Cadence (Nasdaq: CDNS) has announced that Nokia is deploying the Cadence AuraStack AI Super Agent on the Cadence Millennium M2000 AI Supercomputer to advance electronic system design. The platform unifies three layers — an agentic AI layer, Cadence’s engineering software, and GPU-accelerated computing — in a secure, on-premises system. Cadence and Nokia say the deployment accelerates key design and analysis workflows by 5X and compresses design and analysis cycles from months to days, while keeping sensitive engineering data and intellectual property inside Nokia’s own environment.
The announcement is notable less for any single capability than for what it says about where AI is heading in the semiconductor industry. The first generation of AI in chip design was assistive — tools that suggested or checked what an engineer had already decided. The second generation, of which this deployment is an example, is autonomous: long-running agents that pursue a goal, evaluate intermediate results and iterate toward closure, with engineers supervising outcomes rather than executing steps.
The point of the platform, in Cadence’s framing, is to let Nokia automate engineering tasks, explore complex system trade-offs and close multiphysics analysis loops across heterogeneous electronic systems. “Cadence is helping us rethink how we approach electronic system design and heterogeneous integration,” said Derek Urbaniak, Nokia’s head of SoC development. “The integrated platform enables our teams to evaluate complex tradeoffs and close multiphysics loops in a fraction of the time, while keeping sensitive engineering data and IP under our control.”
The three-layer platform
Cadence describes the system as three layers of AI-driven engineering. The agentic AI layer is the AuraStack AI Super Agent, built with NVIDIA Nemotron open models, which automates and coordinates design tasks including schematic capture, placement, routing and multi-objective optimisation. The computational layer is Cadence’s EDA and system-design software, which uses GPU-optimised NVIDIA CUDA-X libraries — including cuDSS — and NVIDIA Omniverse to accelerate power integrity, thermal analysis, electromagnetic characterisation and circuit simulation. The hardware layer is the Millennium M2000 itself, powered by NVIDIA Blackwell GPUs, providing a shared, scalable foundation that can run multiple agentic engineering workloads in parallel.
GPU-accelerated Cadence applications available on the platform include the Celsius Thermal Solver, the Clarity 3D Solver, the Quantus FS and EMX parasitic extractors, the Spectre X circuit simulator and the Voltus IC Power Integrity solution. The architecture matters because it addresses the central worry about autonomous agents in engineering: trust. An agent that reasons over a design hierarchy is only useful if its suggestions can be checked against signoff-accurate results, so Cadence keeps the agents coupled to the same physics-based solvers — thermal, electromagnetic, power-integrity and circuit-simulation engines — that engineers already rely on. The claim is that AI-directed actions stay grounded in proven computational models rather than plausible-looking guesses. In other words, the agents are not floating free: they sit on top of the physics-based solvers that engineers already trust for signoff-accurate results.
Why on-prem matters
The distinctive part of this deployment is not the agent but where it runs. Agentic design touches a company’s most valuable intellectual property, and the Cadence platform runs entirely inside the customer’s environment rather than through a public cloud. “AI gives engineering teams the speed to explore more designs and the insight to make better tradeoffs, while keeping proprietary data and expertise under their control,” said Tim Costa, NVIDIA’s vice president and general manager for industrial and computational engineering.
For the industry, the on-premises model is a meaningful data point. Much of the AI boom runs in public clouds, but the most sensitive engineering workloads have resisted that shift because chip and system designs are among the most valuable trade secrets a company holds. A vendor that can deliver frontier agentic capability on hardware inside the customer’s firewall removes a common objection — and it is why the same “keep the data home” logic appears in telecom, defence and regulated-industry AI deployments.
That is a governance and security argument as much as a performance one, and it is the same logic now driving “sovereign AI” across the enterprise: the most sensitive workloads stay on infrastructure the customer controls. For a systems company like Nokia, which designs networking silicon and heterogeneous systems, keeping design data and models in-house is a precondition, not a preference. The modular Millennium architecture, Cadence adds, lets Nokia expand capacity as requirements grow.
The agentic-EDA race
The deployment lands in a category moving quickly. Cadence and NVIDIA have expanded their partnership steadily through 2026, combining agentic AI, physics-based simulation and digital twins; Cadence extended its ChipStack AI Super Agent, aimed at RTL design and verification, to what the company calls “Level-5” autonomy, and NVIDIA has adopted Cadence’s AgentStack orchestration flow in its own design work. The Millennium M2000, built on NVIDIA’s accelerated computing, reportedly delivers up to 80X greater throughput and up to 20X lower power than a CPU-based predecessor.
Cadence is not alone. At DAC 2026, Synopsys announced its own fully autonomous, long-running agentic workflows — a design-verification agent claiming up to 50X faster time-to-validated RTL, an autonomous thermal-analysis workflow, and more than 20 GPU-accelerated EDA and multiphysics products — also built with NVIDIA Nemotron and secured by the NVIDIA OpenShell runtime. The two largest EDA vendors are, in effect, racing to put long-running agents on top of GPU-accelerated solvers, and both are leaning on NVIDIA’s stack. There is also a hardware dimension to the race. Agentic workflows are compute-hungry — long-running agents that iterate over thousands of simulations consume GPU hours at a scale that only accelerated infrastructure can supply. That is why the vendors’ agentic ambitions are inseparable from their GPU partnerships, and why a customer like Nokia adopting an on-premises supercomputer is as much a compute decision as an AI one.
Competition is Intensifying
The competition is intensifying precisely because the stakes are high. EDA is a duopoly at the leading edge, and both incumbents see agentic AI as the next platform shift — a way to compress design cycles and, in turn, to justify new, GPU-heavy infrastructure spend. That alignment of incentives explains why the vendors are moving fast and why the performance claims, all in the same dramatic register, should be read with care until independent benchmarks appear. The competitive pressure is why a named customer like Nokia matters: in a category where every vendor claims dramatic speed-ups, deployments are the proof points buyers look for.
What to watch
Three things will show whether the deployment delivers. First, whether Cadence or Nokia publish independent benchmarks for the 5X figure, since performance claims in this category are plentiful and comparable data is scarce. Second, whether the agentic workflows reach production signoff or stay in supervised pilots, which is the line between a productivity tool and a redesign of the engineering process. Third, whether the on-premises, IP-protective model proves a durable advantage for agentic AI — or whether buyers come to expect it as standard and the differentiation shifts back to the quality of the agents and solvers.
For readers tracking the category, the practical question is which claims survive scrutiny. A 5X workflow acceleration and a months-to-days compression are the kind of numbers that sound decisive in a press release but depend heavily on the baseline, the workflow and the workload. Until a customer or a third party publishes measured results, the safest reading is that agentic design is promising, is being adopted by serious buyers, and has not yet been shown to deliver its headline gains under independent test.
A fourth signal is cost. On-premises supercomputers are capital-intensive, and the case for them rests on sustained, heavy usage rather than occasional bursts. Whether Nokia’s deployment expands, stays static or shrinks will say as much about the economics of agentic design as any speed-up figure — and it will be the number other buyers watch.
What the deployment does and doesn’t establish
It establishes a named customer deployment of an agentic AI platform on an on-premises supercomputer, a three-layer architecture spanning agents, EDA software and GPU hardware, and a specific list of accelerated applications. It does not establish independent performance results, the breadth of Nokia’s production use, or that agentic design flows are more reliable than the supervised, engineer-in-the-loop approaches they extend.
The honest summary: Cadence has landed a marquee customer for agentic AI in electronic system design, and the architecture is coherent — agents on top of trusted solvers, running on hardware inside the customer’s walls. The performance figures are the companies’ own, and the category is crowded, with Synopsys pursuing the same direction on the same NVIDIA foundation. What Nokia’s deployment shows most clearly is where the EDA industry is placing its bet: that the next productivity leap in chip and system design will come not from a better tool but from long-running agents, grounded in physics, kept close to the IP they help create. It is a bet that the design of the next generation of chips and systems will be done, increasingly, by agents that engineers direct rather than tasks they personally perform.
Editor’s Note: This article draws on the Cadence release and the company’s newsroom materials, plus independent coverage by HPCwire/AIwire, Investing.com and Engtechnica, and context from NVIDIA’s blog and Synopsys’s DAC 2026 announcements. The platform, architecture and availability are as described by Cadence; the 5X and months-to-days performance figures are company claims and are not independently verified, and Nokia reports a deployment rather than measured outcomes.