Top AI News Today: Sept 2 (13 Stories)

September 2 AI Wrap: Pentagon's ChatGPT and Grok Rollout Tops a Heavy News Day

Government and policy news dominated today. The Pentagon flipped the switch on the biggest expansion yet of its internal AI platform, giving more than 3 million military and civilian personnel access to tailored versions of ChatGPT and Grok. Hours later, the European Commission classified ChatGPT as a Very Large Online Search Engine under its toughest digital rules, the first time a generative AI chatbot has been placed in that category. And Anthropic locked in a 35 billion dollar cloud computing deal tied to a Texas data center, on top of 45 billion dollars it committed just days earlier.

On the product side, Runway unveiled a strange new kind of AI that generates entire software interfaces frame by frame instead of running code, Alibaba pushed its video generator to 30 seconds of narrated footage in one pass, and Google spent the day responding to criticism over inconsistent safety advice from its AI Overviews. Here are the 13 stories that matter most from today's AI news, explained in plain English.

The Pentagon adds ChatGPT and Grok for 3 million personnel

The Department of Defense expanded its GenAI.mil platform on August 31, 2026, adding OpenAI's ChatGPT Mil and xAI's Grok for Government alongside the Google Gemini tool that started the platform in December 2025. The rollout gives more than 3 million military and civilian personnel access to commercial grade AI tools built for unclassified defense work, on top of the roughly 1.7 million people already using the

platform.

GenAI.mil works like a secure, government run version of these popular chat tools, so staff can get AI help drafting policy, reviewing procurement documents, and managing logistics without sending sensitive government data through ordinary consumer apps. ChatGPT Mil is aimed at document heavy administrative work, while Grok for Government is pitched at operational tasks like supply chain management and market research for acquisition staff. Both tools cleared Impact Level 5 accreditation, the Pentagon's standard for handling sensitive but unclassified information.

Notably absent from the expansion is Anthropic's Claude, which was part of the same 2025 prototype awards but has not joined GenAI.mil. That gap lines up with the Pentagon's earlier decision to blacklist Anthropic as a supply chain risk, a designation a federal judge struck down as illegal just days before this rollout, though the Defense Department has said it still plans to finish removing Claude from its systems by the end of September regardless of the ruling.

The EU classifies ChatGPT as a Very Large Online Search Engine

The European Commission designated ChatGPT a Very Large Online Search Engine under the EU's Digital Services Act on September 1, 2026, the first time the bloc has placed a generative AI chatbot in that category. The designation applies because ChatGPT reported roughly 159 million average monthly users in the EU, well above the 45 million user threshold that triggers the rule.

The Commission's reasoning is specific: ChatGPT counts because it can search the web and return information gathered from online sources, not simply because it is a popular AI product. That capability based logic matters beyond OpenAI, since it gives regulators a template they could apply to Google's Gemini, Anthropic's Claude, or Perplexity once those tools cross the same usage threshold inside the EU.

OpenAI now has about four months to comply with the stricter rules that come with the designation, including annual risk assessments, independent audits, stronger protections for minors, and giving vetted researchers access to how the system works. Failing to comply can cost up to 6 percent of a company's global annual revenue, and the EU has already fined other large platforms hundreds of millions of euros under the same law, giving OpenAI a clear sense of what enforcement could look like if it misses the deadline.

Anthropic signs a 35 billion dollar cloud deal tied to a Texas data center

Anthropic signed a 35 billion dollar computing deal with Lambda, a cloud provider backed by Nvidia, on August 31, 2026, to secure GPU capacity at a new Texas data center. The facility, in Nueces County, is being built by Hut 8, a company that switched from bitcoin mining into AI data centers, and covers roughly 350 megawatts of computing power that Anthropic plans to use for training and running its Claude models.

The deal has an unusual structure worth understanding: Nvidia supplies the chips, has invested directly in Lambda, and reportedly holds the actual lease on the data center building itself. That means Nvidia is playing three roles at once in a single transaction, as chip seller, cloud investor, and landlord, which is exactly the kind of arrangement that has raised eyebrows among investors worried about circular deals where the same companies both sell computing power and help finance the businesses buying it.

This Lambda agreement comes on top of a separate 45 billion dollar deal Anthropic struck just days earlier to rent capacity from Nscale's West Virginia data center campus, bringing Anthropic's new infrastructure commitments to roughly 80 billion dollars in about a week. As a company preparing for what is expected to be a large IPO, Anthropic is racing to lock in enough computing power to meet growing demand for products like Claude Code, even as questions swirl about whether this level of interlinked AI spending is financially sustainable.

Runway unveils Solaris, software that generates itself frame by frame

AI video company Runway introduced Solaris on August 31, 2026, the first model in what it calls the Interface World Model family. Instead of writing code that a browser or app then renders, Solaris generates a working software interface frame by frame in real time, responding directly to a person's clicks, drags, or voice commands.

Ordinary software works in stages: a designer sketches a layout, a developer turns it into code, and the resulting app stays fixed until someone ships an update. Solaris collapses all of that into one step by rendering the interface itself as a continuously generated video, similar to how Runway's existing video generation tools work, so the

buttons, menus, and screens a person sees are produced live rather than pulled from a pre-built template.

Runway says Solaris can already handle things like dragging a shirt onto a photo of yourself or rearranging furniture in a room, and that it performs better than leading language models at recreating a website's look from a single screenshot. The system is still an early research preview with real limits, including error prone text rendering and no support yet for screen readers, and Runway has not announced pricing or a public launch date.

Alibaba's WAN 3.0 generates 30 seconds of narrated video in one pass

Alibaba's video generation model WAN 3.0 can now produce up to 30 seconds of 1080p video with synchronized audio in a single generation pass, according to pricing and access details that appeared on the fal platform this week. The model is priced at 0.05, 0.10, and 0.20 dollars per second depending on quality tier, with an accelerated Prime version available at roughly a third less.

Thirty seconds may sound short, but it is a meaningful jump for AI video generation, where most competing tools still top out at just a few seconds per clip before needing to be stitched together manually. Generating audio in the same pass as the video, rather than adding it afterward as a separate step, also means the sound and visuals are created together, which tends to produce more natural timing between what is shown on screen and what is heard.

Alibaba has not published independent benchmark comparisons for WAN 3.0 yet, so how it stacks up against video tools from Runway, Google, or OpenAI on quality remains untested outside the company's own demos. Still, the combination of longer clips, built in audio, and Alibaba's usual aggressive pricing puts pressure on Western video generation tools that have generally charged more for shorter, silent output.

Google works to fix AI Overviews after inconsistent safety warnings

Google spent the past two weeks responding to criticism after users found that its AI

Overviews search feature gave dramatically different safety advice depending on which nationality, religion, or ethnicity was mentioned in an otherwise identical prompt. Some versions of the prompt triggered emergency style warnings telling people to leave, find a safe location, or call authorities, while other, nearly identical prompts produced ordinary, friendly social advice.

Google publicly acknowledged on August 20, 2026 that the results were not what they should be and said engineering teams were actively working on improvements. Reporters who retested the prompts after that statement found that many nationality based versions began returning similar, harmless answers, though some unusual variations of the prompt reportedly still triggered odd safety warnings even afterward.

The episode is the latest in a string of complaints about AI Overviews producing inconsistent or unreliable answers, following earlier concerns about the feature giving incomplete health information without accounting for basic factors like age or sex. Because AI Overviews sits at the top of Google's search results and is seen by a huge share of everyday users, even occasional inconsistent answers can reach far more people than a similar mistake buried in a single chatbot conversation.

A security researcher shows how an AI coding agent can be tricked into running code it should not

A security researcher published findings this week describing a technique that can get an AI coding agent's most autonomous mode to execute code it was never supposed to run, without asking for permission first. The writeup describes a multi step method that steers the AI away from its normal, more cautious tools and toward a path where it ends up running attacker supplied code.

Coding agents that can act with less human oversight are useful precisely because they can carry out long, multi step programming tasks on their own, but that same independence is what makes this kind of research important: a technique that reliably tricks an agent into an unintended action in a sandboxed demo could, in theory, be adapted to cause real harm if a similar weakness turned up in a production system handling real code or data.

This kind of research is a normal, expected part of how AI labs harden autonomous coding tools before wider release, and responsible researchers typically report findings

like this to the company first so a fix can ship before the details go public. It also lines up with a broader theme this year: as coding agents get more independence, agent security has become one of the most actively discussed topics at AI industry events, including planned sessions at this year's TechCrunch Disrupt conference.

The EU commits 387.8 million euros to build a supercomputer on AMD chips

EuroHPC, the European joint undertaking that funds supercomputing projects, signed a 387.8 million euro contract on August 31, 2026 with Atos owned Bull to build a new supercomputer called LUMI-AI at a data center in Kajaani, Finland. The system will run on AMD's Instinct MI430X GPUs paired with sixth generation EPYC processors that have up to 256 cores each.

LUMI-AI is designed to give European researchers and companies roughly ten times the AI computing capacity and twice the general scientific computing throughput of the current LUMI supercomputer it will sit alongside. Government funded supercomputers like this matter because they give universities, startups, and public researchers access to large scale AI training power without needing to rent it from a private cloud provider, which is especially important for Europe as it tries to reduce its reliance on American and Chinese AI infrastructure.

The choice of AMD chips rather than Nvidia is notable on its own, since it shows a major public AI computing project betting on Nvidia's most direct competitor for high end AI hardware. Finland's Kajaani site already hosts the current LUMI system, one of Europe's most powerful supercomputers, so LUMI-AI extends an existing hub rather than building entirely new infrastructure from the ground up.

AI sales platform Clay raises funding at a 7 billion dollar valuation

Clay, an AI powered sales and marketing platform, is raising a new funding round led by Wellington Management at a valuation of roughly 7 billion dollars, according to sources cited by Axios this week. That is a sharp jump from the roughly 5 billion dollar valuation set in a January 2026 employee tender and more than double the 3.1 billion dollar valuation from a CapitalG led round last summer.

Clay's product uses AI to automate the research and outreach work that sales and marketing teams traditionally do by hand, such as finding the right contacts at a company, personalizing outreach messages, and enriching customer data from many different sources at once. The company has become one of the most talked about examples of what is often called AI native go to market software, meaning it was built around AI automation from the start rather than adding AI features onto an older product.

The rapid rise in valuation, more than doubling in about a year, mirrors a broader pattern this year of AI focused enterprise software companies raising money at valuations that are hard to justify with current revenue alone, betting instead that AI automation will keep expanding into more of how sales and marketing teams work. Investors are clearly willing to pay a premium for that bet, even as public market AI stocks have shown more volatility over the same period.

Clipto raises 15 million dollars for AI powered local media search

Clipto, a three year old startup that indexes local videos, audio, images, meetings, and documents for natural language search, raised 15 million dollars in an all equity round at a 250 million dollar valuation. The round was led by HSG, formerly known as Sequoia China, with participation from several other venture firms.

Clipto's product lets someone search across all of their own media and documents using plain language questions instead of exact file names or keywords, and it integrates directly with ChatGPT and Claude so that AI agents built on those platforms can query a person's or company's own content index as part of completing a task. Founder Henry Kang says the company reached 15 million dollars in annual recurring revenue at the start of 2026 and remains profitable on net income.

With more than 30 million users across over 100 countries, Clipto is part of a growing category of startups that do not build their own foundation models but instead make existing models like ChatGPT and Claude more useful by giving them structured, searchable access to a person's own data. As AI agents take on more real work, this kind of connective tissue between raw personal data and the models that reason over it is becoming its own competitive market.

A confidential filing ties a fatal Tesla crash to its driver assist system

A confidential filing with the National Highway Traffic Safety Administration confirms that Tesla's driver assist system was verified engaged when a Model 3 hit 104 miles per hour on a residential street in Clute, Texas on May 20, 2026, before leaving the road and catching fire. Local police had publicly attributed the crash to a medical episode, but the automation's involvement only surfaced through the federal disclosure.

The gap between the public explanation and what Tesla's own confidential filing shows matters because it feeds directly into an active NHTSA investigation into how Tesla reports crashes involving its driver assistance features. Federal regulators rely heavily on manufacturers to self report this kind of data accurately, so a documented case where the public story did not match internal records raises questions about how completely that reporting process is working.

The case also echoes concerns raised previously by some of Tesla's own AI safety trainers about how reliable the company's public safety statistics really are. Driver assistance and full self driving features remain one of the most consequential real world deployments of AI decision making, since mistakes or misreported incidents can directly affect whether regulators and the public trust the technology enough to let it keep expanding.

Huawei says its AI chip revenue will jump 60 percent this year

Huawei expects its AI chip revenue to reach roughly 12 billion dollars in 2026, a 60 percent jump from 7.5 billion dollars in 2025, based on orders already secured according to a Financial Times report. The surge comes even as Huawei's overall first half profit fell 36 percent, since the company is pouring an unusually large share of its revenue into AI research and chip development.

Huawei's research and development spending rose about 25 percent to roughly 121 billion yuan in the first half of 2026, close to 26 percent of its total revenue, an extraordinary figure for a hardware company. That spending reflects Huawei's push to build a fully self sufficient AI and semiconductor supply chain, reducing its dependence on Nvidia and other foreign chip makers after years of US export restrictions.

Part of the recent surge in orders is tied to DeepSeek's decision to optimize its V4 model specifically for Huawei's Ascend chips, which triggered a wave of procurement from major Chinese technology companies including Alibaba, ByteDance, and Tencent. Huawei is also reportedly planning to roughly double production of its most advanced Ascend 910C chip next year, a sign that China's push toward domestic AI chip production is gathering real momentum even as it comes at a real cost to Huawei's short term profitability.

Anthropic and OpenAI headline the AI stage at TechCrunch Disrupt

Anthropic and OpenAI will both appear on the AI Stage at TechCrunch Disrupt 2026, running October 13 through 15 in San Francisco, organizers confirmed this week. The stage, presented by Google for Startups, will focus on how AI is reshaping business models, exposing new security gaps, and creating entirely new categories of jobs.

Sessions planned for the event include a discussion of what enterprise AI security actually requires as AI systems start making autonomous decisions inside sensitive business systems, and a separate conversation specifically about the challenge of securing AI agents, which were never originally designed with strong security guarantees built in. Both topics line up directly with the kind of agent security research making headlines this week.

Having Anthropic and OpenAI, two of the most closely watched and often competitive AI labs, share a stage at the same startup conference is itself a small signal of how central AI has become to the broader startup and investor conversation, not just as a product feature but as the central question shaping how new companies are built, funded, and secured going forward.

Quick Recap

The Pentagon added ChatGPT Mil and Grok for Government to GenAI.mil for 3 million personnel.

The EU classified ChatGPT as a Very Large Online Search Engine under the Digital Services Act.

Anthropic signed a 35 billion dollar cloud deal with Nvidia-backed Lambda for a Texas data center.

Runway unveiled Solaris, an AI that generates software interfaces frame by frame with no code.

Alibaba's WAN 3.0 can generate 30 seconds of narrated 1080p video in one pass.

Google is working to fix AI Overviews after it gave inconsistent nationality based safety warnings.

A security researcher showed a technique for tricking an autonomous coding agent into unintended actions.

The EU committed 387.8 million euros to build the LUMI-AI supercomputer on AMD chips.

AI sales platform Clay is raising funding at a roughly 7 billion dollar valuation.

Clipto raised 15 million dollars for AI powered local media search that plugs into ChatGPT and Claude.

A confidential filing tied Tesla's driver assist system to a fatal Texas crash.

Huawei expects a 60 percent jump in AI chip revenue in 2026 despite falling profits.

Anthropic and OpenAI will both appear on the AI Stage at TechCrunch Disrupt 2026.

Frequently Asked Questions

What is the top AI news today?

The biggest story is the Pentagon adding ChatGPT Mil and Grok for Government to its GenAI.mil platform, giving more than 3 million military and civilian personnel access to the tools. Close behind is the EU classifying ChatGPT as a Very Large Online Search Engine.

Does the Pentagon use ChatGPT and Grok?

Yes. As of August 31, 2026, the Department of Defense added ChatGPT Mil and Grok for Government to its GenAI.mil platform, alongside the Google Gemini tool that was already there, giving access to more than 3 million personnel.

Is ChatGPT classified as a search engine in the EU?

Yes. On September 1, 2026, the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act, because it can search the web and had roughly 159 million average monthly users in the EU.

How much did Anthropic spend on cloud computing this week?

Anthropic signed a 35 billion dollar deal with Nvidia-backed Lambda for a Texas data center on August 31, 2026, on top of a separate 45 billion dollar deal with Nscale announced just days earlier, totaling roughly 80 billion dollars in about a week.

What is Runway Solaris?

Solaris is Runway's first Interface World Model, an AI system announced on August 31, 2026 that generates a working software interface frame by frame in real time instead of running code.

Recommended Blogs

How to Use Claude AI

How to Use Google Gemini

ChatGPT Free for Beginners 2026

Best AI Tools for Coding 2026

What Is Agentic AI

Learn AI in 5 Minutes a Day

Keeping up with AI does not require reading every release note yourself. Unrot delivers the day's biggest AI developments in a 5 minute daily lesson, written in plain English for beginners, students, and working professionals who want to stay current without the jargon.