Case Study

Open-Source AI vs. Proprietary Models: The Breakthrough Is Real — But Not Where Everyone Is Looking

This isn't an SEO roundup. 9 primary sources — from Stanford's AI Index to Andreessen Horowitz's enterprise survey to DeepSeek's own technical report — evaluated, every claim linked directly to its source.

06.05.2026 · Research Case Study

Industry: AI & Enterprise Software Read time: 12 min Evidence: High (primary sources, benchmarks & market data) 9 sources reviewed

Quick-Check — Executive Summary

Core finding: Enterprise AI adoption jumped from 55% (2023) to 78% (2024) according to the Stanford AI Index 2025 — that alone says nothing about which type of model companies actually chose.
The trend points proprietary: a16z now counts over a third of enterprises with a clear closed-source preference; willingness to host directly with frontier labs rose from roughly 40% to roughly 80%.
Yet open source is already everyday infrastructure: 60% of IT decision-makers use open ecosystems according to IBM, and more than 80% have built at least a quarter of their AI stack on them.
The quality gap is shrinking fast: from an 8% benchmark deficit down to 1.7% in a single year (Stanford) — though at the very top of the leaderboards, LMArena and Artificial Analysis show it's still real.
The real lever is price: DeepSeek-V3 cost 5.576 million USD to train according to its own technical report; leading open models now solve tasks up to 25x cheaper than top proprietary models.
Governance beats cost: DeepSeek bans in Italy and Taiwan, half of Netskope's customer base blocking access — data sovereignty is holding back open-source adoption harder than any benchmark.

The Setup

Open source vs. proprietary: why the debate stays on the surface

Since Llama, Mistral, Qwen and DeepSeek released new generations of models in rapid succession, a narrative has taken hold: open-source AI is closing the gap on OpenAI, Anthropic and Google, and turning into a serious enterprise option. The question this narrative usually skips is simple but decisive: how many companies actually run open models in production — not in a sandbox, but in systems with real users, real data and real compliance obligations? And exactly where does the line run between "technically competitive" and "enterprise-ready"?

For this analysis we evaluated nine primary sources: the Stanford AI Index 2025, Andreessen Horowitz's (a16z) ongoing enterprise AI survey, an IBM study on open-source ROI, Mistral AI's official model announcement, the current live leaderboards from Arena (formerly LMSYS Chatbot Arena) and Artificial Analysis, DeepSeek's official DeepSeek-V3 technical report, Meta's own AI blog on Llama usage, and a Cybersecurity Dive investigation into DeepSeek-related enterprise risk. Every number in this article comes directly from one of these sources and is linked below.

The adoption jump is historically fast — but model-agnostic

The Stanford AI Index 2025 puts organizational AI adoption at 78% in 2024, up from 55% in 2023, and finds active use of generative AI in business functions more than doubled from 33% to 71%. That's a 23-percentage-point jump in twelve months — a speed comparable enterprise technologies have historically taken several years to reach. Critically, the report doesn't break this down by model type: these numbers say AI is being used, not whether it's open or closed. That gap between the adoption curve and the actual model strategy is exactly where this analysis starts.

Source: Stanford HAI — AI Index Report 2025

Enterprises are drifting toward proprietary, not toward open source

Andreessen Horowitz has repeatedly surveyed CIOs and AI leads at large companies about their model strategy since March 2024. The result: preference for closed-source models has risen "steadily" since then, with over a third of enterprises now explicitly preferring closed models. In parallel, willingness to host models directly with frontier labs like OpenAI or Anthropic grew from roughly 40% in March 2024 to roughly 80% today. a16z names three drivers: the pace of quality improvements in proprietary models, limited in-house AI talent, and data-security concerns — the very argument open-source proponents usually make in the opposite direction.

Source: Andreessen Horowitz (a16z) — Enterprise AI Survey

And yet: open source is already everyday reality

The picture flips once you ask about actual usage instead of stated preference. An IBM study from December 2024 found 60% of surveyed IT decision-makers already use open-source ecosystems as a source of AI tooling, and more than 80% say at least a quarter of their AI solutions are built on open source. Companies using open source report positive ROI more often (51%) than those that don't (41%). For 2025, 41% of respondents expected further growth in open-source-based solutions, up from 37% the year before. Open source, in other words, is less a deliberate strategic choice than infrastructure that's already running everywhere — often below the radar of the C-suite that a16z's survey actually reaches.

Source: IBM Newsroom — Open-Source AI ROI Study

The license trap: "open weights" is not "open source"

One reason for the conflicting numbers: the term "open source" gets used very loosely in practice. Mistral AI released Mistral Large 2 in July 2024 (123B parameters, 84.0% on MMLU, 128k context window, positioned by Mistral itself on par with GPT-4o, Claude 3 Opus and Llama 3 405B) with openly available weights — but under the Mistral Research License, which explicitly excludes commercial use and requires a separate, paid commercial license. Llama runs under its own community license with usage caps for large enterprises, DeepSeek's models sit under a permissive MIT license, Qwen under Apache 2.0. For enterprise legal and compliance teams that means every "open-source" decision needs its own license review — a governance cost that almost never shows up in marketing comparisons of "open" versus "closed."

Source: Mistral AI — Mistral Large 2 announcement

The quality gap is closing measurably — a live check confirms the trend

The Stanford AI Index 2025 quantifies the performance gap between leading open and closed models: on selected benchmarks, the deficit for open models shrank from 8% to 1.7% within a single year. A live check on the Arena leaderboard (the successor platform to LMSYS Chatbot Arena, where users blind-rate model responses against each other) confirms the direction but also shows the limit: the current ranking is led by a proprietary model with a clear margin, and the best open model sits noticeably behind it. The gap is narrowing — it hasn't disappeared at the very top.

Source: Arena (formerly LMSYS Chatbot Arena) — Leaderboard

Proprietary still leads at the top — but cost-per-task flips the math

The Artificial Analysis leaderboard compares models not just by an intelligence index but also by cost per task solved — and this is where the real lever for enterprises appears: the current leading open model in the dataset solves tasks for a fraction of a US cent, while a comparably capable proprietary model costs roughly 25 times as much. The quality difference between the two over the same period is measurable but far smaller than the price difference. For high-volume use cases with moderate quality requirements — classification, summarization, internal search — the quality debate turns into a cost-benefit calculation.

Source: Artificial Analysis — LLM Leaderboard

The cost proof from practice: DeepSeek-V3

What this means for training economics is visible in DeepSeek's official DeepSeek-V3 technical report: the Mixture-of-Experts model, with 671B total and 37B parameters activated per token, was trained on 14.8 trillion tokens and needed 2.788 million H800 GPU-hours to do it — at an assumed 2 USD per GPU-hour, that's a total training cost of 5.576 million USD. DeepSeek-AI explicitly notes there were no "irrecoverable loss spikes" or rollbacks during the entire training run — a sign of technical maturity, not just cost optimization. The model positions itself, per the report, on par with leading closed-source models, at a fraction of the publicly discussed training budgets of major proprietary systems.

Source: DeepSeek-AI — DeepSeek-V3 Technical Report (arXiv)

Technical Deep-Dive

DeepSeek-V3 illustrates why open-weight cost savings aren't purely a licensing effect but also an architecture effect. By combining Multi-Head Latent Attention (MLA) with the DeepSeekMoE architecture, the model activates only 37 of its total 671B parameters per token — the rest stay inactive for any given request. That drives down both training and inference cost drastically without reducing effective model capacity. For companies considering self-hosting, that's the actual lever: it's not the license fee that decides economics, but whether the underlying architecture genuinely needs little compute per request. Whoever hosts a dense 70B model pays the full parameter count on every request; whoever hosts a Mixture-of-Experts model of similar overall size but a small active path can run comparable quality far more cheaply. That's also why Artificial Analysis's cost-per-task figures for MoE models like DeepSeek sit so far below those of dense models in a similar parameter class.

Who's actually running open source: Llama in enterprise practice

Meta's own AI blog reports that Llama token usage among major cloud providers more than doubled between May and July 2024, and monthly usage grew tenfold since January 2024. On Hugging Face, Meta counts over 350 million downloads (a 10x increase year-over-year), 20 million of those in a single month alone, plus more than 60,000 derivative models across the ecosystem. Meta names concrete enterprise customers including AT&T, DoorDash, Goldman Sachs, Nomura, Shopify, Spotify, Zoom, Accenture, Infosys and KPMG; Shopify, per Meta, runs 40 to 60 million inferences per day with a Llama-based vision model for product data. This is the most concrete available evidence that open-source adoption at scale isn't hype but production infrastructure — at least from a vendor with a clear self-interest in showing exactly that.

Source: Meta AI — Llama Usage Update (May–July 2024)

The governance brake: when data protection decides the model choice

DeepSeek shows the counterweight to the cost argument. According to reporting by Cybersecurity Dive, security vendor Armis alone blocked access to the DeepSeek chat app for "hundreds" of its corporate clients within a single week; at Netskope, roughly half the customer base blocks employee access by its own account. Italy and Taiwan have banned DeepSeek applications outright, and the US government is reviewing national-security risk. The core problem: user data is stored on servers in China and falls under Chinese law. Cited security experts warn that employees reflexively feed sensitive data into these tools — customer data, source code, intellectual property. For enterprise governance, that means the open-source question is often not a cost question at all, but a sovereignty question — and that gets decided by server location and jurisdiction, not the price per token.

Source: Cybersecurity Dive — DeepSeek Enterprise Security Risks

"The core risk is that employees open tools like DeepSeek and start entering sensitive data — customer data, source code, regulated data, intellectual property."

Security experts, quoted in Cybersecurity Dive — 2025

Open source vs. proprietary: the key metrics compared
Parameter	Baseline	Observed value	Effect	Source
Inference cost (GPT-3.5-level)	20 USD / M tokens (Nov 2022)	0.07 USD / M tokens (Oct 2024)	−280x	[1]
Open-weight performance gap (benchmark)	8% deficit (2023)	1.7% deficit (2024)	−79% relative	[1]
Organizational AI adoption (overall)	55% (2023)	78% (2024)	+23 pts	[1]
Preference for closed-source (enterprises)	Baseline March 2024	over a third now explicitly prefer closed-source	rising trend	[2]
Willingness to host directly with frontier labs	roughly 40% (March 2024)	roughly 80% (current)	+40 pts	[2]
DeepSeek-V3 training cost	2.788M H800 GPU-hours	5.576M USD total cost	a fraction of known closed budgets	[7]

Learnings & Actionable Insights

What companies should take from the 2025 open-source debate

For companies

"Open vs. closed" isn't a global decision, it's a per-use-case one. For high-volume tasks with moderate quality requirements, the cost numbers clearly favor open models; for tasks at the quality frontier, proprietary is often still ahead. Check license type and server location before every decision, not after.

For agencies

Model-stack advisory is becoming its own service line: 81% of enterprises already orchestrate three or more model families in parallel, per a16z. Clients don't need a recommendation for "the one tool" — they need a routing and licensing strategy across multiple vendors.

For developers & creators

Open models like Qwen, DeepSeek and Llama remain the cheapest environment for prototyping and fine-tuning. Understanding why Mixture-of-Experts architectures flip the cost curve leads to better infrastructure decisions than just reading benchmark rankings.

For investors & analysts

The TCO convergence a16z observes weakens simple "cheaper than API" theses for open-source infrastructure. The more durable, cost-independent driver is governance and data sovereignty — that's where standalone market segments form, independent of price per token.

Open Questions

How fast will the quality gap close specifically on reasoning and agentic tasks, where a16z finds proprietary models still clearly preferred?
Will the EU AI Act ultimately act as a driver or a brake on open-source adoption among European companies?
How many companies underestimate the real total cost of ownership of self-hosting — engineering, maintenance, security patching — versus a pure token-price calculation?
Does the current license fragmentation (community license, research license, Apache 2.0, MIT) persist, or does the market consolidate on more uniform standards?
Are companies increasingly decoupling "model quality" from "country of origin" as Chinese models keep closing the gap despite governance concerns?

Verified Sources & Studies

AI Index Report 2025Stanford HAI, April 2025The only independent long-run series on adoption, inference cost and the open-weight-vs-closed performance gap.
Leaders, gainers and unexpected winners in the Enterprise AI arms raceAndreessen Horowitz (a16z)Ongoing first-hand CIO survey on model preference, multi-model strategies and TCO convergence.
IBM Study: More Companies Turning to Open-Source AI Tools to Unlock ROIIBM Newsroom, December 19, 2024Documents actual open-source usage inside IT departments, independent of C-suite preference surveys.
Large Enough — Mistral Large 2Mistral AI, July 2024Official benchmark and license details straight from the model maker, illustrating license fragmentation in detail.
Arena Leaderboard (formerly LMSYS Chatbot Arena)Arena / LMArenaIndependent, community-based live ranking as a counter-check to vendor claims.
Artificial Analysis LLM LeaderboardArtificial AnalysisThe only source that systematically pairs an intelligence index with cost per task.
DeepSeek-V3 Technical ReportDeepSeek-AI, arXiv, December 2024Primary source for architecture, training cost and GPU-hours of a specific open-weight model.
With 10x growth since 2023, Llama is the leading engine of AI innovationMeta AI Blog, 2024The only source with concrete, named enterprise adoption cases for an open model.
DeepSeek surge hits companies, posing security risksCybersecurity Dive, 2025Documents with concrete numbers from Armis and Netskope how governance concerns really slow open-source adoption.

← News