FOMO Buster #033: "Anthropic's S-1: $16.3B in six months, an $80.7B loss, and a $750B compute tab"

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

Anthropic files its S-1, and the numbers are absurd in both directions

Anthropic filed for its IPO on Monday, planning to list on Nasdaq under the ticker ANT, and the filing is the rare document where both the bull case and the bear case are in the headline numbers. Revenue for the first half of 2026 was $16.3B, up from about $1.2B in the same period a year earlier, a roughly 14x jump. The net loss for the same half was $80.7B. Both numbers are real, and the gap between them is the story.

Start with the loss, because it is mostly not what it looks like. About $72B of the $80.7B is non-cash accounting charges tied to revaluing convertible notes and SAFEs as Anthropic’s own valuation climbed, not money spent. The operating picture is far smaller and moving the right way: an operating loss of $693M for the half, improved from $2.1B a year earlier, and one analysis of the filing puts the June quarter itself at its first quarterly operating profit, $974M on $11.56B in revenue. Anthropic separately told investors that third-quarter revenue hit $16.8B, up 45% from the second quarter, meaning a single quarter nearly matched the whole first half. The business underneath the accounting is scaling fast and losing less per dollar as it does.

The numbers that should worry a reader are the commitments, not the loss. The filing points to up to $750B in compute commitments in the coming years, including noncancelable cloud deals reported at $111B with Google, $110B with AWS, and $31B with Microsoft, much of it payable regardless of usage. It also flags that Anthropic struggles to secure enough compute from those three hyperscalers precisely because they increasingly treat it as a competitor rather than a customer, and that about 20% of first-half revenue came from just two customers. And the governance: public buyers get the economics, but a single Class F share held by an LLC of the seven co-founders carries 50.1% of the vote, and the Long-Term Benefit Trust elects most of the board. Reports peg the expected valuation at $1.5T to $2T, which would make this one of the largest IPOs ever.

The read: this filing is the AI boom’s balance sheet in one document. Revenue compounding at a pace few companies have ever matched, a headline loss that is mostly the company’s own rising valuation flowing through the accounts, and a future that has already been rented from three landlords who are also rivals. The question for the public market is not whether Anthropic can grow. It is whether a company can outgrow $750B of promises it has already made.

Signal: Read past the $80.7B. The operating loss shrank 3x while revenue grew 14x, but the compute commitments are real, payable regardless of usage, and owed to competitors. That tension, not the loss, is the investment case. Sources: Anthropic S-1 filing (SEC) · TechCrunch: Anthropic officially files · Dealroom: draft S-1 numbers

02 · Lab Watch

OpenAI x Fanatics: Shop the Look, tested on 30,000 MLB products

OpenAI’s newest customer story, posted today, has Fanatics using OpenAI to help fans complete game-day outfits with Shop the Look. Testing across roughly 30,000 MLB products showed 10% higher engagement. Small numbers, but a concrete one: this is what agentic shopping looks like when it works, a fan finishing an outfit rather than searching for one piece at a time.

Google Beam comes to 20 USO locations

Google’s newest post, also from today, brings Google Beam, its 3D video calling system, to 20 USO locations to connect deployed service members with their families. Beam is the rare Google hardware story that is about presence rather than productivity, and the USO is a telling first deployment: the value of a call that feels like a room is highest when the room is thousands of miles away.

Anthropic’s own news feed posted nothing new; the S-1 arrived via the SEC filing and Hacker News, not the company’s blog. The rest of the five lab feeds are unchanged since yesterday’s run: OpenAI’s Decisions API and storage-scaling part two, the frontier-training security post, Meta’s Muse Wonder models, Mistral’s Voxtral 3, Anthropic’s independent-evaluators update, and Google’s magic-state quantum post are all covered in earlier issues. The full feed lists are in the appendix, and the moment a lab posts something new it becomes a card here on its own merits, trending or not.

03 · Open Weights

DeepSeek V4.1 Flash still leads trending, and the packaging wave rolls on

DeepSeek-V4.1-Flash-1017 still tops Hugging Face trending (score 1,092, 95,889 downloads, 3,118 likes), and the ISTA-DASLab GSQ-RCO repack from yesterday’s card is still trending too (891,147 downloads and climbing). The new climbers behind it: RecursiveSI’s flint (860), Nanbeige4.3-33B (537), XHToken’s Spark-X2.6-4B (531), and black-forest-labs’ flux-3-image-klein-comfy (466, with 700,652 downloads). Lightricks’ LTX-2.5 (409 trending, 9,401 likes) and MiniMax-M3.1-Flash (362) round out the top of the board.

Trending vs. gravity wells: Qwen3.8-27B, the most-liked model on the platform at 18,436 likes, is still trending at #8, holding its ground. The cooling wells are the older flagships: FLUX.1-dev (15,761 likes, #2 all time) and DeepSeek-R1 (14,410 likes, #3) are both absent from the trending top 30. The pattern from the last two weeks holds: the runnable build wins distribution. A base model trends once; the ComfyUI repack, the GGUF quant, and the single-file package trend for weeks.

04 · Money Flow

GitHub ditched its LLM for a small classifier, and cut secret-scan costs by 98%

GitHub’s October 13 engineering post, surfaced in today’s feeds, describes replacing the LLM pipeline behind generic secret detection with a two-stage system: regex filters cast a wide net for credential-like candidates, then a fine-tuned ModernBERT classifier reads each candidate together with its file path, secret type, variable names, and surrounding code, and decides whether it is a real secret or a lookalike. The reported results: more than 98% lower serving cost than the LLM pipeline, 86% faster median post-push scan time, 13% higher precision and 19% higher recall offline, and an 11.5% improvement in owner-weighted alert acceptance in a customer A/B test. The speed gain also let GitHub extend generic secret detection to push protection, blocking credentials before they enter a repository instead of only alerting after the commit.

The mechanism is the lesson. The hard cases in secret detection are hard negatives: a password-shaped value in a test fixture, a localhost connection string in a setup guide, a checksum that looks like a token. More regex rules can suppress some of them, but each exception risks hiding a real secret. A small classifier trained on exactly this distinction, with context as its input, is the right tool, and it costs roughly 2% of the LLM it replaced. This is the second datapoint in a week for the same thesis: for a narrow classification job, a small purpose-built model does not just match the generalist, it embarrasses it on cost. The LLM-everywhere bill is getting audited, one pipeline at a time.

Signal: When the job is classification, not generation, the small model wins on cost by 50x. Audit every LLM call in your stack that is really just sorting things into buckets. Sources: GitHub engineering post · Help Net Security coverage

05 · Leaderboard

OpenRouter: Lemonade keeps sliding, Command Code keeps leading

By tokens, Command Code still leads the board by a wide margin at 3.6T tokens and 43M requests, followed by Hermes Agent at 1.66T, pi at 1.41T, and Claude Code at 1.11T. Kilo Code holds the middle at 571B tokens. The story is still Lemonade: the novelty entry that spiked to #3 two days ago fell to #22 yesterday and now sits at #24 with 57B tokens. One-day spikes in this ranking have now reversed three times in two weeks, which is the pattern to remember: a single day’s novelty entry is noise until it holds for three. The rest of the board is stable, which is itself a signal. The agent-harness market has a top four, and the churn is all below it.

06 · The Argument

Two bugs in softmax and cross-entropy, hiding in every training run

A new paper from Jonathan T. Barron, Dan Gnanapragasam, and Michael Tschannen, submitted October 22, finds two numerical bugs in the standard implementations of softmax, log-softmax, and cross-entropy, the functions at the bottom of nearly every training run. The first, Softmax Collapse, rounds the target logit’s gradient to zero when that one logit is much larger than the others, leaving gradients that no longer sum to zero. The second, Softmax Cancellation, fires when student logits are close to teacher logits, and produces non-zero gradients even when the student and teacher logits are exactly equal. The authors present simple fixed implementations of all three functions.

The fixes do more than tidy up arithmetic. They eliminate the previously observed “Slingshot” phenomenon, speed up grokking with weight decay, prevent accuracy collapse without weight decay, and stop student networks from drifting away from the teacher during distillation. And they produce a clean theoretical payoff: after memorization, logit growth is bounded by an equilibrium between weight decay and Adam’s epsilon hyperparameter. The argument here is about leverage. A bug in a loss function is not one team’s bug. It is a tax on every model trained with that function, paid in wasted compute and unexplained training phenomena, and the fix is a few lines of code. The most important papers are often not the ones that add a capability. They are the ones that remove a tax everyone had stopped noticing.

Signal: If your training runs show unexplained drift, collapse, or slingshot behavior, check the loss function before the architecture. The bug may be underneath everything you have already tuned. Sources: The paper (arXiv)

07 · Papers

• SceneBloom: Agentic Indoor Environment Generation with Realistic and Diverse Layouts for Embodied Learning: language-guided indoor worlds for training embodied agents, with object-level reasoning and simulation-ready layouts.
• RobustRLlib: A Unified Library and Benchmark for Robust Reinforcement Learning Algorithms: one library and benchmark for RL policies that survive distribution shift.
• DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection: collecting dexterous manipulation demos without a full robot setup.
• Flownomaly: Single-Step Conditional-Mean Recovery for Few-Shot Anomaly Detection: industrial defect detection from a handful of clean images, without stored exemplars.
• Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning: identical weights and BF16 tensors can still diverge when inference and training round in different places, and aligning them stabilizes RL. Pairs with today’s Argument card.
• Monetizing Intelligence: Advertising and Subscription Design on LLM Platforms: how much capability ad revenue should fund, and what users should pay to avoid ads. Timely, with ChatGPT ads expanding to 26 new countries last week.
• Green Code or Costly Code? Evaluating the Energy Costs of Iterative LLM-Based Code Optimizations: faster code is not always cheaper code, once energy is counted.
• Closing the Modality Gap via Simplex-Constrained Representations: a fix for the persistent gap between CLIP-style embeddings across modalities.
• If at First You Don’t Succeed, Backtrack: Lightweight Recovery for Frozen VLAs: recovering a vision-language-action model once its own actions carry it off its training distribution.
• SemantEq: Faithful Bidirectional Equivalence of Formal Statements: as provers get stronger, the burden shifts to checking that the formal statement actually says what the mathematician meant.

08 · GitHub Trending

Skills are the week’s theme on the trending board. Nutlope/hallmark leads daily (+2,235 today, 35,435 total): the anti-AI-slop design skill for Claude Code, Cursor, and Codex. debpalash/VoiceStudio (+2,292, 88,254) is the fully-local ElevenLabs alternative: voice cloning, dubbing, dictation, and audiobooks in 646 languages. pbakaus/impeccable (+1,551, 88,654) is the design language that makes AI harnesses better at design. TauricResearch/TradingAgents (+1,376, 115,501) is the multi-agent LLM trading framework. makecindy/cindy (+1,268, 5,908) is the open-source agent that works out of the box. JetBrains/thinkrail (+1,156, 2,440) is JetBrains’ vibe-coding IDE built around pi. NVIDIA/SkillSpector (+203, 21,826) scans agent skills for vulnerabilities, prompt injection, and data exfiltration before you install them, the security-week thread in repo form. The rest of the board:

• pollen-robotics/microduck (+844, 12,203): a tiny biped duck robot
• affaan-m/ECC (+853, 289,387): the agent harness performance optimization system
• Panniantong/Agent-Reach (+723, 103,081): give your agent eyes to see the whole internet, one CLI, zero API fees
• nightly-labs/openbot (+86, 1,190): a local-first desktop workspace for persistent AI teammates

Weekly: cloudflare/security-audit-skill (+15,072 this week, 48,021) still leads the week; stablyai/orca (+9,622, 106,670), Graphify-Labs/graphify (+6,903, 135,840), anomalyco/opencode (+5,921, 221,805), Comfy-Org/ComfyUI (+4,790, 143,163).

09 · Product Hunt

Today’s Product Hunt leaderboard fetch failed: the scout run reported the feed as stale, and the data on hand is October 24’s leaderboard, which already ran in full in yesterday’s issue. Rather than re-print the same list as if it were new, this section is a miss, and yesterday’s top 10 stands: zCAD AI (294 votes), myPip (247), 1Password for Agents (226), AgentValley (152), Canva Craft (129), LinkIntel (119), Many Displays (101), Penta (99), Byter (95), and Basedash Scheduled CSV Exports (84). See yesterday’s issue for the full list.

10 · Hacker News

• Anthropic S-1 (553 points, 484 comments): the filing behind The One Thing, and the thread is doing the math on the $750B compute tab.
• Dolly Parton’s Imagination Library (607 points, 102 comments): the book-gifting program, and a thread about what scales in the nonprofit world.
• Switzerland. Block by block. At 1:1 scale (333 points, 150 comments): the whole country, rendered block by block.
• Show HN: TitleArt, make text like it’s 1997 (324 points, 61 comments): retro text art, done properly.
• Why I Am a Programmer (282 points, 87 comments): a personal answer to a question the front page keeps asking.
• The State of Context Engineering (229 points, 58 comments): where the context-engineering conversation stands, from a practitioner who has been in it all year.
• Show HN: A Truly Modular Keyboard (223 points, 71 comments): the keyboard as a set of swappable modules.
• PaperPilot, zero-cloud, pure-Rust PDF engine with WASM and MCP (197 points, 73 comments): PDF tooling that never phones home.
• Websites Should Be Printable (127 points, 110 comments): the case for pages that survive the printer, and the thread has opinions.
• The new iPad mini (68 points, 36 comments) and the store page (89 points, 96 comments): Apple’s biggest iPad mini update, with an OLED display in the mix.
• Fixing Tool-Call Accuracy in vLLM (55 points, 8 comments): a small fix with an outsized effect on agentic workloads.
• Brown student maps the hidden global labor powering AI (28 points, 17 comments): the data workers behind the models, mapped.

11 · OpenRouter Apps

Command Code’s lead is in requests as well as tokens: 43M requests against 15M for Hermes Agent and 9.9M for Claude Code. Hermes Agent, the open-source self-improving agent from Nous Research with persistent memory across sessions, holds second on tokens. pi, the minimalist coding agent that starred in yesterday’s MemoryTax card, is third. Kilo Code, the open-source agent that works across VS Code, JetBrains, and CLI, holds the middle. DeepSeek Harness, Freebuff (coding models funded by ads), OpenClaw (the messaging-app agent), and Cline round out the token board. The app to watch is not a mover but a model: Freebuff’s ad-funded coding models are the subscription-optional shape of this market, and 222B tokens says the shape has demand.

12 · Watch

• Dwarkesh Patel: “What will computer chips look like in 2045?” (Oct 22): the chip roadmap beyond the transistor era, from the channel’s technology series. The freshest on-topic video in the queue that clears the bar.
• TheAIGRID’s unseen uploads stay skipped for clickbait framing, and Dwarkesh Patel’s history series is excellent and the wrong newsletter.

13 · Listen

No new episodes in the three tracked feeds since yesterday’s issue: Lenny’s Podcast, Latent Space, and Hard Fork all published nothing fresh in the last 24 hours. Yesterday’s pick still stands as the newest listen: Lenny’s Podcast, “An 86-minute guide to finding PMF” with Rob Snyder (Oct 25).

14 · Read

• Anthropic’s S-1 filing (SEC): the document behind The One Thing, all 200-plus pages of it.
• TechCrunch: Anthropic officially files for its public debut: the cleanest summary of the filing’s numbers and governance.
• A Fix for Softmax and Cross-Entropy (arXiv): the paper behind The Argument, short and worth reading in full.
• The State of Context Engineering: the HN thread’s source post, and a good map of the field.
• OpenAI: Securing our frontier model training environment: the DNS-escape incident, the training pause, and the no-internet policy, in OpenAI’s own words.

15 · Safe to Ignore

• The five lab feeds, otherwise unchanged: Decisions API, storage-scaling part two, the frontier-training security post, Muse Wonder, Voxtral 3, the Anthropic evaluators update, and magic-state quantum are all covered in earlier issues; today’s genuinely new lab posts are the Fanatics and Beam items in Lab Watch.
• Kimi K3.1 Pro in today’s AlphaSignal newsletter: covered October 24. Weights are still due by November 4, and there is nothing new since.
• Nemotron 3.5 Lightning resurfacing in the RSS feed: an August 11 release, not news, and the “1M context on one GPU” line needs its asterisk (NVIDIA’s own single-H100 guidance is 256K in BF16).
• GPU unlock: covered yesterday, both as an HN card and in GitHub Trending.
• Today’s newsletter sponsored blocks: the Unblocked event and the AWS webinar are partner promos, not editorial signal.
• The rest of the AlphaSignal RSS tail: RealVisXL V5.0, CyberTiel, the MiniMax H3 repack, Pomerado, Sarvam’s Knowledge Fabric, Hy-SimulMT, Pi Pocket, OpenMuse, and the Lyra C neuron explorer all lack a verifiable primary announcement in today’s fetch; skipped until a source exists.
• TheAIGRID’s clickbait framing: unseen and staying that way.
• Dwarkesh Patel’s history series: excellent, wrong newsletter.

16 · Build on This

Pre-push secret scanning for the rest of us

GitHub just proved the pattern: a small classifier, trained on one narrow job, beats a general LLM at that job and costs about 2% as much to run. What GitHub did not do is give that win to everyone. Its version sits behind Secret Protection licensing, and the developers most likely to leak a key are solo builders and small teams shipping fast with AI help, the people with no security team and no enterprise plan.

Build the scanner they can actually buy. It runs locally, before push, not after: regex candidates plus a small classifier that learns from the repo it guards, so the test fixture that looks like a password stops flagging after one dismissal and the real key never does. It covers the formats that actually leak in AI-written code: API keys pasted into examples, tokens in .env files that were never gitignored, credentials in notebook outputs. It ships as a one-time purchase with a per-repo model that improves with use, not a per-seat subscription. The moat is not the classifier, which anyone can train. It is the false-positive memory: every dismissal makes the next scan quieter, and a quiet scanner is one developers leave turned on. Start with the ten secret formats that account for most real leaks, add the next ten from user reports, and let the dismissal data become the product.