Thanks so much for this analysis - it is super helpful. Quick question regarding where you got the information for the specs. You mentioned that Kunlunxin uses a VLIW architecture instead of GPU design, GDDR7 instead of 7nm, and SMIC 7nm manufacturing. I thought Samsung was Baidu's foundry partner but, I could be wrong. I have found it difficult to find specs on the P800.
Good catch. GDDR6 is correct, not GDDR7. That was my error.
On Samsung: they fabbed first-gen Kunlun on 14nm. Gen 2 moved to TSMC 7nm. For P800, after TSMC and Samsung were restricted from supplying Chinese AI chipmakers in late 2024, current production likely uses SMIC's 7nm.
P800 specs are hard to find. Most comes from Baidu conferences, Chinese securities research, and industry newsletters. No official datasheet exists.
On architecture: XPU isn't a GPU design, and calling it "VLIW" oversimplifies it. It's Baidu's proprietary many-small-core architecture.
The idle chips framing is genius because it flips the conventional wisdom on waste. When people see $14M sitting unused they assume dysfunction, but the three-layer structure you mapped shows it's actually strategic patience. I ran into similiar dynamics working with overseas expansion teams where compliance costs can dwarf efficency losses short-term. One thing worth tracking is whether non-telecom external revenue breaks 20% becuase that's when we'll know if the market actually wants sovereign compute or just needs it.
Thanks so much for this analysis - it is super helpful. Quick question regarding where you got the information for the specs. You mentioned that Kunlunxin uses a VLIW architecture instead of GPU design, GDDR7 instead of 7nm, and SMIC 7nm manufacturing. I thought Samsung was Baidu's foundry partner but, I could be wrong. I have found it difficult to find specs on the P800.
Good catch. GDDR6 is correct, not GDDR7. That was my error.
On Samsung: they fabbed first-gen Kunlun on 14nm. Gen 2 moved to TSMC 7nm. For P800, after TSMC and Samsung were restricted from supplying Chinese AI chipmakers in late 2024, current production likely uses SMIC's 7nm.
P800 specs are hard to find. Most comes from Baidu conferences, Chinese securities research, and industry newsletters. No official datasheet exists.
On architecture: XPU isn't a GPU design, and calling it "VLIW" oversimplifies it. It's Baidu's proprietary many-small-core architecture.
Thanks for the technical rigor.
The idle chips framing is genius because it flips the conventional wisdom on waste. When people see $14M sitting unused they assume dysfunction, but the three-layer structure you mapped shows it's actually strategic patience. I ran into similiar dynamics working with overseas expansion teams where compliance costs can dwarf efficency losses short-term. One thing worth tracking is whether non-telecom external revenue breaks 20% becuase that's when we'll know if the market actually wants sovereign compute or just needs it.