Here's the evening roundup of AI news for October 27, 2026. As a follow-up to Anthropic's S-1 filing this morning, the outlook — an IPO two weeks out at a $2 trillion valuation — and figures like 1400% growth are making the rounds now that the numbers have been laid out. NVIDIA has published a technical report on Nemotron 3.5 Bolt, a VLM for local AI that runs in 24GB of memory. From Japan, we have a large-scale experiment on AI-generated spear phishing, HIS automating corporate travel quotes, and a look at where the company that banned AI in new-hire training stands a year later — plenty of material for thinking through the impact on real-world work.
[Update] Anthropic's IPO Is Two Weeks Away at a $2 Trillion Valuation — 1400% Growth, 48% of Revenue Through the Cloud
Following up on our morning coverage of Anthropic's S-1 filing, the [Latent Space AI News newsletter](https://latent.space/p/ainews-anthropic-s1/) has broken down the numbers in the document.
According to the newsletter, the IPO is two weeks away with a target valuation of $2 trillion. Growth has reached 1400%, and 48% of revenue flows through cloud partners. Margins stand at 85%, and the company recorded $1 billion in operating income in Q2 — a figure adjusted for $1 billion per quarter in stock-based compensation, making it the company's first profitable quarter ever. Cash on hand reportedly reaches $134 billion, while cumulative burn since founding has stayed at just $10 billion.
True to its billing as a 350-page "singular S-1 for the history books," the runaway growth rests on deepening ties with cloud providers. How the market prices a business structure in which nearly half of revenue comes via cloud partners after the listing will likely serve as a reference for future fundraising across the AI industry. Note that the detailed figures here are based on the newsletter's summary, so check the original SEC filing for the precise breakdown.
NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB for Local AI — 84 TPS on RTX Spark
NVIDIA has published a technical report on arXiv for Nemotron 3.5 Bolt 74B-A9B 24GB, a vision-language model optimized for agentic use on local hardware ([arXiv:2610.33039](https://arxiv.org/abs/2610.33039)).
Built on a hybrid Mamba-Transformer architecture derived from the Nemotron 3.5 Super VL 120B-A12B family, it combines Star Elastic compression with 2-bit quantization-aware distillation (QAD), a design that optimizes speed and intelligence per gigabyte even in memory-constrained deployments. It is trained for reasoning, tool use, and web and terminal operation, and NVIDIA has simultaneously released LiLiCorr, a speculative drafter for hitting inference speed targets.
Compared with existing models in the same memory and speed class, it achieves strong accuracy on GDPVal v2 (general agent capability), BrowseComp (web browsing), and Terminal-Bench 2.1 (terminal operation). On an RTX Spark it pushes out 84 TPS — speed on par with models that have far smaller active parameter counts — which is considered sufficient for long-running agentic workloads.
Also in local AI news, a [comparison of inference engines on a single RTX 5090](https://www.reddit.com/r/LocalLLaMA/comments/1xgkzb9/qwen38_27b_on_one_rtx_5090_six_inference_engines/) was posted to Reddit's r/LocalLLaMA. It runs six engines — llama.cpp, vLLM, SGLang, NInfer, velox-inferer, and TensorSharp — in an identical environment, comparing decode speed, long prompts, concurrent requests, image input, and tool loops. What stands out is the caveat the author attached themselves: because each engine uses different quantized checkpoints, drafters, and KV formats, they explicitly note that it is "not a comparison isolating the performance of the runtime code alone." Publishing a self-benchmark without hiding the differences in conditions sets a good example for anyone reading comparison posts in the local LLM space.
AI-Generated Spear Phishing Pulls Roughly 2.7x the Click Rate — Large-Scale Experiment with Over 7,000 Targets
According to [ITmedia](https://www.itmedia.co.jp/news/article/2610/27/2000002550/), researchers affiliated with TU Berlin and others have published "A Large-Scale Study of Personalized Phishing using Large Language Models," a paper validating the effectiveness of LLM-generated personalized phishing.
In the experiment, AI looked up targets' email addresses and generated a phishing email tailored to each individual. Across more than 7,000 targets, the click rate reportedly reached roughly 2.7 times that of conventional attacks.
Until now, targeted attacks were held back by the sheer labor of research and writing. If AI can collapse that effort in one stroke, defenders will have no choice but to plan for polished, individually addressed emails — not just training against spray-and-pray campaigns. With the labor cost of attacks falling and detection turning to AI as well, an "AI vs. AI" contest is starting to feel very real.
HIS Automates Corporate Travel Quoting with AI Agents — 100 Hours Saved Per Month
An [ITmedia article](https://www.itmedia.co.jp/enterprise/articles/2610/26/news033.html) covers HIS's automation of its corporate travel quoting operation.
According to the piece, HIS brought AI agents and robots into quoting work, cutting 100 hours of labor per month. The company had been experimenting with AI-driven efficiency since around 2023 but had not reached the accuracy it was aiming for, and the article digs into the initiative that turned things around.
Quoting looks like a simple task on the surface, but once combinations of flights, hotels, and booking conditions enter the picture, it leans heavily on human experience. It is hardly rare for generative AI deployments to miss the accuracy hoped for in the early going, and as a case of reaching production quality after repeated iterations, this one offers a useful reference for companies now considering weaving AI into their operations.
One Year After Banning AI in New-Hire Training — How Did いえらぶGROUP's Policy Turn Out?
Another [ITmedia article](https://www.itmedia.co.jp/news/article/2610/27/2000002585/) catches up with いえらぶGROUP, which adopted an AI ban aimed at new-hire training.
Roughly a year ago, the policy — that junior staff should not use AI until they have built solid fundamentals — sparked debate over whether it was out-of-touch gatekeeping. A year in, the article checks how field proficiency and the policy's day-to-day operation have changed.
Whether to guard against juniors' over-reliance on AI or to prioritize mastering the tools — this is a question facing many Japanese companies. Read alongside this morning's piece on Itochu Interactive's training shift from "using AI" to "delegating work to it," comparing where these two companies, which took diametrically opposite approaches, stand after a year offers food for thought on your own policies.
Can Ponytail Rein in Claude Code's Over-Writing? Compared on Sonnet 5.5 and Opus 5.5, Plus Applying /usage Fix Suggestions with the F Key
Qiita has a pair of Claude Code operations posts this time.
First up: [Can Ponytail Curb Claude Code's Over-Writing?](https://qiita.com/morphox_ai/items/5856dceca60f8a7c7270). The author put Ponytail, a popular skill for curbing the tendency to write wasteful amounts of code during vibe coding, through its paces on both Sonnet 5.5 and Opus 5.5. Ponytail's GitHub repo has topped 177,000 stars as of writing, and since it changed substantially not long after the author's June beginner's guide, this verification against the current setup is well worth having.
Second: [Auto-Applying the Fix Suggestions in /usage with the F Key](https://qiita.com/leomarokun/items/f9ce01e535a31588738b). As of Claude Code v2.1.305, the usage fix suggestions (Recommendations) shown by the /usage command have become a single line naming the cause, and can now be applied by pressing the f key. The post walks through how to read the screen on a Pro plan environment (v2.1.310).
With so many Claude Code users bumping into plan limits, model selection and waste reduction translate directly into cost. That both the skill side (Ponytail) and the core side (/usage) are moving toward controlling over-writing and overuse may be a sign of where the focus of development cost management lies in the agent era.