Anthropicat$45BARR,CerebrasIPOResetFrontierAIComps

◆ DEEP DIVES

1. 01

   Anthropic's 80x Growth Rate and the PE Bloc War That Followed It

   The Revenue Print That Changes the Stack

   Anthropic went from $9B to $45B annualized revenue in approximately five months, which is the sort of number that either rewrites the competitive map or gets revised in a footnote nobody reads. OpenAI needed roughly two years to cover a similar absolute dollar range. Anthropic covered it in a quarter and change, running at 80x annual growth against an internal plan of 10x, and is now raising at $1 trillion. That would make it the most valuable private company ever created. The number is the punchline.

   Three potential $1T+ AI IPOs — Anthropic, OpenAI, SpaceX at $1.75T — each individually exceed the entire 1999-2000 venture IPO market (~$45B raised at $270B aggregate in real terms).

   The velocity matters more than the headline because it answers a question the market had been pricing as genuinely open: which frontier lab is actually winning enterprise. The moat questions around OpenAI — no unique technology, limited engagement stickiness, no network effect — now sit next to a competitor growing fivefold in the time it takes to negotiate a single enterprise contract.

   ---

   PE Has Picked Sides

   What makes this structural rather than episodic is that the deployment blocs are forming in public. The lineup:

   Dimension	OpenAI Bloc	Anthropic Bloc
   PE Partners	TPG, Advent, Bain Capital, Brookfield (+15 others)	Blackstone, Goldman Sachs, Hellman & Friedman
   JV Capital	$10B target, $4B+ committed	Undisclosed, >$1B anchored
   Compute Alignment	Microsoft ($280B revenue commitment)	Google Cloud ($200B commitment)
   Distribution	Tomoro acquisition (150 FDEs) + Big 4 partners	Goldman JV + Anthropic direct enterprise

   The fund-positioning implication is immediate, and slightly uncomfortable for anyone mid-process. Independent sponsors chasing AI services targets now have approximately two quarters before the lab-backed JVs start bidding with balance sheets no independent can match. The 6-month window thesis making the rounds this week is probably aggressive on timing and directionally correct on everything else. Once both labs are visible acquirers, seller expectations reprice upward and stay there.

   ---

   The Dispersion Trade

   The OpenAI-versus-Anthropic spread on secondaries is the most actionable pair in private AI, or rather, the most actionable one that does not require guessing what Microsoft does next. Anthropic's revenue velocity, enterprise traction, and compute leasing behavior (taking capacity from Musk's Colossus, which is a sentence that would have read as satire eighteen months ago) all point to operational momentum that OpenAI's metrics — despite higher mindshare — do not match. OpenAI at $852B post-March on roughly $20B revenue is about 42x. Anthropic at $1T on $45B ARR is about 22x forward. The faster grower is the cheaper multiple.

   The counter-thesis deserves stating: OpenAI's distribution through Microsoft, its consumer brand, and the $10B deployment JV create switching costs that Anthropic's API-first posture does not replicate. The bear case on Anthropic is that it becomes the better product in a market that pays for distribution. That case has historically lost in enterprise software. It has not yet been tested at this scale.

   The Telco Bear Case

   Benedict Evans' parallel deserves real IC time rather than a polite nod: telecom traffic grew several thousand-fold over twenty years while the stocks went flat, $1T of revenue against $200B of capex. If AI runs the same curve, model-layer equity at 100x+ revenue is structurally mispriced regardless of how fast anyone is growing this quarter. Size positions so you survive that scenario and have dry powder to buy it.

   Action items

   • Widen the discount on OpenAI secondaries relative to Anthropic in your book; model exit scenarios at current tender marks versus IPO hold-through
   • Screen 10-15 AI services targets with 50+ FDEs and F500 logos for accelerated outreach before lab JVs sweep them
   • Add 'telco commoditization' downside scenario to every model-layer investment memo at next IC
   • Build a deployment-services PE thesis: identify targets where 20-40% of cost base is AI-addressable (BPO, claims, legal ops) using Long Lake/Amex GBT $6.3B buyout as template

   Sources:Anthropic at $45B ARR + PE deployment vehicles: the AI services alpha is now · The argument going around this week is that OpenAI and Anthropic standing up private equity joint ventures · Cerebras has repriced itself twelve and a half times over · AI infra concentration risk just hit systemic levels

2. 02

   Procurement-as-Equity: The New Structural Risk Hiding in AI Infra Term Sheets

   What OpenAI Just Invented

   On Christmas Eve 2025 two deals closed within hours of each other: Nvidia agreed to buy Groq, and OpenAI signed a $20 billion, 750-megawatt supply agreement with Cerebras carrying equity warrants that take OpenAI to eleven percent of the company. Cerebras goes public Thursday at a $50B+ valuation, which is twelve-and-a-half times what it was worth eighteen months ago at four billion, and it will raise over five billion dollars in the largest semiconductor IPO ever recorded.

   The multiple is not the story. The structure is. OpenAI has figured out that it can pay for chips in a currency the chipmaker values more than cash — or rather, the more interesting version, credibility at IPO — and extract terms a normal customer would never get. The contract is described as 'a substantial portion of revenues for several years' and is terminable at OpenAI's discretion. That is procurement wearing a cap table as a costume.

   What is going public on Thursday is a $20B OpenAI contract trading at roughly a 125x revenue multiple, and the trade worth having is the repricing of every private inference deal already in the pipeline.

   ---

   Why This Is a Template, Not a One-Off

   The innovation — compute buyer takes equity in supplier while keeping walkaway rights — gets copied because it works for both seats at the table. The supplier gets a credibility anchor and an IPO backstop. The buyer gets below-market terms and optionality it could not otherwise price. Anthropic, Google, and Meta will replicate this template because the economics beat arms-length procurement on every axis that matters.

   For anyone holding AI infra equity, this produces a form of structural dilution that standard term sheets do not yet contemplate. Pro-rata on infra companies courting frontier labs is now subject to customer-equity deals that route around traditional financing rounds entirely. The provisions to demand:

   • Anti-dilution protections specifically against customer-equity conversions
   • Right of first refusal on any equity-linked procurement agreement
   • Information rights on customer termination optionality
   • Revenue concentration covenants with equity-conversion triggers

   ---

   The Private Comp Repricing

   Cerebras at fifty billion sets the public anchor. Every independent inference chip startup — Tenstorrent, Rain, SambaNova, d-Matrix, Etched — now prices against that comp, and the math cuts both ways:

   Scenario	Comp Effect	Action
   Cerebras trades up post-IPO	Private marks expand 20-40%	Lead at 15-25x forward in current rounds
   Cerebras gaps down within 2 quarters	Private inference stack compresses 30-40%	Best entry points look embarrassing at the time — fund them
   Cerebras trades flat, OpenAI exercises termination	Customer-concentration discount becomes permanent	Only back companies with diversified customer books

   This is probably wrong, but the independent inference scarcity premium is real. With Groq now inside Nvidia and Cerebras effectively OpenAI-captive, the roster of genuinely independent inference startups with hyperscaler optionality is shrinking in a countable way. Any asset with a clean cap table and a non-OpenAI anchor attracts a scarcity premium that did not exist a quarter ago.

   Three data points matter more than the tape post-IPO: quarterly capacity utilization (the prospectus itself admits to planning challenges), customer count disclosure (if OpenAI remains dominant through 2026, concentration becomes a permanent discount), and the depth of the AWS partnership, which is the only visible diversification story on the page.

   Action items

   • Require anti-dilution and ROFR provisions against customer-equity deals in all active AI infra term sheets before next closing
   • Build a ranked watchlist of independent inference startups (Tenstorrent, Rain, SambaNova, d-Matrix, Etched) by non-OpenAI anchor customer, capital runway through 2027, and M&A optionality
   • Request meetings with Cerebras IR post-IPO to pressure-test the $400M→$8B revenue ramp and the $3B 2026-27 cash burn against actual buildout timelines
   • Model Cerebras post-IPO gap-down scenarios and their read-through on sector multiples for public AI infra exposure (Nvidia, AMD, hyperscalers)

   Sources:Cerebras has repriced itself twelve and a half times over · Cerebras doubled in ninety days · Nvidia's China door narrows as Cerebras prices Wed · Cerebras $4.8B IPO + Huang snub · Inference is splitting into two markets

3. 03

   The Barbell Effect: Mid-Tier AI Model Economics Just Died — Reprice Your SaaS Book

   The Pricing Data

   GPT-5.5 and Opus 4.7 shipped with major price hikes, which is OpenAI and Anthropic telling you they think they have pricing power. They might. DeepSeek V4 Flash, meanwhile, undercuts the frontier at roughly five times lower cost, with a quality gap that is narrower than it was three months ago. The middle tier has effectively stopped existing as a gross-margin layer for application companies.

   This is not a thesis. Monday.com is the proof point: stock down forty-eight percent year-to-date, growth decelerated from above twenty-seven percent to nineteen or twenty, flat headcount guidance running alongside compressing gross margins. Management flagged AI compute costs themselves. The OpEx saved from the hiring freeze is being recycled into inference COGS, which is a less forgiving line item.

   Per-seat SaaS revenue is structurally exposed as AI displaces headcount — the pricing model disruption is larger than the AI feature disruption.

   ---

   The Pricing Model Transition Is Accelerating

   Kyle Poyar's survey puts numbers on what was previously a vibe:

   • Primary outcome-based pricing: five percent today, thirty-one percent by mid-2029 (a six-fold shift)
   • Hybrid models: twenty-five percent in 2025, thirty-seven in 2026, forty-seven projected
   • FedEx's CDIO publicly preferring outcome-based pricing as a trust mechanism
   • ServiceNow's COO calling it unmeasurable, which makes this a binary bet that resolves by FY27

   Benioff committed to outcome-based. ServiceNow's COO rejected it. One of them is wrong, and which one determines the multiple for an entire category.

   ---

   The Two-Front War

   Per-seat incumbents are taking fire from both sides at once:

   Pressure Vector	Mechanism	Timeline
   Seat count shrinks	AI replaces workflow-level human work; fewer users need the tool	Happening now (Monday flat headcount, growth decel)
   Pricing model shifts	Buyers demand outcome-tied or usage-based pricing on remaining seats	12-18 months to majority adoption
   COGS migrate from OpEx to inference	Every AI feature adds variable cost where headcount was fixed	Immediate — visible in Q2 prints

   The Ramp and Prime Intellect result is the interesting puzzle at the model layer: a small RL model beat Claude Opus by four percent at Haiku latency on spreadsheet Q&A. Enterprises sitting on proprietary data now have a credible route around the foundation-model API, which means the token bill is not merely variable, it is contestable.

   Where the Alpha Lives

   Three positions look asymmetric, and naming the opportunity cost matters here — every dollar into a seat-based incumbent is a dollar not into these. (1) Vertical AI agents with hard outcome metrics (closed tickets, resolved claims, booked meetings) can sign outcome contracts without the measurement argument ServiceNow just raised — source aggressively. (2) Outcome-measurement infrastructure is the unsolved problem the same COO named in public, and picks-and-shovels usually get paid before the miners do. (3) Any SaaS portco with more than sixty percent seat-based revenue and AI-displaceable workflows deserves a twenty to thirty percent NTM multiple compression assumption over the next eighteen months. This is probably wrong on one of the three. It is almost certainly right on the other two.

   Action items

   • Run a per-seat revenue exposure audit across the SaaS portfolio this week — flag holdings with >60% seat-based revenue and AI-displaceable workflows for trim/hedge discussion
   • Source 5-10 vertical AI agent companies with contractually measurable outcomes (ticket resolution, claims processed, meetings booked) before the Poyar data becomes consensus
   • Add 'outcome measurability' and 'services revenue percentage' as mandatory diligence gates for all AI software deals effective immediately
   • Stress-test every AI-wrapper deal in pipeline against a scenario where mid-tier model margins compress 40% and only premium or commodity tiers survive

   Sources:SaaS per-seat model is dying · The pitch from every SaaS management team this cycle · AI infra security gap is now investable: Ollama 300K-server exposure + model pricing barbell · Native interaction models shipped this week · Arista just told the market something worth listening to