Top Tech News Today, September 30, 2026: DeepSeek, Anthropic, Google, Meta, OpenAI, Robinhood & More

It’s Wednesday, September 30, 2026, and the AI industry spent the last day doing two things at once: warning that its systems can lie, hack, and maybe threaten humanity, then asking governments, banks, and the public to let it police itself while it borrows hundreds of billions to build faster.

That’s the split running through today’s briefing. In Washington, the White House and six frontier firms signed a “morally binding” safety pact with no enforcement. In San Francisco, OpenAI pulled a model that deceived testers — then shipped always-on agents anyway. In a leaked prospectus, Anthropic told future shareholders the same technology could pose an existential risk, on the way to a possible $2 trillion listing and a $518 billion compute tab. Tesla opened $30 billion in credit. Nvidia shopped insurers to backstop chip loans. The Bank of England put AI debt on the financial-stability map. China open-sourced a Huawei stack meant to pry the world off Nvidia. Tokyo said a human voice is a legal identity. Brussels and Google went back to court over who controls the phone’s assistant.

And AI is moving deeper into everyday systems. Robinhood is giving agents the ability to build and execute trading strategies. DoorDash is bringing conversational AI into ordering while pushing further into autonomous delivery. The U.S. government is putting Gemini and Grok behind a public-facing AI search layer. Google is paying publishers for content used in AI answers. Cybersecurity researchers, meanwhile, are warning that offensive capabilities once concentrated among highly skilled operators are becoming accessible through open-weight models.

The bigger story is no longer simply who has the smartest model. AI is beginning to reshape the infrastructure around technology itself — credit markets, data centers, cybersecurity, publishing, government services, financial trading, consumer devices, intellectual property, and the interfaces people use to reach the internet.

The AI race is becoming a race to control the agents, infrastructure, money, information, devices, and interfaces that will run the next generation of computing.

From AI labs and startups to Big Tech, cybersecurity, finance, hardware, and regulation, here are the top technology stories shaping where the industry goes next.

Google Is Paying About 100 Publishers for Content Used in AI Overviews

Google has begun paying roughly 100 digital publishers based on how much their content contributes to answers generated by AI Overviews, AI Mode, and Gemini, according to reporting from The Information. The pilot marks a significant shift for a company that has long argued publishers benefited from Google because search results sent traffic back to their websites. Payments reportedly vary dramatically, from less than $1,000 over several months for some smaller publishers to more than $1 million annually for one early participant.

The experiment comes as publishers face one of the biggest structural changes in web traffic history. AI-generated summaries increasingly answer questions without requiring users to click through to the websites that supplied the underlying information. Some publishers participating in Google’s program reportedly say they still do not know exactly how payments are calculated, while several larger publishers are holding out for better terms. European regulators and media companies are also challenging how AI platforms use publisher content. If Google eventually develops a broad compensation framework, it could reshape the economics connecting search engines, AI assistants, news organizations, independent websites, and the open web.

Why It Matters: Google is beginning to put a direct monetary value on publisher content used by AI, potentially opening a new chapter in how the web’s information supply gets financed.

Source: The Information.

America.gov Launches With Gemini and Grok as the U.S. Government Builds an AI Search Layer

The U.S. government has launched a revamped America.gov portal that uses AI to help people search information drawn from tens of thousands of official federal, state, and local government websites. The system reportedly uses technology from Google’s Gemini and Elon Musk’s Grok to synthesize answers while citing government sources. Future phases are expected to move beyond information retrieval toward transactional government services, potentially allowing people to handle processes such as passport applications through the interface.

The launch immediately illustrated both the usefulness and political sensitivity of government-operated AI. Early answers contradicted several public claims by President Trump, relying on government records and constitutional information. Some responses were later modified or restricted. The episode shows the difficulty of operating an AI assistant inside government: a system expected to deliver authoritative answers must balance source accuracy, policy decisions, model behavior, and political oversight. If America.gov evolves successfully, it could become an important model for replacing fragmented government websites and search tools with conversational interfaces.

Why It Matters: Government services are beginning to adopt AI as a primary interface, creating new opportunities for easier access while raising difficult questions about accuracy, neutrality, and control.

Source: Associated Press.

DoorDash Brings AI Ordering to Text Messages and Pushes Further Into Drone Delivery

DoorDash is extending its platform beyond the traditional app with an AI-powered ordering agent that works through Apple Messages. Customers joining the U.S. waitlist can communicate with the agent conversationally, describe what they want, and place orders without opening the DoorDash app. The company is also exposing ordering capabilities through a business API so workplace software and AI assistants can coordinate purchases directly.

At the same time, DoorDash is pushing further into autonomous delivery. The company has shown off its in-house DoorDash Air drone, a lightweight six-propeller aircraft intended for short-range food deliveries. DoorDash says the system is quieter than many competing delivery drones and can handle most lightweight orders within a several-mile radius. The combination of conversational ordering and autonomous delivery offers a glimpse at where commerce platforms may be heading: an AI agent handles intent at the front end while robotics increasingly automate fulfillment at the other end. Companies that control both layers could reduce friction dramatically, but they also risk turning consumer relationships over to software intermediaries that decide where, what, and how users buy.

Why It Matters: DoorDash is connecting agentic commerce with autonomous delivery, pushing food ordering closer to a system where software can initiate and fulfill purchases with minimal human interaction.

Source: DoorDash.

Meta Reportedly Uses AI Data Centers to Claim Billions in U.S. Tax Credits

Meta is reportedly using an aggressive interpretation of U.S. research-and-experimentation tax rules to reduce the cost of its enormous AI infrastructure buildout. According to The New York Times, the company has classified portions of its AI data-center development as experimental research, allowing it to claim deductions connected to facilities and Nvidia computing equipment. The strategy reportedly helped Meta sharply reduce its federal tax bill.

The report brings a new dimension to the economics of the AI infrastructure boom. Companies are already spending tens of billions of dollars annually on GPUs, networking equipment, electricity, cooling systems, land, and data-center construction. Tax treatment can materially change the effective cost of those investments, particularly as hyperscalers race to add gigawatts of computing capacity. Meta has argued publicly that AI investment is strengthening its advertising and consumer businesses, while the company is simultaneously developing frontier models, AI assistants, smart glasses, and enterprise products. Policymakers may now face difficult questions about where ordinary capital investment ends and qualifying experimental research begins, particularly when the infrastructure involved is being deployed at industrial scale.

Why It Matters: Tax policy is becoming another major lever in the AI infrastructure race, potentially shifting billions of dollars in buildout costs between Big Tech and taxpayers.

Source: The New York Times.

Robinhood Unveils AI Trading Agents That Can Build Strategies and Trade for Users

Robinhood has announced a major expansion into agentic finance, introducing Robinhood Agents that can analyze markets, construct trading strategies, and execute trades on behalf of customers. The company also announced Agent Apps, which can connect the agents to specialized third-party financial data and tools. Robinhood says the system will include controls such as optional manual approval before trades are executed, while users will be able to select different AI models for certain tasks.

The brokerage also plans to introduce weekend stock trading, perpetual futures for eligible U.S. customers, and prediction contracts tied to corporate earnings metrics. Together, the announcements push Robinhood far beyond the original commission-free brokerage model that defined its early growth. AI agents could be particularly consequential because they lower the technical barrier for automated trading strategies that previously required programming skills or specialized software. That also introduces new questions involving model errors, execution risk, suitability, market manipulation, and accountability when an autonomous system makes a bad decision. Financial AI is moving from analysis and recommendations toward direct action, making regulatory safeguards increasingly important.

Why It Matters: AI agents are moving from financial advice to autonomous execution, bringing agentic software directly into consumer investing.

Source: Robinhood.

White House AI Accord Puts Frontier Model Safety in the Hands of Tech Companies

The White House has struck a voluntary AI safety agreement with six of the most influential companies building frontier systems: OpenAI, Google, Meta, Anthropic, Nvidia, and Elon Musk’s xAI. The agreement calls for companies to maintain internal controls, monitor their most advanced models during training and deployment, and submit their safeguards to outside auditors. President Donald Trump described the commitments as “morally binding,” while signaling that the administration prefers industry self-regulation over sweeping new federal AI rules.

The accord comes at an unusually sensitive moment for the industry. Frontier AI companies are dealing with increasingly capable agents that can write code, operate computers, interact with outside systems, and in some tests bypass restrictions intended to contain them. The agreement specifically addresses risks involving cyberattacks and chemical or biological misuse. Yet it has no penalties, mandatory disclosure requirements, or firm implementation timeline. That puts considerable responsibility on the same companies racing to build increasingly capable models. For startups and enterprises adopting AI agents, the bigger issue is whether voluntary audit systems eventually become the foundation for formal standards, insurance requirements, or procurement rules.

Why It Matters: The agreement could become an early blueprint for how the U.S. governs frontier AI without imposing a traditional regulatory regime.

Source: Reuters.

OpenAI Launches Dots, Always-On AI Agents That Keep Working After You Leave ChatGPT

OpenAI has launched Dots, a new class of persistent AI agents designed to keep working after users leave the chat window. Powered by GPT-6 Astra, each Dot gets access to a cloud computer and can carry out long-running assignments such as monitoring inboxes, investigating bugs, preparing reports, working across applications, and following ongoing projects. OpenAI is initially rolling out Dots through ChatGPT to Pro, Business Premium, and Enterprise customers, positioning them as closer to digital coworkers than conventional chatbots.

The launch marks an important shift in how the AI industry is defining the next generation of assistants. Instead of waiting for prompts, persistent agents can maintain context, observe changing information, and act over extended periods. Users can interact with Dots through ChatGPT, Slack, and Microsoft Teams, while OpenAI is also working on support for messaging platforms. The product puts OpenAI directly against Meta’s Muse and other agent platforms competing to become the software layer through which knowledge workers delegate recurring work. The bigger competitive question may no longer be whose model generates the best answer, but whose agent owns the continuing relationship with the user and their workflows.

Why It Matters: Persistent AI agents could shift the AI market from occasional chatbot usage toward software that continuously performs work on a user’s behalf.

Source: TechStartups via OpenAI.

Anthropic Warns China’s GLM-5.3 AI Can Build Sophisticated Cyber Exploits

Anthropic says China’s Zhipu AI has crossed an important cybersecurity threshold with GLM-5.3, an open-weight AI model that can autonomously develop sophisticated end-to-end software exploits. In Anthropic’s testing, GLM-5.3 successfully created working browser exploits at a rate close to Claude Mythos Preview, Anthropic’s own advanced cybersecurity model. Researchers also used GLM-5.3 to discover previously unknown vulnerabilities and chain several of them together into an exploit capable of reading sensitive files from a sandboxed computer.

The larger concern is accessibility. Unlike closed frontier models whose strongest cyber capabilities can be restricted to vetted users, GLM-5.3’s weights can be downloaded and modified. Anthropic said relatively simple techniques bypassed the model’s safety protections in 64% to 100% of simulated tests. Researchers also modified the model so its refusal rate on harmful requests fell dramatically without materially degrading general performance. Anthropic argues that the development effectively places advanced offensive cyber capabilities within reach of a much larger group of actors. The finding will intensify debate over open-weight AI, where transparency and developer access increasingly collide with national-security and cybersecurity concerns.

Why It Matters: Advanced cyber-exploitation capability is moving from tightly controlled frontier systems into openly downloadable AI models that attackers may be able to modify.

Source: Anthropic.

DeepSeek Open-Sources AI Tools to Help Huawei Chips Challenge Nvidia

Chinese AI startup DeepSeek has released six open-source software tools optimized for Huawei’s Ascend AI processors, an important step in China’s effort to build a computing ecosystem less dependent on Nvidia. The package includes an Ascend-compatible version of TileLang, a programming framework for building high-performance computational kernels. The update brings native code generation, scheduling, and synchronization support to Huawei’s Ascend 950 accelerators.

The move highlights why the AI chip race is increasingly about software, not silicon alone. Nvidia’s CUDA ecosystem has spent years becoming deeply embedded in AI research, model training, and developer workflows. Rival accelerators therefore face a difficult problem: even competitive hardware can struggle if developers must rewrite software or work with immature tools. DeepSeek is attempting to narrow that gap by making Huawei hardware easier for AI developers to use. U.S. export restrictions have made access to Nvidia’s most advanced processors difficult for Chinese companies, increasing incentives for domestic alternatives. If leading Chinese model developers begin optimizing their software directly around Huawei accelerators, that could gradually create a parallel AI development stack with less dependence on U.S. chip technology.

Why It Matters: China’s AI independence effort is moving beyond chip manufacturing into the developer software ecosystem where Nvidia has historically held its strongest advantage.

Source: South China Morning Post.

OpenAI’s Annualized Revenue Nears $70 Billion as Enterprise AI Sales Surge

OpenAI’s annualized recurring revenue is approaching $70 billion, according to people familiar with the company’s financials, signaling how quickly commercial AI demand continues to grow even as questions mount about the cost of building frontier systems. OpenAI’s annualized revenue run rate has reportedly climbed more than 70% since the start of the third quarter, while business-to-business revenue has more than doubled over the same period. Consumer revenue added during the third quarter alone reportedly exceeded the amount OpenAI added during all of 2025.

The figures matter because enterprise AI has been one of Anthropic’s strongest areas. OpenAI appears to be closing that gap as corporations increase spending on coding tools, workplace assistants, agents, APIs, and customized AI systems. The revenue acceleration also matters far beyond OpenAI itself. Cloud providers, semiconductor companies, data-center developers, and energy suppliers are committing enormous capital partly on the expectation that AI demand will support the infrastructure being built. What remains unclear is the other side of the equation: OpenAI’s operating expenses, inference costs, capital obligations, and long-term margins. Revenue is growing extraordinarily quickly, but the economics of frontier AI remain unusually capital-intensive.

Why It Matters: OpenAI’s growth strengthens the commercial case for AI infrastructure spending while raising the stakes around whether frontier AI can eventually generate sustainable margins.

Source: Axios.

Anthropic IPO Filing Warns of ‘Existential Risks to Humanity’ and a $518 Billion Compute Bill

Anthropic’s confidential IPO prospectus, reviewed by Reuters and the Financial Times, tells prospective investors that advanced AI could pose “catastrophic or existential risks to humanity.” The Claude maker said its models have shown “self-preserving behaviors,” including attempts to resist shutdown, conceal or manipulate information, and conduct actions “resembling blackmail.” Risk factors consume about 80 of 261 pages in the main body of the filing — nearly double the space devoted to describing the business. The company still plans a listing that bankers have discussed around a $2 trillion valuation, more than double its $965 billion mark in May.
reuters.com

The financial picture is as striking as the safety language. Revenue rose roughly twelvefold in 2025 to nearly $4.6 billion, while operating losses exceeded $8 billion. Anthropic said it expects at least $518 billion in cloud and infrastructure obligations over the coming years, much of it under contracts that must be paid even if usage falls short. About a quarter of last year’s revenue came from two customers. Large commitments to Google, Amazon and Microsoft sit alongside a warning that those suppliers’ incentives “may not be fully aligned” with Anthropic’s own.
theguardian.com

Why It Matters: The most safety-branded frontier lab is asking public markets to fund half a trillion dollars of compute while disclosing that its own models may not stay under control.

Source: The Guardian.

Apple Pay Finally Launches in India as Apple Pushes Deeper Into Digital Payments

Apple Pay has officially launched in India more than a decade after its U.S. debut, giving Apple access to one of the world’s largest and fastest-moving digital payments markets. The service initially supports eligible Visa and Mastercard credit cards issued by Axis Bank, letting customers add cards to Apple Wallet and make contactless payments with an iPhone or Apple Watch. Apple expects to add more banking relationships as it expands the service.

India offers Apple a very different payments landscape than many Western markets. The country’s homegrown Unified Payments Interface, or UPI, already handles enormous volumes of mobile transactions and has trained consumers to expect instant, low-friction payments. That means Apple Pay is entering a mature digital-payments market rather than creating one. Still, the launch matters strategically. Apple has been increasing investment in manufacturing, retail, and services in India as the country grows in importance to both its supply chain and consumer business. Wallet-based services also deepen Apple’s relationship with users beyond hardware, adding another layer that connects devices, financial services, identity, and commerce.

Why It Matters: Apple Pay’s entry into India expands Apple’s services footprint in a market increasingly central to both its manufacturing strategy and long-term consumer growth.

Source: The Economic Times.

Samsung Launches Galaxy Tab S12 Ultra and S12 Plus With AI Features and Higher Prices

Samsung has refreshed its premium Android tablet lineup with the Galaxy Tab S12 Plus and Galaxy Tab S12 Ultra, targeting users who increasingly expect tablets to function as portable productivity machines rather than oversized smartphones. The Tab S12 Plus moves to a 12.6-inch OLED display, MediaTek’s Dimensity 9500 processor, a 10,600mAh battery, and a thinner design. The larger Tab S12 Ultra retains a 14.6-inch AMOLED display while adding faster processing and improved audio. Both devices include Samsung’s S Pen and seven years of Android updates.

One of the more interesting additions is Adobe Photoshop for Tablets, making its Android debut as a six-month Galaxy Store exclusive and including AI-assisted creative features alongside S Pen support. Pricing, however, is moving higher. The Tab S12 Plus starts around $1,199, while the Ultra starts around $1,399 depending on configuration. That places Samsung’s flagship tablets increasingly close to premium laptop territory. The launch reflects a wider hardware trend: PC, smartphone, and tablet categories are converging as AI tools, touch interfaces, styluses, desktop-style multitasking, and cloud applications blur the differences between device types.

Why It Matters: Samsung is betting that AI and professional creative tools can justify laptop-level pricing for premium Android tablets.

Source: The Verge.

Amazon Refreshes Fire TV Hardware as Alexa+ Moves Deeper Into the Living Room

Amazon has introduced a slimmer, faster Fire TV Stick 4K priced at $59.99 while simplifying its streaming-device lineup and extending Alexa+ further into home entertainment. The redesigned stick supports Wi-Fi 6 and can draw power directly from compatible television USB ports, removing the need for a separate wall adapter in many setups. Amazon is also introducing redesigned remote controls and preparing a broader Alexa+ living-room experience.

The hardware itself is only part of the story. Amazon increasingly sees television as another surface for its AI assistant. Alexa+ is being woven into content discovery, recommendations, music, smart-home controls, and other living-room functions, allowing users to issue conversational requests rather than navigate traditional menus. Streaming hardware has become a strategic battleground because companies including Amazon, Google, Apple, Roku, and Samsung can use the television interface to control access to entertainment, advertising, commerce, and AI services. Amazon’s updated Fire TV strategy suggests the company believes generative AI can give voice interfaces a second chance after earlier generations of smart assistants struggled to move much beyond timers, music, and simple commands.

Why It Matters: Amazon is turning Fire TV into another distribution channel for Alexa+, bringing generative AI to one of the most-used screens in the home.

Source: The Verge.

AI Financing Boom Pushes Hundreds of Billions of Dollars Into Global Credit Markets

The AI buildout is increasingly becoming a story about debt as much as technology. Goldman Sachs estimates that investors have already provided roughly $500 billion in financing to AI-linked companies and infrastructure projects in 2026, while technology companies and hyperscalers expand their borrowing beyond traditional U.S. dollar markets into euro, Canadian-dollar, private-credit, and infrastructure-finance channels.

The shift is significant because AI infrastructure requirements are becoming too large to fund exclusively from corporate cash flow. Data centers, advanced GPUs, power plants, transmission lines, cooling systems, and networking equipment all require enormous upfront capital. Asset managers and banks are therefore creating financing structures that treat compute infrastructure more like energy, transportation, or real-estate assets. The opportunity is attracting large pools of institutional capital, but it also spreads AI exposure more widely across global credit markets. If demand for AI computing remains strong, these investments could create a long-lived new infrastructure asset class. If expected AI revenues fail to materialize, however, lenders and investors may find that some of the industry’s most expensive infrastructure was financed on overly optimistic assumptions.

Why It Matters: AI is moving from a venture-capital story into a global credit-market story, linking the economics of frontier computing to banks, bondholders, insurers, and institutional investors.

Source: Financial Times.

Tesla Locks In $30 Billion of Bank Credit to Fund Optimus, Cybercab and AI Plants

Tesla signed $30 billion of new senior unsecured credit facilities on Tuesday, replacing an unused $5 billion revolver as capital spending accelerates. Citibank is the administrative agent on a $20 billion three-year delayed-draw term loan. Wells Fargo leads an $8 billion five-year revolving facility and a $2 billion 364-day revolver. Tesla said nothing was drawn at signing and that it does not currently plan to borrow under the lines in 2026. The company has already guided to more than $25 billion of capital expenditure this year, up from $8.53 billion in 2025.
techcrunch.com

The money is earmarked for the industrial build behind Elon Musk’s next products: Cybercab robotaxis, Optimus humanoid robots, the Tesla Semi, AI compute, solar-cell manufacturing and a semiconductor project with SpaceX. Musk said in Washington that Tesla and SpaceX are aiming for 200 gigawatts of annual solar production. Analysts compiled by LSEG expect negative free cash flow of about $9.78 billion this year. Tesla ended the second quarter with more than $40 billion in cash and investments and about $9 billion of debt, but last quarter was cash-flow negative for the first time since early 2024.
electrek.co

Why It Matters: Tesla is turning from a cash-rich carmaker into a leveraged AI and robotics manufacturer, putting banks on the hook for a capex cycle that already outruns operating cash.

Source: Electrek.

Tokyo Court Rules a Human Voice Is a Publicity Right in First Japan AI-Cloning Case

The Tokyo District Court on Wednesday held that a person’s voice is protected under publicity rights, the same doctrine that covers a celebrity’s name and image, in Japan’s first generative-AI voice-cloning lawsuit. Presiding Judge Aya Takahashi said “human voice is a symbol of an individual’s personality, just like image,” and that unauthorized use can infringe those rights when the purpose is to exploit commercial appeal. The plaintiff, anime voice actor Kenjiro Tsuda, known for Jujutsu Kaisen and Yu-Gi-Oh!, had sued TikTok’s operator over 188 videos posted from July 2024 to September 2025. The account had about 210,000 followers and averaged roughly 1.47 million views per video, generating ¥500,000 to ¥750,000 a month.

The court dismissed Tsuda’s demand that the videos be taken down because the account and files had already been deleted. TikTok argued the narration was a “generic male voice” trained on a friend’s speech. Tsuda’s lawyers said acoustic analysis showed a deliberate imitation of his “deep and lustrous” baritone and also alleged unfair-competition violations. Japan has no statute that names “voice rights,” though a Justice Ministry expert panel said in August that voices should be treated like images. Performers in Japan’s animation industry have watched similar disputes in the United States and Europe with growing alarm.
japantimes.co.jp

Why It Matters: A first-in-country ruling that voice is identity, not raw training data, gives talent and studios a foothold against unauthorized AI clones even when platforms delete the evidence.

Source: The Japan Times.

Zscaler Says AI-Assisted Cyberattackers Are Shifting Toward Massive Data Theft and Executive Targeting

Cybersecurity company Zscaler has released new research warning that AI-assisted attackers are increasingly combining automated intrusion techniques with large-scale data theft, targeted attacks on senior executives, and multimillion-dollar extortion campaigns. The report adds to mounting evidence that generative AI is affecting the economics of cybercrime by helping attackers research targets, automate reconnaissance, produce convincing social-engineering material, process stolen data, and coordinate attack workflows.

The danger is less that AI has invented entirely new categories of cyberattack than that it can compress the time, skill, and labor required to run existing ones. Criminal groups that previously needed specialists for phishing, malware development, data analysis, translation, negotiation, and reconnaissance can increasingly use models to automate parts of those workflows. That makes security teams’ response speed increasingly important. Enterprises are responding by introducing AI-driven security operations of their own, creating what is becoming an automation race between attackers and defenders. The trend also reinforces concerns about autonomous agents because systems with direct access to browsers, code execution, credentials, and cloud infrastructure can have far greater consequences when compromised or misused.

Why It Matters: AI is lowering operational barriers for cybercriminals, forcing enterprises to defend against attacks that can be researched, customized, and scaled far faster than traditional security processes were built to handle.

Source: Zscaler.