This suggests they are bypassing prefabricated Huawei systems.
First confirmation that DeepSeek is targeting to build infrastructure from megawatt to *gigawatt* range. And again, as in other listings: it just does not look like they plan to buy Huawei's prefab 950 pods. They are designing their own systems.
Some users praise DeepSeek's first-principles approach to gigawatt-scale data center hiring for AI buildout, while others call GPU shortages for Chinese firms a tragedy.
DeepSeek posted a new job on its website: IDC Design Planning Engineer. This is the person who designs a data center before it gets built. Site selection, power systems, liquid cooling, rack layout, construction specs.
When an AI model company starts hiring to build its own physical infrastructure, the signal is clear. My read is DeepSeek is planning something at GW-scale. That is a different category entirely.
And here. @zephyr_z9 what's this about Wanka? Sugon scaleX640? Or is this a translation error? 万卡集群 https://app.mokahr.com/social-recruitment/high-flyer/140576#/job/83e14658-d30e-42ae-b7ec-f263caed55d6
First confirmation that DeepSeek is targeting to build infrastructure from megawatt to *gigawatt* range. And again, as in other listings: it just does not look like they plan to buy Huawei's prefab 950 pods. They are designing their own systems.
@Spacew3asel I think Huawei GPUs are the most likely substrate, which makes it interesting that they apparently believe default server configurations are not optimal for them.
@teortaxesTex @zephyr_z9 万卡 = ten thousand accelerator cards
So Wanka cluster means a 10,000-GPU cluster
@teortaxesTex planning and designing GW level IDC doesn't necessarily mean no prefeb pods
@teortaxesTex @zephyr_z9 Translation error, 万卡集群 means ~10K GPU cluster
@qjhtc Just how much do they intend to build? But these listings, combined, look like they're building Fire-Flyer 3, not a Huawei campus.
@teortaxesTex Could well be a collab, the interconnect being extremely likely to be made with Huawei components anyways (there aren't that many alternatives and all high performance interconnects are export controlled in some way or another)
@teortaxesTex Huawei！
@teortaxesTex Deepseek 是“第一性原理”的拥护者，即使文峰没有在采访中说过这样的事，但从他的作风我看得出来，他实践的就是“第一性原理”。
@teortaxesTex Building system that enables GPUs to talk fast to each other in a pod turns out to be incredibly difficult. Between pods, not so much.
@teortaxesTex gigawatt range changes the conversation entirely.
huawei prefabs just arent built for that scale anyway.
@poezhao0605 What stands out is the scope of responsibilities — it covers the entire infrastructure stack before construction even begins. That usually only appears when the project is being designed as a systems-level asset, not just capacity expansion. GW-scale planning fits that pattern.
@poezhao0605 I mean this was supposed to happen sooner or later. Deepseek is in it for a long game.
@teortaxesTex Makes the CATL investment even more strategic
@teortaxesTex @zephyr_z9 It is a tragedy... Chinese companies are still struggling to gather enough GPU/NPU to train models.
@teortaxesTex @zephyr_z9 It is just a translation error. wanka is "万卡", which means ten thousand.
Capital can accelerate construction, but grid interconnection operates on a fundamentally different timeline. While data center buildouts typically complete within 1–2 years, power integration often requires 5–8 years of coordination and permitting. In that sense, gigawatt-scale capacity remains a long-duration constraint shaped by energy infrastructure cycles.
@teortaxesTex so the timeline goes to at least 3years later before we see true sota model from deepseek?
This suggests they are bypassing prefabricated Huawei systems.
First confirmation that DeepSeek is targeting to build infrastructure from megawatt to *gigawatt* range. And again, as in other listings: it just does not look like they plan to buy Huawei's prefab 950 pods. They are designing their own systems.
