AI Data Center Semiconductor Market Consolidation: Intel Considers Snapping up Challenger SambaNova Systems

Intel has reportedly signed a non-binding term sheet agreement to acquire SambaNova Systems. While the deal is not yet finalized and still needs to clear regulatory hurdles, it would represent a pivotal step toward production-ready end-to-end inference systems in Intel’s portfolio under its 2025 installed Chief Executive Officer (CEO), Lip-Bu Tan. The deal is reported to value SambaNova at around US$1.6 billion, which is a sizable discount to its highest funding round valuation at around US$5 billion, where Intel participated along with SoftBank and other private equity firms, including Walden International (founded by Lip-Bu Tan). A brief of SambaNova Systems follows:

Technology & Platforms:

• Overview: Marketed as a low-power, full-stack (hardware and software) Artificial Intelligence (AI) solution with enterprise-grade model accuracy and throughput. Scalability, security, and model ownership are central to the proposition.
• Compute: Currently shipping the fourth-generation Reconfigurable Dataflow Unit (RDU) SN40L chip designed for large frontier model inferencing (and some fine-tuning/training workloads) and agentic workloads, with a larger memory footprint than Graphics Processing Units (GPUs). The “dataflow” architecture bypasses costly memory calls inherent to GPUs during AI computations.
• Systems: Data center and enterprise on-premises full-stack AI solution made up of servers and integrated racks (SambaRack), with simple integration into existing air-cooled data center environments.
• Software: An end-to-end software offering with Application Programming Interfaces (APIs), User Interfaces (UIs), open-source models (Llama, Qwen DeepSeek, and OpenAI), and a workload orchestrator.

Commercial & Go-to-Market:

• SambaStack: The core full-stack hardware and software solution with a UI, access to open-source models, and the SambaRack servers and racks.
• SambaCloud: A cloud inference platform with a selection of open-source models accessible via APIs.
• SambaManaged: A modular “turn-key” data center product and managed service offering (or self-service) with a rapid 90-day deployment, Service Level Agreement (SLA), scalable to 100+ racks, and compatible with existing data center infrastructure.
• Target Customers: Lighthouse customers to date have included scientific research labs such as Los Alamos and Argonne National Laboratory for scientific research. Developers and enterprises are also targeted with cloud, on-premises, or hybrid implementations tailored to evolving enterprise needs: Maitai, Hume.ai, and Aion Labs are showcased publicly.
• Sovereign AI Partnerships: A growing sovereign AI cloud proposition with several deals announced in 2025, including:
  • OVHcloud in France (expanding to other European endpoints)
  • Infercom for a German (to be expanded across Europe) Inference-as-a-Service (IaaS) cloud
  • Argyll in the United Kingdom for a renewable-powered sovereign cloud
  • SCX in Australia for a sovereign cloud launching in New South Wales and later South Australia
• Amazon Web Services (AWS): SambaCloud can be accessed through the AWS Marketplace, including its open-source models and inference services, with billing linked to end-customers’ AWS accounts and integration between AWS workloads and SambaNova Cloud.

The deal has the potential to be highly impactful for Intel’s enterprise AI strategy, where low-latency, efficient cost-per-token are top of mind (and end-to-end, systems-level proposition has been lacking).

SambaNova targets the inference wave as the industry moves from large training runs to deploying models of various sizes in production workloads—both in data centers and enterprise on-premises settings. This is largely in line with Intel’s revamped AI strategy, focusing on inference workloads before eventually addressing training. The purchase of SambaNova can be seen as a move to close its capability gap (compared to NVIDIA, AMD, and other inference challengers such as Cerebras) by adding a differentiated inference system to its portfolio. Along with it comes IP, experienced engineers (which have recently left Intel), a software stack, and end-to-end system-level know-how. This is a strategic priority for Intel, which has deprioritized training platforms in favor of inference workloads and fallen behind NVIDIA (and to a lesser extent AMD) on system-level designs that incorporate interconnect, networking, and cooling. Some potential alignment and synergies for Intel include:

• Central Processing Unit (CPU) Cross-Sell: All SambaRack servers require a host CPU, which has previously been AMD and can quickly be replaced with the Intel Xeon portfolio and co-marketed under both brands, amplifying both.
• SambaRack as Intel’s “Deep GPU Xceleration (DGX)”: The fully integrated rack-scale solutions from SambaNova are akin to NVIDIA’s successful and widely adopted DGX supercomputer and can be leveraged by Intel to gain a foothold in end-to-end systems.
• SambaCloud & Intel Tiber AI Cloud: While Intel’s cloud remains more of a developer environment and sandbox for testing Intel hardware, SambaCloud is a production-grade Inference-as-a-Service (IaaS) platform with API access to popular models. SambaCloud users would be funneled toward the rest of Intel’s portfolio.

The acquisition of SambaNova has the potential to boost Intel’s server-based AI inference product roadmap and proposition. For Intel, the following areas are worth considering:

• Carefully Consider Positioning Vis-à-Vis Gaudi: Both the RDU and Gaudi are Application-Specific Integrated Circuits (ASICs) with a value proposition distinct from GPUs, so Intel must consider how to position SambaNova’s RDUs alongside Gaudi to avoid cannibalization or confusion. Both target enterprise inference, and evidence suggests that Gaudi will be deprioritized over time, with the IP repurposed.
• Integration with OneAPI: SambaNova’s software stacks SambaFlow and SambaStudio should slot into OneAPI so that model conversion is zero-touch and simple. Model support should come early (as with Compute Unified Device Architecture (CUDA)) to give developers the confidence needed to invest. Moreover, the number of open-source models should be expanded beyond Meta, DeepSeek, Qwen, and a handful of others to include those from Google and Microsoft, as well as language-specific models with a sovereign tilt, to boost SambaNova’s growing sovereign AI proposition.
• Retain Distinct Systems Messaging: Retain SambaNova’s systems-level branding and proposition rather than focusing on semiconductor performance and metrics. The Total Cost of Ownership (TCO) benefits and applicability to agentic workloads, thanks to the memory architecture of the RDU, should be amplified, rather than solely measured in Tera Operations Per Second (TOPS).
• Benchmark Performance: The highly efficient performance, as reported by Artificial Analysis, over other cloud platforms, should be promoted alongside the TCO benefits and expanded to other prominent benchmarks, such as MLPerf’s data center inference, where some of the most widespread open-source models are run for the transparent tests.
• Vertical Expansion: SambaNova has been popular in government and research lab settings, where data privacy is key, and their end-to-end solution provides control. This should be expanded to other sectors, particularly Fortune 500 companies that are moving into production with Agentic AI systems deployed on-premises in regulated sectors such as healthcare and legal services.