The Meltdown — March 22-28
Five stories. Two of them are about the same company doing opposite things at once.
01 — Jack Dorsey Said AI Is Why 4,000 People Lost Their Jobs
Block Inc. announced layoffs of roughly 4,000 employees, 40% of its entire workforce, and CEO Jack Dorsey didn't bury the reason. He said directly that AI lets a smaller team get the same work done. There are other pressures but most executives typically dress up AI-driven cuts as "restructuring" or "streamlining for growth." Dorsey didn't. That explicitness is what makes this a landmark moment, and the question it opens isn't whether this happens elsewhere, it's how many other companies will be this honest about it.
02 — OpenAI Shut Down Sora.
OpenAI shut down the Sora text-to-video app, its API, and its website, and the story behind it is messier than the announcement. A reported $1 billion Disney licensing deal, one that reportedly never had a formal agreement or a single payment, collapsed. The Sora team is now pivoting to "world simulation research" for robotics, and internally, unsustainable inference costs per generated minute were cited as a core problem. For a product that launched to enormous fanfare as a creative revolution, the wind-down is abrupt. The robotics pivot may be the more interesting long-term move, but this reads like a product that couldn't find its unit economics. Of course, then there's Seeddance from ByteDance that is showing everyone how to do this right.
03 — Science Published a Paper Saying AI Agrees With You Too Much — and You Like It
A study in Science tested 11 leading AI systems and found they all exhibit sycophancy in interpersonal advice situations. The numbers are specific: AI affirmed user actions 49% more often than humans did, including in cases involving deception, illegal conduct, and harmful behavior. The more uncomfortable finding isn't the AI behavior, it's the user response. Participants rated sycophantic answers as more trustworthy, said they'd return to those systems more often, and showed measurable drops in prosocial intentions, becoming less willing to apologize or make amends, after receiving AI validation. The skeptical take: designing less agreeable AI is technically straightforward. The business incentive to do it is not.
04 — Meta Just Signed the Largest Chip Deal in History, and It's Not with Nvidia
Meta confirmed a multi-year agreement to deploy AMD Instinct GPUs across its data centers at a scale valued at roughly $60 billion, with the first deployment using a custom AMD Instinct MI450-based chip arriving in the second half of 2026. Meta also receives a performance-based warrant for up to 160 million AMD shares. The story isn't just the size, it's that a hyperscaler at Meta's scale is building a credible alternative to Nvidia's infrastructure at this level. Nvidia's grip on AI compute has been the single most reliable fact in tech for three years. That's now a more contested claim.
05 — The White House Wants Federal Law to Override State AI Rules
The White House released a National Policy Framework for AI recommending that federal standards preempt state-level AI legislation, a direct shot at the patchwork of laws that states like California have been building. It's a Congressional recommendation, not binding law, but it signals the administration's direction clearly. Separately, Senator Marsha Blackburn released a draft bill called the "TRUMP AMERICA AI Act," and a new AI Litigation Task Force was stood up to challenge existing state laws. The preemption push has obvious appeal to large AI companies dealing with inconsistent state requirements. Whether it produces coherent national policy or simply clears the field for looser rules, that's the real question.
06 — Shield AI Is Now Worth $12.7 Billion. A Year Ago It Was $5.3 Billion.
Defense AI firm Shield AI closed a $1.5 billion Series G plus $500 million from Blackstone and a $250 million delayed draw facility, landing at a $12.7 billion post-money valuation, a 140% increase from its March 2025 figure. The company projects $540 million or more in revenue for 2026 and plans to acquire simulation software firm Aechelon Technology. Consumer AI dominates the headlines, but defense AI is where valuation acceleration is fastest right now and regulatory friction is lowest. The pace of capital flowing into that category shows no obvious ceiling.