AI Industry Daily Briefing — October 15, 2026

The Executive Read

The thread today is who sells the picks and shovels, and how far ahead they are being asked to commit. Two filings from the machinery of chipmaking landed within a day of each other. TSMC, the contract chipmaker that builds chips for Nvidia, AMD, Apple and others, reported revenue of $46.72 billion for the quarter ended September 30, above the top of its own guided range, and guided to $52.3 billion to $54.3 billion for the current quarter. ASML, the Dutch company that makes the machines TSMC and others use to print chips, reported €11.7 billion in sales and said in its release that some customers have expressed interest in placing orders for 2029 and even 2030. Neither company breaks out AI revenue in the documents I read, so the AI link is the companies’ framing and ASML’s CEO’s words, not a measured share. Further up the stack, Modal, a company that rents out computing for running AI software, said in its own blog post that it has passed $1 billion in annualized revenue and raised $750 million. Those are Modal’s figures. And Anthropic published what it says is a sample of how Americans used Claude in the weeks before the midterm elections, along with changes to its safeguards. The numbers there are Anthropic’s, drawn from its own sample. Everything below is labelled by who said it.

Top AI Headlines

TSMC Beats Its Own Guidance and Forecasts a Larger Fourth Quarter

What happened. TSMC, the Taiwan Semiconductor Manufacturing Company, reported third-quarter results on October 15. Revenue was NT$1,494.24 billion, or US$46.72 billion. That is up 41.2% in US dollars from a year earlier and 16.2% from the second quarter. TSMC’s own guidance for the quarter had been $44.6 billion to $45.8 billion, so the result came in above the top of that range. Net income was NT$763.06 billion, up 68.7% from a year earlier. Gross margin, the share of revenue left after the direct cost of making chips, was 66.4%, inside the 65% to 67% range the company had guided. Operating margin was 58.7%. For the fourth quarter, TSMC guides to revenue of $52.3 billion to $54.3 billion, with gross margin of 65% to 67%. It also says full-year 2026 revenue should rise by around 44% in US dollar terms. Chief Financial Officer Wendell Huang said the quarter was supported by strong demand for leading-edge process technologies and pointed to the steep ramp-up of 2-nanometer technology, the newest generation of chip making, as a driver heading into the fourth quarter.

Why it matters. The filing shows where the money is going. Chips at 7 nanometers and below made up 79% of wafer revenue: 33% from 3-nanometer, 28% from 5-nanometer, 10% from 7-nanometer, and 8% from 2-nanometer. The smaller the number, the more advanced the chip. The 2-nanometer share is already 8%, and the CFO calls the ramp steep. Capital spending, as shown on the cash flow slide in TSMC’s presentation, was NT$443.21 billion in the quarter, down about 11% from the second quarter by my calculation but up about 54% from a year earlier. Wafer shipments were 4.476 million 12-inch equivalents, up 3.2% on the quarter. Revenue rose 16.2% in dollars on 3.2% more wafers, which suggests most of the sequential growth came from price and mix, not volume. That is my inference; TSMC does not say so in these documents. TSMC’s results are the closest thing the industry has to a hard read on demand for chips from Nvidia, AMD, Apple and others, because those companies buy from it.

What I could not verify: TSMC’s platform breakdown, which would show high-performance computing (the category that includes AI accelerators) as a share of revenue, sits in a chart image I could not read. Some aggregators have quoted a figure; I am not repeating one I could not trace to the filing. The company does not report an “AI revenue” line in these documents.

Business implication. For anyone buying AI hardware, the fourth-quarter guide implies TSMC expects more demand, not less, and is still pricing and mixing toward its newest nodes. The margin guide of 65% to 67% is close to this quarter’s 66.4%, so TSMC is guiding for growth in dollars without a jump in profit per dollar. Buyers should read that as a supplier that is not cutting prices to fill capacity.

Sources: TSMC press release, “TSMC Reports Third Quarter EPS of NT$29.43” (October 15); TSMC third-quarter presentation, Form 6-K on SEC.gov; TSMC results and guidance, Form 6-K on SEC.gov; TSMC investor relations, Q3 2026 results table.

ASML Says Some Customers Want to Order Chip Machines for 2029 and 2030

What happened. ASML, which makes the lithography machines that print circuits onto silicon, reported third-quarter results on October 14. Net sales were €11.66 billion, gross margin was 57.5%, and net income was €4.96 billion. It sold 113 new lithography systems and 8 used ones. Installed base management sales, meaning upgrades and service on machines already in customers’ factories, were €3.29 billion. For the fourth quarter ASML guides to sales of €14.2 billion to €15.2 billion and a gross margin of around 57%. For the full year it guides to €44 billion to €45 billion in sales at a gross margin of around 56%.

Why it matters. The most important sentences are the CEO’s. Christophe Fouquet said in the release that customers are raising their demand forecasts, “driven primarily by their views on the size and duration of AI-related investments.” He said some customers are expressing interest in placing orders for 2029 and even 2030. He also said the supply ASML has planned for 2027 and 2028 is not expected to fully meet customer demand, and that the company is looking at ways to add capacity in 2028 beyond the 30% increase it has already communicated. This is the company’s own characterization of what customers have told it. Interest in ordering is not the same as signed orders, and the release as I read it did not report a net bookings figure, so I cannot say how many orders were actually placed.

The scale of the year is worth a number. ASML’s full-year guide of €44 billion to €45 billion compares with €11.66 billion in sales in this one quarter. The fourth-quarter guide implies a step up of about 21% to 30% from the third.

Business implication. A machine maker asking its customers to think about 2030 is a sign that chip companies are planning factories on the assumption that AI spending keeps going. That is a bet by those customers, not a fact about the future. Capital that ASML’s customers spend is, in turn, a claim on power, buildings and water. ASML also spent about €4.1 billion on share buybacks in the quarter and set an interim dividend of €1.88 per share, payable November 4. This is informational and not a recommendation.

Source: ASML Q3 2026 results press release, Form 6-K on SEC.gov. I read the release on the SEC site; I did not listen to the earnings call.

Anthropic Publishes Midterm-Election Usage Data and Safeguard Changes

What happened. On October 14, Anthropic published a post on how it is preparing for the 2026 US midterm elections. It says it analyzed a random sample of 250,000 US consumer conversations on Claude.ai from September 21 to 27, using a method that protects user privacy. Its figures: 5.75% touched broadly on politics, public policy or social issues. 0.51% were specifically about elections or voting. About 0.01% asked Claude whom or what to vote for; Anthropic reports the upper end of a 95% confidence interval, which is the highest value consistent with its sample. And 0.002% were requests that violated its election integrity policies, measured in a separate analysis of another 250,000 conversations. Anthropic says it expects political and election questions to rise as Election Day approaches.

On product changes: since April, Claude.ai has shown a banner pointing to TurboVote, a nonpartisan voting resource run by Democracy Works. Later in October, Anthropic says, Claude will draw official voting information from Democracy Works into its answers for US users, such as where to vote and what ID to bring. It says Claude will also show election results on election night and afterward, drawn from official numbers.

Why it matters. Three parts deserve separate treatment. First, bias testing. Anthropic says its Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 models scored 99% on its own political evenhandedness evaluation. That is a vendor’s score on a vendor’s test. The post also describes outside reviews. One, from the Foundation for American Innovation, compared matched left- and right-leaning prompts and found that Claude Sonnet 4.5, an older model, did not always hedge evenly and was more likely to soften right-leaning arguments. Reviewers found Claude rarely refused requests but sometimes added unsolicited caveats, such as inserting an opposing view into an advocacy email. Anthropic lists fixes it plans.

Second, outside red-teaming, meaning deliberate attempts to break the safeguards. The Institute for Strategic Dialogue and the McCain Institute tested protections in the US, Israel and Brazil. Anthropic says Claude reliably refused single harmful requests but could miss malicious campaigns split into many harmless-looking tasks. It has added classifiers, which are automated detectors, for multi-step misuse such as fake accounts and news sites, and for financially motivated election schemes.

Third, observed abuse. Anthropic says a foreign-based operator of a fake news site used Claude to produce US midterm coverage favoring Russian, Iranian and North Korean narratives, and that other actors created fabricated news to attract American audiences and ad revenue. It says it banned the accounts involved. These are Anthropic’s descriptions; the post does not name the operators.

Business implication. Any company deploying a chatbot to the public in an election year now has a template: publish usage data, point to official voting sources, and say where testing found gaps. Campaigns get a guide to permitted uses, such as drafting press statements and fundraising emails, and prohibited ones, such as fake social media personas and content posing as independent news sites. A sample of 250,000 conversations from one week is large, but it is one week of one product’s consumer traffic, and says nothing about API customers or other chatbots.

Sources: Anthropic, “Preparing for the 2026 US midterms” (October 14); Anthropic, “2026 Usage Policy update” (October 8).

Regulation and Policy Watch

Anthropic’s updated usage policy takes effect November 12. Anthropic published the update on October 8 and says most changes clarify existing rules while adding examples for Claude’s longer-running autonomous work. Changes it lists: rules on fake accounts, fabricated news sites and influence operations are consolidated in one section covering political and commercial activity. The elections section is renamed “Do Not Undermine Democratic Processes” and now targets voter deception and election disruption; a blanket ban on personalized vote and campaign targeting was removed, though deceptive targeting and misuse of personal data stay prohibited under other sections. Weapons rules now explicitly cover guidance and control software, components, and arming drones or other autonomous vehicles. Claude may not be used to decide or recommend whom to investigate, arrest or charge, or to build surveillance tools. For Claude connected to physical hardware, a qualified operator must be able to observe and stop it, and the equipment must stay in a safe state if Claude disconnects. A new clause prohibits sustained, needless cruelty toward models, limited to extreme cases. The policy also clarifies enforcement on users in unsupported regions, including entities majority-owned or controlled from them. This is a company’s private rulebook, not law, but it governs how customers may use Claude.

Source: Anthropic, “2026 Usage Policy update” (October 8).

The European Commission’s copyright consultation on generative AI runs to November 3. According to press coverage of the Commission’s announcement, the Commission opened a targeted consultation on September 29 covering the use of copyright-protected content in generative AI, online piracy of live events, performers’ remuneration, and research access. Respondents named include rights holders, AI providers, collective management organisations and national authorities. The Commission has not said which approach it prefers, and coverage notes the consultation is not a legislative proposal and creates no new obligations. One summary says a proposal might arrive in 2027; that timing is a secondary report. I did not reach the Commission’s own portal page, so details here come from coverage, and anyone planning to respond should check the official page.

Sources: LexisNexis LNB News, “Commission launches targeted copyright consultation” (September 30); Brussels Times. Both are secondary; the Commission’s own page was not retrieved.

Emerging Startup Radar

Modal says it raised $750 million at a $15.75 billion valuation and has passed $1 billion in annualized revenue. In an October 14 blog post, Modal CEO Erik Bernhardsson says the Series D was led by Accel, with Kleiner Perkins joining and existing investors including Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures participating. The post gives a $15.75 billion post-money valuation, which is the value after the new money is counted. Modal sells serverless computing for AI, meaning customers run code on rented chips without managing the machines. Per the post, Modal “recently crossed $1B in annualized run rate,” a figure that multiplies a recent period’s revenue out to a year. It says usage has tripled since its Series C and grown more than 15 times year over year. Its Sandboxes product, which gives software agents isolated places to run code, has a $300 million run rate on its own, with close to 3 billion sandboxes launched. The company says it draws on a pool spanning more than 20 clouds and has more than 200 employees in New York, San Francisco and Stockholm.

Named customers and uses, all as claimed by Modal: DoorDash on its AI platform, Cognition, Ramp, Linear and Hebbia on background agents, Runway on video, Physical Intelligence, Generalist AI and 1X on robot training, Chai Discovery and Achira on drug discovery, and Suno on music. It also says a product called TypeSafe AI’s Jev served one trillion tokens on Modal in its first three days. I have not verified any customer relationship with the customers themselves. The post says multi-node training, Modal servers and virtual-machine Sandboxes are now generally available.

What it signals: the AI boom is paying for plumbing as well as models. A company whose pitch is “run your agent’s code safely” says that one product line alone is a $300 million business.

Source: Modal, “Series D: $1B in annualized run rate, new funding, and where Modal goes next” (October 14). Company-reported; revenue is unaudited.

AI Infrastructure and Market Signals

Scaleway picks AMD for its first AI supercomputer outside France. At its ai-PULSE event on October 13, the French cloud provider Scaleway announced an AI supercomputer in Sweden built on AMD Instinct MI355X chips and AMD’s open ROCm software. Scaleway calls it Europe’s first commercial AMD AI supercomputer. Partners named: Bull for integration, racks, cooling and installation; Giga Computing for servers; HPE Networking for Ethernet; VAST Data for storage; Motul for liquid cooling fluid. The announcement builds on a partnership with Tele2. It gives no GPU count, power figure, site or delivery date. Without those, the size of the system cannot be judged. What is stated is the supplier choice: a European provider selling to European startups has chosen AMD instead of Nvidia.

Source: Scaleway, “Scaleway expands AI infrastructure in Europe with new AMD Instinct GPU-based supercomputer in Sweden” (October 13).

ByteDance is reported to be in talks with MediaTek over data-center chips. Bloomberg News reported on October 15 that ByteDance, TikTok’s Chinese parent, is in talks to secure AI chips for data centers from the Taiwanese chip designer MediaTek. Bloomberg’s page was paywalled and I could not read it, so I am relying on its headline and on a syndicated summary that attributes the reporting to Bloomberg and says the talks and terms are unconfirmed. I cannot say the size, the product, or whether US export rules apply. Treat it as a report, not a deal.

Source: Bloomberg News, “ByteDance in Talks to Secure MediaTek AI Chips for Data Centers” (October 15). Reporting is Bloomberg’s; I did not read the article body.

Read-across from the two filings. TSMC’s capital spending of NT$443.21 billion in a single quarter and ASML’s talk of extra capacity for 2028 point the same way: suppliers are building for continued growth, and both say demand is outrunning what they can make. Both are supplier statements about customers’ intentions. Neither is evidence that end demand, meaning people and companies paying for AI services at a profit, will match the capacity being planned.

Public Investment Watchlist

Informational only; this is not investment advice. I did not retrieve share prices or market reactions from an exchange or company source, so none are quoted.

• TSMC (NYSE: TSM; Taiwan: 2330). Q3 revenue $46.72 billion against guidance of $44.6–45.8 billion; Q4 guided to $52.3–54.3 billion; full-year 2026 revenue growth guided to around 44% in dollars. Source: TSMC’s filings.
• ASML (NASDAQ: ASML; Euronext Amsterdam: ASML). Q3 sales €11.66 billion; Q4 guided to €14.2–15.2 billion; full-year €44–45 billion; interim dividend €1.88 payable November 4. Source: ASML’s filing.
• AMD (NASDAQ: AMD). Named as the GPU supplier for Scaleway’s planned Swedish system; no volume or revenue figure was given.
• MediaTek (Taiwan: 2454). Subject of a Bloomberg report on talks with ByteDance; neither company’s confirmation was found.
• Modal (private). $15.75 billion post-money valuation per its own post; not publicly traded.

Watchlist

1. October 19: Abdulrahman Alfozan detention hearing. The OpenAI engineer charged with bugging an executive conference room, per our October 14 edition, is due before Magistrate Judge Alex G. Tse. The charges are allegations.
2. Anthropic’s state-specific voting answers. The company says they arrive later in October; watch whether the official-source feature ships and whether it publishes results on how it performs.
3. Mistral Large 4 open weights. Mistral said on October 6 it would release them by the end of October; no license has been named.
4. November 3: EU copyright consultation closes. Responses from AI companies and rights holders will show where the lines are drawn before any proposal.
5. November 4: ASML interim dividend payable. Also the next date the company’s customers’ order intentions could show up as numbers.
6. November 12: Anthropic’s usage policy takes effect. Enterprise customers deploying Claude with physical hardware should read the new operator and safe-state requirements.
7. TSMC platform data. The high-performance computing share of Q3 revenue should appear in the company’s management report and call transcript; I could not retrieve either.

---

Scope note. Meta’s Muse Wonder models (October 12) were covered in the October 13 edition, and I found no new verified model launch dated October 14 or 15, so there is no Model and Product section today.