Are open source or proprietary AI models winning in the enterprise?

Both, depending on the workload. Menlo Ventures found closed-source models power 87% of enterprise production workloads, while Databricks found 76% of companies using large language models have also adopted an open-source model somewhere in their stack, often alongside a proprietary system.

Which AI vendors hold the most enterprise market share?

OpenAI's share of enterprise LLM usage fell from roughly 50% in 2023 to 25% by Menlo's 2025 mid-year survey, with Anthropic capturing 32% and Google taking 20%, though OpenAI still commanded about 69% of relative market share among companies with a generative AI model actually in production.

Why do enterprises choose proprietary AI models for production?

Andreessen Horowitz's enterprise AI survey found control was cited as the primary driver behind model choice by 60% of executives, while cost was cited by only 10%, showing enterprises prioritize reliability and vendor accountability over cost savings for production workloads.

Where do open-source AI models see the strongest enterprise adoption?

Technology Checker's 2026 domain-level crawl found IT services and consulting firms lead open-source AI deployment at 8.4% of detected companies, followed by technology and internet firms, higher education, and software development, with regulated industries like financial services and healthcare also favoring self-hosted models for data-residency and audit requirements.

Published: July 21, 2026

Open Source vs. Proprietary AI Models: Who's Winning the Enterprise?

Ask a room full of enterprise technology leaders whether open source or proprietary AI is winning, and you'll get two contradictory answers that are both correct. According to Menlo Ventures' 2025 Mid-Year LLM Market Update, closed-source models now power 87% of enterprise production workloads. Yet Databricks' State of AI report found that 76% of companies using large language models have adopted an open-source model somewhere in their stack, often running it alongside a proprietary system rather than instead of one. The real story isn't a contest with a single winner it's a rapidly bifurcating stack where the two approaches are being deployed for entirely different jobs.

The Adoption Numbers Tell a Split Story

The scale of enterprise AI spending makes this split worth taking seriously. Menlo Ventures tracked enterprise LLM spend more than doubling in just six months, from $3.5 billion in November 2024 to $8.4 billion by mid-2025, and total enterprise generative AI investment tripled to $37 billion across 2025 as a whole. Gartner separately forecasts that over 80% of enterprises will have deployed generative AI applications or GenAI APIs by 2026, up from less than 5% in 2023. That is not a market settling into one architecture; it's a market still actively testing which workloads justify which approach, at a pace few software categories have matched.

What's changed is who is capturing that spend. OpenAI held roughly 50% of enterprise LLM usage in 2023; by Menlo's 2025 mid-year survey, its share had fallen to 25%, with Anthropic capturing 32% and Google taking 20%. Enterprises are also not loyal to a single vendor: organizations typically deploy three or more foundation models at once, and while only 11% of teams switched providers outright in the past year, 66% upgraded to a newer model from their existing vendor. That pattern multi-model, but sticky within a chosen vendor describes the proprietary side of the ledger far better than a simple 'closed models are winning' headline would suggest.

Why Proprietary Still Wins the Production Battle

Closed models dominate production because enterprises are optimizing for something other than raw capability or cost. In Andreessen Horowitz's enterprise AI survey, control was cited as the primary driver behind model choice by 60% of executives, while cost was cited by only 10% a striking reversal of what most vendors assume drives adoption. Reliability, managed infrastructure, and vendor accountability matter more to large organizations than shaving inference costs, particularly once a workload moves from a pilot into something customer-facing. That preference shows up directly in the production numbers: in a16z's data, OpenAI alone commanded about 69% of relative market share among companies with at least one generative AI model actually in production, compared with just 13% for Google and 12% for Meta's Llama even though far more companies were testing open alternatives than that gap implies.

Where Open Source Quietly Wins

The open-source story is less about replacing proprietary models and more about who is actually shipping with them. Domain-level detection data from Technology Checker’s 2026 crawl of over 50 million company domains found that IT services and consulting firms lead open-source AI deployment at 8.4% of detected companies, followed by technology and internet firms at 6.9%, higher education at 4.7%, and software development at 4.4% a pattern that fits small, technical teams forking and self-hosting frameworks for client-specific builds rather than large enterprises standardizing on open weights company-wide. That nuance matters: vanity metrics like GitHub stars or Hugging Face download counts overstate real deployment, since roughly half of all models hosted on Hugging Face see fewer than 200 downloads. The open-source advantage isn't cost or hype it's control over data residency, the ability to fine-tune on proprietary datasets without sending them to a third party, and freedom from API rate limits in latency-sensitive applications.

This is also where regulated industries quietly diverge from the broader market narrative. Financial services and healthcare organizations facing strict data-residency and audit requirements are more likely to self-host an open-weight model specifically to keep sensitive records off a third-party vendor's servers, even when a proprietary API would otherwise be the faster path to production. That trade-off slower time-to-deploy in exchange for full data custody rarely shows up in headline adoption statistics, but it explains why open-weight usage persists even in sectors where proprietary vendors otherwise dominate the conversation.

The Real Battleground Is the Hybrid Stack

The more useful question for enterprise buyers isn't which model type is 'winning,' but which workloads are migrating to which architecture. Coding and developer tools alone attracted $7.3 billion in enterprise AI investment in 2025, with half of all developers now using AI daily a category where open-weight models running on internal infrastructure are increasingly competitive with hosted APIs on cost-per-token at scale. At the same time, general-purpose copilots, which demand consistent quality across unpredictable user queries, drew $8.4 billion and remain dominated by proprietary vendors. Enterprises aren't choosing sides; they're routing workloads to whichever architecture matches the specific mix of control, latency, compliance, and cost sensitivity that workload demands.

What This Means for Enterprise AI Strategy

Framing this as a binary contest misses where the market has actually landed. The data points to enterprises running proprietary models for customer-facing, high-stakes production use cases where reliability and vendor support outweigh cost, while reserving open-weight models for internal tooling, regulated data environments, and workloads where fine-tuning control matters more than out-of-the-box capability. For technology leaders building a 2026 AI roadmap, the practical takeaway is to stop asking whether to standardize on open or closed models, and start building the evaluation criteria — data sensitivity, latency tolerance, customization needs, and total cost at projected scale — that determines which architecture each individual workload actually needs.

Frequently Asked Questions

Are open-source AI models cheaper than proprietary models for enterprises?

Not necessarily once total cost of ownership is factored in. Open-weight models can have lower per-token hosting costs at scale, but they require in-house MLOps expertise, infrastructure management, and fine-tuning work that proprietary APIs bundle into the subscription price. a16z's enterprise survey found cost was cited as the primary reason for model choice by only 10% of executives control and customization mattered far more.

Why do closed-source models still dominate enterprise production if open source usage is so widespread?

Because 'usage' and 'production deployment' measure different things. Many enterprises test or use open-source models for internal tools, prototyping, or specific fine-tuned use cases, but reserve customer-facing and mission-critical workloads for proprietary vendors that offer managed infrastructure, uptime guarantees, and dedicated support which is why Menlo Ventures found closed models powering 87% of production workloads even as most companies also touch open source somewhere in their stack.

Which industries lead open-source AI adoption?

Technical, software-building sectors lead by a clear margin. IT services and consulting, technology and internet companies, higher education, and software development firms show the highest detected rates of live open-source AI deployment, reflecting their in-house technical capacity to self-host and customize open-weight frameworks.

Is the gap between open-source and proprietary AI model performance closing?

Open-weight models have narrowed the capability gap considerably, but the deciding factor for most enterprises has shifted from raw benchmark performance to operational concerns governance, vendor reliability, and integration support which is why multi-model, hybrid strategies rather than a single winner-take-all architecture have become the norm.