Why Enterprise AI Is Moving to Open Source
A plain-language reading of Glean founder Arvind Jain on 20VC: cost and “good enough” capability—not old data fears—are pushing enterprises toward open models.
For years, many companies wanted more control over which AI models they run. Few wanted to stay locked to a single vendor forever. The desire was old. What changed, according to Jain, is that open models finally became good enough for a large share of real work—and the bill for closed APIs became hard to ignore.
The main driver is cost
Jain describes a pattern that CFOs know too well: a company sets an AI budget for the year, then burns through it in a month or two. When inference feels like an open tap, finance starts asking harder questions. That pressure, more than ideology about open source, is accelerating the shift.
Some enterprises also need workloads to stay inside private data centers for compliance. That still matters. But Jain’s sharper point is about fear that has mostly faded: the early panic that model vendors would quietly train on customer data. With proper commercial contracts, he says, that anxiety has largely settled. What remains front and center is the price of tokens.
“Good enough” arrived later than the desire for control
Think of it like office software a decade ago: companies wanted alternatives, but only switched when a rival product could handle most daily jobs without constant pain. Jain’s claim is that open models have crossed a similar line for most enterprise use cases—not that they beat every frontier model on every hard task, but that they can carry the bulk of the load.
In the interview he places that capability jump very recently—within roughly the prior month when they spoke—and notes that his team became much more confident routing most workloads to strong open models. The remaining hesitation he hears from customers, he argues, is often not “is open source capable?” but comfort with particular model origins and geopolitics: a psychological and reputation layer more than a pure quality gap.
What this means for the model business
If most routine enterprise work can run on cheaper open options, pure model APIs face sharper pricing pressure. Jain even frames selling models alone as a tougher business than the hype suggested—with more rivals, open alternatives, and rumors of aggressive price cuts among frontier labs.
That does not mean frontier labs become irrelevant. In his telling, their value increasingly sits in products and ecosystems around the models—not only in renting the raw intelligence. For buyers, the practical takeaway is simpler: assume you will multi-home. Design so switching models is normal, and treat open options as the default path for volume work when quality is close enough.
A three-year bet—stated clearly
Asked how much enterprise workload will run on open-source models in three years, Jain answers with unusual certainty: the vast majority. Whether that timeline holds is for the market to prove. What matters for general readers is the direction he is naming—cost discipline plus “good enough” open models are already rewriting who gets the everyday enterprise traffic.