$NVDA Says GPT-6 Astra Ultrafast Delivers Up to 8x Faster Token Generation on Blackwell

$NVDA says OpenAI’s GPT-6 Astra Ultrafast is running on Blackwell infrastructure with token generation speeds up to 8x faster than Astra Standard.
The performance gain highlights how much AI inference is increasingly being optimized at both the model and hardware level.
OpenAI is also using its own models to optimize kernels directly for Nvidia’s Blackwell architecture.
That work is aimed at improving latency, throughput and overall inference cost on $NVDA hardware.
The development shows that frontier AI labs are not simply buying more GPUs.
They are increasingly co-optimizing software, models and low-level compute workloads around specific accelerator architectures.
For $NVDA, that creates an additional advantage beyond raw chip performance.
The deeper AI models become optimized around Blackwell, the more difficult it may be for competing hardware to match the full system-level economics without similar software integration.