LLM adoption is surging across enterprises, with usage up nearly 150% year-over-year in some sectors. Yet the money flowing into infrastructure is slowing dramatically, with new infrastructure investments down by more than half compared to last year.
These changes are part of a broader shift highlighted in Menlo Ventures’ 2025 Mid-Year LLM Market Update. The report takes a grounded look at how the GenAI space is maturing, with momentum moving away from general-purpose models and early agent experiments toward more focused applications, specialized workflows, and greater emphasis on delivering business results.
Menlo Ventures is a venture capital firm focused on AI, infrastructure, and enterprise software. This mid-year update builds on their 2024 market report, offering a detailed look at how investment patterns, product strategies, and enterprise adoption are shifting as the LLM ecosystem moves into its next phase.
One of the biggest shifts Menlo highlights is the cooling of infrastructure investment. In 2024, model training platforms, orchestration tools, and vector databases were attracting heavy funding. Fast forward to mid-2025, and deal volume across these areas has fallen by more than half.
Open models like Mixtral, Claude, and LLaMA 3 have improved rapidly, making it easier and cheaper to build on existing systems instead of reinventing the stack. Model API spending has also surged, up nearly 4x over the past year, further reducing the need for companies to build or operate their own foundation models. As a result, the competitive edge around infrastructure is fading, and the ability to serve or fine-tune a model is no longer enough to stand out.
Instead, value is moving toward companies that plug directly into real workflows, bring their own proprietary data, or focus on solving problems in specific domains. Menlo points out that the most successful players are not trying to rebuild the entire stack. They are using the best tools available and focusing their efforts where it matters most, at the application layer, where users actually see the impact.
As the report explains, “the market is moving from horizontal platforms to vertical stacks.” The strongest startups are “solving problems for a specific user in a specific domain” and bundling “domain-specific UX, workflows, data, and integrations.”
These players are gaining traction faster, showing “stronger product-market fit and more efficient go-to-market motion,” especially when paired with techniques like retrieval-augmented generation (RAG). In this environment, “startups that control distribution or data are far more defensible than those building pure infrastructure.”
The report also notes that agent development is becoming more focused and practical. After a period of early excitement around general-purpose agents, which often promised broad capabilities but lacked reliability, the market is shifting toward tools designed for repeatable tasks. These include use cases like document summarization, lead generation, and structured data extraction.
Menlo says this renewed focus on reliability is shifting how AI startups are being evaluated. Investors are looking more closely at core business fundamentals such as how fast a product delivers value, how strong the margins are, and whether customers actually stay. In response, many startups are narrowing their focus, bundling their services, or simplifying how they sell.
More established software companies are entering the space as well. The report notes that larger platforms are now adding LLM features to products their customers already use. This gives them a built-in advantage because they already have users, trust, and reach that newer AI startups are still trying to build.
Shutterstock
Menlo expects this to lead to more consolidation in the market and a smaller number of companies trying to own the full stack. At the same time, enterprise buyers are maturing in how they adopt GenAI. Many are on their second or third deployment cycle and are now prioritizing solutions that are secure, stable, and manageable rather than experimental.
Even so, the report points to a few areas that are gaining momentum. These include agent observability, systems built for compliance, auto-RAG pipelines, and synthetic data platforms. Menlo sees these as key parts of what will drive the next wave of enterprise GenAI.
That shift is also showing up in how companies are choosing their LLMs. OpenAI is still the most used API, but its share has dropped from 80% to 59% in the past two years. Claude and Mistral are gaining ground, especially in sectors that care more about cost or compliance.
Usage is also becoming more spread out. Instead of relying on a single provider, many teams are mixing and matching based on price, performance, and what fits the task. Claude alone grew from just 3% to 16% of enterprise use in a year.
Menlo also points to growing interest in open source. Open models are getting better fast, and that is giving companies more room to build without locking themselves into one provider. This flexibility matters more as teams start to deploy models in real production environments.
At the same time, expectations are changing. Around 70% of companies say they are already on their second or third LLM rollout. Buyers are no longer chasing flashy demos. What they want now is stability, control, and clear business value. Menlo calls this a sign of growing fatigue in the market. The mid-year update suggests the market is settling into a more practical phase where buyers are making decisions based on real needs, and vendors are adjusting their strategies to meet them.