NVIDIA is tackling one of artificial intelligence's biggest headaches: the staggering amount of electricity data centers need to run AI workloads. The company just announced two major tools designed to squeeze more computing power out of the same amount of energy. The first is Vera, a new processor built specifically for AI agents, and the second is DSX OS, software that helps data center operators manage power consumption across entire facilities. Together, they signal a shift in how the AI industry thinks about energy constraints.

What Is NVIDIA Vera and Why Does It Matter for Data Centers?

Vera is NVIDIA's first CPU (central processing unit) designed from the ground up for AI agents, which are AI systems that can take actions, run code, and evaluate results rather than just answer questions. The processor completes tasks 1.8 times faster than traditional x86 CPUs, the industry standard for decades. This speed matters because faster task completion means data centers can accomplish more work without running their hardware longer, directly reducing energy consumption per unit of work.

The chip is built on NVIDIA's Olympus core architecture and includes 88 processing cores, along with a memory system that can move data at speeds up to 1.2 terabytes per second. For context, that's roughly equivalent to transferring the contents of 300,000 high-definition movies in a single second. Major AI labs and cloud providers are already planning to adopt Vera, including Anthropic (the company behind Claude), OpenAI, SpaceX's AI division, and hyperscalers like ByteDance and Oracle Cloud Infrastructure.

"AI agents will be the largest users of computing. Vera is the first CPU designed for that future, built to run agentic AI at hyperscale with extraordinary performance, efficiency and programmability," said Jensen Huang, founder and CEO of NVIDIA.

Jensen Huang, Founder and CEO, NVIDIA

How Can Data Centers Reduce Power Consumption Without Slowing Down?

NVIDIA's new DSX OS software addresses the core problem facing AI data centers: power is the limiting factor. Even with unlimited computing hardware, most facilities hit a hard ceiling on how much electricity they can draw from the grid. DSX OS solves this by treating power as a programmable resource that can be dynamically allocated across GPUs, racks, cooling systems, and workloads in real time.

The software includes several interconnected tools that work together to maximize efficiency:

• DSX MaxLPS: Dynamically enforces power policies at multiple levels, recovering "stranded" capacity that would otherwise go unused and enabling data centers to run up to 40% more GPUs at peak efficiency within a fixed power budget, with minimal impact on inference workload performance.
• DSX Flex: Connects AI workloads to grid services so data centers can automatically adapt to demand response events, load shedding requests, and renewable energy availability, helping operators respond to grid conditions in seconds rather than hours.
• DSX Exchange: An MQTT-based communication hub that makes facility-level signals such as grid events, thermal data, and power anomalies visible to the software managing the rest of the AI factory, enabling real-time coordination between compute, networking, power, and cooling systems.

The 40% efficiency gain is particularly significant because it means data centers can serve 40% more AI workloads without requesting additional power from the grid. For a facility already operating near its power limit, this is transformative. Partners including CoreWeave, Lambda, and Nscale are already deploying MaxLPS, while companies like ENGIE and the UK National Grid are testing DSX Flex.

What Does This Mean for Europe's AI Infrastructure?

The timing of these announcements coincides with a major expansion of AI data center capacity in Europe. SoftBank Group and Sesterce announced plans to build a 1-gigawatt AI data center campus in Bosquel, France, which will be one of Europe's largest dedicated AI computing facilities. A gigawatt is enough power to supply roughly 750,000 homes, illustrating the scale of energy demand these facilities require.

The Bosquel campus is designed to leverage France's advantages for AI infrastructure, including advanced grid infrastructure, engineering talent, and industrial land availability. Once operational, the facility will create 400 permanent jobs in data center operations, energy systems, security, and infrastructure management. The project also includes a 10 million euro endowment fund to promote AI adoption in local businesses, schools, and universities.

"AI will shape the next era of technology, industry and human progress, and that future will require a new generation of infrastructure. Bosquel is an important step in SoftBank's commitment to helping build that foundation in France," said Masayoshi Son, Chairman and CEO of SoftBank Group Corp.

Masayoshi Son, Chairman and CEO, SoftBank Group Corp.

How Are Data Centers Shifting From Cost-Per-Compute to Cost-Per-Token?

The economics of AI data centers are fundamentally changing. Historically, data center operators optimized for "cores per dollar," meaning they wanted as many processing cores as possible for their investment. Today, the metric that matters is "tokens per dollar," where a token is a small unit of text or data that an AI model processes. This shift reflects the reality that AI workloads are increasingly about throughput and efficiency rather than raw processing power.

Vera addresses this directly by completing agentic workloads, data processing, and orchestration tasks faster and more efficiently than previous CPUs. The processor is optimized for the kinds of work that sit on the critical path of modern AI factories, including code compilation, Python execution, Java processing, and database operations. According to benchmarks from Phoronix, an open-source testing suite, Vera delivered the fastest overall performance across these agentic workloads.

The shift to tokens-per-dollar economics also explains why DSX OS's power optimization features are so valuable. By enabling data centers to run 40% more GPUs within the same power budget, the software directly improves the tokens-per-watt metric, which translates to lower costs for AI model training and inference. This is especially important as AI models grow larger and more complex, requiring exponentially more computing power to train and deploy.

When Will These Technologies Be Available?

Vera systems will be available from system builders and cloud partners starting in the fall of 2026. Major manufacturers including Dell Technologies, HPE, Lenovo, and Supermicro are building standalone Vera CPU systems at scale, along with numerous other system integrators. Cloud service providers planning to deploy Vera include Akamai, ByteDance, Cloudflare, CoreWeave, and Oracle Cloud Infrastructure.

DSX OS components are already being released as open-source software, allowing ecosystem partners to integrate them into existing platforms without rebuilding from scratch. This approach eliminates months of custom development and accelerates the deployment of power-efficient AI infrastructure across the industry.

The convergence of faster CPUs, intelligent power management software, and massive new data center capacity suggests that the AI industry is moving past the acute energy crisis that threatened to constrain growth. However, the challenge remains enormous: as AI models continue to grow and demand for AI services accelerates, data centers will need to find ever more creative ways to maximize computing output per unit of energy consumed.