American scientist Gary Marcus is not letting Larry Ellison have his second. When the Oracle founder not too long ago argued that AI fashions from OpenAI, Google, Meta, and xAI are all turning into commodities as a result of they practice on the identical public web information, Marcus pounced. His message, posted on X, was blunt: he referred to as this precise consequence two years in the past, and Silicon Valley merely wasn’t listening. “i warned these guys of precisely this downside – no moat as a result of everyone seems to be coaching on identical information – two years in the past,” the NYU Professor Emeritus wrote. He added that the trade’s refusal to listen to out anybody past its personal bubble is “gonna price them a very huge sum of money.”
Larry Ellison thinks the subsequent AI goldmine is locked inside personal information
Ellison’s argument got here throughout Oracle’s fiscal Q2 2026 earnings name in December. The CTO laid out a easy thesis: each main massive language mannequin is skilled on the identical scraped web information, so all of them find yourself roughly the identical. “They’re all mainly the identical. And that is why they’re turning into commoditized so rapidly,” he stated. His pitch was that the subsequent massive wave will not be about constructing higher base fashions. It’s going to be about feeding AI proprietary enterprise information whereas conserving it locked down and safe.Oracle is placing actual cash behind that guess. The corporate is projecting round $50 billion in capital expenditures for the yr, up from a $35 billion estimate simply three months earlier. Its argument is that the majority worthwhile personal information already sits inside Oracle databases, and its AI Knowledge Platform makes use of strategies like Retrieval-Augmented Era to let fashions pull from that information in actual time with out safety trade-offs.
Marcus referred to as the worth wars and the ‘no moat’ downside again in 2024
For Marcus, that is vindication he is been ready for. Again round March 2024, he predicted a crowded discipline of GPT-4-level fashions, brutal worth wars, no actual moat, and cussed reliability issues that scaling alone would not repair. A lot of that performed out. Claude, Gemini, Llama, Grok, and open-source challengers like DeepSeek now cluster tightly on benchmarks, whereas API costs have collapsed and income keep elusive for everybody besides chipmakers like Nvidia.Marcus has lengthy pushed alternate options to brute-force scaling, favouring neurosymbolic methods that mix sample recognition with logic and structured reasoning. He is additionally been vocal about what he sees as reckless capital spending, as soon as branding the AI capex spree one of many greatest misallocations of cash in historical past.He is not absolutely writing off Ellison’s thought, although. Marcus concedes proprietary information helps “in some slender use circumstances,” simply not because the broad rescue the trade hopes for. Critics counter that actual moats could lie elsewhere anyway, in distribution, model, security, or the user-interaction information flowing by way of shopper apps quite than static enterprise data.