Anthropic’s planned public offering tests whether customer spending can sustain the computing capacity promised during the AI boom. In this week’s Hard Fork, Max Read and technology reporter Erin Griffith separate rapid revenue growth from the risks of long-term infrastructure commitments. Griffith describes roughly $16.3 billion in first-half revenue and an operating loss of $600 million–$700 million, down from over $2 billion a year earlier. She attributes the much larger headline net loss to an accounting effect involving convertible financing. Those are the episode’s reported figures, not an independently checked financial analysis. The central exposure is the capacity Anthropic has contracted to buy, much of it on terms Griffith describes as noncancelable. Slower growth could mean customers are limiting AI spending, or that cheaper models and competing suppliers are taking market share. The first explanation threatens the wider buildout more directly than the second. Griffith also questions how much public shareholders can change when founders retain voting control and special governance rights. A public price broadens scrutiny, while markets can still reward hype. The later roundtable follows the same spending story into everyday work. Mike Isaac describes an Amazon memo rejecting token consumption and lines of code as engineering scorecards, then asks how managers can fairly assess long-term outcomes during quarterly reviews. The panel’s security discussion separates finding flaws with AI from deploying repairs across thousands of independently operated websites. Microsoft’s offer of buyouts to 60,000 employees becomes another counting problem: eligibility is different from actual departures, and infrastructure spending pressure is different from demonstrated replacement of workers by AI. Jasmine Sun closes with an Anthropic study in which experts reportedly adopted AI recommendations and arguments far more often than they acknowledged its influence. The study details remain attributed to her account, but the accountability question extends beyond it: human approval means little if the reviewer simply repeats the model’s reasoning. Across the conversation, spending, code volume, and a human signature all need to be checked against the results they are supposed to represent.
“If the demand’s not there, then they’re going to be sitting on all these contracts.” — Erin Griffith, around 05:38
“The problem is you can’t really judge long-term outcomes on a quarterly basis when you’re trying to figure out who should still be employed here.” — Mike Isaac, 47:11
“But only 15% of the time did they say, oh, yeah, I was influenced by this AI agent.” — Jasmine Sun, 49:24, describing the study