Enterprise software valuations are collapsing because investors now see AI as a direct replacement for human-powered functions, not just an add-on. Mark Benioff noted on All-In that Salesforce will spend $300 million this year on Anthropic tokens to power coding agents, a move he sees as automating business development. The low-end software market is finished, according to Chamath Palihapitiya. The economic driver is the end of AI subsidies: Microsoft and GitHub are pivoting to usage-based billing as infinite token demand makes per-seat pricing unsustainable. Startups are responding by cutting entry-level engineering and QA positions, trading fixed salaries for variable AI costs.
This shift is already visible in Bitcoin's core infrastructure. Steve Lee from Spiral detailed Project Loupe on Presidio Bitcoin Jam, an open-source AI security scanner for Bitcoin repositories. It uses LLMs to find vulnerabilities, filters the results, and funds the token costs automatically. This tool performs work traditionally done by junior developers or external audit firms. The coding agent isn't a futuristic concept; it's a deployed tool shipping with Bitcoin Core v31.
"Spiral's Project Loupe is an open-source AI security scanner for Bitcoin repos that uses LLMs to find vulnerabilities, filters results, and funds the token costs."
- Steve Lee, Presidio Bitcoin Jam
The technical barrier to building software has collapsed, ushering in what Steve Lee calls 'vibe coding.' He points to Babbel Agent, a live podcast translation tool funded by Lightning payments, built by a non-engineer. These aren't 'Bitcoin apps' in the old sense; they are general utilities that use Bitcoin because it's the only viable micro-payment rail. The 'idea guy' is now the developer. Nathaniel Whittemore argued on The AI Daily Brief that product development is shifting focus from raw models to the 'harnesses' - the interfaces like Cursor's SDK that let anyone deploy an agent.
The economic pressure is structural. On All-In, Benioff framed AI agents as a way to make his own company more efficient by replacing human functions. This logic applies everywhere. When AI inference shifts from a subsidized flat rate to a metered cost, a junior developer's salary looks like a capital expense to be optimized away. The $300 million Salesforce is spending on Anthropic tokens is capital redirected from headcount.
"Salesforce will spend $300 million on Anthropic tokens this year to power coding agents, but believes an intermediary layer is needed to route queries efficiently and avoid unnecessary costs."
- Mark Benioff, All-In with Chamath, Jason, Sacks & Friedberg
The displacement isn't just about writing code. Project Loupe shows AI agents are taking over specialized, high-stakes review work. The next phase is recursive: as AI writes more code, AI will be needed to audit it, creating a self-reinforcing loop that marginalizes the human in the middle. Whittemore noted OpenAI had to audit a bizarre 'goblin' bug that leaked through recursive training, highlighting the new complexity of model hygiene. The tools to manage this complexity are also becoming automated.
This creates a bifurcated future. High-end architects and system designers who manage the agent harnesses will remain. The broad base of early-career engineers who learn by doing maintenance, writing tests, and reviewing code - the traditional apprenticeship path - faces obsolescence. The 'vibe coder' with an idea and a prompt might build the next utility app, but they won't hire a team of juniors to scale it. They'll buy more tokens.
