{"slug": "oumi-larridin-and-runware-build-tools-to-cut-enterprise-ai-bills", "title": "Oumi, Larridin and Runware build tools to cut enterprise AI bills", "summary": "Oumi cofounder and CEO Manos Koukoumidis is building tools to help enterprises cut AI costs by using smaller, task-specific models instead of expensive frontier systems, a market that also includes consultants like Adaptovate and measurement platform Larridin. The FinOps Foundation's 2026 report found that 98% of respondents managed AI spending, up from 63% in 2025, highlighting the growing need for cost control. Oumi, which raised a $10 million seed round in 2024 led by Venrock and Obvious Ventures, provides a platform for building, fine-tuning, and deploying models across local machines, cloud providers, and commercial APIs.", "body_md": "[Manos Koukoumidis](https://www.oumi.ai/company?ref=runtimewire) has found a business in a problem corporate AI buyers created for themselves: they adopted powerful models quickly, then discovered that convenience came with an open-ended bill.\n\nKoukoumidis, Oumi's cofounder and CEO, described the routine use of expensive frontier models for narrow tasks as \"widely irrational and inefficient\" in a [Business Insider report published Aug. 2, 2026](https://www.businessinsider.com/ai-cost-saving-businesses-startups-roi-2026-7?ref=runtimewire). His prescription is to give companies smaller models fitted to specific jobs rather than sending every request to a general-purpose system from OpenAI, Anthropic or Google.\n\nThat argument draws on Koukoumidis' experience inside the incumbent stack. He previously led generative AI work at Google Cloud, according to [Oumi's January 2025 launch announcement](https://www.oumi.ai/press-release?ref=runtimewire). Koukoumidis has a Princeton PhD and a thesis that AI needed an open development platform resembling Linux. Oumi now provides tools for building, fine-tuning, evaluating and deploying models across local machines, cloud providers and commercial APIs.\n\nOumi raised a $10 million seed round in 2024 led by Venrock and Obvious Ventures, with Plug and Play and Ascend participating. That financing and Oumi's January 2025 public launch are established history. The current development is the market forming around Koukoumidis' diagnosis. Consultants, measurement platforms, model routers and inference providers are all trying to become the layer that tells enterprises which AI is worth paying for.\n\n### The bill arrives after experimentation\n\nEarly enterprise AI programs rewarded adoption. Employees received access to chatbots and coding assistants, product groups added model APIs, and executives watched usage as evidence that an AI strategy was taking hold. Cost attribution and verified productivity gains often came later.\n\nThe [FinOps Foundation's 2026 report](https://data.finops.org/?ref=runtimewire) found that 98% of respondents managed AI spending, up from 63% in 2025. Respondents ranked granular tracking of tokens, model requests and GPU use as their most desired missing tooling capability. They also identified cost allocation and determining AI's business value as persistent problems.\n\nThat gap has created room for advisers such as Adaptovate. [Michael Murphy](https://www.adaptovate.com/people/michael-murphy/?ref=runtimewire), an Adaptovate partner who previously led transformation projects at Deloitte, told Business Insider that the work starts with decisions about organizational structure, talent and operating models. Adaptovate's assignments range from helping a food manufacturer introduce AI in its supply chain to helping a 30,000-person professional-services organization redesign teams around generative AI products.\n\nConsultants benefit when AI adoption becomes an organizational redesign rather than a software purchase. Their challenge is proving that a new operating model produced measurable gains after accounting for implementation costs, training and the underlying model bill.\n\n### Larridin wants to measure the return\n\n[Larridin founders Russ Fradin and Jim Larrison](https://larridin.com/about-us?ref=runtimewire) came to the problem from digital measurement and later helped build Dynamic Signal. They built Larridin with Ameya Kanitkar, an engineering and data executive whose prior employers include LinkedIn, Coinbase and Groupon, as cofounder and CTO.\n\nTheir product, [Larridin Scout](https://larridin.com/scout?ref=runtimewire), monitors which AI applications employees use, how often they use them and how workers assess the resulting productivity gains. Larridin says Scout recognizes more than 300 AI tools and refreshes usage data every 15 minutes. Scout also combines passive activity detection with employee surveys, which gives management a view of adoption while leaving part of the productivity calculation dependent on self-reported results.\n\nKanitkar told [Business Insider](https://www.businessinsider.com/ai-cost-saving-businesses-startups-roi-2026-7?ref=runtimewire) that Larridin had a slow start because early buyers were less concerned with spending. He said traction had doubled each quarter since the beginning of 2026 as corporate commitments grew. That growth rate is Larridin's own operating measure.\n\nLarridin launched Scout in 2025 and raised $17 million in seed funding from Andreessen Horowitz, Bloomberg Beta, GV and other investors. Larridin's founders are betting that AI spending will need a measurement layer resembling the analytics infrastructure that followed the commercial web. The comparison is useful, although AI attribution is harder: a click can be counted directly, while time saved or work improved by a model depends on workflow, employee behavior and the baseline chosen by management.\n\n### Builders go after the unit cost\n\nOther founders are attacking the expense before it reaches a finance dashboard.\n\n[Flaviu Radulescu](https://runware.ai/blog/authors/flaviu-radulescu?ref=runtimewire), Runware's founder and CEO, spent more than 20 years building data and compute infrastructure, including systems for Booking.com, Vodafone and Transport for London. He founded Runware in 2023 and designed custom inference clusters spanning storage, cooling and other hardware components around the GPUs.\n\nRunware cofounder [Ioana Hreninciuc](https://runware.ai/blog/authors/ioana-hreninciuc?ref=runtimewire) brought a product and operating background from Homa Games, Huuuge Games and GameAnalytics. Business Insider groups Runware with builders trying to lower AI costs at the infrastructure layer.\n\nRunware sells a unified API for switching and routing work across image, video, audio and language models. On Dec. 11, 2025, Runware [announced a $50 million Series A](https://runware.ai/blog/runware-raises-50m-series-a-to-power-all-intelligent-applications?ref=runtimewire) led by Dawn Capital, with Comcast Ventures, Speedinvest, Insight Partners and a16z Speedrun participating. Runware says its infrastructure had processed over 10 billion generations for more than 200,000 developers at the time. Those figures and Runware's claims of price and performance improvements come from Runware.\n\nJunchen Jiang is pursuing a narrower technical wedge at [Tensormesh](https://www.tensormesh.ai/blog-posts/tensormesh-emerges-from-stealth?ref=runtimewire). Jiang, a University of Chicago computer science professor, co-created LMCache, an open-source project for retaining the intermediate data that models generate while processing prompts. Tensormesh commercializes that research so repeated context can be reused rather than recomputed on each request.\n\nIn May 2026, Tensormesh [announced $20 million in new funding](https://www.tensormesh.ai/blog-posts/tensormesh-raises-20m-launches-inference-platform?ref=runtimewire) from AMD Ventures, CoreWeave, NVentures, Valley Capital Partners and Laude Ventures, bringing its stated total to $24.5 million. Tensormesh claims its caching can cut latency and GPU spending by as much as 10 times, while its own disclosure says results vary by workload, hardware and infrastructure configuration. The funding gives Tensormesh an unusual investor lineup: several backers sell or finance the compute that Tensormesh is trying to make more productive.\n\n### Conifer starts with the cheapest model\n\nAt the earliest end of the category, Michael Jeffords and Charles Muehlberger are building [Conifer](https://www.ycombinator.com/companies/conifer?ref=runtimewire), a three-person San Francisco company in Y Combinator's Summer 2026 batch.\n\nJeffords previously worked on clinical research and machine-learning systems for detecting signs of neurological disease. Muehlberger is a Princeton engineering researcher whose work includes multimodal inference on edge devices and physical neural networks. Their product routes each query from a user's own hardware to an efficient cloud model and finally to a frontier model when the task demands it.\n\nConifer told [Business Insider](https://www.businessinsider.com/ai-cost-saving-businesses-startups-roi-2026-7?ref=runtimewire) it had raised $1.3 million before Demo Day, representing about 20% of its target.\n\nThe founders' routing hierarchy captures the central bet behind this category. Frontier models remain useful for difficult work, while smaller and local systems can absorb routine requests. A buyer that can identify the difference before inference begins has a chance to reduce costs without simply imposing usage limits.\n\nThe durable businesses will need to demonstrate savings against clear baselines and retain customers as model prices fall. Measurement providers must connect activity to business results. Routers need enough workload data to choose models accurately. Infrastructure providers must keep their cost advantage as chips, open models and incumbent APIs improve.\n\nEnterprise buyers once treated maximum model access as the safest AI strategy. Koukoumidis, Kanitkar, Hreninciuc, Jiang, Jeffords and Muehlberger are building for the next phase, when every token, GPU cycle and enterprise license has to earn its place on the bill.", "url": "https://wpnews.pro/news/oumi-larridin-and-runware-build-tools-to-cut-enterprise-ai-bills", "canonical_source": "https://runtimewire.com/article/oumi-larridin-runware-ai-cost-controls", "published_at": "2026-08-03 07:19:09+00:00", "updated_at": "2026-08-03 07:25:34.643274+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-tools", "ai-startups"], "entities": ["Oumi", "Manos Koukoumidis", "Larridin", "Adaptovate", "FinOps Foundation", "Venrock", "Obvious Ventures", "Google Cloud"], "alternates": {"html": "https://wpnews.pro/news/oumi-larridin-and-runware-build-tools-to-cut-enterprise-ai-bills", "markdown": "https://wpnews.pro/news/oumi-larridin-and-runware-build-tools-to-cut-enterprise-ai-bills.md", "text": "https://wpnews.pro/news/oumi-larridin-and-runware-build-tools-to-cut-enterprise-ai-bills.txt", "jsonld": "https://wpnews.pro/news/oumi-larridin-and-runware-build-tools-to-cut-enterprise-ai-bills.jsonld"}}