Adaptation of LLM is increasing in businesses, with some sectors using about 150 % more year year. Nevertheless, the amount of infrastructure flowing is dramatically slowing down, with new infrastructure investment decreases by more than half the previous year.
These changes are part of a broader shift featured in the 2025 -year -old LLM market update of Menlo Ventures. The report takes a look at how the genius is being solidified, with speed of which is more focused on models and early agent experiments, specially emphasizing the emphasis on providing applications, special workflows, and business results.
Menlo Ventures is a venture capital firm that focuses on AI, infrastructure, and enterprise software. This mid -year refreshing is based on their 2024 market report, which is a detailed look at how the LLM environmental system is changing after the investment samples, product strategies, and enterprise adopt after going to its next phase.
Infrastructure investment is to cool down. In 2024, the model was attracting a heavy fund, and vector database heavy funds. In the mid -2025, the dell volume in these areas has dropped by more than half.
Open models such as Maxteral, Cloud, and Lalama 3 have improved rapidly, which makes it easier and cheaper to build on existing systems rather than regenerate the stack. Over the past one year, the costs of the model API have also increased by about 4X, which has further reduced the need for companies to create or run their foundation models. As a result, competitive edge is ending around the infrastructure, and the ability to serve or fix a model is no longer enough to stand.
Instead, the value is moving towards companies that directly plug, bring their proprietary data, or focus on solving problems in specific domains. Menlo said the most successful players are not trying to rebuild the entire stack. They are using the best tools available and focusing on their efforts where it is most important, in the application layer, where users really see the effect.
As the report states, “the market is moving from the horizontal platform to vertical steaks.” The strongest startups are “solving issues for a particular domain in a particular domain” and bundled “Domain -related UX, Workflows, Data and Integration”.
These players are gaining traction rapidly, showing the “strong product fit and more efficient market motion”, especially when made with a recovery technique with a collective breed (RAG). In this environment, “Startups that control the distribution or data are far more defensive than pure infrastructure builders.”
The report also notes that the agent’s development is becoming more focused and practical. After the initial enthusiasm around the general purpose agents, which often promised extensive abilities, but their reliability is lacking, the market is moving towards the tools created for repetitive tasks. These include use issues such as a document summarizing, lead generation, and extracting data.
Menlo says the focus of this new focus on vomiting is how AI startups are being evaluated. Investors are looking more closely at the basic principles of basic business, such as how fast the value of a product is, how strong the margins, and what the Grahs really live. In response, many of the startups are diminishing their attention, making their services bundled, or making it easier to sell how they sell.
Further established software companies are also entering the place. The report notes that major platforms are now adding LLM features to their customers already using products. This gives them a built -in benefit because they are already trying to reach the user, confidence, and that they are still trying to build a new AI startup.
Shutter stock
Menlo expects a small number of companies trying to own more stability and full stack in the market. At the same time, the enterprise buyers are getting stronger on how they adopt the living. Many people are on their second or third deployment cycle and are now preferring solutions that are safe, stable and manageable rather than experimental.
Nevertheless, the report points to a few areas that are pushing. These include agent observation, compliance systems, auto-rag pipelines, and artificial data platforms. Menlo sees them as the key parts of what the enterprise will remove the next wave.
This change also shows how companies are choosing their LLM. Open is still the most used API, but its share has been reduced from 80 % to 59 % over the past two years. Clouds and misunderstandings are getting land, especially in fields that care more about cost or compliance.
Uses are also spreading. Instead of relying on a single provider, many teams are mixing and combining the price, performance, and the cost -based price. In a year, the cloud alone increased the use of enterprise beyond only 3 %.
The Menlo also indicates the growing interest in the open source. Open models are improving rapidly, and it is making companies more space to build themselves without being locked in a provider. This flexibility is important when teams begin deploying models in a real production environment.
At the same time, expectations are changing. About 70 70 % of companies say they are already on their second or third LLM rollout. Buyers are no longer chasing the shiny demo. Now what they want is stability, control and clear business price. Menlo calls it a symbol of growing fatigue in the market. Mid -year updates show that the market is in a more practical phase where buyers are making decisions based on real needs, and shopkeepers are adjusting their strategy to meet them.
Relevant
