Global Bond Yields Hit Multi-Decade Highs as AI Capex and Sovereign Debt Collide
Anthropic is preparing to grant CEO Dario Amodei and co-founders a new class of shares carrying super-voting rights ahead of a public listing expected as soon as late September, according to The Information. The company has filed a confidential S-1 with the SEC. Goldman Sachs, JPMorgan and Morgan Stanley are reported underwriters. If priced near the $1 trillion to $2 trillion range that bankers and backers have discussed, it would rank among the largest IPOs ever conducted.
The governance move addresses an arithmetical problem specific to Anthropic. Dario Amodei owns roughly 2% of the company. Repeated mega-rounds — the Series H raised $65 billion at a ~$965 billion post-money valuation — diluted founders to levels far below those of Zuckerberg or Spiegel when their companies listed. Super-voting rights restore operational control without clawing back equity.
A Hybrid Without Precedent
Anthropic's governance stacks two defensive layers. The first is super-voting stock for founders. The second is the Long-Term Benefit Trust, a five-member independent body of AI safety experts holding no equity. The LTBT controls Class T shares granting phased authority to elect and remove a board majority.
Amazon and Google, which committed enormous capital, hold limited or no voting rights by design; Google is capped around 14–15% with no board seats. Anthropic's Delaware public benefit corporation status adds a structural constraint: directors must weigh shareholder returns against a stated public-benefit purpose.
For public shareholders, the consequence is blunt. They supply capital at multi-trillion-dollar prices while founders and the Trust retain disproportionate say over strategy, spending and board composition. If management overbuilds compute, declines a lucrative contract on safety grounds, or sustains heavy capex through a downturn, ordinary shareholders will have limited recourse.
Revenue Real, Multiples Aggressive
Commercial traction justifies attention. Preliminary Q2 2026 revenue exceeded $11.5 billion, more than doubling Q1. The annualized run-rate crossed $65 billion by end of July, up from ~$9 billion at year-end 2025. Ramp's independent data confirms Anthropic is used by 43.5% of U.S. businesses in its sample, ahead of OpenAI at 39.7%. Anthropic was targeting its first quarterly operating profit — roughly $559 million, a 5% margin during extraordinary growth.
The valuation math demands scrutiny. A $2 trillion IPO implies ~30.8× the current run-rate. Bankers are reportedly underwriting partly against internal projections of $190–200 billion in 2028 revenue — an unusual reliance on forecasts two years out. Reuters notes this dependence as reflecting uncertainty around current margins. The public S-1 has yet to reveal gross margins, compute take-or-pay commitments, customer concentration, stock compensation or final governance terms.
Anthropic has committed more than $100 billion to AWS over ten years, $30 billion of Azure capacity, multi-gigawatt Google/Broadcom builds starting 2027, and all ~300MW at SpaceX's Colossus 1 facility. If the AWS figure alone averaged evenly, it would consume ~$10 billion annually — about 15% of today's run-rate — before other infrastructure and personnel. These commitments must be reserved years before the demand and pricing conditions that determine their returns become clear.
Rent the Intelligence, Own What Surrounds It
Anthropic itself has inadvertently demonstrated the most durable investment thesis — one that runs counter to paying $2 trillion for the model layer.
The company diversifies across three silicon architectures (Trainium, Google TPUs, Nvidia GPUs). It advertises 90% cost reductions from prompt caching and 50% from batch processing. It partnered with Blackstone, Hellman & Friedman and Goldman Sachs to build a services company deploying Claude into mid-market workflows. Each decision reveals the same conviction: the scarce resource has migrated away from raw model access toward orchestration, integration and the proprietary data surrounding it.
Capital allocators should internalize this signal. The sharpest asymmetry available is acquiring conventional businesses priced on current labor economics — companies where 20–50% of operating expense sits in repeatable knowledge work with proprietary data and measurable rework costs — then rebuilding their costliest workflows around agents before that productivity gain appears in the purchase multiple.
A parallel play exists in software: build the vendor-neutral control plane. Model routing, agent security, evaluation, identity and compliance gain pricing power as enterprises adopt multiple models. Ramp confirms this multi-model behavior is accelerating among the most sophisticated buyers.
Anthropic may prove a generational company. At $2 trillion, investors would be paying the model supplier's price for returns that may accrue more reliably to infrastructure owners, workflow operators and the neutral middleware between them. Anthropic's own strategic choices — spreading across clouds, silicon and services — are the clearest evidence of where durable economics actually concentrate.
not investment advice