Compute Oversupply, Illusion of Thinking, and the GPT-Realtime-2 Era — 2026-05-07#

Highlights#

Today’s chatter reveals growing skepticism around the economic realities of AI scaling, underscored by xAI’s surprising massive compute offload to Anthropic and explosive revelations about OpenAI’s shaky infrastructure financing. Meanwhile, as frontier models shift towards local agentic execution and advanced voice capabilities with GPT-Realtime-2, experts like Terence Tao are sounding alarms on the widening gap between algorithmic plausibility and actual veracity.

Top Stories#

• xAI Rents 300 Megawatts of Compute to Anthropic: Elon Musk’s xAI has effectively rented out the Colossus 1 datacenter—boasting over 220,000 Nvidia GPUs, including 150,000 H100s—to competitor Anthropic. Commentators argue this massive supply dump signals that pure scaling will not achieve AGI, and serves as a potential warning sign of an impending compute oversupply that could devastate hyperscalers’ margins. In a striking caveat, Musk reserved the right to reclaim the computing power if Anthropic’s AI engages in actions harmful to humanity. (Source)
• OpenAI’s Precarious Infrastructure Economics: New reports indicate that when OpenAI and Broadcom announced a landmark 10 GW custom AI chip deal, the companies had not actually secured funding to pay for it. Broadcom is currently negotiating to finance the $18 billion first phase of chip production, exposing the massive, potentially unsustainable capital requirements operating behind the scenes at OpenAI. (Source)
• Terence Tao Highlights Plausibility vs. Veracity: Esteemed mathematician Terence Tao warned that modern AI models are optimized merely to maximize the probability of looking right, prioritizing plausibility over truth. He described current architectures as a “convincing mirror” that can confidently deliver dangerous medical or financial advice, identifying the gap between convincing and correct outputs as the most critical AI safety variable currently unsolved. (Source)
• OpenAI Drops GPT-Realtime-2 in the API: OpenAI released GPT-Realtime-2, bringing “GPT-5-class reasoning” to voice agents capable of collaborating and solving complex problems in real-time. Sam Altman noted a generational divide in this adoption, pointing out that younger users heavily prefer dumping context via voice, whereas older demographics continue to prefer typing. (Source)
• Perplexity Launches Local “Personal Computer” Mac App: Perplexity has launched a new Mac application featuring “Personal Computer,” an agent purpose-built to run tasks across local apps and files natively. Paired with a Mac mini, the tool enables users to operate a 24/7 always-on personal AI agent remotely. (Source)

Articles Worth Reading#

The Uncomfortable ROI of the AI Boom (Source) Goldman Sachs analysts have issued a stark warning that the current dynamics in the semiconductor sector are both “unprecedented and unsustainable”. Currently, almost every enterprise heavily investing in AI infrastructure is losing money, while semiconductor firms reap all the financial rewards. If application-layer companies cannot begin generating actual profits from their AI tools soon, a drastic scale-back in capital expenditure will become inevitable.

The Psychological Limits of AI Chatbot Advice (Source) A fascinating new psychological study reveals the limits of leveraging LLMs as personal counselors or life coaches. Researchers found that while participants frequently followed the advice provided by AI chatbots regarding their health, careers, or relationships, they exhibited no sustained improvement in well-being weeks later. This highlights a dangerous paradigm where LLMs wield significant influence over real-world personal decisions without delivering any measurable psychological benefits to the end user.

OpenAI’s Governance Unraveling in Court (Source) The ongoing Musk v Altman trial continues to surface damaging internal details about OpenAI’s governance and its original mandate. Former employee Rosie Campbell testified that OpenAI’s core mission does not fundamentally require building AGI, a statement critics argue completely undermines the justification for creating its for-profit subsidiary in the first place. Concurrently, recently revealed depositions highlight executive inaction, with Helen Toner stating that Mira Murati intentionally sat on the fence during Sam Altman’s ouster, waiting “to see which way the wind would blow”.