Infinity raises $15 million seed round for AI chip inference software

Infinity Inc. has raised $15 million in seed funding to build software that prepares AI chips to run inference workloads across different processor designs. The round gives the one-year-old AI infrastructure research company a $100 million post-money valuation and targets a specific industry bottleneck: alternative AI chips often arrive before the software stack needed to use them well.

The company said it will spend the capital on engineering hires, expansion of its automated research platform and work with chipmakers, including d-Matrix Corp. Infinity also said it is already producing millions of dollars in annual recurring revenue from chip design partnerships, though it did not disclose a specific ARR figure, customer count or headcount.

Touring Capital LLC was described as a significant participant in the round. Principal Venture Partners LP also invested, alongside unnamed executives and angel investors from OpenAI LLC, Anthropic PBC and chip companies. Infinity did not name those individuals.

Software as the chip bottleneck

Infinity’s product is called Ignition. The company describes it as an AI agent that creates, tests and tunes the low-level tools required to run models efficiently on different processors. That includes compilers, debuggers, profilers, kernels and orchestration for inference workloads, according to founder and Chief Executive Jeremy Nixon.

The pitch is aimed at one of Nvidia Corp.’s strongest advantages. Nvidia’s CUDA ecosystem gives developers a mature path to production on its GPUs. Infinity says Ignition can help other chip vendors create comparable software support in hours or days, rather than waiting for teams of engineers to hand-code and tune the stack for each architecture.

That is a large claim, and the company’s evidence is still company-reported. Kernel development, memory layout, compiler behavior and numerical accuracy are areas where small failures can break performance or correctness. Infinity said human engineers define the tools Ignition should build, while the system handles the implementation work.

For proprietary instruction sets, Infinity said it builds decompilers that convert executable code into higher-level source code that can be modified with standard programming languages. The company also said it uses checks for bit-level accuracy, memory overwrite detection and required numerical behavior.

Early performance claims

Infinity said Ignition improved inference throughput for Alibaba’s Qwen3-8B model by 34% versus the vLLM framework after one day of automated optimization. Nixon said the test ran on a single Nvidia H100 GPU and has not been independently validated.

In work with d-Matrix, Infinity said its software reached up to 92% of the theoretical peak performance of the Corsair accelerator within 10 hours. The company said the result came from spreading matrix multiplication work across all 32 compute units. Infinity also said it had Qwen3, Qwen3.5 and Gemma4 running fully on the chip within 10 days.

The company describes Ignition as a self-improving system that records prior attempts, captures successful techniques and uses a representation of solved problems to speed later optimization runs. That framing is consistent with current AI infrastructure marketing, but the useful test will be whether the system keeps producing gains across more chips, models and deployment constraints.

Infinity’s business model is also tied to performance claims. Nixon said its optimization agreements typically give the company about 20% of savings on additional compute purchases, based on throughput gains measured against an agreed baseline, rather than conventional software license fees.

Nixon previously worked as a Google Brain researcher and co-founded AGI House Labs Inc., described as a hybrid community hub, applied AI lab and venture capital firm. Infinity said the funding will also support talks with more chip companies.

This story draws on original reporting from SiliconANGLE.