Today in AI — August 11, 2026 (2nd edition)

The buildout became a market Tuesday: GPU rental futures are coming to CME, Anthropic hit the IPO roadshow, and CoreWeave and Super Micro posted blowout quarters — while agents moved into business workflows, Gemini crossed a billion users, and the trust wars over watermarks, surveillance and poisoned data kept escalating.

Models & Research

Sergey Brin took direct oversight of Google's Gemini team as reports say the long-delayed Gemini 3.5 Pro has been shelved to clear the runway for a Gemini 4 push — a code-red for a lab that missed three shipping deadlines and has lost Shazeer, Jumper and Dean. Google has not commented.
Nvidia shipped Nemotron 3.5 Lightning, a 30B-parameter MoE built for always-on agent workloads, plus NeMo Switchyard, an open-source library that routes each step of an agent run to the best model by quality, cost or latency — a tollbooth play on the agent economy. The post-training recipes ship with the model so enterprises can re-tune it on their own data.
Google's AMIE medical AI matched or beat 30 primary care physicians in real-time video consultations, the first expert-level result for audio-visual medical AI, per a new paper — though doctors still won on rapport, and real-world studies are pending.
Trajectory, a startup of former DeepMind, Apple, OpenAI and Meta staffers, raised $40 million at a $300 million valuation for continual learning — models that keep learning after deployment instead of freezing or forgetting.
IBM Research showed agents can learn from their own mistakes at roughly a seventh of the tokens of a rival memory approach, hitting 89.3 percent task completion on AppWorld — how memory is served, not just what's remembered, is becoming the cost battleground.
Cactus released Needle 2, an open 45M-parameter tool-calling model that ships as a single 14MB binary and runs a full agent session in about 28MB of RAM — on-device agentic AI as a size-class competition.

Industry

CME Group will start trading GPU rental futures on October 5, turning the cost of AI compute into a standardized, tradable commodity for the first time — H100 and B200 contracts tracking hourly rental prices, listed on NYMEX pending regulatory review.
Anthropic is meeting investors ahead of a fall IPO that the Wall Street Journal says could be the biggest ever, fielding tough questions about cheaper Chinese rivals and whether the infrastructure boom justifies a near-trillion-dollar valuation.
Brad Lightcap, OpenAI's longest-serving business executive, left after eight years "to start something new" — the biggest name yet in a leadership exodus that also claimed ethics lead Chloé Bakalar, who exited less than a year after joining.
The compute buildout's public markets lit up: CoreWeave's Q2 revenue jumped 112 percent to $2.58 billion with a $104 billion backlog, while Super Micro's fiscal Q4 revenue rose 93 percent to $11.1 billion with above-consensus guidance — though both still carry profit questions.
Anthropic reportedly struck a $9.1 billion, 20-year deal with Bitcoin miner Riot Platforms for 191 megawatts of Texas compute — the biggest crypto-to-AI pivot yet, and a sign the lab will buy power from whoever can plug it in.
Igor Babuschkin's River AI raised $1 billion, led by General Catalyst, to build user-owned, trainable AI computers for homes — the highest-profile bet yet on local, personal AI as a third way beyond the cloud oligopoly.
Kevin Weil, OpenAI's former chief product officer, is seeking $150 million at a valuation of at least $750 million for an AI-for-science startup, Business Insider reports — the second senior OpenAI leader this week to announce a next act.
Manus said it will soon operate as an independent company again, closing out the $2 billion acquisition Chinese regulators ordered Meta to unwind — with a data wipe for some users scheduled for August 23–24.
Amazon quietly stripped product names from its order confirmation emails, a shift that keeps purchase data out of rival AI shopping agents' hands just as Google pushes agentic shopping through Gmail.
ByteDance's Doubao assistant began charging merchants a roughly 12 percent commission on AI-booked hotel orders — the first time an AI assistant's referral traffic has been priced as its own channel.
Mistral launched regional inference endpoints, a service-level priority tier, and a coalition of European industrial giants — ASML, CMA CGM and Amadeus among them — committed to securing up to 1 gigawatt of compute by 2030.
Gemini crossed 1 billion monthly users, Google's fastest-growing product ever and its 14th to the mark — roughly even with ChatGPT less than a year after visibly trailing.
Spotify will label AI-generated artist identities with "AI Persona" badges and exclude their music from recommendations by default, judging identity rather than process — platforms are moving from labeling content to labeling the entities behind it.
Ant Group led a strategic round in tactile-sensing startup Daimon Robotics — its first bet on robot "fingertips" — as Chinese capital shifts from walking humanoids to hands that can actually manipulate the world.

Policy

Shanghai published its 15th Five-Year Plan targeting 4 trillion yuan ($560 billion) in software industry scale by 2030, explicitly calling for deep integration of domestic chips with mainstream large models and funding non-Transformer architectures, world models and physical intelligence.
Anthropic began embedding invisible watermarks in text from every Claude model launched on or after August 2, with signed C2PA metadata on files — rolling the EU AI Act's transparency code out to users worldwide, not just Europe.
Live facial recognition expanded on two fronts: British Transport Police took its trial into London Underground stations, while Western Australia Police's Perth deployment has scanned more than 130,000 faces — one-to-many matching with watchlists critics call opaque.
A Kremlin-linked unit posing as a human rights group has been poisoning the sources chatbots learn from — in Russian-language tests, models repeated the propaganda in up to 86 percent of cases.
The FBI's terror watchlist office is shopping for predictive AI that would score people as threats before any crime, per procurement documents obtained by Reason — a request civil-liberties groups are calling pre-crime policing.
More than 40 digital-asset organizations, led by the Bitcoin Policy Institute, asked frontier labs for vetted defender access to their most capable models, arguing AI-assisted attacks are outrunning defenders. No lab has publicly responded.
Mark Zuckerberg published a 6,500-word manifesto arguing AI's real risk is concentrated control, with distribution and competing labs as the check on a centralized superintelligence — the sharpest counter-narrative yet from the industry's most prominent optimist.
North Korea-linked hackers are running AI models on their own hardware — local Ollama setups, agent frameworks and Cursor — to mine stolen documents and draft phishing campaigns without touching Western clouds, per a South Korean report.

Tools

OpenAI shipped a ChatGPT desktop app for Linux in preview, bundling ChatGPT, ChatGPT Work and Codex — including desktop control — native at last on the OS where much of the world's AI development happens.
Wix launched Symphony, a multi-agent platform that runs booking, CRM, marketing and operations for small businesses, with an orchestrator named Maestro dispatching specialized agents and a daily agent "stand-up" summarizing completed work.
SpaceXAI launched Grok Bot in beta, an agent app that completes work tasks across applications, bundled with SuperGrok Heavy and Cursor premium tiers — the first concrete link between Grok and Cursor's paying base.
Modular shipped Mojo 1.0, the first stable release of Chris Lattner's AI-native language, promising an "epoch of stability" for Python-like code that compiles through MLIR to any accelerator.
OpenAI's Daybreak models are now available on AWS, extending the lab's distribution beyond its Microsoft partnership.
Unsloth shipped a native desktop app for its no-code fine-tuning Studio, dropping the local-model bar from a pip install to "download an app" — with one-click export for llama.cpp, Ollama and vLLM.
Zhipu's coding agent ZCode crossed 1 million users with new autonomous features, a sign Chinese labs are competing on the harness around the model, not just the model itself.
The creator of Redis ported MiniMax's H3 multimodal model to a Mac, a single-developer, no-framework inference engine that renders a 512×512 image in about 17 seconds on an M5 Max.
Security researchers cracked a critical Zoom vulnerability using fewer than 20 prompts on off-the-shelf AI models, compressing what they call nation-state work into a day — Zoom has patched it.

GPU rental futures, a near-trillion-dollar IPO roadshow and agents running small businesses — is the AI buildout getting more real or more fragile? Write to us via our contact page.

Sources: antirez/h3.c (GitHub)