May 15, 2026

The fourth pillar

Engineering compensation had three pillars: cash, equity, and benefits. The fourth — AI tokens — arrived in 2026 and has no system of record. That's a problem worth $50B by 2030.

The phrase “fourth pillar” came from Tomasz Tunguz in late March 2026, written up in his investor letter. The framing was instantly intuitive because every line he wrote was something everyone already knew separately and nobody had stacked into one shape.

Engineers running Claude Code at Anthropic itself were posting $150,000/month token bills. Top firms set $150-$250/engineer/month budgets, then watched them get blown. Uber burned its 2026 AI budget in four months. The conversation went from “should this be on the corporate card” to “this is part of comp” in roughly one quarter.

What changed

What changed wasn’t the price of tokens. Token prices have fallen consistently. What changed was who pays for what. The default in 2024-2025 was that the company paid for AI tools and engineers used them inside a vague envelope. The default in 2026 is that AI tools are a comp line: this engineer gets $X/month in token budget, like they get $Y/year in equity.

Once you frame it that way, every other compensation system suddenly has gaps:

Payroll doesn’t track it.
HRIS doesn’t allocate it.
The cost-management platform sees it as cloud spend.
The AI observability tool sees it as API calls.

Nobody owns the seat where the four facts meet: this person, this allowance, this consumption, this outcome.

What gets built

The system of record for the fourth pillar has the same shape as the system of record for the third (Carta for equity). A multi-tenant ledger. A policy engine. HRIS connectors. Tax / benefit-in-kind metadata. Vesting. Portability on departure. An audit trail tight enough that the next downturn doesn’t kill the program out of opacity.

We’ve built the first version. It’s called TknBudget. The product surface for vesting is called TokenForge. The architecture commits to that future in the schema today — forge_grants sits alongside budgets and ledger. The UI lands in v4. The data is locked from day one.

Why now is the only window

The window for building this is short. Inside 18 months, either an HRIS incumbent (Rippling, Deel) will bolt on a thin tokens module, or one of the FinOps incumbents (Vantage, Finout) will rebrand around AI. Either path produces a mediocre version of the system of record and locks the rest of the market into a worse default.

The opportunity to ship the right shape — built from scratch on the right data model, with HR as the first-class owner — closes faster than people think.

The fourth pillar is here. The system of record isn’t yet. That’s the work.

Subscribe for new essays: tknfrg.com/subscribe