The Investor

The Event

Anthropic filed an S-1 confidentially as a Public Benefit Corporation, which makes it the first frontier lab to walk into the public-market pipeline. In roughly six months the private AI comp set stops being a narrative and becomes arithmetic: audited revenue, gross margin, customer concentration, capex.

The numbers leading in are not subtle. The latest Series H printed at $965B on $47B ARR (~20x forward), with $15B from hyperscalers and memory OEMs — Micron, Samsung, SK Hynix all taking equity. Revenue went from $400M to $4.8B in twelve months, a 12x YoY growth rate that compressed the OpenAI gap from 4:1 to 1.25:1. The CEO conceded 80x growth against a 10x plan, which is the kind of admission that sounds like a brag and reads, in an S-1, like a forecasting problem.

Why This Reprices Everything

The filing forces three disclosures the private market has been guessing at for two years:

• Compute COGS — the actual gross margin of a frontier model business. Consensus guesses 40-55 percent. The marks being used assume the 75-80 percent SaaS benchmark.
• Enterprise concentration — how much of the $47B ARR is Amazon and Google commitment versus diversified enterprise demand.
• Capex-to-revenue trajectory — whether $219B of industry spend is closing the gap to revenue or widening it.

Every private AI mark in the book — frontier-lab secondaries, application-layer companies using Anthropic multiples as a comp — gets sorted into defensible or indefensible on these three numbers.

The Simultaneous Chinese Pricing Attack

In the same week, MiniMax M3 launched at $0.12 per million input tokens against Anthropic Opus at $5. That is a 97.6 percent discount with claimed near-parity on coding. DeepSeek V4 and Kimi K2.6 sit in the same direction. xAI's Grok Build entered at $1/$2 per million tokens through Cursor, OpenRouter, and Vercel.

This isn't a feature war. It's a structural margin event for any provider whose moat depends on token-pricing power.

Open-weight has closed the gap on the highest-value workload, agentic coding, and the API tier is being undercut from below. Any fund marking proprietary-LLM positions at 40-60x ARR without stress-testing a 30-50 percent gross margin compression is carrying stale paper.

Cross-Source Tension

This is probably wrong, but the sources disagree productively. The bubble framework shows 1 of 5 indicators red, with revenue doubling every 0.73 years, which argues sub-bubble with eighteen months of runway. The Chinese pricing data argues the opposite for proprietary model bets specifically. The reconciliation, or rather the more interesting version of it: bullish for picks-and-shovels and verticalized applications, bearish for AI-wrapper SaaS at 50-80x ARR and standalone frontier model bets without data moats.

The Structure

Apollo and Blackstone built a $36B SPV to buy Google TPUs and lease them to Anthropic, which is either an interesting piece of financial engineering or — to use a more honest description — the exact template that turned aircraft leasing into a standalone credit category in the 1970s. The thesis is that this is not a one-off. Within twelve months, Apollo, Blackstone, Ares, KKR, and Brookfield each stand up a dedicated AI-infra credit pod. The thesis could be wrong if the depreciation curves on accelerators turn out to be too steep for off-balance-sheet treatment to clear, but that is a footnote, not a counter.

The amplifier was Anthropic putting memory OEMs on its cap table. Micron, Samsung, and SK Hynix took equity, which suppliers do not do when their order books clear. They do it when a purchase order is no longer sufficient to guarantee allocation. Combined with ten gigawatts of locked compute across Amazon, Google/Broadcom, and SpaceX's Colossus clusters, the Series H is better read as supply-chain financing dressed as equity.

The Neocloud Signal

Neoclouds picked up 600bps of AI capex share, twelve percent to eighteen, in roughly nine months. The hyperscaler monopoly is fragmenting in real time. Baseten raising one billion at eleven billion sets a fresh comp for inference infrastructure at premium multiples, landing the same week NVIDIA confirms Vera Rubin's supply chain is 2x Grace Blackwell scale.

The SoftBank Anchor

SoftBank committed €75B/5GW to French AI data centers, with a forty-five billion euro, 3.1GW first phase alongside Schneider at Dunkirk. Largest European AI infra deal on record. SoftBank itself crossed three hundred and five billion dollars to pass Toyota as Japan's most valuable company, purely on AI exposure. Non-US capital is voting that AI capex is the new industrial base, which is a strong claim and also the one the price is making.

Every dollar of frontier-lab valuation is a vote for compute, memory, and power inputs. The $36B SPV just told you how those votes will be financed.

The Investable Categories

Three layers of the AI compute stack are now separately financeable, which is the part allocators should be reading twice:

1. Credit layer — chip-leasing SPVs with off-balance-sheet treatment and aircraft-leasing economics, which is to say lender returns with operator risk priced out
2. Equity layer — neoclouds with secured allocation, where the 600bps share gain is reproducible if the supply constraint holds
3. Infrastructure layer — power, cooling, interconnect, substation suppliers sitting on multi-year contracted demand, the dullest tier and probably the best risk-adjusted one

What Changed

For three years the consensus has been that CUDA's moat is unbreachable, with Doug O'Laughlin's 'three-headed hydra' of hardware, networking, and software cited as the reason to underwrite Nvidia at more or less any multiple anyone cared to print. This week is the first hard evidence the third head is bleeding.

DeepSeek V4 trained at scale on Huawei Ascend. AIGCode posted 65% MFU on MoE pre-training, which they claim is roughly twice the industry average. Huawei open-sourced CANN's runtime and compiler in December 2025. Chinese banking deployments are landing on Ascend in production, not in slideware. Wall Street is calling it 'another DeepSeek moment,' which is the sort of thing Wall Street says.

The Bifurcation Thesis

The AI compute market now runs as two parallel ecosystems, and the interesting question is which side of the wall any given line item in the book actually sits on:

The awkward part, or rather the more interesting version of it: US export controls funded this transition. AIGCode adopted Ascend because GPUs were unavailable, not because anyone enjoyed the tooling. Forced adoption produced an unwilling design-partner base, which matured into a contributing community now shipping upstream optimizations. That is a hard outcome to engineer on purpose, and Commerce did it by accident.

Portfolio Implications

The Commerce Department's shift to headquarters-based enforcement, which closes the offshore-subsidiary loophole, accelerates the bifurcation rather than slowing it. Any company whose moat or TAM assumes CUDA lock-in in Greater China deserves a re-read. That includes Western AI infra, MLOps, and inference startups underwriting Chinese enterprise revenue lines they are unlikely to collect on.

CUDA's moat just got its first credible attacker, and the only investors who get repriced are the ones still pretending it didn't happen.

Diligence Gaps to Close

This is probably wrong in places, but the maturity looks real and uneven rather than uniformly real. Qiuwu Chen, who waited four months for a Huawei support ticket in early 2024, now describes CANN as moving from 'infancy' to 'youth.' ChinaTalk's May 2025 finding of low organic developer engagement is a year old, not a decade. The distance between '4M community members' and actual collaboration is exactly where managed metrics hide. Primary-source diligence is non-negotiable: real MFU benchmarks, production-versus-POC ratios, and lived support SLA experience.

The Playbook

The Bending Spoons model — buy bloated vertical SaaS, strip headcount, rebuild core with agents, raise prices, and license the workflow data to frontier labs — is going institutional. The acquisition logic has flipped from revenue multiples to workflow-data extraction value. Frontier labs are quietly paying for domain-specific workflow telemetry that they cannot generate synthetically.

This converges with the organic traffic collapse data. 60% of desktop searches now end without a click, 83% of AI Overview searches suppress clicks entirely, and 73% of page-one brands are absent from AI answers. Mid-market SaaS companies that acquired customers through SEO face a double squeeze: eroding acquisition economics and a reframing of their value from ARR to data.

The Roll-Up Economics

The trade works because of a valuation arbitrage:

• Buy at: Distressed SaaS multiples (3-6x ARR for sub-$200M companies with compressed growth)
• Rebuild with: Agent-driven ops at 50-70% lower headcount
• Exit at: Rebuilt product value + data-licensing annuity to model trainers

The window closes when the first publicly-marked $1B exit validates the playbook and multiples re-rate upward for the entire target pool.

The Enterprise Renewal Cliff

CFOs are entering a cycle where they'll ask enterprise AI line items what the measurable outcome was. A non-trivial number of AI portcos will not have a clean answer. This creates a barbell: companies that own a measurable customer outcome compound; thin model-wrappers face renewal compression as labs absorb the bottom 30-40% of generic agent workloads.

The honest read: labs eating 30-40% of generic workloads collapses the moat for thin AI wrappers, while companies with deep workflow data and distribution become acquisition targets at premium data-licensing multiples.

Where This Intersects

The convergence of (a) SEO-dependent SaaS losing its organic acquisition channel, (b) frontier labs willing to pay for workflow data, and (c) agents enabling 50-70% headcount reduction in acquired companies creates a PE-style opportunity with venture-style returns. The target list: 15-25 vertical SaaS companies with rich workflow telemetry, sub-$200M ARR, and compressed multiples.