SI / IDC procurement / public tenders / AIDC operators
In-stock chassis sale
You want countable, deliverable chassis — not another deck. We purchased 278 AMD Instinct MI308X eight-GPU servers. Single-node acceptance travels with the inquiry pack. Token talks are credit only: Zhipu, Kimi and MiniMax; MiniMax is in the signing process, and a formal contract has not been signed.
Reply with unit count, delivery site and payment (wire / LC). The costing band is runnable Python after NDA.
Value to you
- For an SI or a tender desk, the first question is whether the kit exists. 278 chassis are purchased and can be released in lots of 1 / 4 / 8 / 16.
- Acceptance is single-node (KIMI-K27-ACPT-2026-001). We do not call that a fully commissioned 2,224-card cluster.
- Acceleration is our FX line: +29–40% throughput versus local NVMe on the stated load (R2/R3). That is not a 2× NVIDIA claim.
- Downstream talks are real and not overstated: MiniMax is in signing; a formal contract has not been signed.
Agenda for a first formal meeting
- Count and lot size.
- Delivery, acceptance and payment (wire / LC).
- Optional FX array bundle — not mandatory.
- When serials and tax papers for that lot will be shown.
What this means in a term sheet
- 278 chassis / 2,224 cards purchased; value on tax documents.
- Lots are available; no RMB list price on this page.
- MiniMax is in token-supply signing; a formal contract has not been signed.
- Single-node acceptance only; no claim that the full fleet is commissioned.
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
- The assets are purchased and countable.
- Acceptance and signed storage reports sit in the evidence library.
- The delivery checklist and LOI are reusable text.
- We do not upgrade talks into a signed offtake.
Next actions and dates
D0 (today)
Use the contact form, choose purchase, and name the unit count.
48 hours after NDA
Delivery checklist, one-pager and the reproducible model.
After that lot’s papers are filed
SPA. Until then, inquiry only.
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).
Contact Wang Limin · +86 13161818898 · mingxin@agentmail.to
Other counterparties
Licensed financial lessors
Sale-and-leaseback
Title sits with the lessor; repossession, not a mortgage fight.
Banks, factors, SCF platforms
Supply-chain finance and factoring
The payer is a foundation-model company; confirmation does not travel through a trade chain.
IDC, GPU cloud, bare-metal, OEMs
Long-term GPU chassis rental
We arrive with models and traffic; assets stop waiting for a customer.
Strategy / IR / industrial capital / IT distribution
Listed-company and industrial collaboration
Joint test first: about ten weeks, G1–G4, stop-loss, no cheque at that stage.
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.