Anthropic IPO
TL;DR Highlight
Anthropic opened its IPO roadshow, disclosing Q3'26 revenue of $17.0B and other figures.
Who Should Read
A developer who uses the Claude API or Claude Code at work and wants to understand Anthropic's finances and supply stability.
Core Mechanics
- Q3'26 revenue was $17.0B
- The filing says revenue grew more than 12x year over year.
- Customers spending $100K+ per year topped 11,900, also up more than 12x year over year.
- Compute efficiency up over 4x
- Up more than 4x versus Q1'25, per the filing.
- It compares revenue divided by compute and infrastructure expenses.
- Net loss was $23.3B
- It includes a $21.9B non-cash charge from fair value changes in convertible notes, derivative liabilities, and SAFEs.
- Operating margin is listed as (3)%.
- IPO dates and ticker are set
- The roadshow starts November 10 (EST).
- Pricing is expected November 18 and trading November 19, under Nasdaq ticker ANT.
- Customer stories center on agents and coding
- Rakuten says time to market for new features fell from 24 to 5 days.
- Salesforce says an API migration took 13 days instead of 231 days of work.
Community Reactions
- Comment summarizing price range and size
- A commenter listed a $70–78 price range and an $85.3B gross deal size including a $26B private placement.
- They also cited ~$968B market cap at $74 and $900.9B in compute commitments through 2045.
- Opinion questioning why they list now
- One commenter argued that if models were capable of recursive self-improvement, VCs would delay the IPO until dominance was clear.
- They wondered why $100B is needed today.
- Experience that switching models is easy
- One commenter said most apps don't need frontier models and switching to cheaper ones is easy.
- Another said open-weight models trail the frontier by a few months at far lower cost.
- Mention of the Mythos GPU port claim
- A commenter said the roadshow video claims Mythos ported Claude from NVIDIA to AMD GPUs in 3 days with almost no human intervention.
- They read it as weakening NVIDIA's moat if true.
How to Apply
- Build a Claude dependency checklist
- Spend 30 minutes today listing every place your service calls Claude.
- Note for each whether it could be swapped to another model.
- Abstract your model calls
- Move model names and API keys into config.
- Make it possible to run the same prompt against another model.
- Separate tasks that need a frontier model
- Compare cheaper-model and top-model outputs per task.
- Move tasks with a small quality gap to the cheaper model.
Try It
# Paste into Claude or ChatGPT (the source is an IPO page, so this is a prompt-level lightweight check)
The features in our service that call an LLM are:
{{feature list: name, model used, monthly call volume}}
For each feature, give me a table with:
- Whether a frontier model is truly needed (Y/N) and why
- Quality metrics to check if switched to a cheaper model
- Code locations to decouple so the provider can change
Next step: pick one N item and compare both models' outputs on the same input.Terminology
Related Papers
Gemini 4 Neon
Gemini 4 Neon cuts cost per task 32% versus Carbon, at $3 input and $15 output per million tokens
AI is killing All About Berlin
After Google AI Overviews, a Berlin guide site lost 70-75% of its traffic.
Not Everything Needs an LLM
PostHog used small decision models to skip 98.7% of expensive LLM calls
GPT‑6.2 Astra and Sol
GPT-6.2 Astra and Sol cut costs sharply, but over-engineering and early-stop reports are mixed
Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
One trn2.48xlarge serves GLM-5.3-Flash through an OpenAI-compatible API using NKI megakernels.
Know When to Zoom: How Small Is Too Small for Vision-Language Models?
Splitting large photos into tiles raised Claude and Gemini accuracy by 11.1–14.5 points
Understanding Positional Embeddings in Modern LLMs
A walkthrough of how LLMs encode token position, from GPT-2 to RoPE
Show HN: DeepSeek V3 in ~200 lines of pure Jax
DeepSeek V3's core design in 190 lines of JAX, trained on a MacBook Air in 2.6 hours