The GPU giant's latest software stack delivers up to 20x more tokens per dollar on existing Blackwell hardware, a move that matters for every crypto project betting on AI infrastructure.
Nvidia just proved that sometimes the best hardware upgrade is better software. The company revealed on July 1 that its optimized inference stack can slash token costs by up to 5x for DeepSeek V4 running on Blackwell systems, with throughput improvements reaching as high as 20x compared to baseline configurations on the exact same chips.
Those gains didn’t come from a new chip announcement or a fresh silicon architecture. They came from roughly one month of software engineering on open-source frameworks like vLLM and SGLang, plus a cocktail of techniques including disaggregated serving, NVLink expert parallelism, NVFP4 precision, and multi-token prediction.
What the numbers actually mean #
Nvidia is increasingly pushing two metrics it wants the industry to obsess over: tokens per dollar and tokens per watt. Both measure how much useful AI output you squeeze from a given investment in hardware and electricity.
Enterprise inference provider Baseten, collaborating with Nvidia on the rollout, demonstrated a 50% increase in tokens per second using TensorRT-LLM on DeepSeek V4 Pro systems. That’s a more conservative number than the headline figures, but it reflects real-world production conditions rather than lab benchmarks.
All of these improvements maintain the strict latency targets required for agentic AI applications, which is the category of AI that can autonomously execute multi-step tasks.
Why crypto should pay attention #
Projects like Render, Akash, and io.net have built their value propositions around the idea that GPU compute is expensive and scarce, making decentralized alternatives attractive. When Nvidia can deliver 5x more output from existing centralized hardware through software alone, the competitive moat for decentralized compute networks narrows.
The rise of agentic AI is one of the hottest narratives in crypto right now. Cheaper inference makes those agents dramatically less expensive to run, but it also means the infrastructure moats some crypto-native AI projects claim to offer may be thinner than investors assume.
Nvidia’s CUDA ecosystem has been locking in developers and enterprises for over a decade. Every time the open-source community builds a framework like vLLM or SGLang, Nvidia optimizes it for its own hardware, creating a flywheel that’s extremely difficult for competitors, centralized or decentralized, to break.
What this means for investors #
The shift toward cost-per-token economics validates a thesis that many DePIN projects have been quietly building around: that the real value in AI infrastructure isn’t the raw GPU, it’s the software layer that orchestrates workloads efficiently. The one-month timeline Nvidia cited for achieving these gains suggests the pace of software-driven improvement is accelerating.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our