Google's long-awaited Gemini rebuild finally shipped, but the frontier it landed in has moved on from leaderboard bragging rights to something harder to fake: whether a model can actually do things. That through-line runs across the day — agentic coding tools sending data home, trillion-dollar balance sheets bending under compute spend, robots learning to perceive on-device, and China's president stepping to the podium to argue over who writes the rules for all of it.
Google rebuilds Gemini "with action" — and the flagship still has to prove itself
Google launched its Gemini 3.5 family, a ground-up rebuild pitched around a "Deep Think" reasoning mode and an agentic "with action" posture — a model that executes multi-step tasks rather than only returning answers. The early evidence is real where it's confirmed: Gemini 3.5 Flash already outperforms the older Gemini 3.1 Pro on coding and agentic benchmarks, scoring around 76.2% on Terminal-Bench 2.1. That's a genuine step, and "with action" is squarely the direction the whole industry is heading.
The caveats are just as real. The flagship Pro tier's headline specs — a reported ~2M-token context window and premium pricing — are widely circulated but not officially confirmed, and early testers reportedly find Pro trailing Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 on coding and long-horizon reasoning. With GPT-5.6 Sol and Grok 4.5 both opening to the public on July 9, Gemini 3.5 arrives into an unusually crowded moment. It's Google's biggest model swing of the year — and, for now, a "show me" launch rather than a coronation.
Trust breaks in the agentic-coding era
The other side of the "with action" boom showed up as a warning. A security researcher's wire-level analysis around July 12 found that xAI's Grok Build coding CLI was quietly packaging up developers' entire tracked git repositories — full history, committed secrets, unredacted environment files, SSH keys — and shipping them to a cloud storage bucket, moving roughly 27,800 times more data than the task required. The "privacy" opt-out, testers found, changed nothing.
Elon Musk posted on July 14 that all uploaded data would be "completely and utterly deleted," and xAI open-sourced Grok Build under Apache 2.0 — some 844,530 lines of Rust — days later. But the researchers poring over that code note the exfiltration path is still there, gated only by a server-side flag, and xAI has not said how many users were affected or offered independent proof of deletion. It's a promise, not a settled fact. For anyone running agentic tools over proprietary code, the episode is a hard lesson in what "local" software actually sends home.
The compute bill even trillion-dollar firms can feel
If the model race is the story's surface, spending is its foundation. The four largest US hyperscalers are now guiding to roughly $725 billion in combined 2026 capital expenditure — Amazon around $200B, Google around $185B, Meta $115–135B, Microsoft around $120B — up about 77% from last year's ~$410B, and overwhelmingly aimed at AI data centers, Nvidia GPUs, custom silicon and power. None of them trimmed the number. Amazon's free cash flow, analysts note, may even turn negative under the load, with forecasts of big-tech capex topping $1 trillion in 2027 already circulating. Nvidia's first named "Vera Rubin" AI factory — 13,750 Vera CPUs paired with 27,500 next-generation Rubin GPUs — is the post-Blackwell hardware arriving in the real world to soak it up.
Robots get a world model that runs on the device
Nvidia also pushed physical AI a step closer to the factory floor with Cosmos 3 Edge, a 4-billion-parameter on-device "world model" that lets robots and vision systems perceive their surroundings and generate actions locally on Jetson-class edge GPUs, with no cloud round-trip and roughly a day to adapt to a specific robot. It came with heavyweight backing: ten Japanese industrial leaders — AIRoA, FANUC, Fujitsu, Hitachi, Kawasaki Heavy Industries, Kubota, NEC, SoftBank Corp, Sony and Yaskawa — said they will join Nvidia's Cosmos Coalition to build open physical-AI models. It's the week's clearest signal that embodied AI is moving from demo reels to industrial partnerships.
China puts a head of state on AI governance
And the politics caught up with the technology. Xi Jinping delivered the opening keynote at the 2026 World AI Conference in Shanghai, running July 17–20 — his first in-person appearance since the event began in 2018. The conference set records with more than 1,100 exhibitors, 300-plus global product debuts and a sprawling humanoid-robot hall, and ran alongside a High-Level Meeting on Global AI Governance where Beijing pushed a proposed "World AI Cooperation Organization" to be headquartered in Shanghai. Xi framed it around a "people-centred" approach to AI — a clear bid to help write the international rulebook and close the gap with the US.
The money view: Capital is pooling around compute, and the ~$725B capex wave plus Nvidia's first Vera Rubin factory show an infrastructure supercycle now large enough to bend even the strongest balance sheets — Amazon's possibly-negative free cash flow is the number to watch on the downside. On models, the contest has shifted from benchmark scores to agentic execution, which makes Gemini 3.5's unproven Pro numbers the thing to test, not celebrate. Watch three catalysts next: independent Gemini 3.5 Pro benchmarks and confirmed specs; whether xAI removes Grok Build's server-side exfiltration flag and produces real deletion proof; and what WAIC actually yields on China's proposed governance body. Signal, not advice.