In AI discussions people often compare open models to classic open-source projects like Linux. Those analogies are helpful but limited: a fully open-source language model — meaning one that publishes the full training recipe, data and code — is most comparable to an operating system. Open-weight models, which publish only trained weights and inference code, are more like specific software releases that you install on top of an existing platform.
Model lifetimes and reuse Model weights can be transient on average, but they often have long useful lives; many companies still run workflows based on Llama 3. Building a fully open training recipe is resource-intensive. Modern examples include the Olmo models (which the article’s author helped build at the Allen Institute for AI) and earlier work such as EleutherAI’s Pythia. Any organization can pick up those recipes, modify them and run training to produce new weights; ideally, community improvements to data or training code feed into subsequent models.
Nvidia’s near-open strategy Nvidia is investing heavily in a near-open approach: for their Nemotron models they disclose as much training data and code as they legally can. Their aim is to enable many teams to build “token machines,” reducing the risk that intelligence becomes monopolized while generating wide demand for inference across companies that will want Nvidia hardware. Press reports state Nvidia is spending roughly $26 billion on this effort.
Two possible futures The author outlines two main scenarios:
-
Success path: If Nvidia’s approach works — if releasing near-open training assets creates substantially more demand for its chips and services — the investment could pay off and sustain an open-weight ecosystem with profitable feedback for participants.
-
Long-tail path: If the financial feedback loop fails to materialize, open-weight models could diverge from leading closed models and evolve into a long-tail ecosystem focused on efficiency, modifiability and specialization. In that scenario, closed, well-capitalized players would likely retain the most valuable domains (knowledge work collaboration, drug discovery, software engineering), while open-weight solutions serve enterprise-specific, on-prem and repetitive business tasks.
Economic reality and shifting technical practices Building top-tier models is extremely capital intensive, which challenges the long-term accessibility of competitive training. Although training has remained more accessible in industry than many expected, there is a default assumption among some that training is prohibitively expensive and open recipes lag behind. A few exits from training — companies such as Databricks and 01.ai are mentioned — look more like exceptions than a clear industry trend.
Training complexity and the rise of post-training work Training is becoming more complex and abstracted. The present open-model ecosystem has been sustained by a boom in post-training activity: teams take base models (examples cited include DeepSeek V4 Flash, Inkling Small, GLM 5.X) and fine-tune them for agentic tasks using popular finetuning APIs like Tinker. But the ability to train a base model into a general agentic reasoner is becoming increasingly opaque, similar to how large-scale pretraining practices once were. This shift may even change the commonly used pretraining / midtraining / post-training lexicon toward something like pretraining, reasoning training, and post-training.
Licenses, revenue-sharing and financial experiments Fewer teams are investing in training entire models, and open-model builders are experimenting with revenue-share licenses tied to downstream product or inference use. These are attempts to create viable financing for near-frontier open-weight models. The success of those experiments will be pivotal for whether Nvidia’s demand-growth strategy around open-source assets can be sustained — those open-model builders must succeed for the approach to persist.
Hyperscalers and flooding the market with tokens Another category of actor monetizes AI indirectly: large hyperscalers with deep balance sheets. If Meta were to release a strong model like Muse Spark 1.2 as open weights, it could significantly slow revenue growth for token-selling competitors such as Anthropic and OpenAI by commoditizing complementary assets. Nvidia, by contrast, prefers to enable many teams to “fish for tokens” so that the broader ecosystem generates the inference demand that sustains its hardware business.
Conclusion The near-term future of open-weight models points in two directions: either major investments and platform-like feedback loops (for example Nvidia’s reported $26 billion effort) succeed in generating persistent demand for training and inference, or open-weight models settle into a long-tail niche emphasizing efficiency and vertical specialization. In the coming years, licensing experiments, revenue-sharing mechanisms and the strategies of hyperscalers will determine which path becomes dominant.
