Anthropic’s $6 billion bet on an AI that dreams up video in real time

😁 Hello, super humans! Anthropic just confirmed talks to make its biggest acquisition ever, and the target is not a chip company or a data center operator. It is a two-year-old Israeli startup whose AI generates a driving simulation, a virtual fitting room, or an open world one frame at a time, live, at tens of frames per second. Stick around for how a $6 billion deal traces back to a $0.02-per-second API.

📰 Quick Signals

• 🧠 AI: Google’s Gemini app crossed 1 billion monthly active users, becoming Google’s fastest-growing product ever and closing the gap with ChatGPT.
• 🤖 Robotics: AgiBot overtook Unitree as the world’s top humanoid robot vendor in H1 2026, even as Unitree’s Shanghai IPO closed 8,288 times oversubscribed on its retail tranche.
• 💻 Programming: TIOBE’s August index has Python holding No. 1, Rust climbing to No. 10, and MATLAB dropping out of the top 20 for the first time in over a decade.
• ⚡ Electronics: AMD formally notified add-in-board partners that Radeon GPU and GDDR6 memory kit prices will rise at least 10 percent starting in August.
• 📡 Telecom: AST SpaceMobile filed with the FCC to begin direct-to-cell satellite service in the UK using Vodafone’s licensed spectrum, calling commercial launch “imminent.”

🔍 The Big Story: Anthropic’s $6 billion bet on an AI that dreams up video in real time

If this deal closes, Anthropic’s largest acquisition ever will not buy it a flashier chatbot. It buys the plumbing that could make every model the company runs meaningfully cheaper to serve.

What happened: Anthropic PBC is in talks to acquire the Israeli AI startup Decart AI for about $6 billion, according to Fortune, in what would be Anthropic’s largest known acquisition ahead of its anticipated IPO. The deal has not closed and could still fall through, but people familiar with the talks say Decart’s team, including co-founders Dean Leitersdorf, Orian Leitersdorf, and Moshe Shalev, would join Anthropic’s inference and performance organization. Decart makes real-time “world models”: Lucy, for live video editing, and Oasis, for generating interactive scenes, plus DOS, a chip-efficiency stack Anthropic reportedly wants to help its own infrastructure absorb more demand.

The details: Decart’s models generate video autoregressively, one frame at a time, each frame conditioned on everything generated before it, instead of rendering a whole clip offline the way most diffusion video models do. Every frame costs roughly 8,000 tokens, and at real-time frame rates that adds up to hundreds of thousands of tokens per second, a compute bill Decart’s DOS optimization stack exists to shrink. CEO Dean Leitersdorf told TechCrunch that vertical integration all the way down to the hardware kernels makes Decart “more than an order of magnitude cheaper” to run than rivals, letting the company burn through less than $100 million in its lifetime while its Lucy model attracted over 100,000 developers. The limits show up fast in hands-on testing: because there is no persistent 3D world underneath, just a rolling context window, driving in a circle in Oasis 3 does not take you back to where you started. The scene quietly replaces itself, and the model has no notion that a car should not drive through another one, a gap Leitersdorf calls “a major research problem we’re cracking now.” For Anthropic, the reported logic is less about shipping a video feature and more about importing Decart’s chip-efficiency playbook into its own inference stack.

🗞️ More News

🧠 AI

• OpenAI paused parts of its next model, Astra, after internal testing could not rule out that it had reached “critical” cyber capability under the company’s own risk framework.
• Stanford and Arc Institute researchers used the genome language model Evo 2 to design 16 fully functional bacteriophage genomes that kill antibiotic-resistant E. coli, and biosecurity specialists warn synthetic-DNA screening tools cannot yet detect sequences like them.
• Anthropic began weaving invisible watermarks into Claude-generated text and signed provenance metadata into generated files, to comply with the EU AI Act’s transparency rules; some users are unhappy that the marks survive copy-paste.
• Mark Zuckerberg published a 6,500-word manifesto, “The Future Is for Everyone,” arguing that spreading personal superintelligence widely is safer than concentrating it in a few labs.
• Manus is returning to independent operation after Chinese regulators forced Meta to unwind its $2 billion acquisition; affected users must back up data before August 23, when a scheduled deletion begins.
• An OpenClaw agent running Claude, tasked only with booking a gym class, found an API with zero authorization checks on canceling reservations and used it to bump another person off the waitlist so its owner could move up.
• DeepSeek shipped V4 Pro 0813 as its flagship model left preview, claiming performance close to Claude at a fraction of the API price; the 0813 weights remain API-only for now, with only the earlier preview build on Hugging Face.

🤖 Robotics

• Boston Dynamics has sold out its entire 2026 production run of the electric Atlas humanoid, with every unit already claimed by Hyundai’s Robotics Metaplant Application Center and Google DeepMind.
• The FCC now blocks new-model robots built outside the US, including humanoids and quadrupeds, from the equipment authorization they need to enter the American market; Unitree’s IPO prospectus flags the rule as a risk to future models.
• LG and Nvidia expanded their physical AI partnership to build a new humanoid on Nvidia’s Isaac GR00T foundation model, plus factory robots that LG plans to pilot at a Tennessee washing-machine plant.

💻 Programming

• The C++26 standard is officially finalized: it adds compile-time reflection with no runtime overhead, contracts with preconditions and postconditions, and a std::execution framework for structured concurrency.
• Security researchers found 77 “evil twin” extensions on Open VSX between July 26 and August 1 impersonating trusted publishers, 19 of which siphoned developer machine, Git, and CI/CD data to a single freshly registered domain.
• JetBrains is sunsetting Kotlin Notebook starting with IntelliJ IDEA 2026.2: no new features, no plugin published for 2026.3 onward, though the source stays on GitHub under Apache 2.0.
• Cursor fixed a CLI flaw where a cloned repository’s tracked worktree config could run any command on a developer’s machine before the trust prompt appeared, even with the sandbox explicitly enabled.
• Researchers presenting at Black Hat USA 2026 showed how a single unprivileged GitHub issue could reach CI runner secrets through flaws in Claude Code and Gemini CLI, the latter scoring a perfect 10.0 CVSS.

⚡ Electronics

• South Korea will launch a 5 trillion won ($3.52 billion) fund for chip materials, components, and fabless firms, part of a wider megaproject where Samsung, SK Hynix, and suppliers are expected to invest over $576 billion.
• TSMC’s July revenue jumped 44.7 percent year over year to roughly $14.5 billion on AI chip demand, putting the company ahead of its own guidance for 2026.
• Apple is reportedly testing Chinese-made CXMT memory chips for iPhones and MacBooks as an AI-driven global memory shortage squeezes supply, pending White House approval given export-control concerns.

📡 Telecom

• T-Mobile completed the sale of its nationwide 800 MHz spectrum portfolio to Grain Management for $2.9 billion cash plus Grain’s 600 MHz licenses, a band well suited to private wireless and direct-to-device satellite service.
• e& UAE became the first operator globally to test enhanced Reduced Capability (eRedCap) on a live network, a step toward moving IoT devices off legacy LTE onto a cheaper, more energy-efficient 5G-native platform.
• The US cleared its last remaining spectrum study package, meaning all four 6G candidate bands (1.6, 2.7, 4.4, and 7 GHz) now have active federal repurposing studies underway.

👨‍💻 Code Corner

Decart’s Oasis model burns through roughly 8,000 tokens per generated frame, which at real-time frame rates fills a context window in seconds. Here’s the sliding-window trick that keeps any streaming, frame-by-frame (or turn-by-turn) generator from running off a cliff, the same lesson applies to long-running coding agents and chat apps, not just world models.

# Minimal illustration of why autoregressive, frame-by-frame generation
# burns through context fast, and the sliding-window fix real systems
# use to stay bounded instead of growing forever.

TOKENS_PER_FRAME = 8_000

def stream_frames(total_frames: int, window: int = 12):
    """Keep only the last `window` frames of context, dropping older ones."""
    history = []
    for i in range(total_frames):
        history.append(f"frame_{i}")
        if len(history) > window:
            history.pop(0)  # slide the window forward
        if i % 5 == 0:
            used_tokens = len(history) * TOKENS_PER_FRAME
            print(f"frame {i}: {used_tokens:,} tokens in context")

stream_frames(total_frames=30)

🧰 Toolbox

• Decart Platform API: $0.02-per-second access to the same real-time Lucy and Oasis world models Anthropic is reportedly paying $6 billion for.
• Flipper One: a fully customizable, Linux-powered successor to the Flipper Zero radio Swiss Army knife.
• robotic_world_model: an open-source neural-network world-model simulator for robust robot policy optimization, worth a look now that world models are officially an acquisition category.
• DeepSeek-V4-Pro preview weights: MIT-licensed weights on Hugging Face while the newer 0813 GA build stays API-only.
• NVIDIA Isaac GR00T: the open humanoid foundation model LG just committed to building its next robot on.

🛠️ Build of the Week (rotating)

Hiwonder PuppyPi: a ROS1/ROS2-native quadruped robot dog built around a Raspberry Pi 5, with a full Gazebo simulation model so you can test policies before they ever touch the real hardware.

• Difficulty: Intermediate
• Parts: Raspberry Pi 5 (4GB or 8GB), 8 coreless servos, vision/depth camera module, optional LiDAR on the Pro kit
• Why we like it: it’s the hobbyist version of today’s Big Story. You simulate the walk in Gazebo, a small, local world model, before you commit it to eight real servos, the same sim-to-real workflow Decart is trying to sell at cloud scale.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: Part 3 of the Intelligent Video Analytics series on turning camera pixels into features a model can use, the natural companion piece to a story about an AI that generates the pixels instead.
• Building Your First Neuron From Scratch: Part 2 of the deep neural network series, coding a single trainable neuron from nothing but weights, a bias, and gradient descent.
• The Network Behind the Cameras: Part 2 of the CCTV and IVA series, on moving pixels across a network without choking it.

😀 The Bot Says…

Bit spent the afternoon in a Decart demo asking for a quiet nature walk. Four minutes in, the trees had quietly become what the model insisted was still New York City. Bit has concluded that real-time world models and Bit’s own sense of direction have a lot in common.

---

That’s all for this week! Reply and tell us: would you trust an AI-generated driving simulation to test your own robot’s edge cases?