Back to News An editorial illustration of a secure vault wrapping an open model core, a stream of approved chips crossing a border line, and a glowing desktop window steered by an unseen agent
November 20, 2026 Deals Chips Agents Security Policy

Palo Alto Networks Buys Applied Compute for $4.5 Billion as Congress Reveals 879,868 Approved AI Chips for China and Gemini Takes the Wheel on Mac

Palo Alto Networks agreed to buy Applied Compute, a platform for post-training and serving open-weight models, in an all-stock deal valued at $4.5 billion. House documents show the Commerce Department approved 879,868 Nvidia H200 and AMD MI325X chips for Chinese companies, and Nvidia says it has stopped H200 shipments. Google gave Gemini on Mac computer use. Plus: ChatGPT for Accounting, GPT-Live-1.1, an AI-assisted attack on the European Commission and Nvidia’s reported Inception talks.

Share

Palo Alto Networks Agrees to Buy Applied Compute for $4.5 Billion

Cybersecurity giant Palo Alto Networks said on Wednesday that it has signed a definitive agreement to acquire Applied Compute, a platform for post-training and running open-weight AI models. According to Palo Alto Networks’ acquisition announcement, the all-stock transaction is valued at $4.5 billion, subject to adjustments, and is expected to close in the company’s fiscal third quarter of 2027 after regulatory clearance. Applied Compute’s co-founder and CEO, Yash Patil, will become Palo Alto Networks’ Co-Chief AI Officer once the deal closes.

The pitch is built around a bet on hybrid AI. Palo Alto Networks says a large share of enterprises will run open-source and open-weight models tuned to their own data, cost and control needs, and that no large provider is currently delivering those models securely. The combined company plans to let customers train, adapt and deploy open models in the public cloud or in sovereign environments. Applied Compute will keep serving AI-native companies, and its research team will also help train Palo Alto Networks’ own large and small models. Chairman and CEO Nikesh Arora said the company saw the platform firsthand as a prospective customer:

“Every enterprise is going to run a mix of open and closed models, and every one of those models needs to be secured. Applied Compute has built the best post-training and inference platform we've seen, and its research team is exceptional.”

The company’s Form 8-K filing on the merger agreement adds detail: the shares issued at closing will be priced at $337.16, the 15-day volume-weighted average closing price ending November 17, and completion depends on the Hart-Scott-Rodino antitrust waiting period. The deal was announced alongside the company’s fiscal first-quarter results.

SEN-X Take

This is the clearest sign yet that security vendors see model customization, not just model access, as the next control point. For companies weighing open-weight models for cost or data residency reasons, the practical question is shifting from “can we run it?” to “who audits the training and serving stack?” Expect more security suites to bundle fine-tuning and inference, and plan procurement reviews accordingly.

Congress Releases Data Showing 879,868 Nvidia and AMD Chips Approved for Chinese Buyers

The House Foreign Affairs Committee voted on Thursday to release export licensing data it received from the Commerce Department’s Bureau of Industry and Security. The committee’s ECRA disclosure summary says that as of May 11, 2026, BIS had received applications to sell 3,444,160 Nvidia H200 and AMD MI325X chips to Chinese companies and had approved 879,868 of them across 21 license applications. The approvals include 75,000 chips for Alibaba, 150,000 for Tencent and its subsidiaries, 43,000 for Baidu and 4,000 for BYD, all of which appear on the 1260H list of Chinese military companies. Alibaba and its subsidiaries had sought 948,950 chips, and the summary notes a separate December 19, 2025 Nvidia application to export 445,000 H200s to Tencent.

The document also says almost every Nvidia application listed Maginfra, Beijing New Matrix Information Technologies, ZTE Kangxun Telecom or Lenovo as companies that may install the systems, and that every Nvidia and AMD application described the diversion risk as “low.” As CNBC’s report on the committee release notes, committee chair Brian Mast called Maginfra and New Matrix suspected “smuggling fronts,” while Bernstein analyst Stacy Rasgon said the approval rate works out to “about one out of every four.” Nvidia told CNBC:

“The review and licensing process spanned more than half a year, with approvals subject to strict conditions and limited volumes. Despite receiving some licenses, we have since stopped shipping H200s to China”

CNBC adds that Nvidia reported $2.4 billion in quarterly revenue from H200 and related chip sales in China on Tuesday, and that former CEO Jensen Huang is scheduled for a closed-door transcribed interview with House committees on December 1.

SEN-X Take

The numbers matter less than the paperwork trail. Every application rated diversion risk as low, yet lawmakers now question several of the named installers. Companies that buy, resell or host advanced accelerators should expect end-use checks to tighten and should keep their own records on who touches the hardware, not just who signs the purchase order.

Gemini on Mac Can Now Click, Type and Work Through Your Apps

Google added computer use to the Gemini app for Mac on Thursday. In Google’s announcement of Gemini computer use on Mac, Group Product Manager Michael Friedman says users can describe a multi-step task across local files, desktop apps and the web, and Gemini will handle the menus, typing, clicking and forms. Google’s examples include batch color-grading photos, pulling weights and addresses from local PDFs into shipping quote forms, and building a Google Slides deck from live dashboards. The key design choice is where the agent works:

“Gemini works inside an isolated background window with a live preview right in your chat. Your main screen and cursor remain completely free, so you can keep typing, designing, or taking calls without interruption.”

The feature is rolling out in English in the U.S. for Google AI Plus, Pro and Ultra subscribers on personal accounts, and it requires an Apple Silicon Mac running macOS Sequoia 15.0 or later. Google’s help page for using Gemini with Mac apps explains that users must grant Accessibility and Screen & System Audio Recording permissions, can keep lists of allowed and blocked apps, and should mention @Computer use in a Task. Gemini tries to make temporary file backups that are kept for 24 hours or until the next prompt, and it asks for confirmation before sensitive actions. Work and school accounts are not supported yet.

SEN-X Take

Desktop agents are moving from demos to consumer subscriptions, and the background window is a smart answer to the “don’t touch my mouse” problem. IT teams should note the personal-account limit: staff may try this on work machines with personal logins. Now is the time to set a policy on which desktop agents are allowed and which apps they may control.

OpenAI Launches ChatGPT for Accounting With Partner Plugins

OpenAI has opened a dedicated accounting destination in ChatGPT. OpenAI’s ChatGPT for Accounting page describes an Accounting plugin that adds guidance for invoice processing, reconciliation and reporting, with connections to Intuit QuickBooks, Xero, BILL, FreshBooks, NetSuite, QuickStart, Rillet and Campfire, and computer use for apps without a plugin. ChatGPT drafts entries from invoices and receipts with suggested account coding, gathers records for reconciliation, and flags possible duplicates or missing information for review. Users can save templates and review checks as reusable skills for each month’s close.

OpenAI’s help article on accounting work says Business members can install the plugin without admin approval, while Enterprise admins must approve it first, and it warns users to check the accounting app before retrying an entry with an unclear status to avoid duplicates. Partners announced their own pieces. Xero said its integration now supports month-end close reviews, client portfolio reviews and meeting preparation, currently in the U.S. From Xero’s post on the expanded OpenAI integration:

“When the initial review points to something that needs doing, you can move from insight to action back in Xero, where your ledger lives and where every change stays traceable.”

By contrast, FreshBooks’ ChatGPT for Accounting announcement says its connection is read-only: it can retrieve live data to answer questions but cannot change anything in a FreshBooks account.

SEN-X Take

Accounting is a good test of whether agents can be trusted with write access. Note the split: some partners allow actions in the ledger, others stay read-only. Finance teams should start with read-only reviews and reconciliations, require a human approval step before any posted entry, and keep the source document linked to every AI-drafted line so auditors can follow it.

GPT-Live-1.1 Reaches the API With a Cheaper Standby Mode

OpenAI released GPT-Live-1.1, an update to its full-duplex voice model, to developers on Thursday. The OpenAI API changelog entry for November 19 says the model improves pronunciation, accent consistency and transcription in multilingual conversations, and that developers can now configure standby for GPT-Live-1 and GPT-Live-1.1 WebRTC sessions so inference pauses during inactivity and resumes when someone speaks.

The GPT-Live-1.1 model page lists pricing of $0.05 per minute of active session time and $0.02 per minute in standby, with backend model and tool usage billed separately. The model works only through the Live sessions endpoint, supports function calling, and has its own concurrent-session limit separate from GPT-Live-1. OpenAI’s session guide says developers set an idle timeout of 5 to 60 seconds, after which the session returns to standby while the connection stays open.

SEN-X Take

Standby pricing changes the economics of always-on voice assistants. Kiosks, in-car helpers and customer lines spend most of their time waiting, and paying 60% less per idle minute makes those deployments far easier to justify. Teams already running GPT-Live should test the idle timeout carefully, since the first reply after standby can be slower.

The European Commission Confirms an AI-Assisted Cyberattack

The European Commission confirmed on Thursday that it was hit last month by a cyberattack involving a publicly available AI tool. According to The Record’s report on the Commission’s confirmation, spokesperson Thomas Regnier said the first signs were detected on October 22 and described the incident as an “AI-generated cyber campaign.” POLITICO first reported the attack and said CERT-EU had identified the use of ARTEX, an open-source AI hacking framework released in July that uses commercial AI models to automate finding and exploiting vulnerabilities.

The attack reportedly disrupted the Commission’s personnel systems for more than 10 days, including payroll, leave requests, medical reimbursements and translation services. Regnier said investigators had found no evidence of a broad data breach affecting staff and knew of no ransom demand, and that the investigation remains open. The Record notes that the FBI warned last week that ARTEX was being used by criminals and nation-states, with some intrusions reaching privileged access in under three hours.

SEN-X Take

A major government body losing payroll for more than a week to an off-the-shelf AI hacking framework should end any debate about whether these tools are practical. The fix is unglamorous: shrink the internet-facing attack surface, patch edge devices fast, and rehearse how HR and finance systems keep running when they must be taken offline.

Nvidia Is Reportedly in Advanced Talks to Buy Diffusion Model Startup Inception

Nvidia is in advanced talks to acquire Inception, the two-year-old startup behind the Mercury family of diffusion language models, for more than $500 million plus earnout awards for employees, The Information reported on Thursday. Tech Funding News’ summary of The Information’s report says the deal would bring co-founders Stefano Ermon, Aditya Grover and Volodymyr Kuleshov to Nvidia, and that the technology could support Nvidia’s Nemotron open models. Inception has more than 30 employees and raised a $50 million seed round in November 2025 that included Nvidia’s venture arm.

Instead of writing text one token at a time, diffusion models refine many tokens in parallel; Inception claims Mercury runs 5 to 10 times faster than speed-optimized models from major labs. The talks reportedly began before Jensen Huang resigned as CEO on November 12. Neither company has confirmed a deal, and the terms could change.

SEN-X Take

If it happens, this would be a cheap way for Nvidia to own a genuinely different approach to text generation, one that favors speed and cost per token. For buyers, diffusion models are worth testing on latency-sensitive work such as autocomplete and routing, but anyone building on a small vendor’s API should keep a fallback ready in case of an acquisition.

OpenAI Shows How Codex Agents Keep ChatGPT Fast

OpenAI also published a detailed look at its internal Performance Factory, a cloud system that uses Codex agents to find and fix slowdowns in ChatGPT. In OpenAI’s Performance Factory engineering post, the company explains that its Regression Workflow opens a task when a performance alert fires; one agent investigates and proposes a fix for an engineer to review, and after deployment another agent checks whether performance has returned to baseline. The team packaged its investigation process into a shared skill called perf-forensics, drawing on more than 30 internal performance tools.

“In just over three months, the Regression Workflow tracked more than 1,700 investigations and root caused over 600 cases.”

OpenAI says the current bottleneck is often waiting for an engineer to review and merge the fix, and that it is exploring whether agents can handle that step themselves.

SEN-X Take

This is one of the more useful agent case studies we have seen because it describes the pattern, not just the result: encode your experts’ playbook as a skill, trigger agents from existing alerts, and keep a human at the merge step. Operations teams can copy that structure today for incident triage without handing agents production write access.

Quick Hits: Tencent Opens Hy4, Gemini API Adds Tool Retrieval and Moonshot Drops a Chip Deal

Tencent has published open weights for its flagship Hy4 model. The Hy4 model card on Hugging Face describes a mixture-of-experts model with 770 billion total parameters, 49 billion active per token and a 1 million token context window, supporting text, image and video understanding, and it acknowledges a notable gap with top proprietary models on the hardest benchmarks. Google launched Tool Retrieval in preview in the Gemini Interactions API; Google’s Tool Retrieval post says agents can defer loading tool schemas until needed, with no separate charge, on Gemini 3.6 and later models. And The Jakarta Post, citing a report by The Information, says Kimi developer Moonshot AI scrapped a $150 million deal to rent Nvidia B300 chips from an Indonesian provider after the new U.S. remote access rule.

SEN-X Take

Tool Retrieval is the item most teams can use right away. Agents connected to dozens of tools waste tokens and pick the wrong function more often as the list grows. Deferring rarely used schemas is a small code change, and the same idea applies across vendors: keep the everyday tools loaded and let the agent fetch the long tail.

What We Left Out

Several leads did not make the cut. Reports that Meta’s Muse surfaced other people’s private messages and that Apple is building its own computer-control agent rest on unnamed sources or job listings, so we are waiting for confirmation. We could not load Legora’s November benchmark table to check Meta’s claim that Muse Wonder leads it. Microsoft’s seven new MAI models were announced earlier in the week, GPT-6.2’s reasoning changes were part of Tuesday’s launch, and Genie 4’s arrival in the Gemini API follows yesterday’s coverage of Genie 4.

Need help navigating AI for your business?

Our team turns these developments into actionable strategy.

Contact SEN-X →