What this is
NVIDIA released two things this week: cuObject and the SCADA Server SDK. In plain terms: until now, when an AI model needed to read data from storage, the data had to pass through the CPU as a "relay station" before reaching the GPU for computation. NVIDIA has now opened a "direct channel" that lets the GPU read storage directly, bypassing the CPU.
The specific pain points are these: AI training, fine-tuning (using extra data to continue training a smaller model), and inference-time data fetching (essentially, the AI looking up references before answering a question) — in all these scenarios, data volumes keep growing and the CPU relay keeps slowing down. NVIDIA claims the new approach can boost data throughput by up to 7x.
Industry view
Our editorial take: this is a signal — the AI industry's bottleneck is shifting from "not enough compute" to "can't feed the data in fast enough."
The bullish camp argues that storage-direct paths will further compress AI inference costs, making on-prem large-model deployments (company-owned servers, not in the cloud) cheaper and faster. This is a new variable for the storage businesses of cloud players like AWS, Alibaba Cloud, and Tencent Cloud.
Skeptical voices are also plentiful. First, cuObject is part of NVIDIA's proprietary ecosystem, locked to CUDA (NVIDIA's GPU programming framework — essentially the "operating system" of the graphics card). Once enterprises adopt it, they become further locked into NVIDIA, weakening their bargaining power against other GPU vendors. Gartner has repeatedly warned about the risk of "excessive vendor concentration" in AI infrastructure.
Impact on regular people
For enterprise IT: Over the next 12–18 months, prices for on-prem AI services should continue to drop, and tech vendors will proactively push "GPU + storage" bundled offerings.
For working professionals: No short-term impact. But if your company is evaluating "should we adopt AI, and which vendor," treat "how much data we have, and whether on-prem deployment is needed" as a decision dimension — don't just look at model benchmarks.
For consumer markets: Cloud-based AI assistants and AI search will feel less laggy, and their ability to serve more users without slowdown will improve.