20VC: Anthropic Files to Go Public | Token Budgeting Panic Hits Corporate America | Cognition Raises $1BN at $26BN Valuation | Apollo Warns PE Software Returns Will be Disastrous | The 9-9-6 Work Ethic: Performative Theatre or Startup Reality?

What It Covers

Harry Stebbings, Rory O'Driscoll, and Jason Lemkin analyze Anthropic's $65B valuation and IPO filing, Cognition's $1B raise at $26B, the SaaS sector's partial recovery, corporate America's token budget crisis, PE software return warnings from Apollo, and whether AI agents will replace engineering headcount by end of 2025.

Key Questions Answered

• •IPO Timing & Capital Grab: Anthropic, OpenAI, and SpaceX are all accelerating public market timelines simultaneously, representing an estimated $300–400B in equity issuance. Google's $80B raise signals that even the most profitable companies are front-loading capital before the AI infrastructure race intensifies. The pattern reflects a strategic shift: staying private is no longer considered advantageous when CapEx requirements are scaling at this pace.
• •Seed Investment Bar Has Shifted: Jason Lemkin now requires a credible path to a $1B fund position — not just a $1B company outcome — before taking a meeting. This implies the underlying company must realistically reach $10B+ to survive dilution. Founders pitching sub-scale TAMs, average CTOs, or complaint-heavy cultures will not get meetings from top-tier seed investors regardless of other merits.
• •Token Budget Crisis Is Real and Immediate: CFOs across corporate America discovered mid-Q1 2025 that AI token spend ran 10x over accrual estimates after Claude's pricing model shifted to pay-as-you-go in early 2025. Uber responded by capping individual spend at $1,500/month. This is not a signal to stop AI adoption — it validates a category worth $500B–$1T — but forces structured budget allocation processes that did not previously exist.
• •Tokens vs. Headcount Trade-off by Year-End: Engineering and product leaders will face explicit budget choices in 2026–2027 planning cycles: maintain headcount or reallocate salary budget to tokens. QA teams, customer success roles, and mid-tier engineers are the most vulnerable. One portfolio company already spends more on tokens than engineering salaries. The ratio of token spend to engineer salary — currently around 10–33% — is the single most consequential number in AI infrastructure modeling.
• •SaaS Recovery Is Selective, Not Broad: The WisdomTree Cloud ETF recovered roughly 25–30% from April lows but remains flat year-to-date, while the Nasdaq is up 21% and semiconductors are up triple digits. Companies with genuine AI attachment — Twilio up 57%, Datadog up 100%, Okta up 56% — outperformed dramatically. Per-seat human license businesses remain structurally challenged; Salesforce explicitly split its business into AI-driven and legacy segments, projecting single-digit growth for the latter.
• •PE Software Returns Face Structural Math Problem: Apollo's warning on PE software returns reflects a straightforward valuation compression problem: firms that acquired SaaS companies at 10x revenue in 2021 now face public market comps of 3–5x revenue. With debt at 5–7x EBITDA and equity below that, even modest growth cannot overcome the entry price. The likely outcome is not total loss but 1.2–1.3x returns over 10-year hold periods — well below target fund performance thresholds.
• •Multi-Model Architecture Is the Cost Containment Strategy: Leading AI-native platforms like Replit now automatically route tasks across models — using Claude Sonnet for initial builds and OpenAI Codex for review — without user awareness. This dual-model approach catches errors on every pass while managing per-token costs. Founders building developer tools or AI-heavy applications should architect for model routing from day one rather than single-provider dependency, as frontier model pricing continues rising while prior-generation models deflate.