The Read

Google shipped the top-ranked model on the industry's reference index and watched its shares shed 9.6% after the close the same day. Gemini 4 Carbon took first place at an introductory $2 per million input tokens, and it is going to cyber defenders before it goes to developers — a frontier model gated as a dual-use asset rather than released as a product. Hours later the numbers landed, and they were mixed: Alphabet grew revenue 25% but missed on both lines and burned $9.07 billion of free cash flow, while Microsoft beat on revenue, operating income and earnings, spent about $5 billion less on capex than the street expected, and was rewarded with a 4% pop. The market read Microsoft's underspend as discipline and sold the neoclouds down four to six and a half percent on it. Read Amy Hood's actual sentence and you get the opposite story: the capacity Microsoft paid for arrived late. Capability is not the constraint and demand is not the constraint — delivery is, and that is a far better problem than the one everyone thinks they are looking at.

🌊 Tide

No shift. One confirmation on cost-collapse, and an unusually clean one: Google moved nine points up the same index in fifteen days without changing the price card by a cent.

Nine index points in fifteen days, at exactly the same price

On October 13 Google made Gemini 4 Argon broadly available at an introductory $2 input / $10 output per million tokens through November 30, going to $4/$20 on December 1, with cached input discounted 95%. Fifteen days later Gemini 4 Carbon arrived on an identical price card — $2/$10 introductory, $4/$20 after, 95% off cached input. Scored on the same yardstick, version 5.0 of the Artificial Analysis Intelligence Index, which Artificial Analysis announced on October 27, Argon sits at 50 and Carbon at 59, first place, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Nine points on one scale, from one vendor, inside a fortnight, with nothing moving in the price column. Most confirmations of this tide have been a dollar figure coming down; this one is the same dollar figure buying materially more, which is the version of the curve that is easy to miss precisely because the price looks unchanged. Carry the caveat lightly: Nathan Lambert spent the same day arguing that Artificial Analysis earns its money verifying results for the labs it ranks, and that any single index is a worldview with a number attached. Treat 59 as one reading rather than the reading — and note that comparing a model's old score to a new model's score across a rebased index is exactly the error the rebase invites. The price card is not a reading. It is published.

Sources: Google — Gemini 4 Carbon: our next frontier model · VentureBeat — Google makes Gemini 4 Argon broadly available with limited-time low API pricing of $2/$10 per 1M token · Artificial Analysis — Announcing Artificial Analysis Intelligence Index v5 · Interconnects — Artificial Analysis v5 and the benchmarker's dilemma

🌊 Waves

Two of the biggest AI spenders reported within hours of each other, and only one of them got credit for the spending

Alphabet grew revenue 25% to $127.6 billion, grew Google Cloud 103% to $30.71 billion, grew Search and other ads 12% to $63.35 billion, said year-over-year query growth was the highest in five years, and disclosed that its first-party model APIs now process nearly 32 billion tokens a minute — up 40% on the quarter and more than 4x on the year. It also missed on both lines: revenue came in against a $127.9 billion consensus and EPS at $2.12 against a $3.01 estimate, net income fell 25% to $26.17 billion from $34.98 billion, and free cash flow was negative $9.07 billion, far short of the $2.7 billion analysts expected. Full-year capex guidance stayed at $195–205 billion. Shares fell 9.6% in after-hours trading. Microsoft, reporting the same evening, beat on revenue, operating income and earnings: $92.28 billion of revenue (up 19%), GAAP net income of $36.63 billion, non-GAAP EPS of $4.89 against a $4.73 consensus, and Azure revenue disclosed in dollars for the first time — $33.17 billion, growing 48%, with next-quarter guidance of 49–50% at constant currency. Its capex including finance leases came in at $45.2 billion, up about 30% year over year but roughly $5.5 billion under what the street had modelled. The stock rose about 4%.

So what: Roadmap implication: both companies are spending at historic scale and only one of them got credit for it, and the difference was not ambition — Alphabet missed revenue, earnings and cash flow at once while Microsoft beat all three. The lesson is narrower and more useful than "the market turned on AI capex": capex now gets graded inside the same quarter it is spent, against the revenue and cash it produced that quarter. If your 2027 plan assumes investors will fund the gap between spend and revenue on narrative, reprice that assumption now and build the quarterly evidence trail that closes it.

Sources: Google — Q3 2026 earnings call: Remarks from our CEO · 9to5Google — Alphabet reports Q3 2026 revenue of $127.6 billion · SiliconANGLE — Microsoft's stock rises on solid earnings and revenue beat, with Azure growing 48% · Microsoft — Microsoft Announces First Quarter Results · SiliconANGLE — Alphabet's stock sinks after shocking investors with top-and-bottom line miss

The neoclouds sold off on Microsoft's capex miss, and Microsoft's own CFO said the miss was third-party capacity arriving late

CoreWeave fell 5.8%, Nebius 6.3%, IREN 5.3% and Applied Digital 4.3% in overnight trading after Microsoft's $45.2 billion capex number landed under the roughly $50.67 billion estimate. The inference the tape made was that the largest buyer of AI capacity is pulling back and its suppliers are the first casualty. That is not what Microsoft said. CFO Amy Hood's explanation was that "some third-party capacity deliveries we expected in Q1 were delayed" — the money was committed, the racks did not arrive on schedule — and she guided next-quarter capex above $50 billion with calendar-year spending around $170 billion, explicitly because of the slippage. So the group that got sold is the group whose late deliveries created the shortfall, and the shortfall is a timing difference, not a demand signal.

So what: Roadmap implication: for anyone buying compute, the scarce thing in 2027 is a delivery date you can hold someone to, not a dollar commitment. Write capacity contracts with dated milestones and real remedies rather than annual totals, and treat a provider's on-time record as the diligence item it has become. For anyone selling into this market, the opening is operational: the differentiator has quietly moved from price per GPU-hour to the ability to energise and hand over racks when promised.

Sources: Stocktwits — CRWV, NBIS, IREN, APLD Stocks Drop After Microsoft Capex Falls Short On Delayed Third-Party Capacity · SiliconANGLE — Microsoft's stock rises on solid earnings and revenue beat, with Azure growing 48%

Google gave its strongest model to cyber defenders before it gave it to customers

Gemini 4 Carbon is being released first to a subset of cyber defenders through Google's Fairwind Program, with more trusted partners added gradually and availability for developers, businesses and consumers promised only "as soon as possible" — no date. Google also says it took part in the US government's voluntary pre-release testing process. The security numbers are why: 88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym, both claimed state-of-the-art, 71.0% on CWE-bench for patching, and a reported 0.0% attack success rate on the Gray Swan indirect prompt-injection benchmark against 0.7% for Gemini 4 Argon. Google says high-severity misalignment alerts in internal live coding environments dropped by almost 80%, that unsupported claims were cut roughly in half, and that the Chrome team found it more than 10x better than another lab's large frontier model at generating high-quality patches. Cisco, the same day, open-sourced Project Quarry, a behaviour-detection engine for spotting AI agents conducting attacks. The pattern to notice is the release order: capability that cuts both ways now ships to the defensive side of the market first and to the commercial side second.

So what: Roadmap implication: assume a staged, defender-first rollout is the new default for frontier launches, which means your access to the best model is no longer a function of your contract size but of what you do with it. Two moves follow. If security is anywhere in your product, get into these early-access programmes now — the queue is forming and it is not priced. And stop planning on general availability dates for frontier capability; plan on a capability band and a lag.

Sources: Google — Gemini 4 Carbon: our next frontier model · The Register — Cisco open sources tool for detecting AI agents on hacking sprees

Agents now open more than 40% of GitHub's pull requests, and GitHub is adding four million cores to survive them

GitHub's Octoverse report put hard numbers on something everyone has been estimating: 17.9 billion pushes in the twelve months to September, more than 2024 and 2025 combined; 34.3 billion commits, more than triple two years ago; 1.3 billion merged pull requests. More than 40% of pull requests were agent-started as of September, up from 17% in March — about 110 million in September alone, roughly six times March's volume. The quality signal is better than the volume one: 90% of agent-involved pull requests were merged within two weeks, against about 86% for those without detected agent involvement. Repositories created this year that carry an agent configuration file averaged eight times as many commits as those without one. On the same day The Register reported the bill: 7.4 billion commits in September alone, five times a year earlier, outages in February, April and August, 90-day availability reconstructed at about 84% on May 7, and a roadmap that now includes four million additional CPU cores, 160 petabytes of high-speed storage, and a re-architecture separating compute from storage that GitHub expects to deliver 35x more throughput, completing its multi-region Azure migration early next year.

So what: Roadmap implication: the agent is no longer an assistant to your development process, it is the median path into your repository, and your bottleneck has moved downstream to review, CI capacity and merge policy. Budget CI compute as a first-class line item for 2027, and put a named owner on agent-authored change review the way you would on a release process. The eight-times-commits finding on repos with an agent config file is the cheapest experiment on this list — it says the returns are sitting in configuration most teams have not written yet.

Sources: The GitHub Blog — Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents · The Register — GitHub powering up four million extra cores to stop its service going down

🌊 Ripples

The Information reports DeepSeek and Moonshot leaned heavily on GPT-6 Astra outputs to train their latest models

According to people with direct knowledge cited by The Information, DeepSeek and Moonshot AI both relied heavily on outputs from OpenAI's GPT-6 Astra when training their newest models, which were released this month. Astra has seen strong demand since its early-September release, and one of the growing customer bases for it is Chinese labs extracting its capabilities. Distillation, in The Information's words, "continues to be one of the biggest contributors to the progress made by Chinese open-weight models" — which is the context for Anthropic saying on October 12 that its new safeguards cut the data harvested in distillation attacks by roughly 90% to 99%, with extracted reasoning tokens down ten- to a hundredfold.

So what: Do this now: if you have been treating the open-weight catch-up curve as independent evidence that frontier training is getting easier, separate the two. The curve is real and useful to you, but part of its slope is borrowed from the closed frontier, which means it is a derivative of someone else's capex and is defensible against. Check your own API terms and output-logging posture while you are in there.

Sources: The Information — OpenAI's Astra Model Becomes Top Distillation Target for China's AI Labs · Reuters — DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train latest models, The Information reports · Reuters via Yahoo Finance — Anthropic says new safeguards sharply reduce AI distillation attacks

New York's chatbot law for minors can order the deletion of models, not just data

Governor Hochul signed New York's AI companion-chatbot child-safety law. For users under 18, or users a company has not confirmed to be adults, a chatbot may not claim to be a real person with human emotions, claim a personal, romantic, professional or authority-figure relationship, ask unprompted emotional questions, reuse a child's private health or personal information from earlier conversations, encourage secrecy or isolation from parents and professionals, promote self-harm or substance use, or produce sexually explicit content. Companies must verify age, must offer at least one verification method that does not rely entirely on government ID, and must delete the age data afterwards. The attorney general enforces it, with civil penalties up to $25,000 per violation — and, notably, remedies that include deleting unlawfully collected data and the models trained on it. It takes effect six months after signing.

So what: Do this now: if you operate any consumer-facing conversational product, the model-deletion remedy is the clause to take to your general counsel this week, because it changes the downside of a sloppy data pipeline from a fine into a retraining event. Build the age-verification path with a non-ID option now — six months is one planning cycle, and New York will not be the last state to copy this text.

Sources: Engadget — New York governor signs AI chatbot child safety law

China's memory champion is shipping HBM3E samples at a 20–30% yield and going to production anyway

Reuters, cited by TrendForce, reports that CXMT has delivered HBM3E samples to several domestic AI chip designers — both a standard design and a revised version aimed at improving manufacturability — and expects limited production within two months. Yields are 20–30%, against roughly 70% for SK hynix, which makes the parts more expensive to produce than imported equivalents, and in Reuters' testing the chip performed closer to SK hynix's previous-generation HBM3 than to its HBM3E. It is going ahead regardless, because domestic high-bandwidth memory is the binding constraint on China's AI accelerators. Products could reach the market as early as 2027.

So what: Do this now: read this as the memory side of the same substitution story as domestic accelerators — worse economics accepted in exchange for supply security. If your supply plan assumes Chinese AI capacity stays gated on imported HBM, put a date on when that stops being true and stress-test the 2027 case. The near-term read for everyone else is simpler: HBM stays tight, so price it that way.

Sources: TrendForce — [News] CXMT Reportedly Ships HBM3E Samples, Eyes Production Within 2 Months Despite 20–30% Yields

A paper arguing that most agent memory systems are worse than plain retrieval topped the research radar

"Memory Tax: Dissecting the Net Value of Agent Memory," from a UIUC group including Jiawei Han and ChengXiang Zhai, ran a large-scale evaluation across multiple benchmarks and model backbones and found that many dedicated memory systems underperform naive RAG — a negative net value. The framing is the useful part: measure memory as benefit minus tax, where the tax is the evidence a memory layer removes or obscures when it reorganises raw history into facts and preferences. It surfaced on Hugging Face's daily papers on Wednesday with 139 upvotes. The day's most-upvoted paper, at 540, went the other way: "The Trinity," from a National University of Singapore-led group, is a recursive self-improvement framework that evolves the agent harness rather than the model and reports mean gains of 34 points across 28 training benchmarks and 26 points across 18 unseen ones, including two real-world embodied tasks.

So what: Do this now: before you buy or build an agent memory layer, run the naive-RAG baseline and make the vendor beat it on your data. That is a one-week test and this paper says it fails more often than the category admits. Note also where both of the day's top papers put the gains — in the harness around the model, not the model — which is the cheapest place you can still compete.

Sources: Hugging Face — Memory Tax: Dissecting the Net Value of Agent Memory · Hugging Face — The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents

An Anthropic engineering manager stopped running coding interviews, and said so out loud

Felix Rieseberg, an engineering manager at Anthropic, said he has dropped traditional coding interviews for his team: "There's no more point in giving you a little React challenge than there is in quizzing people about assembly" — "Claude is now better at both." He added that "our most impactful builders are successful regardless of how good they are at these tests" and that he does not think the coding interview is coming back. He has not replaced it with a process — he says he hasn't "quite gotten there yet" — and is instead scanning résumés and portfolios for small prototypes that show taste, empathy and principled thinking. This is one manager's practice, not an Anthropic hiring policy, and the industry is moving the other way: ZipRecruiter's October survey of 1,500 employers found 70% of information-sector employers have shifted to proctored or in-person assessments specifically to deter AI cheating.

So what: Do this now: pick one role you are hiring for and replace the screening exercise with a review of something the candidate actually built and a conversation about how they drove the agent that helped them build it. You are hiring for judgement and taste in a world where the typing is free, and the industry's instinct — proctoring harder — is re-founding nothing. This is a day-zero problem wearing an HR costume.

Sources: The Register — Anthropic engineer declares the coding interview dead

Read this edition and the full archive at excelsiorgroup.ai/insights/signal