The race to bring artificial intelligence agents at an industrial scale enters a new phase: Fujitsu and NVIDIA are expanding their strategic partnership to co-create full-stack AI infrastructure, combining Fujitsu’s Monaka CPUs with the leading GPUs from the acceleration giant, interconnected via NVLink Fusion. The aim is to develop semi-custom systems for industries such as healthcare, manufacturing, and robotics, initially targeting Japan with global deployments in a second phase.
The initiative has two main axes: a rack-scale hardware co-design and a platform for AI agents built on Kozuchi, Fujitsu’s cloud environment launched in October 2024. The partnership also integrates Fujitsu’s AI workload orchestrator with NVIDIA Dynamo, a modular inference framework designed to simplify adoption, reduce latency, and contain costs in enterprise environments.
“The industrial revolution of AI has begun, and we must build the infrastructure to power it—not just in Japan, but worldwide. Fujitsu is a pioneer and trusted leader in supercomputing, quantum research, and enterprise systems,”
explained Jensen Huang, founder and CEO of NVIDIA, stating that both companies are “connecting and extending” their ecosystems for the AI era.
“Our strategic collaboration with NVIDIA will accelerate AI-driven transformation across industries and administrations,”
added Takahito Tokita, CEO of Fujitsu, highlighting that the first vertical will be manufacturing, “where Japan holds a global leadership position,” with plans to expand into HPC and quantum computing.
What’s new? NVLink Fusion opens doors to “semi-custom” architecture
For years, the NVIDIA stack has been perceived as closed; however, the introduction of NVLink Fusion partially breaks this inertia by enabling non-NVIDIA accelerators—specifically, Fujitsu’s Monaka CPUs—to be integrated factory-installed with NVIDIA GPUs in rack-scale solutions. The NVLink bus provides the bandwidth backbone that mitigates bottlenecks between chips, which is critical for AI agents that reason, plan, and act in near real-time.
Practically, the message to clients with mixed workloads (dense computation, shared memory, demanding I/O) is clear: less manual glue, more sustained performance, and predictable latencies. Most importantly, it offers greater freedom to orchestrate CPU+GPU with profiles tailored to sector and application.
Kozuchi: the “upper layer” was already ready
Fujitsu is not starting from scratch on software. Kozuchi, its cloud AI platform, offers AI services and an agent that combines multiple models to extract knowledge from business data. It originally tackled negotiation and profitability cases, and has expanded into production management, legal matters, and other horizontal enterprise tasks.
With this partnership, Kozuchi becomes the front-end of an infrastructure combining Monaka+GPU at the backend, NVIDIA Dynamo for modular inferences, and Fujitsu’s orchestrator to assign workloads to CPU or GPU based on cost, performance, or data protection. For customers, this translates into a single “control plane” that hides complexity while exposing APIs and agent templates centered on processes.
Why “agents” and not just “models”? Three business reasons
- Supervised autonomy. An agent doesn’t just respond: it can decide steps, call tools, read systems (ERP, MES, CRM), draft, and execute actions with controls and rules. For example, in manufacturing, an agent might balance production, reschedule orders, and notify suppliers, without burdening human teams with repetitive tasks.
- Composition. Agents enable chaining models (vision, language, time series) with optimized total cost: reserving GPUs for tasks that must run on GPU, and delegating logic and pre/post-processing to CPU when sufficient.
- Context and compliance. With layered orchestration, it’s possible to restrict what data agents can see, log decisions, and audit outcomes—crucial in healthcare, finance, and administration.
Japan as a launchpad: a natural fit
The initial deployment will focus on Japan, a market with a rich supercomputing tradition, high industrial density, and a national commitment to responsible AI. Fujitsu has been a long-standing provider of HPC and a partner of research centers there, which facilitates pilots involving sensitive data and lengthy validation periods. Building on these references, the plan is to scale out: use cases such as assembly lines, predictive maintenance, plant co-piloting, clinical triage, and warehouse robots are easily adaptable for other regions with appropriate local customizations.
What does each part contribute?
- Fujitsu: Monaka CPUs, Kozuchi (platform and agent), workload orchestration, sector-specific deployment (manufacturing, healthcare, robotics), relationships within the Japanese ecosystem, and a long history in HPC and quantum systems.
- NVIDIA: GPUs, NVLink Fusion (high-bandwidth interconnection for semi-custom designs), Dynamo (modular inferences), along with an ecosystem of libraries and frameworks that already dominate industry.
The resulting full-stack solution provides clients with ready-to-integrate agents compatible with SaaS and on-premises data, supported by a validated reference architecture from both companies.
How does this differ from just “buy GPUs and go”?
Anyone who has transitioned from a PoC to production knows that 80% of the work occurs after the demo: unexpectedly high latencies, skyrocketing token costs, overwhelmed queues, slow data transfer, and compliance/audit friction. The Fujitsu-NVIDIA approach targets exactly those issues:
- Rack-level CPU+GPU coupling with NVLink Fusion → fewer hops, more throughput.
- Orchestration that selects routes (CPU vs. GPU) based on latency/€ → predictable, tunable.
- Dynamo for cutting and recombining inference pipelines without reengineering the stack.
- Kozuchi as an agent layer with enterprise connectors and sector-specific templates.
The promise: faster time-to-value and more controllable TCO than piecemeal DIY solutions.
From factory to hospital: verticals with immediate traction
- Manufacturing: agents that plan production in response to demand shifts, re-calibrate lines, and enrich data with vision to detect defects. Japan’s strong industrial base and culture of continuous improvement make it fertile ground.
- Healthcare: assistants that summarize histories, propose clinical paths, and manage appointments/referrals while respecting privacy and regulations.
- Robotics: copilots integrating sensors, navigation, and reasoning for warehouse, plant, or hospital tasks.
The challenge is not only technical: it requires AI governance, safety metrics (hallucinations, drift), retraining cycles, and training teams.
What it will mean for the market
- NVIDIA’s “semi-custom” approach: NVLink Fusion enables partners like Fujitsu to bring their CPUs onto the same footing as GPUs, reducing “one-size-fits-all” perceptions and giving integrators more control.
- Fujitsu offers AI “agency” hardware co-designed: shifting from just services to including optimized architecture with a core market component.
- Customers gain a smoother path to production: fewer standalone parts, clearer SLAs, and support from two mission-critical oriented companies.
Questions that remain
- Cost: How will the TCO of a Monaka+GPU+NVLink Fusion stack compare to CPU x86 + GPU or GPU-only options?
- Portability: To what extent will Dynamo + Kozuchi enable migration between clouds and on-premises without redoing everything?
- Sovereignty of data: What deployment modes will be offered for sectors with strict regulatory restrictions?
- Partner ecosystem: What ISVs and certified partners will support the initial pilots?
Signal directions: HPC and quantum on the horizon
Both companies anticipate that their collaboration will expand into high-performance computing and quantum realms, areas where Fujitsu has a long-standing presence — and NVIDIA is actively building bridges using their software for simulation and quantum acceleration. The rationale is clear: more capable AI agents will need physical backends involving physics, optimization, and simulation, surpassing purely linguistic approaches.
In summary
- What: Fujitsu and NVIDIA will jointly develop full-stack AI infrastructure for agents, integrating Monaka CPUs and GPUs with NVLink Fusion, along with Kozuchi and Dynamo software components.
- Purpose: For healthcare, manufacturing, robotics, and other sectors where agents can plan and execute tasks under supervision.
- Location: Starting in Japan, with global expansion planned afterward.
- Importance: It’s an industry-ready pathway to move from PoC to production, with predictable performance and embedded AI governance.
The partnership underscores a broader trend: agents will evolve from lab demos to digital operators anchored to real-world processes. Achieving this requires less “magic” and more systems engineering, exactly what this collaboration promises.
Frequently Asked Questions
What is NVLink Fusion and why is it relevant for AI agents?
It extends the NVLink ecosystem, allowing the construction of semi-custom, rack-scale solutions that connect non-NVIDIA accelerators (e.g., Monaka) with GPUs. For agents, this translates into wider CPU↔GPU bandwidth, lower latency, and a smoother pipeline.
What role does Kozuchi play in the Fujitsu-NVIDIA partnership?
It’s Fujitsu’s AI platform, providing agent services, data management, and system integration. It orchestrates, together with Dynamo and Fujitsu’s orchestrator, which workloads run on CPU or GPU based on cost/latency.
Which sectors will be first to see “production-level” agents?
The initial focus is on manufacturing in Japan, with plans to extend to healthcare and robotics, enabling agents to automate decision-making (replanning, triage, maintenance) with controls and audits.
How can companies start simply?
Using validated agent templates and reference architectures (CPU+GPU), along with Dynamo for modular inference, and Fujitsu’s orchestrator for deployment on cloud or on-premises, supported by clear SLAs and cost metrics.
via: sdxcentral