KPT-B200-SXM6-192G
B200 HGX
Newly rackedSJC-02 in stock
- Memory
- 192 GB HBM3e
- Node
- 8 GPUs / 8U
- In-node fabric
- NVLink 1.8 TB/s
- Draw
- ~14.3 kW
- Minimum
- 8 GPUs
$8.90 / GPU-hr
Outright $392,000 / 8-GPU node
Enterprise AI compute · three U.S. regions
Buy it outright,
or rent it by the hour.
Racked in three days.
We supply enterprise GPU nodes, accelerators and whole racks, two ways to take them: buy them and we manage them for you in a U.S. facility, or pay by the GPU-hour and stop whenever you like. Same hardware, same facility, same serial numbers in the same system.
112 nodes in stock, schedulable now8-GPU minimumNo commitment threshold
Live inventoryUpdated 2 minutes ago
KPT-B200-SXM6-192G
SJC-02 · Santa Clara, CA
24nodes
KPT-H200-SXM5-141G
IAD-01 · Ashburn, VA
64nodes
KPT-H200-SXM5-141G
DFW-01 · Dallas, TX
14nodes
KPT-L40S-PCIE-48G
IAD-01 · Ashburn, VA
46cards
Stock and lead times are public. See the full catalogue →
148nodes
Under managementAcross three U.S. regions, every one serial-tracked.
72h
Order to rackedMedian for in-stock SKUs, including 72 hours of burn-in.
99.9%
Power & coolingDual-fed PDUs, N+1 cooling, written into the contract.
14days
Return windowGrace period after a rental ends; beyond it, the day rate applies.
Own or rent
Same machines,
two ways to take them.
In this industry those are usually two companies doing two things: one sells hardware, the other rents compute. We do both, out of one inventory, one facility and one serial-number system. So you do not have to guess between “what if it sits idle” and “what if renting gets expensive” — both cost curves are plotted below.
Kepter Supply
Buy outright + managed
You put up the capital and keep the asset. We put up the facility, the power, the hands and the spares.
- The asset is on your books — Invoices, serial numbers, depreciation and residual value are yours — deductible, financeable, resellable.
- Unit cost thins out over time — Run for 36 months, total cost lands at about 59% of renting the same configuration by the hour (the arithmetic is in figure 06 below).
- Capacity is certain — The machines are yours. They are not scheduled into a shared pool, so they never queue behind someone else at peak.
- You can move them out — When the managed contract ends we de-rack, pack and ship them out. We do not hold hardware back.
- The managed fee is fixed — Billed per rack and per kW, written into the contract, not settled against a floating power price.
Fits: steady workloads, an 18-month-plus horizon, teams that need the asset on the balance sheet.Ask for a purchase plan
Kepter Flow
Rent by the GPU-hour
Operating expense only. Pay for the hours you use; stopped machines do not bill.
- Available today — For in-stock SKUs, connection details land within hours of provisioning.
- No long contract — Hourly settlement, no minimum term, no early-exit penalty.
- Scale without paperwork — Eight GPUs to start; adding more is confirmed live against remaining stock.
- Hardware risk is not yours — We swap failed units. The replacement time counts against our SLA, not against your bill.
- Convertible to ownership — At any point in a rental you can convert the machines you are on to a purchase, with 60% of rent already paid credited against the price.
Fits: training peaks, pilots, workloads that have not settled into a shape yet.See hourly rates
Not sure which? Rent first, decide after three months.
This is not a sales line: at any point in a rental you can convert the machines you are using to a purchase, with 60% of the rent you have paid credited against the price. Same serial numbers — no migration, no reinstall, no downtime. Three months of real utilisation figures beat any procurement model.
Deployment
Order to training job:
a 72-hour median.
That figure includes 72 hours of burn-in — the machine has already run at full load for three days before it reaches you, and the report ships with it. We publish the duration of every leg, because a date is the only promise worth anything in this business.
01 · Commercial
Select and lock
Tell us the GPU model, the count, the region and roughly how long for; a plan comes back within 48 hours. Two hours after you confirm, stock is locked and serials are assigned on the spot.
You get: quote · serial list
02 · Logistics
Pick and ship
Asset-tagged, ESD-packed, re-checked before racking. The waybill is trackable throughout, and cross-region transfers move on our own fleet.
You get: waybill · live location
03 · Facility
Rack and cable
Assigned a cabinet, wired to both power feeds and to 400G. You will be told which cabinet, which U positions and which switch this batch hangs off.
You get: rack position · topology
04 · Acceptance
Burn in and deliver
Seventy-two hours at full load with temperature, draw, memory ECC and NCCL bandwidth logged throughout. Passing is what gets it delivered; failing gets it swapped and re-run.
You get: burn-in report · access credentials
Inside the rack
What you buy — or rent —
looks exactly like this.
Most compute suppliers give you a GPU model and a number. We have drawn a whole cabinet apart: which U positions your machines sit in, which switch they hang off, which power feed they draw, where the tag goes. All of it lands in your topology document on delivery day.
Hardware
Models on sale and for rent.
Four main product lines, outright price and hourly rate written side by side. The price is not hidden behind “contact sales” — what needs discussing is terms and volume, not the unit price.
KPT-H200-SXM5-141G
H200 HGX
In stock, all regions
- Memory
- 141 GB HBM3e
- Node
- 8 GPUs / 8U
- In-node fabric
- NVLink 900 GB/s
- Draw
- ~10.4 kW
- Minimum
- 8 GPUs
$3.42 / GPU-hr
Outright $248,000 / 8-GPU node
KPT-H100-SXM5-80G
H100 HGX
IAD-01 in stock
- Memory
- 80 GB HBM3
- Node
- 8 GPUs / 8U
- In-node fabric
- NVLink 900 GB/s
- Draw
- ~10.2 kW
- Minimum
- 8 GPUs
$2.28 / GPU-hr
Outright $164,000 / 8-GPU node
KPT-L40S-PCIE-48G
L40S PCIe
IAD-01 in stockSingle GPU
- Memory
- 48 GB GDDR6
- Node
- 1–8 GPUs / 4U
- In-node fabric
- PCIe Gen5
- Draw
- ~3.1 kW
- Minimum
- 1 GPU
$1.15 / GPU-hr
Outright $11,400 / per GPU
Rates are on-demand and exclude commitment discounts. Rent continuously for 3 months and the rate drops 20% automatically; at 12 months, 35%. No contract to sign, no application to file — it shows up on the bill. The outright price includes the first year of managed ops; from year two the managed fee is billed per rack and per kW. All prices are illustrative; the quote governs.
U.S. footprint
Three regions,
one private backbone.
All hardware is inside the United States. Between regions we run dark fibre we lease ourselves, so cross-region traffic never touches the public internet and is never billed as egress — which is a real line item when you are training data-parallel across regions, not a decoration on a spec sheet.
Data stays in country
All three regions are inside the United States; hardware, backups and logs do not leave it. We will supply the facility locations and a device list on request.
Power is in the contract
Dual-fed PDUs, N+1 cooling, and 99.9% power and cooling availability written into the SLA. Miss it and we credit a share of that month’s managed fee.
Come and look
Outright customers can book a quarterly site visit and see their own cabinet and their own tags. It is the plainest way to verify that the asset is in your name.
Serial tracking
Every machine carries
a complete history.
The serial is assigned the moment a unit arrives and does not change until it is retired. Whose hands it passed through, what it ran, how burn-in went, how many repairs, which U of which cabinet it is in right now — scan the QR code on the tag and it is all there.
What outright customers get
- A per-device invoice cross-referenced to serial numbers
- The 72-hour burn-in report as a PDF, with NCCL bandwidth and ECC logs
- Rack position and topology, refreshed quarterly
- The right to book a site visit each quarter
- De-racking, packing and outbound shipping when the managed contract ends
What rental customers get
- The serial numbers of the units you are on (not “an H200 somewhere”)
- The same burn-in report — rentals are not a different batch
- Hour-by-hour billing detail, reconcilable to a single device
- A record of every swap and how long it took, counted against the SLA
- A return inspection report confirming the data was wiped
Pricing
When to buy,
and when to rent.
There is only one honest answer: it depends how long you will run it. The chart below runs the real price list for one 8-GPU H200 node out to 36 months; the curves cross in month 16. Renting wins below 16 months, owning wins above it — we do both, so we have no side to argue for.
Hourly rates
| SKU | On demand | ≥ 3 months | ≥ 12 months | Minimum |
|---|---|---|---|---|
| B200-SXM6-192G | $8.90 | $7.12 | $5.79 | 8 GPUs |
| H200-SXM5-141G | $3.42 | $2.74 | $2.22 | 8 GPUs |
| H100-SXM5-80G | $2.28 | $1.82 | $1.48 | 8 GPUs |
| L40S-PCIE-48G | $1.15 | $0.92 | $0.75 | 1 GPU |
Figures are USD per GPU-hour. Discounts apply automatically, accrued by continuous usage, with nothing to sign or request. Stopped machines do not bill, but storage kept for a retained instance is charged at $0.04 / GB-month.
What we do not charge for
| Item | Us | Common in the industry |
|---|---|---|
| Inter-region traffic | $0 | $0.02–0.09 / GB |
| Egress (first 50 TB/month) | $0 | Metered |
| Starting and stopping instances | $0 | Minimum billing increment |
| Early exit | $0 | 30–100% of the remaining term |
| Downtime during a swap | Not billed | Billed as normal |
| Technical support | Included | Charged by tier |
The “common in the industry” column reflects general practice in published price lists and is not aimed at any one vendor. We have folded these into the base rate — the unit price does not look like the cheapest, but there is no second line on the bill.
Console
Bought and rented,
in one console.
Whether a machine is your asset or a rental, it looks identical in the console: utilisation, temperature, draw, billing, serial number, rack position. The only difference is whether the tag in the corner reads Supply or Flow.
KEPTERCOMPUTE/ fleetSupply + Flow
78.4%Queue utilisation · all regions
112Ready
14Racking
19Term ending
3Fault
214,880Billed GPU-hours this month
$734kThis month’s invoice
3Regions
IAD-01832 GPU · 81% usedSJC-02576 GPU · 74% usedDFW-01288 GPU · 62% usedSLA99.94% MTD
Per-device telemetry
Utilisation, memory in use, temperature, draw and ECC counts, sampled every 15 seconds and kept for 13 months.
Bills that reconcile to hardware
Every line on the invoice opens up to show which serial, which window and which rate produced it.
Complete API
Anything the console does, the REST API does: provision, scale, query stock, pull telemetry, export invoices.
Term warnings
Three reminders before a term ends — 14 days, 7 days, 48 hours — with automatic renewal or release available.
FAQ
Frequently asked
If the answer is not here, just ask. Our sales team does not work from a script: what they can answer they answer on the spot, and what they cannot, they tell you when they can.
Ask something specificAre rented machines the same batch as the ones you sell?
Yes. Same inventory pool, same 72-hour burn-in standard, same facility. The only difference is the commercial relationship: bought units are transferred into your name and taken out of the shared pool, rented units remain ours. Every unit you rent has a serial number, and the console shows its full history — including who used it before you, and for how long.
Once I have bought them, can I take the machines away?
Yes, whenever you like. The hardware is your asset. When the managed contract ends we de-rack, ESD-pack and ship it out, charged as a one-off relocation service. We do not hold hardware back in any form, and there is no “minimum managed term” clause locking your asset inside our facility.
If I rent and then want to buy, what happens to the rent I paid?
You can convert to a purchase at any point in the rental, with 60% of the rent already paid credited against the price, capped at 40% of that price. What converts is the machines you are already using — same serial numbers, no migration, no reinstall, no downtime, effective the same day.
What is the minimum, and is there a minimum term?
HGX models start at eight GPUs (one whole node, because an NVLink domain cannot be split); L40S starts at one. There is no minimum term — billing is hourly and stops when the machine stops. There is no minimum monthly spend and no committed volume either.
What if a machine fails? Whose time is the downtime?
Rental customers: we swap the unit, the swap window is not billed, and the time it took counts against our SLA. P1 faults get a response within 4 hours and a replacement within 24.
Outright customers: Kepter Care covers parts and labour on the same response times. Parts beyond warranty are billed at cost, with no markup.
Do you run your own workloads on these GPUs?
No. We do not train models, we do not serve inference, we do not run any business that competes with a customer, and we do not use idle capacity for mining or for internal jobs. Idle means idle — which is what lets us publish per-cabinet utilisation.
What about data security and compliance?
All three facilities are inside the United States; hardware, backups and logs do not leave the country. Rented instances are wiped on return and a re-inspection report is issued. We do not look inside your instance: telemetry collects hardware-level metrics only — temperature, draw, memory in use, ECC — never memory contents or disk data.
Get started
Tell us the GPU model, the count, and roughly how long for.
Within 48 hours you get a plan naming specific serial numbers and arrival dates, with the arithmetic for both buying and renting. If it is not a fit we will say so — we do not chase quotes.