PyyanThe AI Index

Pyyan / Compare / TPU v7 Ironwood vs B300 Blackwell Ultra vs B200 vs MI400 vs Ascend 950DT

TPU v7 Ironwood vs B300 Blackwell Ultra vs B200 vs MI400 vs Ascend 950DT

5 of 5

GPU & AI Chips · verified 24 Sept 2026

×TPU v7 IronwoodGooglecurrent
×B300 Blackwell UltraNVIDIAcurrent
×B200NVIDIAcurrent
×MI400AMDcurrent
×HAscend 950DTHuaweicurrent
SpecificationTPU v7 IronwoodB300 Blackwell UltraB200MI400Ascend 950DT
SummaryInference-oriented, Google Cloud only.15 PFLOPS dense FP4.The current volume Blackwell part.Double NVIDIA's memory capacity.144GB on Huawei's own HBM, and the part China is building gigawatts around.
Memory192GB HBM3e288GB HBM3e192GB HBM3e432GB HBM4144GB HiZQ 2.0
Bandwidth7.37 TB/s8 TB/s—19.6 TB/s4 TB/s
Price—~$40–50K (verify)$30–50K—Above ¥250,000 (about $37K), reported
VendorGoogleNVIDIANVIDIAAMDHuawei
CategoryGPU & AI ChipsGPU & AI ChipsGPU & AI ChipsGPU & AI ChipsGPU & AI Chips
OfficialGoogle ↗NVIDIA ↗NVIDIA ↗AMD ↗Huawei ↗

Highlighted rows are where these differ.

TPU v7 Ironwood

  • Requires JAX
  • Not purchasable or on any marketplace

Best for GCP inference.

Full spec sheet →

B300 Blackwell Ultra

  • 8 TB/s memory bandwidth
  • Per-GPU pricing widely misquoted — one source cites rack pricing

Best for frontier training.

Full spec sheet →

B200

  • Sold out through mid-2026
  • ~3.6M unit backlog

Best for general training.

Full spec sheet →

MI400

  • 19.6 TB/s bandwidth
  • 40 PFLOPS FP4
  • $7.2B in commitments

Best for large-model inference.

Full spec sheet →

Ascend 950DT

  • 4 TB/s memory bandwidth and a 2 TB/s interconnect, 2.5x the Ascend 910C
  • HiZQ 2.0 is Huawei's own memory standard rather than a licensed HBM3 or HBM3e part
  • Indicated price rose 20% to 50% in two months to September 2026, on a memory shortage
  • Successor Ascend 960DT pulled forward to Q1 2027, Huawei said on 17 September

Best for serving models where Nvidia cannot be bought.

Full spec sheet →