you decide who uses your hours.
Won payments go straight to HYOR AI Tech Ltd.'s Korean bank account. USDC payments go through a payment processor. Our system has no feature for passing hours between accounts.
Prices
Two ways to buy
Try it by the hour. Commit for less.
Every business starts on hourly terms. A business that wants every hour for two years at the lower price passes our business check and signs the two-year agreement, which it may negotiate clause by clause.
| Hourly | Two years | |
|---|---|---|
| Price | ₩26,000 or $20 per fleet-hour plus VAT | ₩20,000 or $15 per fleet-hour plus VAT, fixed for the term |
| Commitment | None | Every hour of two years from signing: 17,520 hours, or 17,544 when the term contains 29 February |
| Models | Any offered model on the Fleet List, chosen for each reservation | The models you choose in the agreement stay available for the whole term; the others while they are listed |
| Business check | Not needed | Required before signing |
| Hours held | Up to 168 at a time, or 720 after a business check; each lasts 12 months | Each month's paid hours, reserved when you choose within that month: 12 in each 12-hour window a day until you have prepaid the whole term |
| Payment | Before use: a won bank transfer against our electronic tax invoice (세금계산서) for Korean businesses, or a USDC invoice | A won bank transfer against our electronic tax invoice, a USD wire from your own account abroad (zero-rated for VAT), or a USDC invoice. Nothing is due in the first 30 days; you pay any amounts up to three deadlines, on days 30, 60 and 90 |
| Reservations | Final. A two-year customer may take your reserved hour until 6 hours before it starts; the hour goes back to your balance | Final, and never taken by anyone |
| Downtime | A reserved hour with more than 15 minutes of downtime comes back in full, automatically, from our own records | |
| Refunds | None, unless we stop hourly service or close your account through no fault of yours | None, except paid hours not yet used if the agreement ends early (clause 11.3). If no payment is made by day 30, it ends with nothing owed |
| Terms | Hourly Terms시간제 이용약관 | Two-year agreement2년 예약 계약서 |
VAT is charged at 10% on every payment except a zero-rated wire. Any amount charged as VAT that turns out not to be owed is an overpayment, and we refund it. The Korean text of each contract prevails.
Fleets and models
Fleet List
One balance, every fleet.
Your hours work on every fleet we operate. A fleet's rate is the number of hours one hour on it takes from your balance. The Fleet List is the list the terms refer to: fleets, rates, models and their licenses.
| Fleet | Machines | State | Rate |
|---|---|---|---|
| Fleet 1 | 16 NVIDIA DGX Spark, 2 TB of memory in all, 100 Gb/s RDMA | Serving | 1 |
| Fleet 2 | Apple Mac Studio, M5 Ultra | Under construction | 1 |
| Fleet 3 | 16 NVIDIA DGX Spark | Under construction | 1 |
| Core model on Fleet 1 | License | Context window | One request | Eight requests | Prefill, 8K prompt |
|---|---|---|---|---|---|
| Kimi K3Moonshot AI · kimi-k3 | Kimi K3 License | 1M | 16.2tok/s | 64tok/s | 590tok/s |
| GLM 5.3Z.AI · glm-5.3 | GLM-5.3 License | 1M | 28.9tok/s | 74tok/s | 1,162tok/s |
| GLM 5.3 FlashZ.AI · glm-5.3-flash | MIT | 128K | 55.0tok/s | 176tok/s | 876tok/s |
| DeepSeek V4.1 FlashDeepSeek · deepseek-v4.1-flash | MIT | 1M | 42.5tok/s ¹ | 151tok/s ¹ | 1,290tok/s ¹ |
| Korean model on Fleet 1 | License | Context window | One request | Prefill, 7K prompt | File rewrite, speculation |
|---|---|---|---|---|---|
| K-EXAONE 2.0LG AI Research · k-exaone-2.0 | Apache 2.0 | 256K | 27.3tok/s | 324tok/s | 190.9tok/s |
| A.X K2SK Telecom · a.x-k2 | Apache 2.0 | 256K | 23.8tok/s | 1,274tok/s | 298.0tok/s |
| Solar Open 2Upstage · solar-open-2 | Upstage Solar License | 256K | 61.2tok/s | 1,025tok/s | 513.7tok/s |
Context windows as served. Decode without speculation, measured on 2026-10-18 with prompts of 7 to 8K tokens. ¹ DeepSeek V4.1 Flash was measured on 2026-10-24 after a prompt of about a million tokens; its prefill figure is that prompt's average. The Korean models were measured on 2026-10-26 with prompts of 6,850 to 7,396 tokens. File rewrite: a file of about 3,800 tokens rewritten whole with one function renamed, with speculation drafting from the request's own text; the output is identical to decoding without speculation, and the rate includes the prompt's prefill (2026-10-26). All nineteen installed models, their licenses and speeds are on the Fleet List; you must follow each model's license.
Responsibility
Who does what
We run the fleet. You run your work.
Capacity only
You alone decide what runs during your hours. We do not monitor, review or control your workloads, data or outputs, and we are not responsible for them or for how they are used.
Your users
You may let your own users, or anyone you choose, use your hours through the key and access tokens you control. That happens outside our system: it keeps no register of it and takes no part. You stay responsible for everyone who uses your hours.
Lawful use
Every use must follow the law, including sanctions and export control, personal data and intellectual property law, and each model's license. We may suspend use that a law or authority requires us to stop, that breaks the terms, or that threatens a fleet.
AI law
If you offer generative AI to your own users, you are their provider under Korea's AI Basic Act and carry its notice and labelling duties.
Your data
You keep every right in your data and outputs. We keep no backups of it. Cached request state is deleted within 24 hours of its last use. The fleets and our servers are in Seoul.
Businesses only
Fleet-hours are sold to businesses for their business, never to consumers. Korean law governs both contracts, and disputes go to arbitration in Seoul under the Korean Commercial Arbitration Board's rules.
Technology
Open source
Sixteen small boxes, one frontier model.
Decoding a large model is limited by how fast its weights can be read. SparkPipe splits every weight and every cache entry across the sixteen nodes, so each reads only its own sixteenth, then merges the partial results over a 100 Gb/s RDMA mesh in 16 to 38 microseconds. The engine is written from scratch in C and CUDA and published under the Apache 2.0 license.
of the model and of every conversation's cache on each node. Kimi K3's 1.56 TB of weights are held whole across the fleet.
to answer the next turn of a 137,700-token conversation from the cache, against 299 s from cold.
tokens per second for Kimi K3 rewriting whole files with speculation. On new prose the gain is 1.0 to 1.2 times.
Speculation drafts tokens from the request's own context, and from a draft model on a separate GPU, and the model verifies every draft, so the output is the model's own. All figures were measured on Fleet 1 through 2026-10-24; methods and receipts are in PERFORMANCE_STATUS.md. Source: github.com/sparkpipe/sparkpipe.
Company
Documents