$10,000 in ARM and PLTR at their IPOs became roughly $40,000 and $200,000, setting up the core claim: public-market re-rating can be explosive, but the post-IPO period is when lockups, earnings, and competition expose the real tradeoff. Anthropic filed to go public as A N on NASDAQ the week of November 16, with a possible ~$100 billion raise at a $1.5 trillion to $2 trillion valuation, and the notes argue the business is exceptional but now faces the same public-market scrutiny that hit ARM, PLTR, and CBRS. The bull case is revenue growth, 85% gross margin in 1H26 on a reported basis, and enterprise dominance in coding; the bear case is 23 unlock dates, heavy compute commitments, and the risk that agentic usage and competition squeeze margins. The stated plan is to wait for Anthropic’s first public earnings report early next year and watch whether compute costs keep falling as a share of revenue before buying.
$10,000 in ARM at its IPO 3 years ago is almost $40,000 today; $10,000 in PLTR at its 2020 IPO is close to $200,000, framing how early public listings can rerate.
Anthropic’s AI models are said to dominate the enterprise market, and Anthropic has filed to go public.
Pre-IPO private companies can pitch billions in future revenue without showing last quarter’s results; after IPO, quarterly earnings, audited books, fixed-schedule risk disclosure, and SEC enforcement start.
Most IPOs pop first, then the market starts pricing market share and direct competition; insiders are typically locked up for around 6 months, and when lockups expire billions of dollars of stock can hit the market.
PLTR priced at $10 on September 30, 2020, traded at 39 four months later, then fell 44% from its peak by late March after about 80% of shares became sellable when its lockup ended on February 18.
ARM relisted on September 14, 2023 at $56, gave back all first-day gains within a week, was down almost 25% by late October, and later traded above $200, up around 300% from the IPO despite a recent pullback.
CBRS started trading this past May at $350, its lockup ends on November 9, and it trades around $110 today, about 60% below its first-day price.
Anthropic’s IPO is expected to start the week of November 16 under ticker A N on NASDAQ, with an expected raise of around $100 billion at a $1.5 trillion to $2 trillion valuation; that would top SpaceX’s June stock raise.
Anthropic’s S1 lists 23 separate unlock dates: two trading days after the first public earnings report, executives and certain employees can sell another 35%; the rest unlock after the next earnings report or 180 days public, whichever comes first; the seven co-founders can sell up to 29% of their shares.
All seven Anthropic co-founders previously worked at OpenAI, and CEO Daario Amade says they left over differences in how to build, release, and scale models while keeping them safe.
Anthropic says it can improve models faster than scaling laws alone predict via 'compute multipliers,' and says Claude helps researchers find those multipliers faster.
Claude is built in three stages: pre-training predicts the next token across text and code; post-training turns that capability into an agent that can plan and execute tasks; reinforcement learning raises high-scoring outcomes and lowers low-scoring ones.
Anthropic says its library of training tasks is a durable advantage, and post-training is where much of its safety work happens, including constitutional AI that checks answers against a public set of principles.
About 80% of Anthropic’s revenue is built by the token, almost all cost of revenue is infrastructure, and inference is where the money is made.
Claude is designed to run on Amazon Tranium, Google TPU, Nvidia GPU, and AMD GPU chip families, letting Anthropic shift workloads to the best price-performance supplier instead of relying on one source.
Anthropic says it has an edge in enterprise work, especially coding, while open weight and open source models — including Meta’s Llama and 'Deep Sea Car 1' — pressure pricing by offering cheaper enterprise alternatives.
Anthropic reported revenue of $41 million in 2023, $386 million in 2024, $4.6 billion in 2025, and $16.3 billion in 1H26; Q2 revenue was stated as 11.6 $6 billion, putting the annualized run rate near $46 billion.
Customers paying over $100,000 a year rose from almost 1,000 to almost 12,000 in 12 months, and no single customer accounts for more than 10% of revenue.
Gross margin was about 85% in 1H26 on a reported basis, but only about 45% including all compute costs; the bull case is that each new 'quad' model earns more revenue per dollar of compute, while Q2 revenue grew 1368% year over year and compute costs 300%.
Anthropic says it must commit to compute 12 to 18 months ahead, and as of June it had nearly $600 billion of compute and data center deals, almost 13x current revenue run rate, with some contracts lasting 20 years.
Anthropic’s backlog is only $10.8 billion, less than one quarter of revenue, while 80% of revenue is token-based and many customers can increase or cut usage quickly.
Anthropic’s S1 warns that agent products use more compute than other products, so agentic growth could worsen compute efficiency; the watch item is total compute costs versus total revenue.
Anthropic reported an $80.7 billion net loss gap in 1H26, with almost 90% of that non-cash and mostly tied to convertible notes revalued higher as the company’s valuation rose.
Anthropic posted a billion of operating profit in Q2, about an 8% operating margin, generated $3.9 billion of operating cash flow in 1H26, and ended June with $131.7 billion in cash and short-term investments.
The stated trade plan is to wait for Anthropic’s first public earnings report early next year, then see whether compute costs keep falling as a share of revenue even as agents scale, with the timing framed around insider selling 2 trading days after that report.