NVIDIA, AMD and Intel have all appeared on the cap table of RadixArk after the company closed a $100 million seed round at a reported $400 million valuation. The startup is known for SGLang, an open-source inference engine.
What SGLang is and who uses it
According to reports, SGLang has been deployed across more than 400,000 GPUs and is used by major cloud and AI vendors including Google, Microsoft, xAI, Oracle, NVIDIA and AMD. The engine focuses on improving runtime efficiency for pre-trained models.
How RadixArk improves inference
RadixArk’s software targets several operational improvements:
- reusing repeated context,
- more efficient memory management,
- batching requests more effectively,
- extracting more work from the same GPUs without locking users into a single hardware vendor.
These techniques aim to reduce compute cost for inference or increase the throughput available from existing hardware.
Why chipmakers invested
The three chipmakers appear to have distinct but complementary incentives to back a neutral inference layer:
- NVIDIA: to increase GPU throughput and make existing GPUs more efficient;
- AMD: to develop an escape route from CUDA dependency;
- Intel: to create an opening for Gaudi and other alternatives.
Backing a neutral software layer reduces the risk of being sidelined by a solution that benefits only a competitor’s hardware.
Implications for AI infrastructure
The investment signals a shift in AI infrastructure from raw hardware competition toward compute control and software-driven efficiency. Neutral inference layers that can run across hardware types are becoming strategically important to both cloud providers and silicon vendors.
While this deal doesn’t imply long-term unanimous platform support, it underlines that missing a neutral layer capable of maximizing expensive silicon would be a strategic disadvantage for any major chipmaker.
