Economy May 5, 2026 08:06 AM

RadixArk Secures $100M Seed to Trim AI Inference Costs

Startup's SGLang engine sits between models and hardware, reusing short-term memory to reduce compute needs

By Ajmal Hussain
RadixArk Secures $100M Seed to Trim AI Inference Costs

RadixArk has raised $100 million in a seed round at a $400 million valuation to advance software that reduces the compute required to run AI models. The round was led by Accel and Spark Capital, with participation from Nvidia’s NVentures, Broadcom CEO Hock Tan, and others. RadixArk's open-source engine, SGLang, operates as a middle layer leveraging short-term memory and a Radix tree filing system to reuse KV cache and skip redundant prompt processing, lowering the cost of inference.

Key Points

  • RadixArk closed a $100 million seed round at a $400 million valuation, led by Accel and Spark Capital with participation from Nvidia's NVentures, Broadcom CEO Hock Tan, and others.
  • RadixArk offers SGLang, an open-source middle-layer engine that leverages short-term memory to reduce the total computing power for AI inference.
  • The inference engine reuses KV cache organized in a Radix tree to skip reprocessing matching prompt segments, thereby lowering inference costs.

RadixArk announced a $100 million seed financing at a $400 million valuation, backing its effort to lower the computing footprint of artificial-intelligence inference. The financing was led by Accel and Spark Capital, and included participation from Nvidia's NVentures, Broadcom CEO Hock Tan, and other investors.

The company's core offering is a software engine intended to improve the efficiency of AI workloads. RadixArk publishes an open-source engine called SGLang that functions as an intermediary layer between generative AI models and the underlying hardware. According to the company's description, SGLang employs short-term memory to reduce the total computing power required for model calculations.

RadixArk's inference engine is designed to reuse key-value (KV) cache entries by organizing segments of conversation into a filing structure the company calls a Radix tree. The engine examines new incoming queries to detect whether the start of a prompt corresponds to material that the system has already processed. When the engine finds such a match, it omits that portion of the prompt from reprocessing and moves directly to generating the response. This approach is presented as a method to cut the cost associated with running an AI model by avoiding redundant computation.

The funding round and the engineering focus reflect an effort to address the cost side of AI inference without changing model architectures or hardware. RadixArk's technology is positioned as a software layer that optimizes how models interact with hardware by leveraging cached short-term context and a structured reuse strategy.


Key points

  • RadixArk closed a $100 million seed round at a $400 million valuation, led by Accel and Spark Capital with participation from Nvidia's NVentures, Broadcom CEO Hock Tan, and others.
  • The company offers SGLang, an open-source engine that acts as a middle layer between AI models and hardware, using short-term memory to reduce compute needs.
  • Its inference engine reuses KV cache via a Radix tree filing system to skip reprocessing matching prompt segments, lowering inference costs.

Risks and uncertainties

  • The article does not provide details on deployment scale, customer adoption, or empirical performance metrics for SGLang, leaving uncertainty about commercial impact.
  • Investor participation is described as including "others," but the article does not list all participants or disclose specific terms beyond valuation and amount, limiting transparency around the round.
  • The piece does not report on timelines for integration with existing model stacks or hardware ecosystems, so the pace and extent of operational adoption remain unspecified.

Readers should note that this report focuses on the funding and technical approach as described; it does not include additional performance claims or external validation beyond the company's stated design and the list of participating investors.

Risks

  • The article does not provide details on deployment scale, customer adoption, or performance metrics, creating uncertainty about commercial impact.
  • Investor participation includes unspecified 'others' and the article does not disclose all terms or participants, limiting transparency about the round.
  • No timelines or integration details with existing model and hardware ecosystems are provided, leaving the pace of adoption unclear.

More from Economy

Mozambique’s Financial Strain Deepens as Debt and Development Challenges Converge May 5, 2026 30-year UK gilt yields reach highest level in nearly 30 years as markets price more BOE hikes May 5, 2026 Romanian Government Ousted After Confidence Motion, Raising Questions Over Deficit Reforms May 5, 2026 Swiss inflation accelerates to 16-month peak as energy prices climb May 5, 2026 Finance Minister Says Debt Relief Relaunch Will Not Disrupt Rate-Cut Cycle May 5, 2026