In the next few months, four companies are raising over $350 billion in equity. – Google: $80B of net equity issuance to fund AI capex – SpaceX: ~$75B IPO at near-trillion valuation – Anthropic: raising at $965B post-money, up 53x from late 2024 – OpenAI: ~$100B at private market levels That's a third of a trillion in fresh paper supply from four names alone. Then there's the second wave. An Anthropic employee who joined in late 2024 with $100K of equity is sitting on roughly $5.3 million on paper today. There are thousands like them at Anthropic, OpenAI, SpaceX, Databricks, Stripe, xAI. Every single one is a future seller. At IPO, at lockup expiry, at the next tender. Founders, employees, and the earliest VCs are all trying to convert paper into cash at the same moment in history. When the people closest to the asset are all sellers at the same time, the question isn't whether the technology works. The question is who's left to be the bag holder. This is what an exit liquidity avalanche looks like. It doesn't crash a market in a day. It bleeds it for years.

A thought provoking new graph from @EntelligenceAI is making the rounds. Ostensibly, it looks like most AI tokens being spent for app dev is going to waste. The issue is more complex however, and there is a lot of waste when humans code too. Here’s my take on how it actually breaks down: In “trad” software engineering, the cost of quality and rework numbers are significantly lower than what this graph implies for AI-generated code. But the comparison is nuanced because the graph is measuring token spend rather than labor hours or defect rates. Historically, industry benchmarks — which I’ve tracked for years for norms and baselines for CIOs — for traditional software development generally look like this: • Rework as a percentage of total engineering effort: • Typical mature teams: 15–25% • Average enterprise teams: 25–40% • Poorly governed projects: 50% • Defect fixing after initial coding: • Often estimated around 20–30% of engineering capacity in normal enterprise environments. • Actual “new value creation” reaching production: • Commonly around 30–50% of total effort depending on bureaucracy, testing rigor, regulatory burden, and architecture quality. So compared to the chart: • The “82% never reaches the product” claim seems dramatically worse than traditional norms. • Traditional engineering usually wastes around 40–70% of total effort when meetings, QA, architecture churn, compliance, integration, and technical debt are included. • Elite software organizations historically get far better leverage, sometimes achieving 50% effective value delivery. But there’s a much deeper issue here: AI-generated code changes the economics because generation is nearly free, so orgs will tolerate *vastly* more speculative, low-quality, or throwaway code generation than humans would ever write manually. That inflates: • Bug fixes • Review burden • Merge conflicts • Architectural inconsistency • Duplicate implementations • Context switching • Governance overhead In other words, the whole graph is less an indictment of AI coding itself and more an indictment of immature AI software engineering operating models. A mature agentic development organization should theoretically: • Reduce boilerplate cost dramatically • Compress rework cycles • Automate testing and remediation • Increase shipped-product ratio over time But most firms are currently in the “AI code inflation” phase and are producing a lot… • More code • Code faster • Lower coherence • Rising governance burden Which is exactly why CIOs and CTOs are suddenly obsessed with solutions for: • AI app dev platform engineering • Agent governance • AI SDLC controls • Code provenance • Policy enforcement • Automated reviews/QA • Architecture constraints • Token economics The key takeaway from the chart is far more than “AI coding is inefficient.” It must be: “Unconstrained code generation without corresponding governance risks scaling entropy faster than productivity.” cc @sarbjeetjohal @joemckendrick @DavidLinthicum @ThadOfSphere