Infinity Ignition

Ignition is Infinity's autonomous research agent that generates, tests, debugs and optimises the low-level kernels and inference software a new AI accelerator needs to run models. It is sold to chipmakers and hardware vendors that lack a mature CUDA-style software stack, on a share of the performance gains.

Infinity Inc. is a San Francisco AI infrastructure startup founded in August 2025 by Jeremy Nixon, a former Google Brain researcher and co-founder of the AGI House community. In July 2026 it announced a $15 million seed round at a $100 million post-money valuation, with Touring Capital and Principal Venture Partners plus chip-industry executives and researchers from OpenAI and Anthropic. TechCrunch put headcount at 26. Its product, Ignition, is an agent that writes the inference software stack for AI accelerators: kernels, kernel orchestration, debuggers, profilers, compilers and, for proprietary instruction sets, decompilers. Humans choose what to build; the agent writes code, measures it on the hardware, and rewrites it, logging what worked as reusable tools so later runs go faster. Verification includes bit-for-bit accuracy checks and memory clobber detection. Infinity targets GPUs, SRAM-based chips, phone chips and systolic arrays. Its one named partner is d-Matrix: Infinity reports that on the Corsair accelerator its agents reached roughly 90 to 92% of theoretical peak on tensor-parallel matrix multiplication within 10 hours of hardware access, and that Qwen3, Qwen3.5 and Gemma4 ran end to end within 10 days. On a single Nvidia H100 it reports a Qwen3-8B engine up to 34.3% faster than vLLM after a day of optimisation. All of these figures are company-reported. Infinity charges no upfront licence; it takes about 20% of the savings on additional compute purchases, measured against an agreed throughput baseline, and says this already brings in millions in annual recurring revenue.

Category

Infrastructure & Cloud

Pricing

Contact for pricing, Usage-based

Target Market

Chipmakers, Hardware Vendors, ML Infrastructure Engineers, CTOs

Deployment

Hybrid

Founded

2025

Headquarters

San Francisco, USA

Team Size

11-50

• ✓
  Autonomous kernel generation

  Writes, tests, debugs and optimises low-level inference kernels for a target chip in a measure-and-rewrite loop on the hardware.

• ✓
  Full toolchain authoring

  Builds debuggers, profilers, compilers and kernel orchestration, plus decompilers for proprietary instruction sets, so a stack is not limited to kernels.

• ✓
  Correctness verification

  Checks outputs bit for bit and detects memory clobbering, which matters when generated low-level code must be trusted in production.

• ✓
  Self-improving tool library

  Logs successes and failures and keeps effective techniques as reusable tools, so each new chip or model bring-up starts from prior work.

• ✓
  Cross-architecture targeting

  Aims at GPUs, SRAM-based chips, phone chips and systolic arrays rather than a single hardware family.

• ✓
  Performance-share pricing

  Charges a share of measured savings against an agreed throughput baseline, aligning the vendor's revenue with real speedups.

• •
  New accelerator bring-up

  A chip startup gets matrix multiplication near theoretical peak and full models running end to end within days of hardware access.

• •
  Day 1 model enablement

  Hardware vendors port newly released open-weight models such as Qwen3.5 and Gemma4 without waiting months for hand-written kernels.

• •
  Inference throughput tuning

  Teams optimise an existing model's tokens per second on current hardware, as in Infinity's reported Qwen3-8B result against vLLM on an H100.

• •
  Software for proprietary instruction sets

  Vendors with undocumented or closed ISAs get decompilers and kernels generated automatically, closing the software gap with CUDA.

There is no list price or self-serve plan. Per SiliconANGLE, optimisation agreements give Infinity roughly 20% of the savings on additional compute purchases, calculated from throughput gains against an agreed baseline, with no upfront licence fee. Engagement starts through the hardware partner form on infinity.inc.

This page was written from 5 sources, 4 on domains other than infinity.inc.

1. 1.infinity.inc — infinity.incvendor
2. 2.techcrunch.com — inference startup infinity raises 15m from touring capital o
3. 3.siliconangle.com — infinity raises 15m run ai inference chipset
4. 4.pulse2.com — infinity raises 15 million seed funding at 100 million valua
5. 5.runtimewire.com — infinity inc raises 15 million ai chip inference software