Nvidia's new chips let one AI coding agent run on thousands of GPUs

Nvidia and cloud provider CoreWeave made two new pieces of hardware generally available: the Vera Rubin NVL72 GPU system and the Vera CPU, both purpose-built specifically for AI agent workloads rather than standard AI model training or chat-style inference. Alongside the hardware, CoreWeave launched Forge, an integrated environment for training, evaluating, and continuously improving AI models and agents on that infrastructure.

Why it matters: AI agents, systems that take multi-step actions on their own rather than just answering a single question, need something fundamentally different from a chatbot: thousands of isolated environments running in parallel, long context windows, and very low latency for real-time reasoning. Building hardware specifically around those demands, rather than adapting general AI chips, is a bet that agentic AI is now big enough to deserve its own dedicated infrastructure.

Independent benchmarks backed up the performance claims with real numbers:

• 4.8x higher token throughput on Vera Rubin NVL72 versus the prior GB200 NVL72 generation, measured by Cognition for software engineering inference.
• 1.7x better performance on passing tasks, according to Terminal-Bench testing.
• 3x faster agent sandbox startup times on Vera CPU, per CoreWeave.
• 11,000+ concurrent agent environments supported in a single rack.

Why it matters: startup time and concurrency aren't just abstract benchmarks, they directly determine how many AI agents a company can realistically run at once, and how quickly those agents can be evaluated and improved through reinforcement learning loops that depend on running thousands of test environments in parallel.

Cognition, the company behind the AI coding agent Devin, became the first production customer running on Vera Rubin NVL72, scaling its deployment to thousands of GPUs within nine months. Other early adopters named in the announcement include Canva, Capital One, MasterClass, and Ennoble Care, a healthcare AI inference company.

In real life it's the difference between a car company showing off a concept engine on a test bench versus a real customer already driving that engine cross-country every day. Devin running at that scale is the cross-country drive.

Why it matters: Cognition's own founding team member, Silas Alberti, described exactly why this matters for a workload like Devin's: "Agentic coding is a complex workload: long contexts, high concurrency and token volumes where cost per token decides what we can ship." Real production use at that scale, not just a benchmark demo, is the strongest evidence the hardware actually delivers on its purpose-built claims.

CoreWeave says serverless reinforcement learning on the new infrastructure trains models 1.4x faster at roughly 40% lower cost compared to companies managing their own infrastructure setups. Nvidia VP Ian Buck framed the broader strategy behind that efficiency: "NVIDIA accelerated computing delivers value across generations... infrastructure that keeps earning for years, and the flexibility to put the right GPU on the right workload."

Background: Nvidia described its relationship with CoreWeave as a nearly decade-long partnership, with CoreWeave maintaining strong performance across multiple prior generations of Nvidia's hardware. This launch continues that pattern rather than representing a brand-new relationship built just for agentic AI specifically.

Forge is CoreWeave's new integrated environment specifically for training, evaluating, and iteratively improving AI models and agents. The idea is to close what Nvidia calls the loop between training, evaluation, and real production use, letting a company see how an agent actually behaves once deployed and feed that behavior directly back into further training, all on the same underlying hardware stack.

Why it matters: without that closed loop, companies training AI agents have historically needed to stitch together separate systems for training, testing, and running agents in production, adding complexity and slowing down how quickly an agent's real-world mistakes can be corrected through retraining.

Who's affected: this is the infrastructure layer sitting underneath AI tools that increasingly act on your behalf, coding assistants, customer service agents, task-management tools, rather than just answering a question and stopping. Faster, cheaper, more reliable agent infrastructure tends to filter down over time into AI products that respond quicker, cost less, and make fewer mistakes for the people actually using them.

The scale here is also a useful signal in itself: when a single AI coding agent can scale to thousands of GPUs in under a year, and a single server rack can juggle over 11,000 concurrent agent sandboxes, it's a concrete measure of how quickly the AI agent category has moved from an interesting idea to genuinely heavy industrial infrastructure.

By the numbers: Cognition's growth mirrors the infrastructure scaling described in this announcement. The company's valuation climbed from $350 million in early 2024 to $4 billion by March 2025, then $10.2 billion by September 2025, and reportedly entered funding discussions at $40 billion or more by August 2026. Devin's own annualized revenue reportedly surged from $1 million in September 2024 to $73 million by June 2025.

Background: Cognition also acquired Windsurf, an AI-native coding tool, in July 2025, after a separate deal for Windsurf to join Google fell through. That acquisition roughly doubled Cognition's annual revenue and gave it a combined coding assistant and development environment, rather than just a standalone AI agent.

Why it matters: that trajectory helps explain why Cognition needed purpose-built agent infrastructure in the first place. A company whose core product went from a $1 million revenue run rate to tens of millions within a year, while its own valuation multiplied more than tenfold, needs hardware that can scale at a similar pace, or risks becoming infrastructure-constrained right as demand for its product accelerates.

• This is backend infrastructure, not a product you'll use directly, but it shapes how capable and responsive AI agent tools you do use might become over time.
• Devin's production scaling is a real-world proof point, not just a benchmark. If you use AI coding tools, infrastructure improvements like this are part of what makes those tools faster and more reliable.
• Lower training costs for AI agents could eventually mean more competition and lower prices for AI agent products aimed at consumers and businesses alike.
• Watch for more companies following Cognition's pattern, of an AI agent product scaling rapidly once purpose-built infrastructure like this becomes available to support it.