Demis Steps Down, Apple’s Memory Problem, Microsoft’s Clever Trick

Summary

This episode centered on three big AI-era tensions: Google’s leadership and product execution, Apple’s exposure to the AI hardware supply chain, and Microsoft’s attempt to manage perceptions around massive infrastructure spending. Alex and M.G. said Demis Hassabis’s transition at DeepMind reflects a real strategic divide between LLM-centric product pressure and a longer-horizon scientific bet on world models and other non-LLM approaches. On Apple, they argued that AI-driven demand for memory and components is tightening supply, raising costs just as Apple may have to push iPhone and other device prices higher. On Microsoft, they debated whether the company’s data-center accounting changes materially reduce spending or mainly improve optics, and whether cloud growth at hyperscalers is increasingly being driven by a small number of AI labs recycling capital back into cloud services.

Notes

• Demis Hassabis is stepping down as Google DeepMind CEO and moving to chairman and chief scientist.
• Google is handing all AI model development to Corey Kavicolu, while Jeff Te is leaving to co-found a startup.
• Alex reads the move as a strategic split between scientific AI and commercial LLM execution.
• Demis is portrayed as believing current LLMs are not enough for AGI.
• Google executives were reportedly frustrated that Hassabis focused less on commercial demands.
• Google’s AI breakthroughs are framed as world-model-like systems such as AlphaFold, not chat LLMs.
• Alex says Hassabis once told him he wanted to make an artificial cell, which he says cannot be done with an LLM.
• Alex argues the industry has poured trillions into LLMs, so diversification lowers the risk of an AI winter if one approach disappoints.
• Google is said to be behind in coding agents and lacks a product comparable to Claude Code.
• Apple is facing a component squeeze as AI data-center demand strains global supply chains.
• Tim Cook called the shortages very significant and said Apple had limited options.
• Apple’s guidance implied 9% to 11% current-quarter growth versus 12% expected.
• M.G. notes Apple had recently hit $5 trillion and briefly regained the top market-cap spot.
• Apple’s device costs are rising because the same memory and components are now more expensive.
• Alex says Apple underweighted AI’s importance and failed to buy enough compute early.
• M.G. says Tim Cook should have anticipated peers spending hundreds of billions on AI capex.
• M.G. says annual AI capex across big tech may now exceed $1 trillion.
• M.G.’s base-case forecast is $100 to $200 of device price increases.
• Higher-end Apple devices could rise $250 to $300, he says.
• The iPhone fold or ultra could launch around $2,000, according to M.G.
• Apple Upgrade is described as a monthly-fee program for iPhone, Mac, iPad, and Apple Watch, currently US-only and built with Klarna.
• A $30 monthly fee for an iPhone Pro could soften sticker shock, M.G. says.
• He frames Apple’s subscription and financing push as a path toward a broader recurring ecosystem.
• Microsoft changed data-center lease accounting from a 15-year to a 25-year cycle for some facilities.
• That accounting shift moves some spend from capex to opex without reducing total spend.
• Alex says GPUs do not last 25 years and can be obsolete in as little as three years.
• Alex also says chips can be roughly 50% of data-center build-out costs, making a 25-year assumption hard to defend.
• M.G. clarifies that the 25-year figure refers to the physical facility, not the servers inside it.
• Server replacement cycles were described as roughly three, six, or seven years.
• M.G. says Amazon and Google have used longer useful-life assumptions in some cases.
• He adds that if AI demand shifts, facilities could become stranded or need repurposing long before 25 years.
• Alex points to cloud growth figures of 82% for Google, around 30% for Microsoft, and 43% for Azure.
• Alex suggests much cloud growth may be OpenAI and Anthropic recycling capital back into cloud spend.
• He specifically highlights Amazon’s planned $50 billion OpenAI investment as part of that circularity question.
• M.G. says Wall Street is not really asking how much cloud growth is circular versus organic.
• He says investors are rewarding hyperscalers because they think AI is finally producing a return.
• A slowdown at OpenAI or Anthropic would cascade across AWS, Azure, and Google Cloud, he argues.
• M.G. says Satya Nadella’s “Switzerland” stance reflects a hedge away from dependence on OpenAI and Anthropic.
• He adds that OpenAI and Anthropic will eventually want their own clouds.
• Amazon’s reported $50 billion OpenAI commitment is unusual because the trigger conditions were reportedly unclear or unmet.
• Google’s Gemini 3.5 Pro was promised for June but still had not launched by August.
• M.G. compares the delay to Meta’s old Behemoth model problems.
• He says Google’s internal fiefdoms are fighting over coding-agent strategy and resources.
• He argues Google’s DeepMind and Brain merger did not eliminate factional conflict.
• Google may try to jump from Gemini 3.5 Pro straight to Gemini 4, creating a narrative risk if 4 disappoints.
• Alex says the broader argument is that startups are locked into LLMs while vertically integrated firms can pursue world models.
• The second approach he cites is the world-model approach.
• Alex says startups pursue LLMs because they need near-term profits, while integrated firms can explore more speculative paths.
• Google can afford to back a less profit-sensitive research direction, he says.
• Alex says Google appears to see world models as the path to AGI.
• M.G. agrees the thesis is overstated but useful for understanding Google’s lag.
• M.G. says Hassabis himself seems to prefer world models from the top down.
• He notes Google I/O highlighted Omni, Google’s world-model analogue, while downplaying other work.
• M.G. adds that OpenAI or Anthropic could reach recursive self-improvement first and generate world-model capabilities themselves.