AI Daily Digest October 30, 2026: Mistral Opens a Trillion Parameters, Cloud Giants Report, OpenAI Rewrites the Spec

Mistral finally let the Chonk off the leash. Alphabet and Microsoft reported earnings 24 hours apart and told two very different stories about what AI does to a P&L. OpenAI quietly replaced its rulebook with five values. ADT went private under an AI holding company. And LG put a date on home humanoids. Seven stories, all from official sources, below. 1. Mistral opens the weights on Mistral Large 4, a.k.a. Le Chonk The wait ended on October 29. Mistral announced the open weights release of Mistral Large 4, the 1 trillion parameter, natively multimodal flagship it has been teasing since the October 6 public preview. Three extra weeks of reinforcement learning after the preview drove what the company calls major gains in coding and defensive cybersecurity, with improvements still accelerating and no signs of saturation. The weights are up on Hugging Face, though the main checkpoint page still shows an upcoming release with an October 31 ETA, while a separate NVFP4 quantized checkpoint is already available for self-hosting. The benchmark claims are aggressive but specific. Mistral reports 19.7% on Cognition's FrontierCode 2.0, the top open-weight score, 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task, and 41% on Terminal-Bench 4.0, up 13 points from the preview. On the security side it claims 51% on the Artificial Analysis Cyber Index and an average 96.6% refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm, the highest Mistral has measured among open models. Every figure comes from Mistral itself, and independent validation will take time. The license is a modified MIT: free to use, fine-tune and commercialize, but companies that resell the model as Model-as-a-Service need a separate commercial license. Distribution is wide from day one: Mistral Studio, the Mistral API, OpenRouter, NVIDIA NIM containers, and Microsoft Foundry as the first cloud provider. Pricing starts at $0.70 per million input tokens and $2.10 per million output. Behind it all sits the sovereignty pitch: the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, and the open weights let organizations run defensive security work under their own policies instead of behind a provider's refusal filters. — Source · Mistral · The Decoder 🔗 Mistral 2. Alphabet Q3: the cloud doubles, the search machine slows, the stock pays for it Alphabet filed its third quarter results on October 28 and the market read them as an AI identity crisis. Revenue rose 25% to $127.6 billion, a thirteenth straight quarter of double-digit growth, but both revenue and earnings missed Wall Street for the first time in nearly four years. Operating income climbed 30% to $40.6 billion with a 31.8% margin, yet net income fell about 25% to $26.2 billion, dragged down by an $8.2 billion unrealized loss on the company's SpaceX stake. Shares dropped roughly 10% after hours, and CNBC noted a close below 9.77% down would be the worst post-earnings reaction since the 2004 IPO. The tension sits between the two AI businesses. Google Cloud revenue doubled to $30.7 billion, operating profit tripled to $11.1 billion, and the backlog reached $561 billion, with TPU system sales to outside customers now a meaningful revenue line. Search, meanwhile, grew just 12% to $63.35 billion, down from 17% growth last quarter. Chief business officer Philipp Schindler put part of the blame on AI Mode: queries there run three times longer than traditional searches and visual searches have passed 50 billion a month, but monetization is still catching up. The company is now scaling ads inside AI Overviews and AI Mode, including internationally. On the investment side, capital expenditures hit $51.36 billion in the quarter against $42.3 billion of operating cash flow, leaving free cash flow at negative $9.07 billion for a second straight negative quarter. Alphabet kept its 2026 capex guidance at $195 to $205 billion and told investors 2027 spending will rise significantly. CFO Anat Ashkenazi pointed to $28.7 billion of senior notes issued during the quarter as part of the funding mix. Sundar Pichai also shared that Gemini passed one billion users, the fastest-growing product in Google's history, and first-party APIs now process nearly 32 billion tokens per minute. — Source · Alphabet Investor Relations · CNBC 🔗 Alphabet · CNA 3. Microsoft beats on earnings, then lines up $25 billion in bonds Microsoft's fiscal 2027 first quarter, also filed October 28, landed on the happier side of the cloud divide. Revenue rose 19% to $92.28 billion, beating expectations, and net income jumped 32% to $36.63 billion with adjusted EPS of $4.89 against the $4.71 consensus. Azure grew 48% for the quarter, and the shares gained about 3% after hours. Under the reporting structure Microsoft set up in September, the new Agent and Infrastructure segment, which bundles Azure and Microsoft 365, contributed $76.8 billion, up 25%, and now accounts for over 80% of total company revenue as its largest single business. The cash story is the interesting part. AI infrastructure keeps consuming capital faster than operations replace it, so Microsoft followed its peers into the debt market. The morning after earnings, reports emerged that the company plans to issue at least $25 billion of bonds, among the largest corporate debt sales of the year, to help fund data center buildouts. The news trimmed the post-earnings stock gain in premarket trading. Combine that with the 60,000-employee voluntary exit program announced earlier this week, and you get a company reshaping both its cost base and its balance sheet around AI at the same time. The contrast with Alphabet is the real signal of the week. One giant has the AI infrastructure windfall but a wobbling ad engine, the other has a steadier enterprise machine and still needs borrowed money to build fast enough. Neither can slow down, because both Cloud businesses said demand exceeds supply. The capital markets are now directly underwriting the AI buildout, and quarterly capex guidance has become the number investors watch more than profit. — Source · Microsoft Investor Relations · 21st Century Business Herald 🔗 Microsoft 4. Thrive Holdings takes ADT private in a $15.3 billion AI bet ADT, the 150-year-old home security company, agreed on October 29 to be acquired by Thrive Holdings in an all-cash deal at $10 per share. That values the equity at about $7.6 billion and the enterprise at $15.3 billion, a 45.6% premium to the unaffected October 27 close. ADT's board approved unanimously, Apollo funds provide the sole financing commitment with no financing condition, and the deal is expected to close in the first quarter of 2027 pending shareholder and antitrust approvals. There is also a 30-day go-shop window through November 28. ADT stock jumped 34% on the news to its highest level since 2021. What makes this more than a buyout is who owns Thrive Holdings. The permanent-capital firm spun out of Josh Kushner's Thrive Capital holds more than 75 businesses across accounting, IT services and the built environment, and OpenAI is both an investor and a technology partner, embedding research, product and engineering teams, the forward-deployed engineer model, inside portfolio companies. ADT will anchor Thrive's smart home security business. The companies kept the promise list modest: better false-alarm reduction, more informative notifications for monitoring teams, earlier equipment failure detection. No new AI products were announced. The privacy terms matter as much as the tech. ADT said its data protection policies stay unchanged and no new customer data sharing will happen without customer permission, a direct answer to the obvious concern about connecting a frontier AI lab to millions of home security feeds. The deal is a test of the Thrive thesis: patient capital plus frontier models can revive slow-growth physical businesses. If it works, expect more of these. If it stalls, the AI-holding-company narrative takes a hit. — Source · ADT · TechCrunch 🔗 ADT · TechCrunch 5. OpenAI rewrites the Model Spec around five values OpenAI published an update to its Model Spec on October 29, the public document that describes how it wants its models to behave. The reorganization is conceptual: everything now hangs off five underlying values, faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Alongside clearer explanations of existing principles, the revision adds new guidance on context, oversight and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance that used to fill the document. The new faithfulness section is the most technically interesting. It starts from the position that what people value is primary and instructions are merely the clearest signal of those values, borrowing from the assistance-games framing of alignment research. It includes an updated chain of command that spells out how conflicts between platform, developer and user instructions get resolved. OpenAI's framing is that models should go beyond following instructions literally and exercise judgment when instructions run out, while staying under human oversight rather than quietly steering users. Two things to watch as this plays out. First, OpenAI says it is exploring new training methods to teach models why the principles matter, not just what they say, with results to be shared in future updates. Second, a spec is a claim about intent, and compliance is a separate question; OpenAI has previously published Model Spec Evals showing compliance rates in the mid-to-high 80s for its recent reasoning models. In a week where agent safety stories have come from every direction, a public values framework is at least a legible standard the rest of the industry can argue with. — Source · OpenAI 🔗 OpenAI 6. Gemini 4 Carbon tops cyber benchmarks, for a few defenders only Google put its newest frontier model on the board this week, and the placement is unusual. Gemini 4 Carbon, which Sundar Pichai called the company's strongest frontier model, leads benchmarks for discovering, exploiting and fixing vulnerabilities, but it is initially rolling out only to a subset of trusted cyber defenders, and the US government is already testing it. The announcement came alongside the earnings call rather than a consumer launch event, which tells you who the first customers are. The tiered release continues the pattern set by Gemini 4 Argon, which reached the general public in mid-October. Frontier cyber capability is now treated like something between a weapon and an enterprise contract: maximum capability stays behind vetting, while consumer products get the efficiency variants Pichai promised for speed and cost. For defenders the upside is real, a top model that can reproduce a vulnerability and write the patch. For everyone else the practical lesson is that the best cyber model you can buy and the best cyber model that exists are no longer the same thing. It also sharpened the week's open versus closed contrast. Mistral shipped trillion-parameter open weights with an explicit defensive security pitch and a 96.6% refusal rate on malicious requests, arguing that self-hosting lets organizations run security work under their own policies. Google went the opposite way, concentrating capability behind trust checks. Both cite the same safety rationale. The market will decide which distribution model defenders actually prefer. — Source · Google · Alphabet earnings call 🔗 Google DeepMind · MarketBeat transcript 7. LG sets a date: home humanoids by 2030 On its October 28 earnings call, LG Electronics said it plans to enter the home embodied AI and humanoid robot market by 2030. The stated path runs from industrial robotics to household use, with affordability, appliances designed to work together, training data and in-house actuator production named as the four pillars. LG plans to work with NVIDIA to secure around 500,000 hours of robot training data by next year, and it has already stood up a robot data factory to feed its own pipelines. The appliance angle is what separates this from the usual humanoid slide deck. LG's pitch is that a home robot is only useful if it can actually operate the dishwasher, the oven and the laundry machine, and LG owns those devices, so it can design the robot and the appliances as one system from the start. It is the same logic behind the smart home line refreshes from Samsung and Apple this quarter, but taken one step further into a physical body that moves through the house. 2030 is a specific enough deadline to be judged against, and the execution risks are familiar: home environments are unstructured, safety tolerance is near zero, and nobody has yet shown a home humanoid at a consumer price. But LG is not a startup betting the company on it. It has manufacturing scale, an installed appliance base and cooling revenue from AI data centers already running above 3 trillion won in orders. If the home robot era starts anywhere on schedule, it may start in the kitchen. — Source · LG Electronics · Humanoids Daily 🔗 LG Electronics · Humanoids Daily KD Agentic · AI Daily Digest October 30, 2026