FuriosaAI makes chips and software that run trained AI models in data centers. Its customers include cloud providers, enterprises and AI application companies that need to generate model responses repeatedly at an acceptable cost and speed. This stage is called inference, distinct from the computing used to train a model initially.

RNGD, the company's second-generation accelerator, is installed in a server as a PCIe card. A customer's engineering team deploys a supported model using Furiosa's software stack, connects it to an application and sends requests to the server. The accelerator performs the mathematical operations needed to produce responses, while the software manages how the model and requests use its computing and memory resources.

The chip uses Furiosa's Tensor Contraction Processor architecture, designed to execute common AI calculations directly and reduce inefficient movement of data. It is manufactured on TSMC's five-nanometer process and uses high-bandwidth memory. Those choices are intended to improve useful work per unit of energy, but the economics still depend on the particular model, response latency and how fully the hardware is used.

Software is necessary to make the hardware usable. Furiosa provides a compiler and serving tools, including support for running inference within Kubernetes-managed infrastructure. Its updates add techniques such as caching repeated prompt content and scheduling requests together, so customers can improve throughput without replacing the physical card.

RNGD entered mass production in January 2026. Samsung SDS uses it for an NPU-as-a-Service offering, LG Group companies deploy it in on-premises and hybrid-cloud environments, and Upstage uses it for live AI services. Furiosa expects to produce 20,000 accelerators in 2026 and more than 40,000 in 2027; these figures are production plans rather than completed shipment totals.

The Broadcom partnership addresses the next generation of the system, including chip technology and the connections needed across servers and racks. It is a development roadmap, not evidence that the third-generation product is already shipping.

FuriosaAI sells inference accelerators into infrastructure deployed by cloud providers and enterprise customers. Its software enables those buyers to run supported models on the chips, making the commercial proposition a working computing platform rather than silicon alone. Public announcements do not establish a universal per-chip price or a separate recurring software fee.

The company designs the architecture and software while relying on manufacturing and component suppliers for physical production. Its costs include semiconductor fabrication, memory, packaging, boards, engineering, model optimization and support. Committing to more production can improve availability, but also requires capital and confidence that buyers will deploy the resulting hardware.

For a cloud provider, the accelerator must support a service that customers will pay to use. For an enterprise, it must fit the organization's applications and infrastructure while delivering enough performance and efficiency to justify switching. Power efficiency matters because electricity and cooling contribute to operating costs, although lower chip power alone does not prove a lower total cost for every workload.

Furiosa's software investment is therefore part of value capture. A card that is difficult to program, cannot support a customer's model or sits underused may fail commercially even if its architecture is efficient. Expanding model coverage and deployment tools can reduce the work required for a buyer to move from a demonstration to a production system.

FuriosaAI competes with Nvidia, other accelerator vendors and cloud platforms developing their own AI chips. Buyers compare the complete cost and reliability of running their models, including software compatibility, utilization and engineering effort, rather than chip specifications in isolation.

Nvidia's GPUs benefit from a broad developer ecosystem, mature software and widespread availability through cloud providers and server manufacturers. Customers can use the same general-purpose platform for many training and inference workloads, reducing the need to adopt another toolchain.

Furiosa narrows its focus to inference efficiency. To win, it must show that savings for a customer's actual workload outweigh migration, integration and support costs, while keeping pace with new model architectures and Nvidia's own improvements.

AMD offers GPUs with a broader computing role, while companies such as Groq, Cerebras and SambaNova pursue different architectures and infrastructure models for AI execution. These alternatives compete for customers seeking performance, capacity or economics beyond Nvidia's default platform.

Furiosa offers deployable inference hardware and a supporting software stack for enterprises and infrastructure providers. Its position depends on making that combination useful across enough real applications, rather than relying on an advantage demonstrated only for one model or benchmark.

Large cloud providers develop their own accelerators and can integrate them with existing software, procurement and data-center operations. Their scale can make those platforms an attractive choice for customers already consuming the provider's cloud services.

Furiosa's opportunity includes organizations wanting more control over where models run, such as on-premises or hybrid deployments. That independence can differentiate its offering, but Furiosa must still supply the deployment experience and support that a large cloud platform bundles into its service.

FuriosaAI's production hardware and software provide a foundation for reaching additional inference buyers beyond its initial deployments. The opportunity spans enterprise infrastructure, third-party clouds and national computing projects, with each market requiring more than simply producing additional cards.

LG Group's deployments provide a reference for organizations that want AI models to run within infrastructure they control. Buyers in this market allocate budgets to internal applications, data handling and computing capacity rather than purchasing only public API access.

Furiosa can offer hardware that enterprises install through infrastructure partners, supported by model-serving software. Wider adoption requires compatibility with enterprise systems, dependable updates and proof that the cost and control benefits justify operating specialized hardware.

Samsung SDS's NPU-as-a-Service offering shows how a provider can package RNGD into a service consumed by its own customers. This lets Furiosa reach application developers without having to sell and support an individual physical installation for each one.

The channel depends on cloud partners earning attractive economics and keeping machines utilized. Furiosa must help those partners support the models their customers actually want, while delivering hardware and software reliably enough for a production service.

Furiosa established a Singapore hub in September 2026 to support Asia-Pacific expansion and has described interest from sovereign AI initiatives and customers in the Middle East. These buyers may seek local computing capacity and control over model deployment, creating procurement budgets outside the company's initial Korean base.

A local presence and infrastructure partnerships can help qualify the hardware and support customers across those markets. Interest and announced partnerships should still be distinguished from completed installations, with delivery, software support and project financing determining how much demand becomes sustained business.

Software adoption and model coverage: Customers already invested in Nvidia's ecosystem face costs when moving models to another architecture. Furiosa must keep its compiler, serving tools and supported workloads useful as models change.

Production commitments and supplier dependence: The company relies on semiconductor, memory and manufacturing partners while planning a substantial increase in output. Supply constraints or weaker-than-expected orders could delay deliveries or leave capital tied up in hardware.

System economics versus chip efficiency: Lower energy use does not guarantee the lowest cost for every inference service. Poor utilization, integration expense or new competing products could erode the advantage buyers expect from specialized accelerators.