跳到正文
DEV Community· HIROKI II·· 1 天前AI 評分56

Anthropic 完成總額約 $93.5B 的 IPO,Gemini 推出 macOS Computer Use 模式

AI Daily Digest November 21, 2026: Anthropic Closes a $93.5B IPO, Gemini Learns to Drive Your Mac

AI 導讀

Anthropic 於 11 月 20 日完成 IPO,發售 1,127,000,000 股 A 類股,每股 $83.00;承銷商全數行使超額配股權後,發售所得總額約 $93.5B。

正文

Cover

Anthropic closes its IPO with the greenshoe fully exercised, and the market answers with a sell-off

Anthropic announced the closing of its initial public offering on November 20: 1,127,000,000 Class A shares sold at $83.00 per share, including the underwriters' full exercise of their option to purchase an additional 147,000,000 shares, alongside the closing of a concurrent private placement. After the greenshoe, gross proceeds from the offering alone reach roughly $93.5 billion, up from the $81.3 billion base deal that had already passed SpaceX's June record as the largest IPO in history. The money is now in the bank, and the concurrent private placement closing means the strategic cheques from the reported Google, Microsoft, Nvidia, and AMD participation have landed with it.

The market's verdict on day two was less celebratory. On November 20, ANT opened down more than 9 percent, at one point trading within about five dollars of its $83 issue price before recovering part of the drop, and the wider tape turned red around it: the Nasdaq lost about 2 percent and gave back nearly all of its gains for the year, the Philadelphia Semiconductor Index fell between 3 and 6 percent at various points, Nvidia hit its lowest level since June 2025, and Alphabet briefly overtook Nvidia in market capitalization. Commentators framed it as the moment the IPO landed and the froth left at the same time; short seller Michael Burry used the day to repeat his bubble warnings, comparing the setup to 1929 and 2000.

For builders, the practical takeaway is about information, not panic. A listed Anthropic now reports quarterly, which means Claude's pricing, margins, and compute commitments become public numbers you can plan against, and the rest of the private AI market just received its first live stress test of what happens when a frontier lab's stock meets a risk-off week. The IPO closing also fixes a milestone date for every clause in enterprise contracts that referenced the offering. — Anthropic · The Guardian

🔗 Anthropic · The Guardian

Google gives Gemini on macOS a Computer Use mode that runs in the background

Google announced a Computer Use capability for the Gemini app on macOS, letting the assistant operate the computer directly: it can switch between desktop applications, read and edit local files, and work across web pages to complete multi-step tasks from a natural language description. Google's own examples include batch-editing photos, pulling weights and dimensions out of local PDFs to fill a shipping form, assembling dashboard screenshots into a presentation, and listing items on a resale platform. Desktop apps such as Blender and Autodesk Fusion have been added to the connector list, and you invoke the mode with @Computer use in the prompt.

The design choices matter as much as the capability. Everything runs in a separate background window, so your cursor and main screen stay yours, with a live preview inside the chat window and a stop button that always works. Permissions are granted per application rather than for the whole machine, sensitive steps such as logins and payments pause for the user to handle, and Google says you can scope access to specific folders and require approval for each request. The feature needs an Apple Silicon Mac on macOS Sequoia 15.0 or newer and a Google AI Plus, Pro, or Ultra subscription, and the first wave is US-only, in English.

The timing is pointed. Apple has been tightening Mac permissions for exactly this class of software, splitting device control from data access in the macOS 27.3 betas, adding per-app scoping and menu bar indicators when one application controls another, and warning that full disk access exposes everything on the machine. Computer use is the feature where agent capability and operating system security policy now visibly collide, and the two companies are converging on the same answer from opposite sides: powerful agents, narrowly scoped, with the human watching. — Google · The Verge

🔗 Google Gemini · The Verge

OpenAI describes its Performance Factory, where Codex agents keep ChatGPT fast

OpenAI published a post describing the Performance Factory, an internal system of Codex-powered workflows that continuously hunt for and fix performance problems across ChatGPT. The framing is the interesting part: instead of treating speed as something you check before a release, OpenAI runs standing agent workflows that profile real workloads, isolate regressions, and ship fixes, so latency is treated as a product surface with its own factory rather than an incident response.

This continues a line we have tracked all month. AWS wrote its entire emulator with agent teams, OpenAI medical teams deployed computer-use agents into production at CVS, and now OpenAI is pointing the same machinery inward at its own infrastructure. The pattern in all three cases is that the agents are not asked to write features but to own an operation, with humans defining what good looks like and the agents handling the grinding, repetitive loop of measurement and repair.

For engineering teams, the transferable idea is not the brand name but the loop: pick a metric users feel, give an agent the tools to measure and change it, and let it run continuously with guardrails. Performance work is a good first candidate because the feedback signal is unambiguous, the blast radius of a bad change is measurable, and nobody enjoys the job manually. — OpenAI

🔗 OpenAI

Nestlé wires Microsoft AI and Boston Dynamics robots into its factories, and its coffee machines

The Wall Street Journal reported on November 20 that Nestlé, the world's largest packaged food company, has teamed up with Microsoft and Boston Dynamics to bring AI into its factories and its products, and both technology companies confirmed the partnerships. The anchor is a new Deep Tech Center in Orbe, Switzerland, where Nestlé plans to combine AI, smart sensors, and robotics to lift efficiency across operations, with findings meant to deploy across its global network of plants. Chief technology officer Stefan Palzer told the Journal the technology developed with the partners would be rolled out across Nestlé facilities worldwide.

There is already a working precedent inside the group. Boston Dynamics has published a case study on Nestlé Purina, where Spot robots run hourly thermal and acoustic inspection routes on packaging lines, freeing maintenance technicians for predictive work, with the fleet expanding into processing areas and additional plants. The new agreement lifts that single-brand pilot to group scale, and it extends beyond the factory floor: the AI work reaches into consumer products, with coffee machines cited as the example, a category few people expected to carry a frontier-adjacent software stack.

Food manufacturing has lagged automotive and electronics in robot adoption because hygienic design, washdown environments, and high-mix lines complicate retrofitting, which is why inspection robots came first. Nestlé is the most consequential consumer brand yet to build a supplier structure around AI-era operations, pairing a Swiss applied-AI facility with its Shenzhen hub that sources Chinese technology for global products. The honest caveat is the usual one: global deployment is a CTO sentence, factory retrofits move at production-shutdown pace, and no site count or model list has been published. Watch whether Spot-class inspection spreads from Purina to the coffee, confectionery, and pet-care plants over the next few quarters. — Boston Dynamics · Microsoft · WSJ

🔗 Boston Dynamics · WSJ

Supermicro starts shipping PCIe Gen 6 all-flash storage servers for AI workloads

Supermicro announced at Supercomputing 2026 that the first of its new all-flash storage servers based on AMD EPYC 9006 series processors, previously codenamed Venice, are now shipping, with more configurations to follow as the SP8 CPU variants arrive. The company says these are the first storage products to support PCIe Gen 6 I/O, with sequential read bandwidth up to 100 percent higher than PCIe Gen 5, and the platforms come under its Data Center Building Block Solutions architecture across the CloudDC, GrandTwin, and Hyper families.

The software story is as significant as the hardware. Supermicro validated the platforms with a wide list of software-defined storage partners, including Hammerspace, IBM, MinIO, Nutanix, Scality, VAST Data, VDURA, and WEKA, so buyers can keep their existing storage stack and swap the underlying server. Chief business officer Vik Malyala framed the launch around a familiar bottleneck: AI and HPC workloads have outrun the storage layer, and training and inference clusters need I/O that matches what the GPUs can consume.

Storage is the unglamorous tier of the AI build-out, and that is precisely why it keeps getting funded. Every GPU cabinet shipped this quarter raises the pressure on the parallel filesystem and checkpoint pipeline behind it, and PCIe Gen 6 plus the new EPYC generation gives the storage tier its own catch-up upgrade path without forcing customers onto exotic new interconnects. — Supermicro · PR Newswire

🔗 Supermicro · PR Newswire

Deutsche Telekom invests 100 million euros in Mistral as European AI consolidation continues

Deutsche Telekom announced that its T.Capital tech fund is investing 100 million euros, about 110 million dollars, in Mistral AI, per the company's announcement and Wall Street Journal reporting on November 19. The investment places Europe's largest telecom behind Europe's most prominent AI lab, and it arrives as Mistral pushes its open-weight models into paid products: the Vibe desktop agent shipped earlier this week with Mistral Large 4, the one trillion parameter open-weight release, as the default model for paying users.

For Deutsche Telekom, the stake is a hedge and a supply line. Carriers across Europe are under pressure to offer sovereign AI services that keep customer data inside the region, and an open-weight European lab whose models can be deployed on-premises fits that requirement in a way that US frontier APIs do not. For Mistral, a strategic shareholder with tens of millions of customers is a distribution channel for Vibe and its enterprise offerings, and a signal that European corporates are willing to pay for regional independence rather than only talk about it.

The deal also lands during a nervous week for AI capital markets, with the sector selling off hard after Anthropic's record IPO. In that context, a nine-figure strategic check from an operating company, rather than a venture fund, reads as a statement about what kind of AI investment survives a risk-off environment: the kind attached to contracts, distribution, and regional policy. — Deutsche Telekom · WSJ

🔗 Deutsche Telekom · WSJ

Cognition turns vulnerability response into a product with Devin Security Threats

Cognition launched Devin Security Threats, live at app.devin.ai/security, pairing a managed index of newly disclosed software vulnerabilities with one-click repository scanning and pull-request remediation on top of Devin's Agentic MapReduce engine. Selecting Scan with Devin sweeps the chosen repositories, ranks findings against the organization's threat model, and assigns confirmed issues to Devin to draft reviewable fix PRs. Teams can enable auto-triage with repository lists and severity thresholds so scans fire automatically whenever Cognition publishes a new advisory, or author custom internal threat definitions from bug bounty reports and incident reviews.

The launch ships with its first two threats built in. Cognition published threat definitions alongside open-source plain-Markdown skills, fix-nginx-vuln and fix-java-vuln in the Devin Marketplace Vulnerability Fixes plugin and the Devin CLI, targeting Java4Shell, the OpenJDK runtime RCE disclosed on November 17, and NGINX2Root, the five memory-corruption bugs in NGINX disclosed on November 18. Because the skills are plain Markdown, appsec engineers can read them as human verification checklists or hand them to a different coding agent entirely, which keeps the fix logic portable rather than locked to Devin.

This is the follow-on to a story we covered yesterday: AI tools helped find critical NGINX and OpenJDK flaws within hours and under a couple hundred dollars of compute. Cognition is betting on the other half of that loop, turning advisory triage and patch drafting into an always-on agent workflow. The industry spent the past month demonstrating that AI finds vulnerabilities cheaply; the next contest is who automates the response, and a managed threat feed with open fix skills is a credible first move. — Cognition · Devin Blog

🔗 Devin Security Threats · Cognition

KD Agentic · AI Daily Digest

來源:DEV Community · dev.to