The hottest debate in artificial intelligence isn’t about which chatbot is smartest anymore.
It’s whether open-source AI will eventually make today’s multibillion-dollar spending race unnecessary.
If open models become “good enough,” investors have wondered whether hyperscalers might need fewer Nvidia Corp. (NASDAQ:NVDA) chips, fewer AI clusters and ultimately less capital expenditure.
According to Bank of America analyst Vivek Arya, that conclusion may be premature.
In a research note published Wednesday, Arya indicated that while open-weight models continue improving rapidly, proprietary—or closed—models still dominate where the highest-value AI workloads reside.
More importantly for investors, he believes either outcome ultimately supports semiconductor demand rather than threatens it.
Why Closed Models Still Lead
The growing enthusiasm around open models accelerated after Nvidia CEO Jensen Huang publicly endorsed broader access to AI through open-weight models.
But Arya said the competitive reality looks different.
“We believe frontier leadership still sits with closed models today,” Arya said.
According to the report, models such as Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro occupy the top positions across leading AI evaluations, while the strongest open models, like those from China, still trail them.
On agentic tool-use, the best closed model scores 55.3 against 43.1 for the strongest open-weight alternative.
On a broad intelligence index the spread is 60.7 to 44.
On coding, 78.3 to 68.8.
The firm added the gap widens precisely where enterprise budgets concentrate: agentic tool-use, coding and complex reasoning.
The lead is largest where the money is. That matters because enterprises typically prioritize reliability over cost.
Why Companies Still Choose Closed AI
Arya argues that proprietary AI providers have advantages extending well beyond model performance.
“Closed models still lead most of the industry standards as well as adoption curve,” Arya said.
Rather than managing GPUs, security, compliance and software updates internally, businesses purchasing closed-model APIs receive managed infrastructure, predictable service-level agreements and a single vendor responsible for performance.
Open models, by contrast, often require customers to handle deployment, safety and infrastructure themselves, creating additional operational complexity.
For investors, that distinction matters because enterprise adoption—not consumer experimentation—represents one of the largest long-term revenue pools in AI.
Why Nvidia And Chip Stocks Could Win Either Way
Perhaps the most important investment takeaway isn’t whether open or closed AI ultimately wins.
It’s that semiconductor companies could benefit under either scenario.
If proprietary frontier models remain dominant, training increasingly capable systems will continue requiring massive GPU clusters and heavy infrastructure spending.
If open models proliferate instead, cheaper AI could dramatically increase usage, pushing inference demand much higher through what’s known as the Jevons paradox—the idea that lower costs often expand total consumption rather than reduce it.
“We see AI semis benefiting either way,” Arya said, adding that closed models would likely keep AI “capital-intensive,” while broader open-model adoption could expand overall demand and the industry’s total addressable market.
A Windows Versus Linux Moment?
Arya believes investors shouldn’t expect one model ecosystem to eliminate the other.
Instead, he compares today’s AI landscape with the decades-long coexistence between Microsoft Windows and Linux.
Open models may dominate cost-sensitive workloads, while proprietary models maintain leadership where performance, security and enterprise trust command premium pricing.
That distinction suggests AI infrastructure spending may become more diversified rather than disappear.
For semiconductor investors, the debate over which model architecture ultimately prevails may therefore be less important than a simpler conclusion: both paths continue pointing toward rising demand for chips, memory and AI infrastructure.
Image: Shutterstock
Login to comment