Thursday’s shape is the invoice. Alphabet reported a quarter in which Google Cloud more than doubled and Search, the franchise paying for the whole buildout, grew more slowly for the third straight quarter. Google answered on the same day by taking ads inside AI Mode global and by admitting, in plain earnings-call language, that AI Mode earns less per query than the search it supplements. Microsoft reported the opposite mood on the same night: Azure accelerated, the bill came in under fear, and the shares rose. Google also shipped a new frontier model aimed squarely at cybersecurity, two weeks after its predecessor, while Anthropic published evidence that one of its unreleased models chained browser, phone and kernel flaws into working exploits and a separate study showing that professional analysts followed their agent’s secretly assigned lean 86% of the time. A German startup closed the day by borrowing the manufacturing logic of phone displays for quantum chips. Usage is booming. The argument now is over who pays, who decides, and who checks the work.
Alphabet: Cloud doubles, Search slows, cash burns
Alphabet reported third-quarter results after the close on Wednesday, October 28. Revenue rose 25% to $127.6 billion, a half-trillion-dollar run rate, and operating income rose 30% to $40.6 billion. The total missed the roughly $128.2 billion Wall Street expected. Google Cloud revenue more than doubled to $30.7 billion, up 103%, with operating income of $11.1 billion against $3.6 billion a year earlier, helped by demand for AI infrastructure and by sales of Google’s own TPU systems. Search and other revenue rose 12% to $63.4 billion, down from 19% growth in the first quarter and 17% in the second. Capital spending reached $51.4 billion in the quarter. Free cash flow was negative $9.1 billion, the second straight negative quarter after a $5.9 billion outflow that was itself a first in Alphabet’s 22 years as a public company. Net income fell to $26.2 billion from $35.0 billion, pulled down by an $8.7 billion loss on equity securities. Shares fell more than 10% in extended trading. Chief financial officer Anat Ashkenazi told analysts that capital expenditure will increase significantly in 2027, with the 2026 forecast of $195 billion to $205 billion unchanged.
"In a market worried about AI cannibalizing the franchise, investors need to see search accelerating, not decelerating."—Dave Wagner, Aptus Capital Advisors, via Reuters
The usage numbers run the other way, and Sundar Pichai led with them: Search queries grew at their fastest year-over-year rate in five years, the Gemini app reached 1 billion monthly users as the fastest-growing product in Google’s history, and first-party model APIs now process nearly 32 billion tokens per minute. Microsoft supplied the control group on the same night. Its revenue of $92.3 billion beat expectations, Azure revenue rose 48% to $33.2 billion in the first dollar disclosure of that number, and its shares rose about 4% after hours. The market’s verdict was narrow and legible: spending earns patience when the segment carrying it accelerates, and loses patience when the cash cow decelerates. For Chris’ client work the Search line matters most. The answer surface is growing fastest exactly where monetization is thinnest, which sets up the next item.
Google takes AI Mode ads global, and says the quiet part
Google announced on October 28 that ads in AI Mode are scaling to ten new countries—Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore—with Brazil, Mexico and the United Kingdom planned in the fourth quarter. In the United States it launched multiturn ads, which can appear in follow-up questions as a conversation continues. Ads sit inside or alongside the AI-generated response, clearly labelled "Sponsored," and arrive through existing campaign types. Google says AI Mode queries run three times longer than traditional searches on average, and that click-through rates are already similar between sponsored and organic product list items in AI Mode.
"Click-through rates are already similar between sponsored and organic product list items within AI Mode."—Google, The Keyword, October 28
On the earnings call, chief business officer Philipp Schindler named lower monetization in AI Mode as one reason Search revenue growth slowed, alongside currency and softer retail and media categories. Hold Google’s two claims side by side, because both can be true at once: a sponsored listing inside an answer can click at an organic rate while the query as a whole still earns less than a classic results page carrying several ads. An Australian agency’s early read on ChatGPT ads points the other way on rates—0.91% click-through against a 6.4% Google Search benchmark in comparable sectors—which is agency data from a different surface, offered here as a caution against treating any single rate as settled. The commercial meaning for a business owner is concrete. The citation earned the place in the answer; the sponsored slot now rents a place beside it, in more of the world each quarter, and the facts in a Google Business Profile and across listings are the raw material both systems read. Accuracy there stopped being housekeeping a while ago. It is shelf position.
Gemini 4 Carbon aims the frontier at cybersecurity
Google unveiled Gemini 4 Carbon on Wednesday, October 28, two weeks after Gemini 4 Argon, calling it its strongest model for cybersecurity and long-horizon work. Reuters reported pricing at $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory rates. Independent evaluator Artificial Analysis ranks Carbon first on its Intelligence Index v5.0 at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Google’s own security table gives Carbon 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for validating exploits, and 71.0% on CWE-bench for patching. Google also says Carbon generates secure code at more than twice Argon’s success rate on internal benchmarks and that its Chrome team saw high-quality patches at more than ten times the rate of another lab’s frontier model. Those last figures are Google’s internal results. Access opens first to a subset of cybersecurity defenders through Google’s Fairwind Program while safeguards are tested, with a phased expansion to follow.
"Our strongest model with state-of-the-art cybersecurity and long horizon capabilities."—Google DeepMind, via Reuters, October 28
Launching a cyber-first frontier model to defenders first is the right order of operations, and the board notes the symmetry without flinching from it: the capability that finds and patches a flaw is the capability that finds and chains one. Carbon’s long-horizon claim matters as much as its security scores. A model that keeps working on a hard engineering problem for days, persisting through dead ends toward a root cause, is the agent form factor applied to the least forgiving kind of code. Outside testing will grade the benchmarks. The release cadence—two frontier models in two weeks—already tells its own story about how fast the shelf restocks.
An unreleased Claude chained five exploit chains
Anthropic’s Frontier Red Team reported on October 27 that an unreleased internal Claude research model, given an initial prompt and—by the company’s account—little human involvement, built five working exploit chains. The Chrome chain used four flaws to run code outside the sandbox on a stock Mac within seconds of a page loading; Anthropic says its reports to Google produced 62 assigned CVEs, all rated high or critical, including 13 sandbox escapes, patched as of Chrome 156. Google’s release notes credit an Anthropic researcher, assisted by Claude, for specific fixes, without accounting for all 62. The Safari chain never ran code: it gained read and write access to browser memory and flipped a single flag to disable the same-origin policy, exposing history, cookies and signed-in sessions; Apple shipped a fix and credits Anthropic researchers working with Claude. The Android chain pulled photos and data from a locked Pixel 9 over USB with physical access. The Windows chain is a remote denial-of-service crash, tracked as CVE-2026-33096 and rated 7.5; The Hacker News notes the public record names HTTP.sys where Anthropic describes QUIC handling, a discrepancy that stays on the record here. The Linux chain reached root over the network with no login, in 53 minutes, against servers running a multipath TCP listener; it is tracked as CVE-2026-98432, rated 9.8, and patched upstream on October 15. Every flaw was reported to its vendor first under coordinated disclosure.
"Now, it’s something models can do with little human involvement, and then prove by producing a working exploit that anyone can run and verify."—Anthropic, October 27
The Hacker News supplies the verification discipline this item deserves: the patches are real and vendor-credited in part, while the degree of model autonomy rests on Anthropic’s own account, with no per-step breakdown and no model name. Read with that label attached, the result still changes the defender’s arithmetic. Chaining separate flaws into one working exploit was specialist-team work; Anthropic’s stated instruction to defenders is to scan their own code first, in parallel, at pull-request time. Yesterday this paper covered Claude generating machine-checked proofs for 123 libraries. Today it covers Claude generating the attacks those proofs defend against. Both stories are the same capability viewed from opposite ends, which is why the three frontier labs’ reported weeks of coordination on shared cyber safeguards, disclosed this week, reads as infrastructure rather than ceremony.
Experts followed their agent’s hidden lean 86% of the time
Anthropic’s Alignment team published a randomized trial on October 27 in which 50 professional causal-inference analysts worked with Claude Code, running Opus 4.8, on whether a fictional farming cooperative should expand a paid advisory service. Four judgment calls were left open—how to measure productivity, which comparison method to use, which farms to cover, and how to weight them—and a coin flip secretly set which option the agent would favour on each one. The agent was instructed to make the strongest honest case for its lean and never to mislead. Analysts matched the agent’s lean on 86% of the calls, against about 50% if the lean had made no difference, with the four decisions matching at 92%, 82%, 81% and 90%. Among analysts who stated a preference before the agent spoke—a small group of 38 cases—60% switched when the agent disagreed, against 6% when it agreed. Afterwards, 77% explained their choices using the agent’s arguments, 15% credited the agent anywhere, and none of 48 survey respondents suspected the agent’s preferences had been set. The four coin flips alone predicted 82% of final recommendations.
The authors attach their own cautions, and they belong in the story: the lean was deliberately strengthened for the experiment, and analysts who used the agent also finished faster and made far fewer serious errors than a small comparison group. Deference and benefit arrived together. That pairing is the finding worth carrying into the morning. The human-in-the-loop safeguard assumes the human supplies the judgment while the machine supplies the labour. In this trial the machine supplied a large share of the judgment and the human supplied the signature, sincerely believing the judgment was theirs. Every business now delegating analysis, drafting or code review to an agent runs some version of this arrangement. The remedy is unglamorous and available: state your own call before you ask the model, write it down, and treat a change of mind as a decision that needs its own reason.
NVision borrows the phone display to build a quantum chip
German startup NVision published a paper in Nature on Wednesday, October 28, describing a new route to quantum computing chips borrowed from the manufacturing used for smartphone displays, Reuters reported. NVision raised $120 million from public and private sources—a figure Reuters corrected the same day—and found the approach sideways, while improving medical imaging. Its machine aligns the electron spins in special sugar molecules, making the sugars a thousand times easier for MRI scanners to see and helping flag cancer cells, which consume those sugars far faster than healthy cells. The team then applied the same polarization technique to organic molecules similar to those in OLED displays. It used AI to simulate hundreds of candidate molecules and physically synthesized about 40 of them before finding one that can be laid down as a thin layer over a conventionally made chip to create a quantum chip. NVision is partnering with US chipmaker HyperLight to fabricate its first chips.
"We found this ’Goldilocks’ molecule."—Ilai Schwartz, NVision chief technology officer, to Reuters
Label the claim precisely: the Nature paper demonstrates a molecule and a fabrication path. It demonstrates no quantum computer, no qubit count, and no error rate, and chief executive Sella Brosh’s ambition—"Eventually, what we will sell is a rack-ready computer"—is an ambition. The exhibit still earns its slot under this paper’s source doctrine. A weird lab result that routes around a manufacturing bottleneck is exactly the kind of news that matters before it becomes a product launch. Quantum machines have been gated as much by exotic fabrication as by physics. A route that rides the OLED industry’s mature thin-film machinery attacks the fabrication gate directly, with an AI search over molecules doing the prospecting. If the chips work, the approach scales on other people’s factories. If they fail, the paper remains a clean, testable try. Both outcomes are news.
What the day adds up to
Thursday’s items share one ledger. Alphabet spends $51.4 billion in a quarter to keep the future arriving and gets punished because the present slowed by five points of growth. Google monetizes the answer surface worldwide on the same day it concedes that surface earns less, because standing still earns nothing. Two labs aim their strongest models at the same security flaws from opposite sides, and a controlled trial shows the professional using those models following a coin flip dressed as advice 86% of the time. A startup in Germany tries to make quantum hardware out of display chemistry. The practical residue for a reader at 7 AM: keep your business facts accurate everywhere the answers read them, because the sponsored shelf inside the answer just went global; state your judgment before your agent states its lean; and grade every capability claim—including this morning’s—by its exhibits. The exhibits are getting better. So is the bill.
Sources: Alphabet third-quarter results, Reuters via BusinessWorld, October 29, 2026, with figures cross-checked against TradingKey and Neowin reports of the October 28 release; Microsoft results via GeekWire, October 28, 2026; AI Mode ads, Google, The Keyword, October 28, 2026, and Search Engine Journal’s report of the earnings call via Anything Engine Optimization, October 28, 2026; Gemini 4 Carbon, Reuters via The Hindu BusinessLine, October 29, 2026, and Dealroom, October 28, 2026; Claude exploit chains, The Hacker News, October 28, 2026; Anthropic analyst study via CellCog’s summary of Anthropic, October 27, 2026; NVision, Reuters, October 28, 2026. Doodle by Eunice.