Adoption vs. where the money goes

Survey after survey says the same thing: Most companies are now using open source AI. Eighty-nine percent, sixty-three percent, “near universal” – pick your number, they all point in the same direction. But when you look at where actual usage is trending, measured in tokens processed rather than checkboxes checked, a very different picture emerges. This article contrasts adoption surveys with actual usage data and shows why “we use open source” and “open source is where our AI work actually happens” are not the same claims.

The most important adoption figures

Taken alone, these numbers suggest that open source AI has essentially won – achieving near-universal adoption across enterprise AI. This is the version of the story most often repeated in industry reporting.

The number that tells a different story: where the tokens actually go

Acceptance surveys ask: “Do you even use this?” Token volume measures what actually gets the job done. The gap between the two is the real story.

This is the number that rarely appears next to adoption statistics. Although a large majority of companies report “using” open-source AI somewhere in their stack, about four out of five tokens actually processed in the industry go through closed, proprietary models – models that cost about six times more per token than the available open alternatives.

Reconcile the two numbers

These aren’t contradictory statistics – they describe two different levels of how companies actually use AI:

• Adoption surveys count across the board: A company that runs a small open source model for an internal tool alongside a much larger proprietary deployment for its main customer-facing product counts as “open source usage” in a survey – even if that open source usage represents only a tiny fraction of total activity

• Token volume measures depth: It captures where the real computational work and therefore the real business value and costs are concentrated

• The typical corporate pattern: Open source models for less demanding, cost-sensitive or specialized internal workloads; proprietary flagship models for the most demanding, customer-visible and performance-critical jobs

This leads to the seemingly strange but consistent combination of near-universal “adoption” and a token volume that is still heavily concentrated in closed models.

Market size and growth

The market is growing really quickly and the cost benefit is real and significant – companies that move significant workloads to open models report around a third lower total cost of ownership. That’s a real, measurable incentive to move toward open source over time, even if it hasn’t yet flipped the overall balance between tokens and volume.

The achievement gap is closing, but it is not yet closed

Compared to a few years ago, when open models significantly lagged behind proprietary flagships, the gap has narrowed significantly. But closed models still have a slight edge at the top of the performance rankings and still occupy the majority of the top spots on the leaderboard – a small but real gap that likely explains part of why the most demanding workloads are still skewed towards proprietary models.

What happened in the large open weight laboratories in the first half of 2026?

The open weight landscape itself has changed significantly in 2026, and not just compared to previous years.

• Qwen followed a high-cadence release strategy and delivered frequent variants for different use cases and workload types in the first half of the year

• DeepSeek took the opposite approach and relied on a single major architectural reset – DeepSeek V4, released in March 2026, uses a mixture-of-experts design with 236 billion total parameters but only 21 billion active per inference, achieving performance comparable to GPT-4o-class models at a fraction of the computational cost

• MetaDespite being one of the earliest and best-known open-weight AI labs with its Llama family, it did not ship any new open-weight Llama releases between January 1 and mid-May 2026 – a notable pause on the part of the company most associated with pushing open models into the mainstream

This Meta pause is a particularly interesting data point given the company’s other moves in 2026: Rather than continuing to push Llama as a free, open alternative, Meta has instead launched a paid Meta Model API around its newer Muse Spark models and is building “Meta Compute,” a commercial cloud company. The company that helped legitimize open-weight AI at scale appears to be shifting at least part of its strategy toward a more traditional, revenue-generating model — even as the rest of the open-weight ecosystem (Qwen, DeepSeek) gains momentum.

Why companies say open source delivers better ROI, even with lower token ownership

One statistic complicates the simple interpretation of the token volume data “Closed models win”: Internal benchmarks show that companies that use open source AI models achieve approximately 25% higher return on investment than companies that rely exclusively on closed model APIs.

That’s not necessarily a contradiction. This likely reflects a selection: companies that successfully deploy open source models tend to do so for well-defined, cost-sensitive workloads where the ROI calculation is cheap and easy to measure – while companies that rely entirely on closed APIs may run a broader, messier mix of experimental and production workloads where measuring clean ROI is more difficult. In other words, open source can have better ROI, in part because it is used more consciously and selectively.

What this means for the future

The trajectory of the data suggests gradual convergence rather than a clear winner. The performance gap is shrinking, the cost gap significantly favors open models, and open source AI market growth is outpacing overall AI market growth. But token volume – the number that actually reflects where the industry’s computational and commercial weight lies today – still leans heavily toward closed models, and the shift in that number has been much slower than the shift in adoption survey headlines suggests. Companies seem to be building a truly mixed environment: open models for cost-sensitive and specialized work, proprietary flagships for work where the final percentage points of performance still matter.

Diploma

The story of adopting open source AI and the story of using open source AI are not the same story, and most reporting only tells the first story. Acceptance surveys paint a picture of near-complete open source penetration of enterprise AI, and that’s technically correct – most companies are actually using it somewhere. But token volume data, the number that truly reflects where computing work and business value is concentrated, shows that about 80% of activity still occurs through closed, proprietary models that cost about six times more per token. The real trend isn’t a clean transition from closed to open models – it’s an increasing, more deliberate split, with open models taking on cost-sensitive and specialized workloads while proprietary flagships retain the most demanding tasks, at least until the shrinking 3.3% performance gap closes completely.

Frequently asked questions