RightNow AI, a Y Combinator F26 GPU research lab, has launched RunInfra, a hosted service for open models that lists cached input tokens from $0.01 per million on select models, according to the company's site. The company markets the product as "Optimized down to the GPU kernel."
The pitch rests on a simple observation: AI agents burn through inference budgets, and low-level engineering can shrink the bill.
What RunInfra sells
RunInfra hosts open models "built for agents," per the company's site. Customers get a single API key that works with both OpenAI and Anthropic SDKs, and the service uses session-aware routing, which keeps related requests together so cached material can be reused. Live effective input rates are displayed for each model.
The $0.01 cached-input price applies only to select models. RightNow has published a methodology page, with version 2 dated Sept. 20, 2026, that explains how it calculates those effective rates.
What the company has not offered is any sense of traction. It disclosed no customers, revenue or usage figures for RunInfra, and the reporting behind this article turned up none elsewhere.
The company behind it
RightNow describes itself as a YC-backed GPU research lab building GPU-native AI infrastructure. Its site lists NVIDIA Inception membership, and its Trust Center shows SOC 2 Type 2 verification as of October 2026. Osama Jaber wrote on LinkedIn that the company began as a CUDA editor and is now "a full inference platform that makes AI models run faster and cost less."
Third-party trackers list Jaber Jaber and Osama Jaber as founders, with a team of about two. Those are aggregator figures rather than company statements, and should be read with that caveat.
Funding beyond YC is undisclosed. An Extruct data room on the F26 cohort showed no round beyond YC, and no authoritative announcement had surfaced as of Oct. 31, 2026. NVIDIA Inception, for its part, is a support program, not equity financing.
A research record that predates the product
RightNow's kernel work came before RunInfra. Its AutoKernel paper (arXiv 2603.21331, March 2026) describes an autonomous agent loop that optimizes Triton and CUDA kernels across prioritized PyTorch operations. MarkTechPost covered the open-source release on April 6, 2026, summarizing it this way: "Give it any model before you go to bed, and wake up to faster Triton kernels." The project's GitHub repository had 1.6k stars, with its latest update on Oct. 23, 2026.
A June 2026 paper, AutoMegaKernel, describes a persistent megakernel harness for Llama-family models and claims batch-1 decode performance against cuBLAS, NVIDIA's standard linear-algebra library. That claim sits in a preprint and awaits broad independent replication. Its repository had 159 stars.
RightNow also released System Zero, a hosted decision model announced Oct. 30, 2026, priced at $0.19 per million uncached input tokens and $0.019 per million cached.
Why agents push costs down to the kernel
Agentic workloads are turning inference into the line item that matters. Google Cloud's State of AI Infrastructure report, published July 7, 2026, found that 83% of organizations need to upgrade their infrastructure to support agentic AI.
The spending forecasts are large. Gartner said on Aug. 10, 2026 that spending on AI-optimized infrastructure as a service would grow 96% in 2026 to $42 billion, then reach $66 billion in 2027. IDC raised its 2026 AI-infrastructure spending forecast to $497 billion in a July 2026 blog post. Grand View Research projects the AI inference market at $132.7 billion in 2026 and $421.8 billion by 2033, an 18.0% compound annual growth rate, though that is vendor research and best treated as directional.
McKinsey, in October 2026, described a shift from training to serving at scale, with latency, throughput and cost constraints demanding deep kernel and system co-design.
A crowded field
RunInfra enters a market where price and caching are the main differentiators. Fireworks emphasizes kernel, quantization and caching work, and cites "99.6% cache" on long sessions. Together claims up to 90% lower cost than proprietary APIs, a vendor claim dating to July 2026. OpenRouter operates as a router and marketplace where prices change rapidly.
Closed-model prices are moving as well. Tom's Hardware reported that Anthropic's per-token input pricing on OpenRouter fell about 77% from May to October 2026. That is a secondary-source figure, and buyers should check per-model pages.
Kernel optimization is also drawing larger names. PyTorch published KernelAgent on March 6, 2026, a closed-loop multi-agent workflow guided by GPU hardware signals. Meta Engineering said on April 2, 2026 that its KernelEvolve agent reached a 100% KernelBench pass rate, based on internal benchmarks. At the MLSys 2026 FlashInfer contest, agent teams beat baseline FlashInfer on Fused MoE by about 1.68 to 1.71 times.
Open-source engines keep improving too. vLLM said in March 2026 that "it is possible to achieve state-of-the-art attention performance using a single, portable kernel implementation," and its Triton backend is now the default on AMD. SGLang reported HiCache hybrid-model caching with up to 15.5x input throughput.
What to watch
Against that backdrop, RightNow's edge is unproven. If agents and open engines can optimize kernels for free, a team of about two has to win on execution and price.
Published rates are the first test. The company's live effective-rate panels, last captured Nov. 1, 2026, let buyers compare cache-hit economics against Fireworks, Together and OpenRouter. Whether any funding round follows the YC batch is something RightNow has not said.
