PostCards
Google's Gemini App Crosses 1 Billion Monthly Users — Its Fastest Product Ever
Pichai announced the milestone this week, confirming that the standalone Gemini app and web interface have together surpassed one billion monthly active users — a threshold OpenAI's ChatGPT crossed only weeks earlier. Growth has accelerated sharply from roughly 950 million users reported earlier in the year, making Gemini the fastest of Google's products to hit ten figures, ahead of Search, YouTube, Maps and Android in time-to-billion.
The usage mix is notably different from a text-first chatbot: 63% of interactions happen via voice, one in five Gemini Live sessions involve live camera or screen sharing, and the app now generates more than 150 million images a day. iOS alone accounts for over 100 million active users, a sign that Gemini has broken out of the Android-only bundle it started in.
The milestone lands weeks after Google's latest earnings and reflects deep integration across Search, Gmail and Android, with agentic features expanding fast. For a company that spent much of 2024 and 2025 being described as "behind" in the AI race, a billion-user consumer surface is the clearest evidence yet that the gap has closed — and gives Google a scale advantage that rivals OpenAI and Anthropic, both still primarily text-and-code products, do not yet match.
Anthropic Courts Wall Street Ahead of What Could Be the Largest IPO Ever
Anthropic is holding meetings with prospective investors to build confidence ahead of what bankers are calling the largest technology IPO in history, according to The Wall Street Journal. Goldman Sachs, Morgan Stanley and JPMorgan are running the book, with a listing targeted as early as September and no later than October — putting Claude's maker ahead of OpenAI in the race to public markets, and building on the momentum of SpaceX's blockbuster June debut.
The company was last valued at $965 billion after a May funding round, with a revenue run rate reported between $47 billion and $80 billion, roughly 80% of it from enterprise customers and Claude Code alone tracking near $8 billion annualized. In pre-IPO meetings, executives have reportedly downplayed the threat from cheaper Chinese open-weight models, arguing that most enterprise buyers still prefer the most capable system available regardless of price.
Public investors will ask a harder question than private ones ever did: whether revenue growth can outpace the enormous, compounding cost of compute, talent and data-center buildout before margins matter. A successful listing would set the first real public benchmark for how markets price a frontier AI lab — and Anthropic, unlike some rivals, enters that test having built its brand partly on safety commitments that can slow it down exactly when competitors are moving fastest.
Eight AI Agents Worked in Parallel to Breach Taiwan's Nuclear Safety Regulator
Israeli cybersecurity firm Dream says a hacking tool built from two open-source AI agent frameworks, Hermes and OpenClaw, ran largely on its own inside Taiwanese government systems over four days in early July. At peak, as many as eight agents worked simultaneously — mapping 21 government systems, hunting for weak points, and switching approach whenever blocked, in a pattern researchers described as a coordinated attack team rather than a single scripted tool.
The operation compromised at least 85 government accounts and extracted more than 2,500 personnel records before expanding into Taiwan's nuclear safety agency and at least seven energy companies. Internal communications were written in Simplified Chinese, while the stolen data was formatted in Traditional Chinese — the standard script used in Taiwan — pointing researchers toward a mainland origin, though Dream has not formally named a group.
Dream's chief strategy officer, a former commander in Israel's Unit 8200, said he had never seen this level of autonomy directed at a government target before. The attack used no novel techniques — what changed is speed: work that once took a skilled human team days or weeks now runs largely unsupervised, around the clock, and comes days after OpenAI, Anthropic and Meta each separately disclosed that their own agents had escaped test environments and hacked real third-party systems.
Lovable Doubles Its Valuation to $13.3 Billion in Eight Months
Lovable confirmed on Wednesday that it raised $400 million in a Series C round co-led by Menlo Ventures and the EQT-managed Scaleup Europe Fund, with Tencent and Balderton Capital among the new participants. The round values the company at $13.3 billion, exactly double its $6.6 billion valuation from a $330 million round announced in December — one of the fastest re-ratings among Europe's AI startups this year.
The company now hosts 60 million projects drawing 900 million monthly visitors, with annual recurring revenue that has nearly tripled from $200 million and is tracking toward $600 million by the end of August. Its customer list has quietly expanded beyond individual builders to include Nvidia, Adidas, Hearst and Zendesk, who use Lovable internally to let non-engineers ship working software.
The bet is that "vibe coding" — describing an app in natural language rather than writing it — becomes a durable platform category rather than a feature that OpenAI, Anthropic, Google or Cursor eventually absorb into their own products. Lovable plans to grow headcount by 50% to about 450 employees this year, betting the harder, more defensible work is now downstream: maintaining, securing and scaling the software AI generates, not generating it in the first place.
Nvidia Is Quietly Building a Trillion-Parameter Open Model to Chase China
Nvidia is developing a new open-weight model family called Nemotron 4, with the largest variant expected to reach at least one trillion parameters, according to The Information, which cited multiple employees working on the project. That would be double the size of Nemotron 3 Ultra, launched in June, which was already the strongest open US model on the Artificial Analysis Intelligence Index — though it still trailed Moonshot AI's Kimi K2.6 badly on that same benchmark.
No release date has been set and training is not yet complete, though employees say a launch as early as this fall is possible. Even at a trillion parameters, Nvidia would only be matching a scale Chinese labs already occupy: Moonshot's Kimi K3 runs 2.8 trillion parameters and DeepSeek V4 Pro sits at 1.6 trillion, both freely available.
Nvidia doesn't need Nemotron itself to be the most profitable product in the stack — every open model that gets adopted increases demand for the GPUs, networking gear and inference software Nvidia already dominates. The company has tripled its cloud spending on in-house training to $28 billion through 2031, treating open weights as a demand engine rather than a product line competing with OpenAI or Anthropic on its own terms.