The money and the metal. This week's AI story isn't a new chatbot — it's where the capital and the hardware are moving. Apple is reshoring chips by the billion, Meta is racing to build its own silicon and double its compute, venture dollars are pouring into a shrinking circle of AI giants, and — almost quietly — a French lab taught a robot to find its way with nothing but a camera and a sentence. Here's the through-line.
Apple bets $30 billion on American chips
Apple has committed more than $30 billion to Broadcom to design and produce custom silicon and wireless-connectivity components — a deal expected to yield over 15 billion US-made chips and a $1.5 billion expansion of Broadcom's Fort Collins, Colorado facility. It is Apple's largest domestic-manufacturing commitment yet, folded into a broader $600 billion, four-year US investment plan. The chips themselves aren't AI accelerators — they're the wireless and custom parts inside Apple's devices — but the significance is the direction of travel: in the AI era, even ordinary silicon is being pulled onshore as supply security and packaging capacity turn strategic. For Broadcom, it's a marquee anchor customer; for Washington, a manufacturing win.
Venture capital becomes an AI-only bet
The macro backdrop is stark. US venture funding hit a record $412.7 billion in the first half of 2026 — nearly 30% more than all of last year — and 86% of it, some $355.9 billion, went to artificial-intelligence companies. Two firms did much of the absorbing: OpenAI ($122B) and Anthropic ($95.6B) together took 43% of all global startup funding, and megadeals of $100 million or more made up 87.5% of the total. As Fortune put it, "almost none of it is trickling down." The result is a barbell market: a handful of AI titans at one end, a funding drought for nearly everyone else at the other.
Meta wants to own its stack
Meta, meanwhile, is trying to escape its dependence on Nvidia. An internal memo says the company will begin producing its own MTIA AI chip in September and aims to double computing capacity from roughly 7 gigawatts this year to 14 gigawatts in 2027. It has locked a Samsung memory-supply agreement and broken ground on a C$13 billion, one-gigawatt AI data center in Alberta — its first in Canada, with room to scale to 1.8 gigawatts. With 2026 spending possibly reaching $145 billion, Meta is buying its way toward controlling chips, memory, and power alike — the hyperscaler answer to a compute-constrained world.
A robot that navigates on a single camera
The most futuristic story is also the cheapest to run. Mistral, the Paris-based frontier lab, released Robostral Navigate, its first robotics model — an 8-billion-parameter system that pilots a robot through rooms it has never seen using a single ordinary color camera and plain-language instructions like "walk through the corridor and stop at the second shelf." Trained entirely in simulation, it scores 76.6% on the R2R-CE navigation benchmark, all without LiDAR or a pre-built map. It arrives as Mistral's valuation nears $23 billion, and it signals that a European lab is now competing on the next frontier — embodied, physical AI — where a software model, not new hardware, unlocks the capability.
A quieter change at the top of OpenAI
Finally, a human note. OpenAI's No. 2, Applications CEO Fidji Simo, is stepping back to a part-time advisory role after a longer-than-expected medical leave for a chronic condition; her product and business duties will be split among Greg Brockman, Sarah Friar, and Jason Kwon. It's a reshuffle of who runs OpenAI's revenue engine at a moment when the company is floating an IPO — a reminder that continuity at the top is part of the story, too.
The money view. The week's signal is that AI's center of gravity has moved from models to metal and money: chips reshored, compute doubled, venture dollars concentrated, and the frontier quietly broadening into robots. Watch three things next — whether Google's rumored Gemini 3.5 Pro actually ships (and how big its context window really is), whether Meta's own chip meaningfully cuts its Nvidia orders when production starts in September, and how smoothly OpenAI's leadership reshuffle lands against its IPO ambitions.