한국어판: 한국어로 읽기
NVIDIA and CoreWeave are framing Vera Rubin not simply as a faster accelerator, but as infrastructure for a continuous agentic-AI loop: production inference, evaluation, post-training and redeployment.
On September 30, 2026, NVIDIA and CoreWeave announced broader production availability of NVIDIA Vera Rubin NVL72 systems on CoreWeave Cloud. Cognition, the company behind Devin, was named as the first customer running production workloads on the platform.
The announcement is easy to read as another hardware-performance story. The more interesting signal is architectural: the competitive unit is shifting from a single accelerator benchmark toward an entire production loop that links serving, evaluation, reinforcement learning, post-training and redeployment.
Vera Rubin is being sold as a system, not a chip
NVIDIA’s announcement highlights Vera Rubin NVL72 together with Spectrum-X 102.4T Ethernet networking and the Vera CPU. CoreWeave’s own release describes the same platform as a production environment that lets customers move onto the new generation without rebuilding the operating model around it.
That is strategically important for agentic workloads. An AI agent in production generates new traces, failures and evaluation data. Those outputs can feed post-training, model updates and another deployment cycle. Infrastructure that shortens that loop may matter as much as peak benchmark performance.
The performance claims are vendor-reported
CoreWeave says Cognition is seeing a 4.8x increase in total token throughput and reports 3.8x output-token throughput per GPU for reinforcement-learning workloads at matched interactivity versus NVIDIA GB200 NVL72.
Those figures are meaningful as deployment evidence, but they are not independent benchmarks. They come from CoreWeave and NVIDIA, which are the infrastructure vendors promoting the system. Workload choice, software configuration, latency targets and utilization can materially affect the comparison.
CoreWeave has separately reported much larger efficiency gains on specific workloads, including a 10x tokens-per-megawatt result on one DeepSeek R1 evaluation. That number is even more workload-specific and should not be generalized to all AI inference or training economics.
Why “closing the loop” matters
The phrase “training to production” used to describe a mostly one-way path: build a model, deploy it, then repeat the process in a later release. Agentic systems create pressure for a tighter cycle because behavior in production becomes part of the next training signal.
That changes infrastructure priorities. Networking, data movement, serving latency, evaluation pipelines and post-training capacity become more tightly coupled. A cloud provider that can support the whole loop may create switching costs that are operational rather than purely contractual.
What is not established yet
It is too early to conclude that Vera Rubin on CoreWeave has the best economics across clouds or workloads. The available figures are mainly vendor disclosures. Independent comparisons of total cost, reliability, utilization, software portability and real-world latency remain limited.
There is also a broader competitive question. If the infrastructure advantage increasingly comes from the integration of hardware, networking, orchestration and post-training systems, then customers may compare platforms as full AI factories rather than simply compare GPU generations.
The durable signal
The announcement suggests that frontier AI infrastructure is moving toward a closed operational loop: deploy agents, collect evaluation signals, improve models and return them to production faster.
Vera Rubin’s raw performance matters, but the more durable competitive question is whether an infrastructure stack can make that loop faster, cheaper and more reliable across real workloads. That remains to be proven beyond the vendors’ own early evidence.