This week, venture firm Menlo Ventures released a report (“2025 Mid-Year LLM Market Update: Foundation Model Landscape + Economics”) based on a Jun-Jul 2025 survey of 150 technical leaders at enterprises and startups. Among the findings was the emergence of Anthropic as the new leader in enterprise AI with 32% of the market (in compute), followed by OpenAI (25%), Google (20%), Llama (9%), and DeepSeek (1%). It’s a stark contrast from OpenAI’s 50% share and market leadership at the end of 2023. In Menlo Ventures words, “[E]nterprise dollars are now consolidating around a few high-performing, closed-source models, giving us a new market leader in Anthropic.”

The report highlighted several shifts in how enterprises are spending. Large companies are becoming less focused on model development and spending dramatically more on foundation model APIs from 3rd-party vendors (like Anthropic and OpenAI) – with such spend rising from $3.5B for full-year 2024 to just $8.4B in H1 2025.

Notably, enterprises are increasingly using these APIs for “production inference” – i.e. deploying the models in their business’ production systems. 49% say that most of their compute is now being used for inference – up from 29% in 2024. (The figure is even higher for startups.) It indicates that the market has matured to the point where enterprises are becoming comfortable bringing AI into the heart of their business.

Surprisingly, open-source models are losing ground relative to closed-source models – and largely because of performance. In part, this is because of Meta’s leading role in open-source LLMs and its Llama 4’s disappointing performance. According to the report, only 13% of workloads today use open-source models, down from 19% in Sep-Oct 2024. In the same vein, only 11% of enterprises are using open-source for at least 50% of their workloads. In general, open-source models are trailing closed-source models in performance by 9-12 months. While open-source has advantages in customization, cost savings, and deployment flexibility, these advantages are all overridden by the gap in performance.

Enterprises are being pragmatic in their adoption of Anthropic’s Claude. IT is often a major cost line for large firms, and code generation was perhaps the first “killer app” for AI (dating back to late 2022). Anthropic began pulling ahead of the others in code generation in mid-2024 with Claude 3.5 Sonnet, and has been able to maintain its popularity – despite switching being relatively easy. According to Menlo Ventures, Anthropic has a 42% market share among developers, followed by OpenAI (21%), Google (16%), and Meta/Llama (10%). Claude Sonnet is also the “most admired” model family among developers, per Stack Overflow’s 2025 Developer Survey (May-Jun 2025).

Enterprises tend to be a relatively sticky audience, barring a significant performance gap. Over the past year, 66% opted to upgrade their models within their existing vendor, and an additional 23% stayed with their existing model and vendor – meaning just 11% switched vendors. Even as prices plummeted on older models, enterprises demonstrated willingness to pay more to stay on the frontier.

Anthropic’s pursuit of business-oriented use cases is driving a surge in its recurring revenue. Anthropic’s annualized revenue has grown rapidly from $3B in late May 2025 to $4B in early Jul 2025 to about $5B at the end of Jul 2025. Furthermore, Claude’s popularity for coding among software engineers could give it a leg up in other realms. For instance, just a few weeks ago, Anthropic launched Claude for Financial Services as a “tailored version of Claude” for companies in the financial services industry.

While most enterprises are now focused on using AI models rather than training them, the major AI players continue to push on the avenues available for continued scaling and performance improvement. The Menlo Ventures report highlights, for instance, post-training using reinforcement learning with verifiable rewards (RLVR). RLVR is particularly useful in arenas with clear ground-truth answers – such as coding, math, complex scheduling, or any task with programmatically verifiable success criteria. Other avenues include long memory, data selection and augmentation, more test-time compute for reasoning models, cross-modal reasoning, multi-agent systems, optimized chips, and more.

Ex-Google CEO Eric Schmidt pointed out at TED2025: “You're six months ahead of me, and we're both on the same path for superintelligence. And you're going to get there, right? And I'm sure you're going to get there, you're that close. And I'm six months behind. Pretty good, right? Sounds pretty good. No. These are network-effect businesses. And in network-effect businesses, it is the slope of your improvement that determines everything.” In short, if a player is ahead and being ahead helps them move faster than anyone else, then they may have already won. On the other hand, one stumble can mean your rivals leapfrog ahead.

AI researcher Andrew Ng recently made a similar point with respect to China: “There is now a path for China to surpass the US in AI. Even though the US is still ahead, China has tremendous momentum with its vibrant open-weights model ecosystem and aggressive moves in semiconductor design and manufacturing. In the startup world, we know momentum matters: Even if a company is small today, a high rate of growth compounded for a few years quickly becomes an unstoppable force…China’s hypercompetitive business landscape and rapid diffusion of knowledge give it tremendous momentum.”

Jul 18 2025 (3 Shifts): AI takes aim at financial analysts

Oct 25 2024 (3 Shifts): The recent slate of capable LLMs