Key points
- Meta — 2026, Meta acknowledged in the 2026 sustainability report it published on November 19 that the company is reconsidering the net-zero emissions target it had set for 2030.
- Anthropic — According to CNBC, Anthropic's IPO raised $93.5 billion after underwriters exercised the overallotment option in full.
- Google — The report's assessment: Google's flagship search is turning into an AI experience at remarkable speed, and even long-standing interface elements are being AI-ified — the homepage has gone from a search box to an AI conversation entry point.
Here is your Daily AI Stew, served up by Robin.
Meta Says It Is Reassessing Its 2030 Net-Zero Goal as AI Data Centers Push Emissions Up
Start with what the company says about itself: the power demand of AI data centers has now run into its environmental commitments.
According to an Ars Technica report dated November 20, 2026, Meta acknowledged in the 2026 sustainability report it published on November 19 that the company is reconsidering the net-zero emissions target it had set for 2030. The same report shows Meta’s greenhouse gas emissions rising by a record percentage — an increase larger than at any other major US company racing to build AI data centers.
The passage comes from the report’s introduction and is signed by Rachel Peterson, vice president of data centers, and Blair Swedeen, head of global net zero and sustainability. Their words: “Our long-term goal has not changed, but the operating environment has.” They attribute that shift to “demand for AI infrastructure growing faster than almost anyone predicted.”
The report states that the company is “reassessing timelines, priorities, and interim milestones — including the 2030 net-zero goal and near-term intensity targets,” on the grounds that the plan should reflect current operating reality while progress is published every year and transparency is maintained.
In other words, 2030 has moved from a committed timeline to a timeline under review; the phrase “long-term goal” stays in place.
The report gives no absolute emissions figure and no itemized comparison with peers — only the relative judgment that the increase outpaces the industry.
Turning 2030 from a committed date into an item under review puts decarbonization milestones and the pace of data-center construction on the same ledger. What to watch next: whether Meta offers new interim milestones, and whether its emissions accounting basis shifts to match.
The AI Selloff Closes Out: Nasdaq 100 Posts Its Worst Week Since 2020, S&P 500 Enters Correction
Same AI selloff, and as the week closed the magnitude changed once again.
According to market coverage aggregated by Google News, the selloff driven by AI concerns handed the Nasdaq 100 its worst single week since 2020; the S&P 500 also fell at least 10% from its recent high, putting it in correction territory by definition.
In other words, “correction” is a mechanical threshold: a 10% drop from the most recent high puts an index into a correction phase, regardless of how large any single day’s loss was.
Before the selloff, AI-related stocks had been under pressure for weeks, the report says; Anthropic’s listing did not dispel doubts about how long the AI boom can last. The S&P 500 fell for seven consecutive sessions, its longest losing streak since the COVID-19 outbreak in 2020. The last comparable bout of market turmoil came in early April 2025, when Trump’s tariff plan jolted global markets.
The selling landed on AI-related tech stocks, especially cloud companies with speculative valuations. Per the report, the price war between Anthropic and OpenAI deepened investor skepticism about the entire AI trade.
Yesterday’s item was indexes entering correction territory; today’s addition is the weekly record — a single day’s drop measures sentiment, while a full week’s drawdown measures whether capital still wants to price compute spending at all.
Underwriters Exercise the Greenshoe in Full: Anthropic’s IPO Raises $93.5 Billion
Same listing, a thicker set of books.
According to CNBC, Anthropic’s IPO raised $93.5 billion after underwriters exercised the overallotment option in full. That option, better known as the “greenshoe,” lets underwriters buy additional new shares at the offering price when demand is strong.
The numbers: the IPO initially raised $81.3 billion at $83 per share; underwriters then used the greenshoe entirely, buying 147 million more shares at the same price and pushing the total to $93.5 billion — $12.2 billion more than originally planned.
There is a split to note as well: about $91.6 billion goes to Anthropic itself, with the rest going to selling shareholders.
So the point here is not a valuation multiple but that demand consumed the greenshoe entirely — underwriters were willing to take the added shares at the offering price. The report also notes that the pricing and valuation basis of the offering had already drawn discussion.
The previous issue covered the size of the debut; today’s addition is the extra shares underwriters bought at the offering price — it shows the price was sellable, not that it holds up once trading begins.
Reuters: The SEC Is Investigating Nvidia’s Accounting
This one has a direction, but no scope yet.
According to a Reuters report dated November 20, 2026, the US Securities and Exchange Commission is investigating Nvidia’s accounting. Citing two people familiar with the matter, the report says the focus of the probe is precisely Nvidia’s accounting.
Before that, the Wall Street Journal said it could not determine the scope of the regulator’s investigation. None of the available reporting gives a start date, a specific scope, any conclusion, or possible consequences.
What can be confirmed is the investigation itself: as of now, no findings, penalties, or remedial measures have been disclosed. For Nvidia and its shareholders, that adds another layer of regulatory uncertainty.
One direction, no scope, no conclusion — and the market is already pricing the company’s accounting basis. What to watch next is not rumor but whether disclosure changes in its next earnings report or filing.
Researchers Planted 2 Million Decoys Across the AI Supply Chain and Found Sensitive Data Leaking Widely
The requests from your AI app may not go straight to the model vendor; there may be a forwarding layer in between that you know nothing about.
According to an Ars Technica report dated November 20, 2026, security researchers seeded 2 million digital decoys into a sprawling network of AI services to trace where entrusted sensitive information actually ends up. What they saw: the chain leaks like a sieve.
First, the chain. Many AI applications today do not connect directly to Anthropic or OpenAI; they send requests to “AI routers” such as OpenRouter and Vercel, also called AI gateways. The router takes in prompts from a large number of end users and forwards each request to whichever inference engine is available, picked by lowest price or best fit for the task. Developers use it for convenience: no need to source models themselves.
The experiment: researcher HD Moore prepared a batch of never-used credentials — access keys tied to empty AWS accounts, SSH passwords, and unique URLs, DNS names, and private git repository addresses. He placed each credential into exactly one prompt and sent it to a single provider. Every server that received these credentials was under his own control, so any use of a given credential could be traced to a specific link in the chain.
The decoys are called “canaries,” after the birds once carried into mines: once data leaks, researchers can follow it back to the link that failed. The report groups the risk into three categories: security breaches, prompt theft, and privacy violations.
Put another way, the leak may not happen at the company you talk to directly; it may happen at an intermediary service you did not even know existed.
When procuring services like these, ask one more question: does the request go straight to the model vendor, or through a third-party router? That layer appears neither in the AI company’s announcements nor in your data processing agreement, yet your keys and prompts all pass through it.
Google’s Search Homepage Redesign: The Search Button Is Gone, the AI Entry Point Moves Up
Some interface elements survive more than twenty years, then vanish just like that.
According to an Ars Technica report dated November 20, 2026, the Google homepage redesign tested over several weeks has now gone live: several buttons were added below the search box, all of them leading to AI features, and the dedicated search button was eliminated entirely.
Robby Stein, vice president of Google Search, confirmed the desktop search page rollout on X, saying it shipped after testing succeeded. The three new buttons are: generate images, analyze files, and brainstorm with Gemini. The “I’m Feeling Lucky” button stays; the Google Search button that used to sit beside it is gone.
Per the report, that button was already awkward in a modern interface: as soon as you start typing a query, autocomplete suggestions cover it, so you have to click outside the input field before you can hit it — one step more than just pressing Enter. Removing it therefore does not interrupt many people’s habits.
The report’s assessment: Google’s flagship search is turning into an AI experience at remarkable speed, and even long-standing interface elements are being AI-ified — the homepage has gone from a search box to an AI conversation entry point.
The position of an entry point shapes habits. That button was already hidden behind autocomplete, so almost nobody objected to losing it, but from now on the homepage steers people toward “chatting with AI.” What is worth watching next is how citations and links that appear inside those conversations get counted toward websites.
Inkling 2 Mini Launched: A Looped Transformer Open-Weight Model That Runs on a Laptop
Running models locally has one more option as of today.
According to Thinking Machines Lab’s release information, Inkling 2 Mini is an open-weight model positioned for local operation: it is built around a “looped Transformer” and was trained entirely in NVFP4 precision. A community GGUF release is already out and can be tried directly through Ollama and llama.cpp.
Two keywords, explained. Open weights means the trained parameters are published so they can be downloaded and executed locally, rather than offered only behind a vendor’s API. Looped Transformer means the model reuses several middle layers multiple times, doing more computation per token without a proportional increase in parameter count.
On scale: it is a 26.4-billion-parameter model, designed to run on a laptop with roughly 20GB of memory, with text and image input and a 1-million-token context window. Per the release notes, it runs its middle layers three times per token, trading fewer generated tokens for reasoning capability, aimed mainly at workloads such as agentic coding. Deployment has prerequisites: you need an Unsloth or llama.cpp build that supports looped layers and per-expert NVFP4 scaling.
The release also offers a caution: looping improves reasoning, but it does not make the weights store more facts; its AA-Omniscience score is comparable to models of the same size. In other words, what the extra computation buys is deeper thinking, not more recall.
In discussion, some compare it with Qwen3.8-27B; those benchmark advantages still need verification in real settings, and whether they transfer to actual workloads is the key question. Others predict more open models will adopt similar architectures, pointing to SensaNova Loop-9B as a reference. Where models can be deployed locally, safety and interpretability concerns are raised as well.
Open weights plus local execution means cost shifts from per-call payment to a one-time hardware investment. The real dividing line is not leaderboard rank, but whether you can reproduce those comparisons on your own workload.
Another “Generalized Riemann Hypothesis” Proof: A 940-Page Manuscript, Right After Anthropic
A second major claim appeared on the same day — this time in mathematics.
According to a personal blog announcement dated November 20, 2026, Alessandro Masullo claims to have produced a proof of the generalized Riemann hypothesis, a 940-page manuscript released in a hurry, right on the heels of an Anthropic-related announcement.
Two terms first. The generalized Riemann hypothesis extends the Riemann hypothesis from the Riemann zeta function to the family of Dirichlet L-functions, asserting that all their nontrivial zeros have real part one half. Earlier reporting said that Anthropic’s Levent Alpöge and Claude had delivered a Lean proof of the ordinary Riemann hypothesis, described as roughly 17 million lines.
But the only thing this announcement confirms is one thing: a proof has been put forward. It is a secondhand account; it establishes neither that the manuscript is correct nor that it is complete. Commenters pointed out that a section early in the manuscript appears to be empty. All of this still needs page-by-page expert review, and no review conclusion exists yet.
The community is speculating about two things: whether Masullo shared part of the work with Anthropic, and whether the proof was arrived at independently. Some also question how the manuscript was first presented and the timing of its release. One commenter noted that mathematics usually leaves verifiable timestamps and hashes for work in progress, in order to establish attribution.
For mathematicians and proof-checking teams, the conclusion right now is a single sentence: both the 940-page manuscript and the 17-million-line Lean proof are claims awaiting independent verification, not a solved generalized Riemann hypothesis.
Two claims in a single day, with verification unable to keep pace with announcement — that is itself a side effect of the AI mathematics boom. The next time a claim like this appears, ask first whether the manuscript can be read publicly, whether the formalization runs, and whether anyone independent has reviewed it.
One thread runs beneath today’s items: the bill for expanding compute is being booked somewhere else first — in emissions tables, fundraising bases, accounting bases, and that forwarding layer in between. All of those accounts are still waiting on later disclosure.
🎧 This episode is also available as a podcast: listen to Daily AI Stew · 2026-11-21.
