Enfabrica
Enfabrica provides ultra-high-performance networking chips for AI data centers.
Enfabrica is a fabless networking semiconductor company that designs ultra-high-performance chips for AI infrastructure. Their flagship Accelerated Compute Fabric SuperNIC (ACF-S) delivers 3.2 terabits/second of connectivity between GPUs, CPUs, and accelerators in data centers, reducing GPU compute requirements by ~50% at equivalent performance. The company also offers EMFASYS, a system combining RDMA Ethernet networking with CXL-based memory to provide elastic memory capacity for AI inference workloads. Built by veterans from Broadcom and Google, Enfabrica differentiates through vendor-agnostic architecture that avoids proprietary lock-in.
Problem solved
GPU networking bottlenecks limit AI compute efficiency; existing fabric solutions create vendor lock-in and cannot scale elastically across heterogeneous processor architectures.
Target customer
Hyperscale cloud providers, AI infrastructure operators, and enterprise data center teams deploying large language models and distributed AI compute clusters.
Founders
R
Rochan Sankar
Founder, President & CEO
26-year industry veteran; Senior Director at Broadcom leading Data Center Ethernet switching; developed four generations of Tomahawk and Trident chips; holds BS in electrical engineering from University of Toronto and MBA from Wharton.
S
Shrijeet Mukherjee
Co-Founder & Chief Development Officer
VP of Software Engineering at Cumulus Networks; previously led networking platforms and architecture at Google.
Funding history
Series A
$50M
2022
Led by Unknown
· Unknown
Series B
$125M
September 2023
Led by Atreides Management
· Sutter Hill Ventures, NVIDIA, IAG Capital Partners, Liberty Global Ventures, Valor Equity Partners, Infinitum Partners, Alumni Ventures
Series C
$115M
November 2024
Led by Spark Capital
· Arm, Cisco Investments, Maverick Silicon, Samsung Catalyst Fund, VentureTech Alliance, Atreides Management, Alumni Ventures, IAG Capital, Liberty Global Ventures, Sutter Hill Ventures, Valor Equity Partners
Total raised:
$290M
Pricing
Not publicly available. Customers can pre-order ACF-S systems by contacting sales@enfabrica.net.
Notable customers
Not disclosed. Enfabrica interacted with customers and partners at Supercomputing '23 in Denver.
Integrations
Standards-based interfaces (CXL, RDMA Ethernet); compatible with NVIDIA, AMD, and other GPU/accelerator vendors; no proprietary lock-in.
Tech stack
GSAP (JavaScript frameworks)
React (JavaScript frameworks)
web-vitals (JavaScript libraries)
Microsoft ASP.NET (Web frameworks)
Webpack
Open Graph
Module Federation
Sitecore (CMS)
Magento (Ecommerce)
Google Analytics (Analytics)
Sentry (Issue trackers)
Google Font API (Font scripts)
PHP (Programming languages)
Google Workspace (Email)
MySQL (Databases)
Google Cloud (IaaS)
Google Cloud Trace (Performance)
Website
Competitors
Broadcom
Broadcom's Tomahawk 6 delivers 102.4 Tbps switching (vs. Enfabrica's 3.2 Tbps ACF-S) but creates vendor lock-in; Enfabrica's SuperNIC is processor-agnostic across multiple AI architectures.
Cisco
Cisco's Silicon One G200/G202 target AI networking but lack the multi-vendor interoperability and elastic memory integration that Enfabrica's platform provides.
Marvell
Marvell's enterprise switches deliver up to 51.2 Tbps but are not optimized for disaggregated AI compute fabrics or elastic memory scenarios.
Why this matters: Enfabrica has assembled elite networking talent (Broadcom's ethernet switching architect, Google's networking lead) to build the first vendor-agnostic, elastic GPU fabric chip at a critical moment when AI infrastructure scaling is the primary bottleneck for hyperscalers. With $290M in funding from infrastructure giants (NVIDIA, Cisco, Arm, Samsung) and a 3.2 Tbps SuperNIC shipping in 2025, Enfabrica is positioned to reshape how AI clusters are built.
Best for: Hyperscale AI infrastructure operators and cloud providers building multi-vendor, disaggregated compute clusters who need maximum efficiency and flexibility without vendor lock-in.
Use cases
Large Language Model Training
Enterprises training LLMs can reduce GPU compute requirements by ~50% at equivalent performance by using Enfabrica's ACF-S to eliminate inter-GPU networking bottlenecks. The vendor-agnostic design allows teams to mix NVIDIA, AMD, and other accelerators in the same cluster.
Distributed AI Inference at Scale
EMFASYS provides elastic memory capacity and bandwidth for memory-bound inference workloads, allowing inference servers to dynamically access remote DDR5 memory at low, bounded latency without proprietary interconnects. This reduces the need to provision per-server memory and improves utilization.
Disaggregated Compute Architecture
Organizations building modular data centers with separate compute, memory, and networking resources can use Enfabrica's fabric to elastically connect heterogeneous components at multi-terabit speeds while maintaining standards-based interfaces, reducing operational complexity and avoiding vendor lock-in.
Alternatives
Broadcom Tomahawk Series
Higher total switch bandwidth (102.4+ Tbps) but optimized for Broadcom ecosystem; Enfabrica chosen when multi-vendor flexibility and elastic memory are priorities.
NVIDIA BlueField SmartNIC
Tightly integrated with NVIDIA GPUs but creates vendor lock-in; Enfabrica preferred for multi-GPU-vendor deployments and disaggregated memory architectures.
Marvell Prestera Fabric
Established enterprise networking platform; Enfabrica differentiated for AI-specific, disaggregated architectures requiring elastic memory and true multi-vendor support.
FAQ
What does Enfabrica do? +
Enfabrica designs ultra-high-performance networking semiconductor chips for AI data centers. Their flagship ACF-S SuperNIC delivers 3.2 Tbps of connectivity between GPUs and accelerators, and their EMFASYS system provides elastic memory capacity via CXL and RDMA. Both products are vendor-agnostic and designed to improve AI compute efficiency without lock-in.
How much does Enfabrica cost? +
Pricing is not publicly available. Customers interested in pre-ordering ACF-S systems should contact sales@enfabrica.net. Pricing likely varies based on configuration and volume.
What are alternatives to Enfabrica? +
Broadcom Tomahawk and Jericho3-AI switches offer higher total bandwidth but less multi-vendor flexibility. NVIDIA BlueField SmartNIC is tightly integrated with NVIDIA GPUs. Marvell and Cisco also offer AI networking solutions but lack Enfabrica's disaggregated, elastic memory approach.
Who uses Enfabrica? +
Hyperscale cloud providers, AI infrastructure operators, and enterprise teams deploying large language models. Specific customer names are not publicly disclosed, though Enfabrica engaged customers and partners at Supercomputing '23.
How does Enfabrica compare to Broadcom? +
Broadcom's Tomahawk 6 delivers 102.4 Tbps of total switch bandwidth, far exceeding Enfabrica's 3.2 Tbps ACF-S. However, Enfabrica's SuperNIC is processor-agnostic and avoids vendor lock-in, whereas Broadcom solutions are tightly integrated into Broadcom's ecosystem. Enfabrica also integrates elastic CXL memory, which Broadcom does not offer.
Tags
AI infrastructure
networking semiconductors
GPU fabric
hyperscaler compute
data center
CXL
RDMA
vendor-agnostic