What we'll learn from Anthropic's $2-$3T IPO | E2325

Anthropic is planning an IPO targeting a valuation of $2 to $3 trillion, with the filing expected as soon as October according to Financial Times insiders. Jason Calacanis believes the $2 trillion figure is sandbagging and that Anthropic will price at the top of the range, which would surpass SpaceX's record private valuation of $1.7 trillion and place Anthropic among only five companies in history to reach the $3 trillion mark.

Investors expect Anthropic's ARR to reach $100 to $120 billion by end of year. Ramp data shows that 43.5 percent of US businesses paid Anthropic for subscriptions or tokens in the last month, compared to 39.7 percent for OpenAI and 6.2 percent for Google's AI tools, a reversal Calacanis attributes to Claude going from obscure 18 months ago to now being the leading enterprise AI product alongside ChatGPT.

The IPO filing will make public for the first time Anthropic's revenue ramp, revenue composition, salary and stock-based compensation, and compute costs relative to revenue. It will also clarify where Anthropic sources compute, with AWS, CoreWeave, and Crusoe Cloud named as possibilities, and will contrast Anthropic's rented compute model against OpenAI's strategy of building and committing to its own data centers. Calacanis advises against buying shares at IPO because bankers and CEOs price offerings to perfection and shares typically fall after the initial pop, recommending instead that investors build a position slowly over a ten-year window.

Anthropic is separately in talks to acquire Israeli startup Descartes AI for around $6 billion, which would be its largest acquisition to date. Descartes makes software that helps chips run more efficiently and can boost throughput by up to 8x across Nvidia chips, Amazon Trainium, and Google TPUs. At a $3 trillion valuation the $6 billion price represents roughly 25 basis points and a 50 percent premium over Descartes' prior valuation. One rationale offered for Anthropic going public is gaining a liquid stock to use as acquisition currency, with the observation that acquisitions under $100 billion are relatively small for a trillion-dollar company and that only one in ten acquisitions needs to succeed for the strategy to be accretive. The broader M&A environment has revived, attributed to the Trump administration's regulatory posture, with recent deals including Cursor acquired for $60 billion and OpenAI buying Jony Ive's company for an estimated $5 to $10 billion.

Grit AI, an 18-month-old South San Francisco robotics company, has raised $32 million in total funding including a recent $26 million Series B, and is deploying AI-assisted robots to address a severe labor shortage in solar panel installation. A strong crew of eight workers installs around 800 panels per day and a weak crew around 400, but Grit AI's robotic system targets 3,000 to 4,000 panels per day with the same crew size, aiming for a cycle time of 18 to 19 seconds panel to panel. The company has completed seven sites totaling approximately 35,000 panels and holds contracts covering roughly 6 million panels, representing a pipeline of approximately 2.8 gigawatts to be installed over the next 18 to 24 months.

Construction productivity has fallen 46 percent since 1965 even as agriculture, retail, and manufacturing saw major gains, and 41 percent of the US skilled construction workforce is projected to retire by 2031. Grit AI's founder Puneet noted that 70 percent of the solar labor shortage stems from nutrition issues and 30 percent from workers failing drug tests, compounding the structural workforce decline. Panel prices fell from $128 per watt in 1975 to roughly 26 cents per watt in the US in 2024, following a pattern where prices drop approximately 48 percent each time production doubles roughly every 18 months, and mechanical labor represents about 8 percent of total installation cost, making robotic labor reduction a meaningful cost lever.

Grit AI trains its robots using a combination of real construction site footage and AI-generated simulation data that injects synthetic equipment models into real footage and models environmental conditions including reflections, heat, rain, mold, and dust. Puneet said the goal is to learn a general representation of the construction world first and then apply it to task-specific models, and that after building that general representation, training a robot to tie rebar took approximately two days. Demand pressure is illustrated by one data center customer who demanded a 900 megawatt plant be built in 18 months and then escalated the order to 1.8 gigawatts in half the time.