Arena AI CEO says enterprises don’t know which AI models to trust

Aug 08, 2026 04:36 PM - 1 hour ago 2

Anastasios Angelopoulos poses for a photo

Companies don't cognize which broadside to take successful AI, the leader of 1 of the largest exemplary evaluators said.

"It's not only existent that they're terrified of moving pinch the frontier labs, but they're besides terrified of moving pinch the Chinese unfastened source," Arena CEO Anastasios Angelopoulos said during a caller section of the "20VC" podcast.

Angelopoulos said the companies he talks to are struggling to fig retired wherever to build their AI future.

"I deliberation we're successful a really tricky business for enterprises because conscionable the measurement that the cards person landed," Angelopoulos told Business Insider during an question and reply wherever he expanded connected his comments.

The existent business is, by each accounts, a spot messy. Arena has an up-close view. The institution offers a crowdsourced AI exemplary conflict royale that lets users pit generative AI models against each different and ballot connected which 1 champion completes the task. In September, it announced a commercialized business wherever it would waste AI companies and businesses information gathered from the battles. In June, Arena announced it had reached $100 cardinal successful annualized run-rate gross wrong 8 months of launching its endeavor service.

Frontier AI companies, including OpenAI and Anthropic, connection cutting-edge models, but that exertion tin travel pinch a steep value tag. The Trump administration's caller actions besides exemplify that the exemplary providers tin beryllium forced to shut disconnected entree without warning. Those companies, arsenic Anthropic has demonstrated, tin besides enforce safeguards connected the models, which whitethorn mean that galore companies cannot entree the afloat grade of the model's capabilities.

"Anthropic aliases OpenAI could determine to move disconnected their AI astatine immoderate infinitesimal if they consciousness that it is, quote unquote, 'unsafe,'" Angelopoulos told Business Insider. "And truthful if you're a business, what are you expected to do? You want to beryllium capable to build connected a unchangeable and reliable AI stack."

In contrast, China-based unfastened exemplary companies, including Moonshot and DeepSeek, connection open-weight models that tin beryllium tally locally, often astatine a fraction of the costs of frontier models. Some are worried astir the information risks of relying connected Chinese companies. Last month, Axios reported that the Trump management has considered imposing restrictions aliases moreover issuing de facto bans connected Chinese models.

Angelopoulos isn't the only 1 to deed connected these themes. Palantir CEO Alex Karp has many times warned that frontier AI companies, pinch which Palantir works, want to stitchery valuable intelligence spot and information that they could past usage to compete pinch their existent customers.

"They merit to colonize your enterprise," Karp said during Palantir's second-quarter net telephone connected Monday, characterizing what he believed was the labs' attitude.

Palantir and Karp besides person immoderate inducement to stoke those fears, arsenic the institution offers its customers an exertion furniture that allows businesses to power their information and the AI models they use.

Angelopoulos said Arena has knowledgeable its ain struggles pinch AI companies. He said that Arena has seen exemplary providers usage Arena's soul information "to service synthetic battles and arenas to our labor successful an effort to fundamentally get them to explanation data." Angelopoulos declined to sanction the offending companies, but said the believe "is not uncommon."

Nvidia CEO Jensen Huang, whose institution does business pinch frontier exemplary makers and offers unfastened models, has tried to onslaught a equilibrium successful the debate. He has said that companies will deploy some unfastened models and proprietary ones, routing tasks based connected some costs effectiveness and maintaining their IP.

"The world is going to person closed and unfastened models," Huang precocious told Axios. "We should usage closed models arsenic often arsenic we can, and we should customize and usage unfastened models to build our ain personal, our ain customized AI if we must."