AI Industry Daily Briefing — October 30, 2026

The Executive Read

The third big cloud result is in, and it points the same way as the first two: customers are buying, and the bill for supplying them is enormous. Amazon’s own release puts AWS revenue at $46.0 billion for the September quarter, up 39%, against $77.3 billion of purchases of property and equipment in that single quarter and a trailing-twelve-month free cash flow outflow of $42.4 billion. Add Alphabet and Microsoft, reported yesterday, and the three companies’ filings show roughly $168 billion of capital spending in one quarter. The other thread today is access. Google announced Gemini 4 Carbon, which it calls its strongest model, but is handing it first to a chosen group of cyber defenders. Mistral went the other way and published the weights of Large 4, a model of about 1 trillion parameters that it says is built for coding and defensive security, while also saying it sees “a potential for uplift to conduct real attacks.” Every benchmark number for both models is the vendor’s own claim. ServiceNow, a software company that sells AI inside business workflows, reported 20.5% revenue growth and raised an AI sales target; its filing does not tell us how investors took it, and we found no reliable figure for the share-price move. In Washington, a Reuters report, relayed by The Next Web, says the Justice Department is examining Nvidia’s Hugging Face deal. Nvidia and Hugging Face did not comment on that report.

Top AI Headlines

Amazon Reports AWS Up 39%, and a $77 Billion Quarter of Spending

What happened. Amazon’s results for the quarter ended September 30, furnished to the SEC on October 29, show net sales of $201.9 billion, up 12%. AWS, its cloud arm, had sales of $46.0 billion, up 39%, and operating income of $17.8 billion, against $11.4 billion a year earlier. Amazon’s total operating income was $28.1 billion, a 13.9% margin, above the $22.5 billion to $26.5 billion range the company had guided to. Net income was $20.7 billion, or $1.89 per diluted share, against $21.2 billion, or $1.95, a year ago. The comparison is distorted: the year-ago quarter included $9.5 billion of pre-tax gains from Amazon’s investments in Anthropic.

Purchases of property and equipment were $77.3 billion in the quarter. Over the trailing twelve months they were $215.2 billion gross. Free cash flow, meaning operating cash flow minus that spending, was an outflow of $42.4 billion over the same twelve months, against an inflow of $14.8 billion a year earlier. The release says 2026 cash capital expenditures will be approximately $220 billion. For the fourth quarter Amazon guided to net sales of $240.0 billion to $247.0 billion and operating income of $28.5 billion to $33.5 billion.

CEO Andy Jassy said in the release: “Our AWS investment in functionality and infrastructure is working.” He also projected “revenue growth rates of over 40% both in 2027 and in 2028” for AWS. That is a forecast by the company, not a result.

Why it matters. Three companies have now shown the same pattern in two days. Google Cloud grew 103%, Azure 48% and AWS 39%, but the smaller percentage sits on the biggest base: AWS’s $46.0 billion is above the $33.2 billion Microsoft disclosed for Azure and the $30.7 billion for Google Cloud. Google Cloud’s figure includes Workspace subscriptions, so the three are not like-for-like. Amazon’s capital spending in one quarter was larger than Alphabet’s ($51.4 billion) or Microsoft’s ($39.4 billion), both from their own filings.

Business implication. A cloud customer is buying from a supplier that is, by its own cash flow statement, spending faster than it earns. That has been funded by operating cash and outside capital, and Amazon’s release does not address how the gap is financed. For buyers, the practical question is capacity: the companies keep saying they cannot build fast enough. We did not read the earnings call and do not report what was said on it. We also found no reliable share-price figure for the after-hours session and make no claim about one.

Sources: Amazon, Q3 2026 earnings release, Form 8-K Exhibit 99.1 (SEC); Alphabet, Q3 2026 release (SEC); Microsoft Investor Relations, FY2027 Q1.

Google Previews Gemini 4 Carbon, but Only Cyber Defenders Get It First

What happened. On October 28 Google announced Gemini 4 Carbon, which its blog calls “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” Access starts this week with a subset of cyber defenders, and will widen gradually to more trusted partners through the Fairwind Program. Google says developers, businesses and consumers get it later, after it gathers feedback and refines safeguards. It gave no date. Google also says it is taking part in the US government’s voluntary pre-release model testing process.

The performance figures below are Google’s own and have not been independently checked. Google reports 88.2% on SEC-bench Pro, a test of finding software vulnerabilities; 54.8% on ExploitGym, which tests whether a model can turn a flaw into a working attack; and 71.0% on CWE-bench, which tests fixing flaws. It says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and fourth on WebDev Arena at 1768 Elo. It says its Chrome team found Carbon produced patches more than 10 times better than another lab’s large model, without naming the model or publishing the method. On safety, Google says successful cheating on the Terminal-Bench 4.0 coding test fell to 0.0% and that internal high-severity misalignment alerts fell almost 80% against Gemini 4 Argon, the model it announced on October 13.

The post lists an introductory price of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterwards. Since the model is not yet generally available, we cannot tell who can pay that price today.

Why it matters. A model that is good at finding and exploiting software flaws is useful to defenders and to attackers. Google’s answer is to control who sees it first. The ExploitGym number is the one to watch: it measures exactly the capability that makes staged release a safety decision and not just a marketing choice. Because the scores are self-reported, outside testers will matter more than the headline.

Business implication. Security teams outside the first group will wait. Companies that depend on Google’s cloud for security work should ask their account teams how Fairwind enrolment works. Those evaluating rival tools should note that Mistral released a competing model the same week with a different access policy, covered below.

Sources: Google, “Gemini 4 Carbon: our next frontier model” (October 28).

Mistral Publishes the Weights of Large 4, a 1-Trillion-Parameter Model

What happened. Mistral AI, the French lab, put out the weights of Mistral Large 4 on its newsroom and on Hugging Face. Weights are the trained numbers that make up a model; publishing them lets anyone download and run it. Mistral’s post, titled “Le Chonk: Built for Coders, Trusted by Defenders,” says the model has about 1 trillion parameters in a mixture-of-experts design, meaning only a fraction of them, roughly 50 billion according to the repository name, are used for each word it generates. It is multimodal, supports more than 160 languages, and has three reasoning settings. Large 4 first appeared as a hosted preview on October 6.

The licence is the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. Use, modification, fine-tuning and commercial use are allowed. Companies that want to sell the model as a hosted service need a separate commercial licence from Mistral. That is a limit that “open weights” does not by itself reveal. The model is also available through Mistral’s API, OpenRouter, Microsoft Foundry and Nvidia-hosted endpoints. API pricing is $0.70 per million input tokens and $2.10 per million output tokens.

The following benchmarks are as reported by Mistral, some of them drawn from third-party tests but presented in Mistral’s post: 41% on Terminal-Bench 4.0, 19.7% on Cognition’s private FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index (roughly tied with GLM-5.3, using about 47% fewer tokens per task), 51% on the Artificial Analysis Cyber Index, and 84.5% on BrowseComp, a web-research test. Mistral says that in blind pairwise tests on Surge Code it ranked behind only Claude Opus 5.5 among the models evaluated. We have not seen independent runs.

Mistral says Cisco, Cloudflare, CrowdStrike, Fortinet and Booz Allen provided supportive statements; the post gives no specific results from most of them. It reports a 96.6% average refusal rate on malicious cyber requests, and says that although the model refused black-box attacks against live targets in its tests, it sees “a potential for uplift to conduct real attacks” and is changing its guardrails. A cybersecurity-specialised variant is in preview for vetted partners including Orange Cyberdefense, under contractual safeguards.

Why it matters. Put the two posts side by side. Google holds back a cyber-capable model and releases it to vetted defenders. Mistral releases the base model to everyone and keeps the specialised variant for vetted partners. Both claim defensive benefit; they differ on whether the general model should be public. Once weights are downloaded, the lab cannot recall them, and guardrails added by the lab can be removed by whoever runs the model. Mistral’s own admission about attack uplift is unusually direct for a launch post.

Business implication. For companies wanting a strong coding model they can run on their own hardware, this is a new option with a licence to read closely. The 1-trillion-parameter size means serious hardware. Customers in regulated sectors should test it on their own tasks before relying on any vendor chart.

Sources: Mistral AI, “Le Chonk: Built for Coders, Trusted by Defenders”; Mistral AI newsroom; Hugging Face, Mistral-Large-4-1T-A52B model page.

ServiceNow Raises Its AI Sales Target as Revenue Grows 20.5%

What happened. ServiceNow’s third-quarter results, filed October 28, show total revenue of $4,106 million, up 20.5%, and subscription revenue of $3,988 million, up 21%. Current remaining performance obligations, the contracted revenue due in the next twelve months, were $13.65 billion, up 20.5%; the longer-dated total was $29.7 billion, up 22%. GAAP operating margin, which follows standard US accounting rules, was 9.5%, or $385 million of operating income; on the company’s non-GAAP basis it was 32.5%. Free cash flow was $822 million.

The company raised its 2026 target for annual contract value from AI products to $1.7 billion, up from $1 billion at the start of the year. It also raised its full-year free cash flow margin target by 100 basis points, or one percentage point, to 36%, and guided to 2026 subscription revenue of $15,780 million to $15,790 million. CEO Bill McDermott said: “This was more than another beat-and-raise quarter for ServiceNow, this was the inflection point for ServiceNow.” That is the company’s characterisation.

Why it matters. Investors have worried this year that AI tools will let customers do what workflow software does without paying for it. The filing is the company’s evidence against that fear: a $1.7 billion AI contract target, though the filing does not say how much of that is already signed. We could not find a reliable account of the share-price reaction to this release, and one aggregator’s figure conflicted with the company’s own results, so we report none.

Business implication. Enterprise software vendors are being judged on whether AI revenue is real and separable. ServiceNow’s disclosure of a specific AI contract target is the kind of number buyers and investors can track from one quarter to the next.

Sources: ServiceNow, Q3 2026 results, Form 8-K (SEC).

Model and Product Updates

Gemini 4 Carbon (staged, defenders only) and Mistral Large 4 (open weights) are covered above. Both releases are pitched on cyber capability, and both come with self-reported scores.

Mistral’s newsroom also lists two earlier October releases that may matter to developers who missed them. Voxtral 3, posted October 20, is described as a release for building agents that can listen, reason, speak and act. Timestral, posted October 15, is described as a 209-million-parameter forecasting model that Mistral says outperforms larger models on a standard CPU. Both claims are Mistral’s, and we did not read the underlying posts, so we cannot say more than the newsroom’s own one-line descriptions.

Anthropic’s newsroom shows no new post dated October 29 or 30. Its latest entries are the October 28 science-credits expansion, covered yesterday, and the October 27 government discount.

Sources: Mistral AI newsroom; Anthropic newsroom.

Regulation and Policy Watch

Justice Department reported to be examining Nvidia’s Hugging Face deal. Reuters reported on October 28, citing two people familiar with the matter, that DOJ officials are asking how Nvidia’s $12.9 billion purchase of Hugging Face could affect AI competition, including whether it would let Nvidia “capture a fundamental distribution point for AI.” Hugging Face is the main public host for open models, including, as of this week, Mistral’s. According to The Next Web’s account of the Reuters report, DOJ declined to comment under its long-standing policy, Nvidia declined to comment, and Hugging Face did not respond. The deal is expected to close in the first half of 2027, and Nvidia has promised to keep the platform open. The reporting is Reuters’; we could not open Reuters’ own page, so we link The Next Web’s summary of it. An inquiry is not a finding of wrongdoing, and no formal action has been announced.

House progressives propose a “no AI bailout” resolution. The American Prospect’s David Dayen reported on October 28 that the Congressional Progressive Caucus, led by chair Rep. Greg Casar (D-TX) with Rep. Maxine Waters (D-CA), has drafted a seven-page, nonbinding “sense of the House” resolution. It says Congress should refuse extra funds or special regulatory support to AI-exposed companies or banks facing insolvency, whether through Congress, the Treasury or the Federal Reserve. Failing firms should use existing bankruptcy processes, executives who take unlawful risks should face liability, and government aid should go to workers. Casar told the Prospect, “We shouldn’t be trying to make them whole.” Waters said, “The American people should never again be forced to foot the bill for billionaires’ reckless bets.” Casar wants a floor vote early in the next Congress if Democrats win the House. A nonbinding resolution creates no law. The Prospect says the trigger was Anthropic’s IPO prospectus, which it describes as showing first-half revenue of $16.3 billion and a loss of $80.7 billion; we have not checked that loss figure against the S-1 and it appears to include items we cannot identify. We saw no bill number.

Malaysia seizes 30 servers carrying AI chips. The Associated Press reports that Malaysia’s customs department said on October 28 it intercepted 30 servers with advanced AI chips, worth about 50 million ringgit (US$12.2 million), at an October 15 inspection in the free trade zone of Kuala Lumpur’s main airport. The servers came from an unnamed Asian country, were falsely declared as electrical components, and were headed for re-export to another unnamed Asian country without the permit Malaysia requires for high-performance chips of US origin. Officials believe a syndicate uses Malaysia as a transit hub and link the case to five smuggling attempts, including a June seizure of 72 servers. A Malaysian company that facilitated the shipment has been called in. Customs did not name the chip models.

Sources: The Next Web, “DOJ looks into Nvidia’s $12.9B Hugging Face deal, Reuters reports” (October 28); The American Prospect, David Dayen, “House Progressives Throw Down a Marker: No AI Bailouts” (October 28); Associated Press, via the Atlanta Journal-Constitution (October 28).

AI Infrastructure and Market Signals

Three filings now sit side by side. In the September quarter, purchases of property and equipment were $77.3 billion at Amazon, $51.4 billion at Alphabet (a figure reported yesterday from its release) and $39.4 billion at Microsoft (from its cash flow statement). That adds to $168.1 billion in three months from three companies. The cloud revenue those purchases serve was $46.0 billion at AWS, $30.7 billion at Google Cloud and $33.2 billion at Azure, though Azure’s dollar figure comes from TechCrunch’s reading of Microsoft’s disclosure rather than the release, and the three segments are defined differently.

The cash side is the sharper signal. Amazon’s trailing free cash flow was an outflow of $42.4 billion. Fortune reported yesterday that Alphabet’s free cash flow was negative $9.1 billion. Operating cash flow at Amazon was still $168.3 billion over twelve months, up 29%, so the business funds much of the spending; it does not fund all of it. Amazon says 2026 cash capex will be about $220 billion.

On the supply side, Malaysia’s customs case is a reminder that demand for advanced chips outruns legal supply in some markets. The AP report says only that the shipment was falsely declared and lacked a permit; it does not say who the buyer was or whether the chips reached China, and we make no claim on either.

ServiceNow, in a different part of the market, shows software vendors reporting AI revenue by contract target rather than by usage. Its $1.7 billion figure is annual contract value that the company expects to sign this year, not recognised revenue.

Sources: Amazon Q3 2026 release (SEC); Alphabet Q3 2026 release (SEC); Microsoft Investor Relations; Fortune (October 28); TechCrunch, Aditya Mehta (October 28); Associated Press, via the Atlanta Journal-Constitution.

Public Investment Watchlist

Informational only; this is not investment advice.

• Amazon (AMZN). Reported October 29 after the close. The primary document is the 8-K linked above. The new items are AWS at $46.0 billion, a $77.3 billion capital-spending quarter, and Jassy’s projection of AWS growth above 40% in 2027 and 2028. We found no reliable account of the share-price move.
• ServiceNow (NOW). Reported October 28. The filing shows a raised AI contract target and a raised cash-flow margin target. One aggregator posted a share-price decline of about 16.6%, but it did not give a time frame and contradicted the company’s own results on other points, so we do not rely on it.
• Alphabet (GOOGL) and Microsoft (MSFT). Covered yesterday. Alphabet’s 10-Q, when filed, should settle the 2027 capital-spending figure that Fortune reported and the release did not state.
• Nvidia (NVDA). The reported DOJ inquiry into the Hugging Face deal adds regulatory risk to a deal the company expects to close in the first half of 2027. Neither Nvidia nor DOJ has confirmed an inquiry.
• Anthropic (private; S-1 filed October 26). Fortune reports, from the S-1, that each of the seven co-founders has “pledged to dedicate 80% of their personal equity in Anthropic to philanthropic causes,” that the filing describes the pledges as voluntary and revocable, and that the founders would collectively hold 50.1% of voting power through a special share. Fortune’s estimate of the pledge’s value, about $79.4 billion, is its own calculation. The S-1 leaves the offering price blank.
• Apple (AAPL). An earnings-preview site gives fiscal fourth-quarter results as November 2. We did not confirm the date from Apple’s investor page.

Sources: Amazon 8-K (SEC); ServiceNow 8-K (SEC); Fortune, Tatiana Sataua (October 26); Alphabet 8-K (SEC).

Watchlist

• Independent tests of Gemini 4 Carbon and Mistral Large 4. Every score above is the vendor’s. Outside runs of the cyber benchmarks would show whether the numbers hold.
• Who gets into Fairwind. Google has not said when Carbon reaches developers or consumers.
• How Mistral’s guardrail changes land. The company says it is modifying them after seeing attack uplift potential; whether it publishes details is open.
• Any formal DOJ step on Nvidia and Hugging Face. So far this rests on one outlet’s anonymous sourcing.
• The Progressive Caucus resolution. Watch for a bill number, text and cosponsors; today it is a draft described by one outlet.
• Apple’s results, expected November 2 (date unconfirmed), and Alphabet’s 10-Q.
• Anthropic’s price range and exchange, whenever the S-1 is amended.