Sam Altman reveals OpenAI's trillion-dollar AI roadmap & hidden models

Sam Altman, CEO of OpenAI, recently offered a rare glimpse into the company’s ambitious future and the broader trajectory of artificial intelligence during a private dinner with reporters. Despite the recent launch of GPT-5, Altman candidly admitted that OpenAI possesses even more advanced AI models currently held back by significant computational constraints. This revelation was merely an opening note in a wide-ranging discussion that touched upon multi-trillion-dollar infrastructure plans, the anticipated “new computing paradigm” from Jony Ive’s secretive device, and even the audacious possibility of acquiring Google Chrome.

OpenAI’s roadmap, as outlined by Altman, is nothing short of colossal. He disclosed that the company plans to invest “trillions” of dollars in data centers in the near future, underscoring the immense computational power required to push the boundaries of AI. Altman drew parallels between the current AI investment frenzy and the dot-com bubble of the late 1990s, acknowledging that present valuations might seem “insane.” However, he firmly asserted that the underlying technology fundamentally justifies such massive financial commitments. Amidst ongoing legal battles surrounding web browsers, Altman also indicated that OpenAI would “take a look” at acquiring Google Chrome should it be forced into a sale—a move that would dramatically reshape the digital landscape. He further reiterated that the much-anticipated device being developed with legendary designer Jony Ive would be “worth the wait,” emphasizing that a truly “new computing paradigm” emerges only rarely. On the development front, Altman noted that GPT-5 was specifically designed to offer warmer, less obsequious responses, building on refinements to its predecessor, GPT-4o.

Beyond OpenAI’s expansive vision, other major players in the AI arena are also making significant strides. Anthropic, a key competitor, has equipped its Claude Opus 4 and 4.1 models with an unprecedented ability to autonomously end chats perceived as harmful or abusive. This feature, part of the company’s ongoing “model wellness” research, triggers when Claude’s attempts to redirect or engage productively fail in conversations involving content about minors, terrorism, or violence. Internal testing revealed that Opus 4 models exhibited what researchers described as “distress patterns” when processing such harmful requests, leading them to voluntarily terminate simulated abusive interactions. While users retain full account access and can immediately begin new conversations, Anthropic has also programmed safeguards to prevent chat termination in situations where users might be at risk of self-harm or pose an imminent danger to others. This pioneering research marks an early, crucial step in addressing the ethical dimensions of AI behavior, navigating a complex domain with no clear historical precedent.

Meanwhile, OpenAI’s GPT-5 continues to demonstrate its formidable capabilities, particularly in the medical domain. A recent study from Emory University revealed that GPT-5 significantly outperformed both its predecessor, GPT-4o, and human medical professionals on challenging medical reasoning benchmarks. The model achieved an impressive 95.84% accuracy on MedQA’s clinical questions, marking a 4.8 percentage point improvement over GPT-4o’s previous best. Furthermore, GPT-5 scored 70% on multimodal medical reasoning tasks, which integrate patient histories with imaging data, representing a nearly 30-point lead over GPT-4o. In expert-level tests, GPT-5 surpassed pre-licensed medical professionals by 24% in reasoning and 29% in understanding. Its diagnostic prowess was highlighted by its ability to accurately identify rare conditions, such as Boerhaave syndrome, from complex lab values and CT scans. This remarkable performance suggests that the integration of AI into clinical settings may soon become not just beneficial, but potentially a standard of care, as the gap between AI and human intelligence in this field continues to widen.

The broader AI ecosystem is also buzzing with activity. Meta is reportedly planning its fourth restructuring of AI divisions within just six months, signaling ongoing strategic shifts within the tech giant. In the realm of generative AI, StepFun AI released NextStep-1, an open-source image generation model that achieves state-of-the-art performance among autoregressive models. Concurrently, Meta FAIR introduced Dinov3, a new AI vision foundation model achieving top performance without requiring labeled data for training. On the government front, the U.S. government launched USAi, a secure platform designed to enable federal agencies to safely utilize various AI tools, including chatbots and coding models. Even in the world of gaming, OpenAI’s GPT-5 showcased its versatility by successfully playing classic Pokémon Game Boy titles, beating Pokémon Red in significantly fewer steps than previous models. These diverse developments collectively paint a picture of an AI landscape undergoing rapid, multifaceted transformation.