NVIDIA Blackwell GPUs Slash DeepSeek V4 AI Token Costs by Up to 5x
NVIDIA’s Blackwell GPU platform is gaining major momentum in AI inference, with new software optimizations delivering up to a 5x reduction in cost per token for DeepSeek V4. The improvement arrives just one month after the model’s release, highlighting how quickly NVIDIA is tuning its hardware and software stack for next-generation AI workloads.
Cost per token has become one of the most important measurements in artificial intelligence infrastructure. For companies running large language models, it directly affects total cost of ownership, scaling potential, and real-world profitability. Lower token costs mean AI providers can serve more users, process more requests, and run more advanced models without dramatically increasing expenses.
According to NVIDIA, the latest gains come from continued improvements across its full-stack inference software, designed to extract more performance from Blackwell-based systems such as GB200 and GB300. These optimizations are not limited to raw GPU power. Instead, NVIDIA is combining hardware acceleration, networking, memory management, and inference software into a tightly integrated platform built for high-volume AI deployment.
DeepSeek V4 is one of the models benefiting from these improvements. With Blackwell GPUs and NVIDIA’s inference stack, the model can now run far more efficiently, reducing the cost of generating tokens while improving throughput for demanding tasks such as reasoning, coding, and long-context processing.
Several AI infrastructure and inference companies are already using NVIDIA Blackwell systems to capture these performance gains.
Baseten has used NVIDIA TensorRT-LLM to serve DeepSeek V4 Pro on Blackwell GPUs, targeting reasoning, coding, and long-context workloads. By combining NVIDIA’s open-source inference library with its own runtime optimizations, Baseten reported up to 50% more tokens per second.
Cognition is using NVIDIA Dynamo to manage inference GPUs and scale reinforcement learning workloads. This gives its engineering team a faster way to expand AI training and inference infrastructure without building every layer internally from the ground up.
Deep Infra is relying on NVIDIA’s inference software stack to serve advanced open-source models on Blackwell systems from day one, including DeepSeek V4. The goal is to provide fast, efficient access to frontier AI models while keeping performance high and operating costs under control.
Together AI has used NVIDIA TensorRT-LLM on Blackwell to support production AI endpoints for real-time coding experiences, helping speed up the transition from model optimization to live deployment.
The 5x drop in token cost is the result of multiple layers of optimization working together. NVIDIA describes its inference software stack as a system that connects production operation, application acceleration, and infrastructure access.
At the production level, the platform handles distributed serving, orchestration, autoscaling, and memory management. This allows inference workloads to run across the most suitable compute and storage resources, helping maintain efficiency as demand rises.
At the application level, NVIDIA’s software improves model execution through runtime techniques such as overlapping compute and communication, kernel fusion, and other performance-focused optimizations. These tools give developers room to tune workloads while still benefiting from NVIDIA’s broader software ecosystem.
At the infrastructure level, the stack exposes the capabilities of NVIDIA GPUs, memory, networking, and complete systems without forcing developers to manually manage every low-level device instruction or data-transfer process. This makes it easier for AI companies to deploy powerful models at scale.
Blackwell’s performance gains are also supported by NVIDIA technologies such as NVLink, NVFP4, and Multi-Token Prediction. When combined with system-level software improvements, these features can deliver up to a 20x increase in throughput in certain AI inference scenarios.
For the AI industry, this is a significant development. As models become larger and more complex, simply adding more hardware is not enough. The future of AI infrastructure depends on making every token cheaper to generate and every GPU cycle more efficient. NVIDIA’s latest Blackwell optimizations show how much performance can be gained when hardware, software, networking, and model execution are tuned together.
The result is a stronger position for NVIDIA in the AI inference market and a clearer path for companies looking to deploy large-scale AI models more affordably. With DeepSeek V4 already seeing major efficiency improvements on Blackwell GPUs, the focus now shifts to how quickly these optimizations can expand across more models, more providers, and more real-world AI applications.
