I'm not convinced the AI bubble exists
TL;DR Highlight
A blogger cites Anthropic's S-1 to argue there's no AI bubble; HN pushes back
Who Should Read
A developer weighing LLM API costs and vendor lock-in risk when choosing which model to build on
Core Mechanics
- Data center spend is one-off, author says
- The author argues data center build-out is a one-time cost, so spending will drop once it's done.
- He cites at least $3T committed plus an estimated $1.9T already spent, about $5T total.
- Anthropic earns $2.08 per $1 of compute
- The author cites Anthropic's S-1 for $2.08 earned per $1 of compute.
- His point is that AI is very profitable once infrastructure build costs are excluded.
- Revenue grew fast
- Anthropic revenue went from $41M in 2023 to $4.585B in 2025.
- First half of 2026 revenue is given as $16.3B.
- P/E in the 20s, unlike dot-com
- The author notes Alphabet, Meta and NVIDIA have stable P/E ratios in the 20s.
- He contrasts this with dot-com firms that were unprofitable or had P/E in the hundreds.
- No moat, still profitable
- He argues a company can profit without a moat, like the chicken sandwich war.
- He says brand loyalty and enterprise lock-in partly protect against churn.
- The author admits real risks
- He lists data center backlash, regulation, an autonomous AI hacker incident and complex debt structures.
- He concedes it gets dangerous if revenue growth suddenly slows.
Community Reactions
- GPU depreciation means it's not one-off
- Commenters said GPUs last 3-5 years, so data center spend won't plummet.
- One noted it behaves like OPEX through depreciation.
- Revenue treated like profit
- Commenters criticized comparing the $250B/year debt payment to Alphabet's revenue rather than earnings.
- They said operating costs make repayment capacity much smaller.
- Switching models is easy, one user says
- One commenter said swapping models is a one-line JSON change and they did it 5 times this year.
- They noted most revenue comes from B2B, not consumers.
- Later results reportedly slowed
- A commenter said Q3 results posted after the S-1 showed slower growth and an operating loss.
- They also said the author conflates technical potential with market demand.
How to Apply
- Audit your model switching cost today
- Move model names and prompts into one config file.
- Swap in another vendor's model, run the same 10 tests and record quality differences.
- Put an adapter layer before LLM calls
- Hide each vendor SDK behind one interface.
- Make switching possible by changing an environment variable if prices rise or service stops.
- Build a token cost increase scenario
- Multiply your monthly token usage by price increase scenarios in a budget sheet.
- Mark features that a cheaper open model could replace.
Try It
# Paste into Claude or ChatGPT as-is (a simple check prompt based on the article's argument)
Our service is {{service description}} and currently uses {{vendor/model}} at {{monthly tokens}} tokens per month.
Summarize the impact on us in a table for these two scenarios.
1) Vendor prices double
2) Vendor shuts down service
For each scenario, write (a) monthly cost change, (b) code changes needed to switch models, (c) 3 preparations I can do today.
Next step: implement the easiest preparation from the answer today.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