The Investor

What Just Happened

Anthropic filed a confidential S-1 targeting fall 2026 at a valuation between $900B and $965B on an annualized revenue run-rate of $47B, up from roughly nine billion six months ago. A 5.2x jump in half a year is either the fastest legitimate revenue ramp in software history or a reminder that six-month-old enterprise contracts make ARR a number doing considerable work. Probably both.

The filing landed the same week OpenAI is reportedly preparing its own roughly one-trillion-dollar registration, which creates a dual-IPO window asking public markets to absorb more than fifty billion in primary from two frontier labs in the same quarter. That level of supply has no modern precedent.

Why This Changes Your Book

Until this filing, every late-stage AI mark lived in a fog of private-round comps set by strategically motivated counterparties. Once Anthropic prints, the fog lifts. The implied multiple — roughly 20x forward revenue at 5x growth — becomes the ceiling rather than the floor for every private AI name in the pipeline.

Any private AI company pitching 50-100x ARR is now mispriced relative to a public comp that just emerged. The next 90 days are the window to remark before auditors do it for you.

The OpenAI Spread Trade

Florida's 83-page product-liability suit naming Altman personally, citing the FSU shooter's 270+ ChatGPT messages and two USF student deaths, opens a measurable gap between the two leaders at precisely the wrong moment for OpenAI's roadshow. Sources diverge on the sizing — estimates run from fifteen to twenty-five percent discount warranted for OpenAI versus Anthropic on secondaries. Bill Gurley publicly calling Anthropic a 'mystery' and invoking 'Dr. Frankenstein' is a tier-one VC distancing weeks before an IPO. That is governance risk the public narrative has not priced.

Three Scenarios to Model

Cross-Source Contradictions Worth Noting

Sources disagree on whether $47B is gross or net revenue. One read has Anthropic's figure pre-inference-partner payments while OpenAI's $30B is net, meaning the headline gap overstates Anthropic's lead. The S-1 treatment of compute costs will resolve it, but the distinction matters for multiple derivation. Opus 4.8's split benchmarks — ARC-AGI-3 shows it tripling GPT-5.5, while Datacurve shows it below GPT-5.5 with higher token burn — leave open whether 20x revenue is paid for product differentiation or, more plausibly, the growth rate alone.

The Signal Behind the Number

Alphabet has issued equity for the first time since 2005, and the package is worth reading slowly: $80B total, split into a $10B private placement to Berkshire Hathaway at a discount, $30B in mandatory converts underwritten by Goldman, JPM and Morgan Stanley, and $40B via an ATM beginning in Q3. The company sits on $126B in cash against $81B in debt, a $45B net cash position. So this is not a balance sheet that needs equity. It is a balance sheet telling everyone within earshot that AI compute demand has outrun even Alphabet's free cash flow.

When the patron saint of value investing writes a $10B check on AI infrastructure at all-time highs, the permission slip for every pension and sovereign allocator just got stapled to the cap table.

What Berkshire's Check Actually Means

Greg Abel deploying from a $373B cash pile is not venture tourism, or rather, not the version of venture tourism Berkshire has historically declined to do. It is Berkshire underwriting the TPU thesis as a value play, which is the loudest available way to suggest the market should start pricing Nvidia's GPU position as cyclical rather than structural. Google Cloud has gone from a money-losing 6% rounding error to a $20B/quarter, 33%-margin business in seven years, having crossed profitability only in Q1 2023. Two and a half years of operating leverage is early innings.

The Neocloud Squeeze

Three data points, held together, tell the structural story:

• Anthropic secured meaningful TPU supply from Google while simultaneously paying SpaceX $1.25B/month for marginal compute it could not get elsewhere, structured as a 180-day lease with 90-day cancellation
• Google's custom-silicon stack carries structural cost advantages at the training layer
• Meta's $125-145B capex (more than 50% of revenue), plus the stated intent to resell excess capacity, points at impending inference oversupply

Read together, frontier labs are price-takers on compute and custom-silicon hyperscalers are price-setters. If that holds, neocloud margins compress toward cost of capital while hyperscaler multiples expand, and CoreWeave-style neoclouds trading at 15-25x forward revenue probably compress 20-30% as TPU and Trainium supply ramps.

The Opportunity and Its Risks

This is probably wrong, but the cleanest expression is not "long Google." It is long the compute hierarchy, short the middle tier: custom-silicon hyperscalers at the top, application-layer companies riding the infrastructure without funding it at the bottom, GPU-rental neoclouds squeezed between.

The contrarian watch: whether Google follows with substantial debt issuance. If the package stays equity-only, management is risk-sharing AI ROI uncertainty with shareholders rather than levering its own balance sheet, which is a thing rational managers do when they are not sure of the return.

The Uncomfortable Math

The entire trillion-dollar-plus AI capex thesis currently rests on a single use case. Essentially all of Anthropic's $47B run-rate traces to agentic coding, and there is no second leg at scale yet. Customer-service agents, legal automation, AI-native consultancies — promising, none of them at revenue that moves the needle. The $965B valuation implicitly assumes one of them emerges. The prospectus will have to say so out loud.

And the customers generating that revenue are hitting walls.

The Margin Threat From Below

MiniMax M3 claims Claude Opus 4.7 parity at 40x lower input token cost, which is the first credible margin threat to Western frontier labs from Chinese open-weight models. Take the haircut: assume the real advantage is five to ten times rather than forty. That still caps the gross margin curve Anthropic can underwrite into a prospectus. Anthropic itself cut Opus fast-mode pricing three times this cycle, from $30/$150 to $10/$50, and ElevenLabs cut Music API pricing 50%. Three vendors moved on price in the same quarter, which is difficult to read as promotional behavior.

If the AI book does not survive a coding-only, prices-halving scenario, that is not a thesis. That is beta with a story attached.

Where the Alpha Actually Sits

The bill-shock data creates a category that did not exist as a budget line a year ago: AI FinOps and governance tooling. Every CFO at every Claude- or GPT-using enterprise sees the $500M story this week. The category is forming, with Helicone, Langfuse, Vantage AI, and Pay-i in the frame and no obvious winner. Enterprise model routing is becoming standard infrastructure too: Snowflake has built CoCo internally, Palo Alto Networks ships bespoke routers to its enterprise customers, and OpenRouter is emerging as the neutral layer for everyone else. UiPath achieved 90%+ cost savings via prompt engineering on repeat tasks. Mars forced Google into flat per-seat pricing across 62,000 employees.

Which is the uncomfortable part for the people building the trillion-dollar capex case. Uncapped usage growth at premium prices, the assumption justifying frontier model valuations, is being actively dismantled by the procurement officers who write the checks.

The Number That Matters

Salesforce bought Contentful, a category-leading headless CMS, somewhere between $1-1.5B against a 2021 mark of $3B. That is a fifty to sixty-seven percent haircut on what should have been the good outcome: a willing strategic, a clean Agentforce fit, and a buyer who could actually articulate the synergy in a sentence. If this is the price for the best case, the rest of the 2021-vintage growth SaaS book is wearing the same haircut whether the marks have caught up or not. The marks usually do, eventually.

The Broader Evidence

• 220+ former unicorns now trade sub-billion
• Roughly half of 857 US unicorns have not raised in three or more years
• Salesforce itself is down 21% YTD, and the only catalyst that worked this year was disclosing a roughly five billion dollar Anthropic stake
• SaaS multiples are 50%+ from peaks lower, and the floor people kept calling did not hold

If Anthropic at $965B is absorbing the marginal AI dollar, every public software name pitching an AI story is competing with a private comparable they cannot reach. That is a reallocation problem before it is a multiple problem.

The Reallocation Mechanism

Capital going into Anthropic at $965B or Alphabet's $80B raise is capital that is not going into growth SaaS at any bid the cap tables would accept. Or rather, the more honest version: it is not going in at all. The names that survive this are the ones whose customers were paying for the product rather than the narrative. There are fewer of those than the prospectuses suggested.

Meanwhile OpenAI is quietly walking off with Salesforce's enterprise GTM bench, having poached CRO Dresser and VP Global Partnerships Landsman, which is what a company does when it is about to compete in apps and partnerships rather than sit politely behind an API. Any B2B SaaS position whose moat assumed OpenAI stayed an API company needs re-underwriting this quarter. This is probably wrong as a universal claim. It is not wrong as a base case.

The Gainsight Tell

Gainsight, a $100M+ ARR category leader in customer success, is publicly saying its own SaaS model is insufficient and spinning Atlas as a separate P&L on the thesis that AI-native services address a larger TAM. When the incumbents start cannibalizing themselves on purpose, the comps move with them. Apply a 15-25% multiple haircut to any SaaS holding that cannot tell a credible AI-native P&L story, and be honest about which of yours can.