NVIDIA has moved its next-generation rack-scale AI system, Vera Rubin NVL72, into production. Racks are already running in data centers at CoreWeave, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, backed by a supply chain spanning more than 350 factory sites across 30 countries[1]. On the same day, Microsoft and Mistral announced a European partnership that will draw on thousands of these GPUs[2].
A 10x claim on tokens per megawatt
The metric NVIDIA leads with for Vera Rubin is not peak performance but how many tokens a system can process per megawatt. For AI data centers that face hard limits on available power, that number maps more directly onto expansion decisions.
CoreWeave, an early adopter, benchmarked the platform on DeepSeek-R1 and reported 10x more throughput per megawatt than the previous-generation Grace Blackwell NVL72[1]. NVIDIA states that compared with GB200 NVL72, Vera Rubin NVL72 delivers up to 10x more tokens per megawatt and one-tenth the cost per million tokens[1].
The context is that agentic systems can consume up to 15x more tokens than traditional AI applications[1]. If consumption grows while unit cost stays flat, operating budgets hit their ceiling before compute capacity does.
All of these figures come from NVIDIA and its partners. The comparison conditions and model serving configurations have not been published, so there is no guarantee they carry over to other workloads. Independent verification is still pending.
Seven chips designed as a single system
NVIDIA says Vera Rubin was engineered as one system from the start, spanning seven chips and five rack trays, rather than assembled from separate off-the-shelf parts[1].
At its center is the new NVIDIA Vera CPU. Built around a custom core called Olympus and aimed at AI factories in the agent era, it delivers 2x single-threaded performance, 3x core-to-core bandwidth and 40 percent lower memory latency versus competing chiplet designs[1]. Preprocessing, scheduling and agent control loops that run alongside the GPUs are exactly the workloads where single-threaded performance shows up.
On the networking side, sixth-generation NVLink handles scale-up with more than 2x the throughput on complex workloads and 3x lower latency[1]. Scale-out combines 102.4Tbps Spectrum-6 switch systems with 1.6Tbps ConnectX-9 SuperNICs, which NVIDIA says yields 1.6x higher RDMA bandwidth than off-the-shelf Ethernet[1]. Spectrum-XGS extends this across locations with 1.9x multi-site throughput[1].
NVLink Fusion also opens the NVLink scale-up stack to third-party XPUs[1]. Rather than locking the platform to NVIDIA silicon alone, it leaves a door open for outside accelerators.
No cables or fans in the tray
The most visible physical change is that the compute tray has no cables, fans or hoses. NVIDIA credits three generations of rack-scale co-design and says tray assembly time drops from hours to one minute[1]. At the scale of tens of thousands of units, that difference lands directly on deployment schedules.
The cooling design changed as well. A 45 degrees Celsius liquid cooling inlet temperature allows chiller-free operation using dry coolers alone[1]. For new AI factories, pairing that with a closed-loop liquid cooling system saves millions of gallons of water per megawatt annually[1]. Water access is becoming a siting constraint for data centers, which makes this an argument separate from raw performance numbers.
In Europe, it becomes Mistral's compute base
On July 21, Microsoft and Mistral announced an expanded strategic partnership[2]. The multibillion-dollar agreement (roughly several hundred billion yen) focuses on expanding AI infrastructure in Europe, with Mistral adding thousands of Vera Rubin GPUs to its capacity and offering a shared platform for training, inference and large-scale deployment[1][2]. Microsoft will use Mistral's expanded European GPU infrastructure to support delivery of its own cloud and AI services[2].
※1 USD = 162 JPY (as of July 21, 2026)
On the model side, Mistral Medium 3.5 and OCR 4 are now available in Microsoft Foundry, and Mistral Medium 3.5 is now in Microsoft Copilot Studio[1][2]. Through Azure Local and Foundry Local, customers can run the same models and tools across cloud and customer-controlled environments[1].
The target is regulated sectors such as finance, healthcare and manufacturing[2]. In domains where data residency and operational procedures are governed by regulation, a capable model alone rarely wins the deal; what matters is whether the vendor can offer everything from public cloud to fully disconnected environments. For NVIDIA, supplying compute to Europe's open-model push secures one clear outlet for its production ramp.
Summary
Vera Rubin NVL72 has entered production and is already running at four major cloud providers. The key points are a supply chain of more than 350 sites in 30 countries, CoreWeave's benchmark of 10x throughput per megawatt, one-minute tray assembly and a chiller-free cooling design. In Europe, the platform underpins a multibillion-dollar Microsoft and Mistral partnership. Every performance figure here comes from the vendors involved, so independent verification remains worth waiting for.
