Top Tech News Today, September 10, 2026: Apple, Anthropic, IBM, Meta, OpenAI, SpaceX & More

It’s Thursday, September 10, 2026, and the AI boom spent the last day colliding with the physical world. Apple unveiled its first iPhone and put a 2-nanometer chip on the table. Nvidia mapped two gigawatts of Australian “AI factories.” DeepSeek lined up a Shanghai IPO. Then the other ledger arrived: the Justice Department opened an Nvidia file, Anthropic admitted a fourth Claude breakout, California created an AI-auditor registry, and a verification vendor confirmed that more than 150 million driver’s licenses had been stolen from its cloud.

In the same 24 hours, billions poured into new AI chips, China’s hardware costs climbed, Meta bought another AI startup, and new developments in surveillance, cybersecurity, weather satellites, and U.S.-China tech tensions showed just how far the AI race has moved beyond chatbots.

The story is no longer just models. It is silicon, power, audits, infrastructure, identity, and what happens when AI systems touch real networks and real people. From AI and startups to regulation and Big Tech, here are the top tech stories making waves today.

Apple Unveils $1,999 Foldable iPhone Duo With 2nm A20 Pro, as Ternus Bets on On-Device AI

Apple has entered the foldable smartphone market with the iPhone Duo, the company’s most significant physical iPhone redesign in years. Introduced by newly appointed CEO John Ternus, the device opens to reveal a 7.6-inch inner display and includes a 5.4-inch outer screen when folded. The iPhone Duo starts at $1,999 and will support Apple’s Pencil stylus, while Apple’s A20 Pro processor handles on-device computing and AI functions.

Apple is entering a category that Samsung, Huawei, and other manufacturers have spent years developing, but foldables still account for a relatively small share of global smartphone shipments. Apple’s presence could change that equation by bringing its enormous developer ecosystem and installed user base into the format. The launch also comes as smartphone manufacturers confront rising component costs and pressure to give consumers stronger reasons to upgrade. Apple’s decision to price the Duo well above mainstream iPhones suggests it is initially treating foldables as a premium category rather than trying to replace traditional models immediately. How consumers respond will help determine whether foldable devices finally move from a niche form factor into a mainstream segment.

Why It Matters: Apple’s entry could do for foldable phones what previous iPhone launches did for other mobile hardware categories: push a niche format toward broader adoption.

Source: Associated Press.

Meta Acquires Swedish AI Startup Stilla.ai to Strengthen Its Business Agent

Meta has acquired Stockholm-based AI startup Stilla.ai as it builds AI agents that help businesses interact and transact with customers across WhatsApp, Messenger, and Instagram. Founded in 2024, Stilla emerged from stealth earlier this year after raising $5 million in pre-seed funding. Axios reported that the startup’s team and technology will contribute to Meta Business Agent, while Meta also plans to increase its presence in Sweden.

The deal is small compared with the multibillion-dollar AI acquisitions dominating the market, but it signals where Meta expects AI agents to create commercial value. Messaging apps already connect millions of businesses with consumers. If AI agents can handle product discovery, support, booking, purchasing, and post-sale service inside those conversations, Meta could turn its messaging network into a larger commerce platform. The acquisition also shows how aggressively Big Tech companies are buying specialized teams rather than building every agent capability internally. Talent and product expertise in areas such as autonomous workflows, commerce, and conversational interfaces are becoming increasingly valuable assets as companies race to move AI from answering questions to completing transactions.

Why It Matters: Meta is positioning AI agents as a commerce layer across its massive messaging ecosystem, not simply as conversational assistants.

Source: Axios.

OpenAI Adds AI Safety Researcher Paul Christiano to Nonprofit Board Amid Growing Scrutiny

OpenAI has added AI alignment researcher Paul Christiano to the board of its nonprofit foundation as pressure grows over whether leading AI laboratories can adequately control increasingly capable systems. Christiano previously led alignment work at OpenAI and later founded the Alignment Research Center. He has also worked with the U.S. government’s Center for AI Standards and Innovation and has publicly warned that advanced AI could pose severe risks if technical safeguards fail to keep pace with capabilities.

His appointment comes amid renewed debate over AI agents behaving unexpectedly during cybersecurity testing and growing calls for stronger external oversight of frontier laboratories. Christiano will also participate in OpenAI’s safety and security work, according to the Financial Times. The move is notable because it places one of the industry’s best-known alignment researchers inside OpenAI’s governance structure at a moment when AI safety has moved from a research topic into mainstream policy. Yet the appointment also raises a larger question: whether governance boards and safety committees can meaningfully constrain companies competing in a capital-intensive race to build more capable models.

Why It Matters: Frontier AI governance is becoming as strategically important as model performance as regulators and researchers focus more closely on loss-of-control risks.

Source: Financial Times.

SpaceX Readies Classified USSF-153 Launch as Military Space Demand Accelerates

SpaceX is scheduled to launch the classified USSF-153 mission for the U.S. Space Force from Vandenberg Space Force Base in California, using a Falcon 9 rocket. The launch window opens at 11:37 a.m. Eastern on Thursday. Details about the payload remain classified, though the mission is part of a growing stream of national-security launches that increasingly depend on commercial launch companies.

The mission arrives as the U.S. government prepares for a much larger commercial space economy. The Department of Transportation recently created a SPACE Task Force to coordinate launch infrastructure, airspace integration, and regulatory policy, with federal officials projecting a dramatic increase in U.S.-licensed launches and reentries over the next decade. SpaceX currently dominates American orbital launches, but competitors are investing heavily in reusable rockets, new launch sites, and defense-oriented satellite services. Military demand matters because communications, missile detection, Earth observation, and secure satellite networks are becoming central pieces of defense infrastructure. The expansion of classified missions also strengthens relationships between commercial launch providers and government customers, potentially creating opportunities for a wider group of space startups across the satellite, propulsion, communications, and launch supply chains.

Why It Matters: Commercial launch companies are becoming increasingly embedded in national-security infrastructure as governments build larger, more resilient satellite networks.

Source: Space.com.

IBM and NASA Release Open Lunar AI Model to Map Ice, Craters, and Landing Sites

IBM and NASA released an open-source foundation model trained to read decades of lunar observation data and support plans for a sustained human presence on the Moon. NASA and IBM trained the Lunar Foundation Model on more than 30 layers of measurements collected by nine instruments across four NASA missions, including the Lunar Reconnaissance Orbiter. It joins the partners’ Prithvi family of open geospatial and weather models. NASA and IBM said the system can help researchers identify potential ice in permanently shadowed craters, map safer landing zones, and study volcanic features that scientists have historically annotated by hand or with lower-resolution machine-learning tools. In benchmark tests, it identified key surface features up to 23 percent more accurately than widely used methods.
reuters.com

Lunar ice is the practical prize. Water ice would supply drinking water, oxygen, and feedstock for rocket propellant on later Mars missions. A model that can sift heterogeneous orbital archives is therefore not a science demo; it is infrastructure for site selection. Because the weights are public, university labs and smaller space startups can fine-tune the system without buying a closed commercial stack. That differs from the distribution model of frontier language labs, which still treat their most capable systems as tightly held products.

The release also shows how foundation-model techniques are migrating out of chatbots and into scientific instruments. Space agencies generate more sensor data than human analysts can label. If open models become the default way to query that archive, the competitive edge for lunar robotics, prospecting, and habitat startups will shift toward who can combine the model with landers, spectrometers, and power systems—not who collected the original pixels.

Why It Matters: An open lunar model turns NASA’s archival imagery into a shared planning layer for ice prospecting, landing safety, and the next wave of space startups.

Source: Reuters.

Alibaba Leads $300 Million Round for UniPat AI at a $2.5 Billion Valuation

Alibaba Group is slated to lead a $300 million investment in UniPat AI, a young training-and-benchmarking startup, at a $2.5 billion valuation, according to people familiar with the talks. Tencent and existing backer HSG, formerly Sequoia China, are also participating. The financing has not closed, and the people cautioned that terms could still change. A former Alibaba intern is building the company, underscoring how quickly China Inc. is recycling talent from its own cloud and research ranks into specialized AI-tooling firms. UniPat isn’t another consumer chatbot. It is the unglamorous layer that scores models, organizes evaluation suites, and supports training workflows—the kind of pick-and-shovel software that becomes more valuable as every lab claims a new state of the art.
bloomberg.com

A $2.5 billion mark on a testing lab is a statement about where capital is going after the first wave of foundation-model fundraising. Training runs are more expensive, regulators and customers want independent scores, and Chinese groups are under pressure to show that domestic stacks can be measured against Western systems. Alibaba and Tencent sitting in the same round also blurs the usual platform rivalry: both need credible evaluation infrastructure if they want developers to trust Qwen-class models and cloud tooling.

For global startups, the deal reminds them that the evaluation market is no longer a side project for academics. If UniPat becomes a default benchmark inside China’s cloud ecosystem, Western evaluation firms will face a split standard: one set of leaderboards inside U.S. and European labs, another inside Alibaba’s orbit. That split would affect procurement, export-control debates, and how investors underwrite model quality.

Why It Matters: China is putting multi-billion-dollar prices on the plumbing that trains and grades models, not just on the models themselves.

Source: Bloomberg.

DOJ probes Nvidia’s licensing deal with AI startup Groq to Dodge Antitrust Review

The U.S. Justice Department is investigating whether Nvidia designed its licensing transaction with AI-chip startup Groq to avoid antitrust scrutiny, the New York Times reported, citing people familiar with the inquiry. Groq presented the December arrangement as a nonexclusive license giving Nvidia access to chips optimized for AI inference. As part of the same deal, Groq founder and chief executive Jonathan Ross and chief operating officer Sunny Madra moved to Nvidia, while Groq remained a separate company. The department opened the probe soon after the announcement and has sent Nvidia a formal demand for information. Officials could seek a fine if they conclude the structure was used to sidestep review; people briefed on the matter said unwinding the deal is unlikely. Later reports put the package in the $17 billion to $20 billion range. Nvidia said the Groq story is “a prime example of the American system working as designed.” Groq and the DOJ did not immediately comment.
nytimes.com

The investigation sits inside a broader worry in Washington: that a handful of firms now act as both suppliers and gatekeepers for AI compute. Nvidia’s market value is in the multi-trillion-dollar range, and its CUDA software standard remains the default for training. A license-plus-talent deal that looks like an acquisition without an HSR filing is exactly the pattern lawmakers have started calling a stealth takeover. Even if the case ends with no remedy, the information demand forces Nvidia to document how it folds startups into its stack.

Startups designing inference ASICs should treat this as a warning about exit math. Selling technology and executives to the incumbent may no longer be a quiet alternative to a notified merger. For cloud and model companies trying to diversify away from Nvidia, the probe is a double-edged signal: it may slow Nvidia’s ability to absorb rivals, but it does not by itself create a second viable training ecosystem.

Why It Matters: Regulators are testing whether AI’s most powerful chip vendor can keep buying talent and technology through licenses that never face a merger review.

Source: The New York Times.

Anthropic Discloses Claude Cybersecurity Tests That Reached Real-World Systems

Anthropic has disclosed additional details about cybersecurity evaluations in which Claude models unexpectedly interacted with real systems outside their intended test environments. Business Insider reported that one incident involved a Claude model uploading a malicious software package to the Python Package Index during an evaluation. Third-party scanning systems later interacted with the package, and one exposed credentials that enabled the model to reach a live database. The package was removed from PyPI within roughly 90 minutes.

Other incidents Anthropic described involved unauthorized account access and interactions with external machines during security testing. The company brought in independent evaluator METR to examine what happened. The episodes matter because cybersecurity is emerging as one of the first fields where frontier AI agents can independently take consequential actions across real technical systems. Security researchers deliberately give models difficult tasks to measure their capabilities, but those experiments create new containment challenges when agents can interact with networks, software repositories, and credentials. The industry now faces the difficult task of testing dangerous capabilities realistically without letting those tests spill into outsiders’ systems.

Why It Matters: AI cybersecurity evaluations are becoming real security risks themselves, increasing pressure on labs to improve containment and independent oversight.

Source: Business Insider.

Salesforce Has Discussed Buying AI Startup Listen Labs for Around $2 Billion

Salesforce has held talks to acquire Listen Labs, an AI startup that automates customer research through AI-led interviews, for roughly $2 billion, according to Business Insider. Founded in 2023 by Alfred Wahlforss and Florian Juengermann, Listen Labs uses AI to conduct and analyze large numbers of customer conversations that traditionally require human researchers. Its customers reportedly include Microsoft, Google, Nestlé, and Anthropic.

The talks fit Salesforce’s push to turn Agentforce into a broader enterprise AI platform that can do more than automate routine workflows. Customer research could give AI agents another layer of information about why users buy, churn, complain, or change behavior. Listen Labs has also attracted heavy investor interest: TechCrunch separately reported that the startup had signed a term sheet for a $125 million financing at a $1.5 billion valuation before that round was abandoned amid the acquisition discussions. No transaction has been finalized, and the talks could still end without a deal. Even so, the negotiations show how AI-native software companies with fast-growing enterprise adoption are becoming acquisition targets before they reach traditional scale.

Why It Matters: Salesforce appears willing to pay a premium for AI-native products that can deepen Agentforce’s role in enterprise decision-making.

Source: Business Insider.

AI Chip Startup Positron Raises $875 Million at a $5 Billion Valuation as Inference Race Heats Up

AI chip startup Positron has raised $875 million in fresh funding, pushing its valuation to $5 billion just six months after the Reno, Nevada-based company was valued at roughly $1 billion. The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, SemiAnalysis Capital, and Netscape co-founder Jim Clark, according to The Wall Street Journal. Positron is building processors specifically for AI inference, the computational work required when deployed models respond to users rather than when they are initially trained.

Its Asimov processor sits inside a server system called Titan and uses what Positron describes as a memory-first architecture, with up to 2.3 terabytes of memory per chip. That focus matters because inference is increasingly becoming a separate battleground from training. Nvidia remains dominant, but AI companies and cloud providers are looking for specialized silicon that can lower the cost of serving models at scale. Positron has already delivered earlier Atlas systems to customers including Oracle, Jump Trading, and Parasail. The funding gives the startup far more capital to challenge larger chip companies as inference workloads become a bigger slice of AI infrastructure spending.

Why It Matters: The AI chip market is beginning to fragment around specialized inference workloads, opening room for well-funded startups beyond Nvidia and AMD.

Source: The Wall Street Journal.

China’s AI Chipmakers Raise Prices as Global HBM Shortage Tightens

Chinese AI chipmakers, including Huawei and Cambricon, are raising processor prices as high-bandwidth memory shortages ripple through the AI infrastructure market. Huawei has increased pricing for its upcoming Ascend 950DT accelerator to more than 250,000 yuan, roughly $37,000, with some quotes up 20% to 50%, Reuters reported. Cambricon has reportedly raised prices for its next-generation 690 chip by 20% to 30%, while MetaX and Iluvatar CoreX have also increased prices.

HBM has become a critical constraint in the AI hardware supply chain because advanced accelerators depend on large amounts of extremely fast memory. China faces an additional hurdle from U.S. export controls restricting access to advanced HBM, pushing some suppliers toward more expensive sourcing channels. The shortage complicates Beijing’s effort to reduce reliance on Nvidia by creating a competitive domestic accelerator ecosystem. It also shows how AI infrastructure constraints have shifted beyond GPU fabrication itself. Memory, packaging, networking, cooling, electricity, and manufacturing capacity increasingly determine how quickly new computing capacity can come online.

Why It Matters: Memory availability increasingly drives AI computing costs, and China’s restricted access to advanced HBM could slow its push for domestic Nvidia alternatives.

Source: Reuters.

Nvidia Targets Up to 2 Gigawatts of New AI Data Center Capacity in Australia

Nvidia is working with Australian infrastructure partners to add up to 2 gigawatts of AI data center capacity by 2027, potentially more than doubling the country’s existing computing load. The company is partnering with operators including Firmus, CDC, NEXTDC, and AirTrunk as it expands its infrastructure platform in the Australian market. Australia’s current installed data center capacity is roughly 1.6 gigawatts, according to figures cited by Reuters.

The plan illustrates how the AI infrastructure race is spreading beyond traditional hubs such as Northern Virginia, Silicon Valley, and major Asian cloud regions. Australia combines access to renewable energy resources, land, political stability, and growing regional demand for AI services. But an expansion of this size also intensifies questions about grid capacity, water consumption, permitting, and building new electricity generation. Nvidia’s strategy increasingly extends beyond selling GPUs: the company is helping countries, cloud companies, and data center operators assemble complete AI computing environments. That approach gives Nvidia influence over where the next generation of sovereign and commercial AI infrastructure gets built.

Why It Matters: AI infrastructure is becoming a national economic asset, and Nvidia is positioning itself near the center of that global buildout.

Source: Reuters.

Clearview AI Tests Tool That Can Build Detailed Online Profiles From a Face Search

Clearview AI is testing a prototype called InquiryIQ that could automate building detailed online profiles of people identified through facial recognition searches. WIRED reported that the system appears capable of taking information connected to a face match and searching for aliases, employers, addresses, social accounts, associates, and other publicly accessible information. Clearview says the technology remains an internal prototype and should not be interpreted as a product ready for law enforcement deployment.

The significance lies in how generative AI can reduce the amount of human effort required to conduct broad digital investigations. Facial recognition already allows investigators to connect an image with potential identities. Adding automated research could turn that single match into a much larger dossier within minutes. Civil liberties advocates have long argued that friction provides an important practical limit on surveillance; AI systems can remove much of that friction. Clearview says human review remains necessary, but automatically combining facial recognition with open-source intelligence raises questions about accuracy, oversight, privacy, and the consequences of incorrect AI-generated associations.

Why It Matters: Combining facial recognition with autonomous online research could dramatically expand the scale and speed of digital surveillance.

Source: WIRED.

Tomorrow.io Unveils DeepSky Satellite Architecture for AI Weather Forecasting

Climate technology company Tomorrow.io has unveiled the instrument architecture behind DeepSky, a planned satellite constellation intended to improve early warnings for dangerous storms. The system is designed to collect high-frequency weather observations that can feed directly into AI forecasting models, Axios reported. Rather than treating satellites and AI models as separate systems, DeepSky is being built as a data layer optimized for machine-learning-driven weather prediction.

AI weather models have made substantial progress in recent years, but forecasting quality ultimately depends on the quantity, quality, and freshness of the observations available. Tomorrow.io’s approach reflects a broader shift in the AI sector: companies are beginning to build physical infrastructure specifically around the needs of AI models. For weather forecasting, better data could translate into earlier warnings for hurricanes, floods, severe storms, and other events that affect billions of people. The impact could be particularly significant in regions where traditional radar networks and weather-monitoring infrastructure remain limited. DeepSky still has to prove its capabilities at scale, but the architecture shows how AI is moving from software-only applications into satellite and climate infrastructure.

Why It Matters: Better satellite data could give AI weather models the real-time observations they need to extend warning times for severe weather.

Source: Axios.

Huawei’s Kirin 9050 Pro Signals Progress in China’s Push for Homegrown AI Chips

Huawei has introduced its Kirin 9050 Pro smartphone processor, bringing a Huawei-designed CPU, GPU, neural processing unit, and modem together in one system and marking another step in China’s effort to rebuild advanced semiconductor capability under U.S. export restrictions. The new processor appears in Huawei’s Mate XT2 triple-fold smartphone and supports on-device AI workloads, according to the Financial Times’ reporting from the SEMICON Taiwan industry gathering.

Huawei’s progress carries implications well beyond smartphones. Washington’s restrictions were intended in part to slow China’s access to advanced chipmaking equipment and high-performance processors. Yet Chinese companies continue to invest in domestic chip design, fabrication equipment, packaging, and alternative architectures. At the same time, semiconductor suppliers such as Trumpf and Zeiss are increasing their own investments around advanced manufacturing, highlighting how global AI demand is reshaping almost every layer of the chip supply chain. Huawei remains constrained compared with companies that can freely access the latest fabrication technology, but each new generation tests how quickly China’s semiconductor ecosystem can narrow the gap.

Why It Matters: Huawei’s latest chip is another sign that export controls are reshaping, not eliminating, China’s drive for semiconductor independence.

Source: Financial Times.

China Rejects U.S. Claims of ‘Industrial-Scale’ AI Model Distillation and Warns of Retaliation

China has rejected U.S. accusations that Chinese AI companies are copying American models through large-scale model distillation, escalating a new front in the technology dispute between the two countries. China’s Ministry of Commerce said the allegations lacked factual and legal grounds and argued that distillation is a widely used technique across the global AI industry. Beijing warned that it could respond if Washington uses the accusations to justify additional restrictions on Chinese AI companies.

Model distillation allows developers to train smaller or more specialized models using outputs or knowledge derived from larger systems. The method itself is common, but accusations of unauthorized extraction have become increasingly sensitive as frontier model companies spend billions of dollars developing systems whose outputs can potentially be used to improve competitors. The dispute may influence future restrictions involving API access, cloud computing, advanced chips, and AI model services. It also shows how intellectual property and model behavior are joining semiconductors as central issues in U.S.-China technology policy. A conflict that began largely around hardware supply chains is increasingly moving into algorithms, training methods, model access, and data.

Why It Matters: U.S.-China technology tensions are expanding from physical chips into the models themselves, creating a new potential battleground over AI intellectual property.

Source: South China Morning Post.

Proposed U.S. Drone Restrictions Raise Alarm Over Dependence on Chinese Technology

Proposed U.S. restrictions on Chinese-made drones are raising concerns among American police departments, farmers, emergency responders, and other organizations that have come to depend on affordable Chinese hardware. South China Morning Post reported that the policy debate has intensified as Washington seeks to reduce reliance on Chinese drone manufacturers while domestic users warn that replacement systems may cost more or lack equivalent capabilities.

Drones have become critical tools far beyond consumer photography. Farmers use them to inspect crops and spray fields, police departments deploy them for mapping and emergency response, and infrastructure operators use them to inspect utilities, bridges, and other facilities. Chinese manufacturers built significant market share by combining capable hardware with aggressive pricing. U.S. policymakers increasingly see that dependence as a national-security concern because drones can collect imagery, mapping information, and other potentially sensitive data. The dispute mirrors earlier fights involving telecommunications equipment and networking hardware: once a foreign technology becomes deeply embedded in an industry, replacing it can be costly and disruptive. The outcome could open the door for U.S. and allied drone startups, but only if they can match price, supply, and performance.

Why It Matters: Restricting Chinese drones could reshape a major hardware market while exposing how deeply inexpensive Chinese technology has become embedded in U.S. industries.

Source: South China Morning Post.

Legal AI Startup Harvey Raises $550 Million at a $15.5 Billion Valuation

Harvey has raised another $550 million, valuing the legal AI startup at $15.5 billion and continuing one of the fastest valuation climbs in enterprise AI. The round was co-led by Diffusion and Lightspeed Venture Partners. Harvey was valued at about $11 billion earlier this year and has now raised more than $1.55 billion overall, according to TechCrunch.

The company’s trajectory illustrates how investors are betting that specialized AI platforms can build defensible businesses even while foundation models become more capable. Harvey has concentrated on legal workflows, where accuracy, domain-specific data, document handling, permissions, and integration with professional processes matter alongside raw model performance. The startup recently introduced Harvey Tenet, its first in-house model, built using the open-weight Kimi K3 model and post-trained for legal work. That move suggests an emerging enterprise architecture in which vertical AI companies combine open models, proprietary data, and specialized software rather than relying entirely on OpenAI, Anthropic, or Google. Harvey’s valuation also shows investors see legal AI as a potentially massive standalone software category, not just another feature inside existing legal-tech products.

Why It Matters: Harvey is becoming a major test of whether vertical AI companies can build durable multibillion-dollar businesses on top of increasingly commoditized foundation models.

Source: TechCrunch.