IDC, GPU cloud, bare-metal, OEMs

Long-term GPU chassis rental

You want idle racks, power and chassis turned into load that is working when the machines boot — not another tenant who only asks for a quote. We purchased 2,224 MI308X cards; incremental rental absorbs traffic, it does not fill a missing first cluster. We bring models, channels and our all-flash array. You keep title and lift idle-hour utilisation. Mid-range cards are in scope.

Please quote available chassis, SKU, lead time and settlement (monthly chassis, GPU-hour, off-peak). Two hard conditions: 2U rack space for our array, and staff access for deployment.

Value to you

For an IDC, GPU cloud or OEM, utilisation is the first sensitive variable. We are not shopping for a first cluster. We bring a deployment and acceptance flow that already ran on 278 purchased chassis / 2,224 cards, and we rent your spare windows for incremental traffic.
We bring three things: deployable open-weight models, token-sales channels in progress, and our own FX array (signed throughput +29–40%, reports R2/R3). KV cache sits on our array, so your local disks can be modest.
You keep title. Settlement can be monthly chassis, GPU-hour or off-peak. After a deal, downstream token-supply / offtake may be co-signed as joint suppliers, so you become a party to the customer contract, not only the landlord. This is a prospective condition; a formal contract has not been signed.
About 600 chassis in the prospect pipeline (intent, not held) means load can be added after the purchased batch runs — the relationship is not a one-shot.

Agenda for a first formal meeting

Available kit: chassis count, SKU (including mid-range), lead time, power and network.
The two hard conditions: 2U for our array; staff access for deploy and watch.
Settlement: monthly / GPU-hour / off-peak; when the clock starts; how idle time is treated.
Operations: fault response, remote and on-site rights, data-residency or MLPS constraints.
Whether offtake is co-signed as joint supply (a formal contract has not been signed), and priority on later racks.

What this means in a term sheet

The purchased 278-chassis (2,224-card) deployment and acceptance flow already runs; incremental rental absorbs traffic growth, not a missing first cluster.
KV cache sits on our all-flash array — local disks on your chassis can be modest. We need 2U for that array.
Settlement is open: monthly chassis, GPU-hour, or off-peak pricing.
We are in commercial talks and joint tests with Zhipu, Kimi and MiniMax; MiniMax is in the token-supply signing process, and a formal contract has not been signed.

Why we can deliver this

Five reasons this is not another deck. Figures carry a denominator or a report ID. Economic numbers are engineering-model estimates, delivered as runnable Python after NDA.

Purchased kit, countable
278 eight-GPU chassis / 2,224 cards on the books; value on tax documents. About 600 chassis are a prospect pipeline (intent, not held) and are not added into stock.
Acceleration is our line
FX arrays are in production; KV tiering is source-adapted by us, not a resale. Signed throughput +29–40%, TTFT −26–32% (R2/R3 vs local NVMe).
Pre-production acceptance is done
8×MI308X, 1T-parameter INT4 MoE, 3,210 samples across four benches, official delta inside the 2% red line.
Downstream progress is not overstated
Talks and joint tests with Zhipu, Kimi and MiniMax; MiniMax is in token-supply signing, and a formal contract has not been signed. Communication is not collection.
Verify first, stop-loss built in
Joint test first, decisions second: gate-based acceptance with built-in stop-loss. G1 Week 2 / G2 Week 4 / G3 Week 7 / G4 Week 10. Full model after NDA.

On this track specifically

We already purchased 2,224 cards and need a long-term increment. This is not a price-check RFQ; the deploy path ran on our own batch.
Acceleration is our product: arrays in production, modest local-disk need, 2U plus access.
Channels are not empty: talks and joint tests with Zhipu, Kimi and MiniMax; MiniMax is in signing, and a formal contract has not been signed.
The gated test already has weeks: G1–G4 at weeks 2 / 4 / 7 / 10; stop-loss if a gate is missed. You need not sign a year first.
After NDA the acceptance scripts are reproducible, so utilisation is not a slogan.

Next actions and dates

D0 (today)
Reply with chassis, SKU, lead time, and whether the 2U and access conditions can be met.
48 hours after NDA
We send the bring-up list, array power/network needs, and the gate table.
Week 1
Align rack location, power, remote accounts; sign access and confidentiality.
Week 2 G1
Delivery acceptance: All materials delivered and inspected; each array passes power-on self-test with all 24 drives recognized.
Week 7 G3
Main joint-test gate: KV tiering acceleration: TTFT reduction ≥25% and throughput gain within the measured +29–40% band; 8 nodes sharing one array approach the NIC ceiling on read-back.
Week 10 G4
Stability & acceptance: 72-hour continuous stress with no interruption; single-drive / single-link fault injection transparent to the workload; signed joint-test report issued. Volume settlement only after the gate.

Verifiable progress

Purchased: 278 × 8-GPU AMD Instinct MI308X chassis (2,224 cards); equipment value to be shown on tax documents; mid-term plan 5,000-card class.
Pre-production acceptance on a single 8×MI308X node for a 1T-parameter INT4 MoE model: 3,210 samples across four accuracy benches all passed (official delta −0.8% to −1.5%, inside the 2% red line).
Deployment adaptation and performance tests with MiniMax are complete; token supply is in the commercial-signing process; a formal contract has not been signed.
Signed storage-acceleration result: throughput +29–40%, TTFT −26–32% (Qwen3-Coder-480B-FP8 long-context cold recovery vs local NVMe; reports R2/R3).

Evidence library →

Send an inquiry All five tracks

Contact Wang Limin · +86 13161818898 · mingxin@agentmail.to

Other counterparties

This page is a public description of one-to-one B2B cooperation. It does not constitute an offer, an investment offer, or any return promise. Figures are the 2026-09-15 external baseline; economic numbers are engineering-model estimates, not signed purchase or sales prices.