Oumi, Runware, Larridin and others cash in on AI cost cuts executives are finally demanding

AI cost cutting just got a whole new industry behind it. Oumi AI CEO Manos Koukoumidis says a common enterprise pattern is “widely irrational and inefficient”: using frontier models like Anthropic, OpenAI, and Google, for niche, everyday tasks. The punchline is that this approach drives “staggering AI bills,” and executives are now paying for experts who can stop the bleeding and prove what is actually working.

That shift is showing up fast in funding, products, and go-to-market. Oumi, founded in 2024, raised $10 million in its 2024 seed round, with a pitch focused on helping individuals “vibe-code” niche AI models in minutes so the system is tailored to the task. Meanwhile, Larridin, a San Francisco measurement layer, told Business Insider it’s seeing traction double every quarter since the start of the year, after a slow early period when companies were not yet concerned about AI spending. The new demand is turning AI spend from a vibes-based line item into something that can be audited, optimized, and engineered.

So what’s actually driving this? The source points to the end of the tokenmaxxing era, where employees were encouraged to burn as many AI tokens as they could. That trend created an internal incentive mismatch: if the goal is “more usage,” then costs rise regardless of whether the output improves decisions, speed, quality, or revenue. As companies move away from that behavior, the next question is inevitable: Are we getting value proportional to the compute bills we are signing?

Enter the “coaches.” These are the consultants and advisors trying to reshape how enterprises adopt generative AI, before the model spend becomes a permanent fixture. Sydney-based Adaptovate, operating for more than eight years and with over 100 consultants across offices worldwide, has spent roughly the last three years focusing specifically on helping clients get the most bang for their AI spending buck. Michael Murphy, a partner at Adaptovate, describes work that starts with strategy, including how decision-making changes, how the talent model changes, and how org structure shifts when scaling AI across the full organization. In other words, the thesis is that cost problems are often process problems first, and technology problems second.

Then there are the “measurers,” who treat AI economics like a dashboardable system. Larridin’s CTO Ameya Kanitkar told Business Insider that the company acts as a measurement layer across clients’ AI tools, employees, agents, and spending. One product focuses on a very specific question: how employee productivity varies with token spend, so companies can locate the “sweet spot” for token budgets. Kanitkar also frames why this category exists now. Companies did not care much about AI spending in the early days of adoption, but after they started pouring “millions” into the resource, measurement became the only way to answer whether those hundreds of millions of dollars are justified. Larridin raised $17 million in seed funding from Andreessen Horowitz, Bloomberg Beta, Google Ventures, and others, and launched the platform in 2025. It now serves clients from data center construction firms to biosciences and financial services.

Finally, the “builders” are going after costs in infrastructure and model execution. The source calls out Runware, cofounded by serial entrepreneur Ioana Hreninciuc, which provides inference infrastructure so AI models run quickly, cost-effectively, and at scale, reducing the risk of outages as companies expand AI products. For teams that buy their own GPUs, Tensormesh is another cost-focused bet. Cofounder Junchen Jiang points to Key-Value Cache, which he says “never appears in people’s token bills” but still affects costs. Tensormesh’s caching system, per the source, helps businesses “accept more queries on fewer GPUs.” The company raised $20 million from hardware and investors including AMD and NVentures, Nvidia’s venture capital arm.

And even more products are in the pipeline. Conifer, still participating in Y Combinator’s summer 2026 batch, had raised $1.3 million, around 20% of its desired raise, before Demo Day. Its cofounders Charles Muehlberger and Michael Jeffords describe a product meant to synthesize queries, split them into pieces, and feed those pieces to different models, including local ones. Conifer’s stated positioning is directly cost-focused, with Jeffords saying the “main customer” is someone focused on cost.

Across these categories, the advice converges. The source summarizes their common belief: stop using frontier AI models for every menial task. Start using lighter, open-source models, and, when it makes sense, build your own models to reduce costs and reliance on the big AI labs. There is also a more CFO-friendly reframe. Kanitkar from Larridin argues AI costs should be treated like capital expenditures rather than operating expenses, and that ROI should be expected over a year or two instead of immediately. That framing matters because boards and finance teams tend to demand quick payback from OPEX, while CAPEX can justify longer implementation cycles.

Zoom out one more layer and you see the governance issue hiding in plain sight. Hreninciuc from Runware warns that incentives currently push people to claim AI spending is paying off, and because nobody wants to be the “canary in the coal mine,” there is “no accurate view of the market.” If true, it means the cost-saving wave is not just about efficiency. It is about restoring honesty to enterprise AI ROI, forcing measurement, and making sure the next wave of AI budget is engineered for outcomes, not just consumption.