a16z-shared Hebbia CEO essay: humans are now cheaper than software

On July 15, a16z amplified a long essay by Hebbia CEO George Sivulka, sparking wide discussion. The piece pushes back on the linear narrative that AI will immediately replace human jobs at scale, arguing that under today's cost structure "humans are now cheaper than software" in some scenarios, and that companies are more likely to first reorganize roles while adding both compute and headcount than to cut staff outright. Its significance is in shifting the conversation from model capability to usage cost, organizational design, and how adoption actually happens.

Core claims and figures

As relayed by @a16z AI, @FinanceYF5, and @smtabatabaie, Sivulka's central claim is that AI is not simply eliminating work: the jobs it creates may outnumber the ones it removes, and for the first time "humans are cheaper than software." @rohanpaul_ai adds that the most aggressive AI adopters are not just substituting AI for people but adding more people and more compute at the same time; the essay cites that by May 2026 the top 1% of enterprises were already spending about $90,000 per year on AI, with the figure implied to keep growing at roughly 1% per month. @luisdans relays a more aggressive forecast: by the end of 2026, per-employee AI token spend could exceed the average software engineer's annual salary. @soumitrashukla9 notes the essay's title, "you just hired a million bad employees," aimed at the blind spot of treating AI as a direct labor substitute.

Discussion and pushback

Focusing on the cost lens, @bibryam frames it as a team-level decision: when AI inference and call costs exceed engineer costs, which tasks go to AI and which stay human becomes a real trade-off. @frog_omo argues from the essay's evidence that the narrative of "AI-driven mass layoffs in software engineering" is badly overstated, and that what companies call "AI-driven cuts" is often cost pressure, investor demands, or pullback after over-expansion — AI is more likely to reshape engineers' workflows and division of labor than to replace jobs outright. @zephyr_z9 frames the long-term contest as "intelligence per watt," where lasting advantage depends not only on stronger models but on whether per-unit compute cost holds up.

2026-07-15 ~ 2026-07-15 · 9 related posts