One day, four stories, and a common thread: the fight for AI is moving into hardware and talent. Apple put a 2-nanometer chip in a $899 desktop and 512GB of memory in a Mac Studio, OpenAI showed its first in-house chip beating Nvidia's Blackwell, Washington put a six-figure price tag on hiring foreign engineers, and Meta is about to sell an AI agent that orders your food and clears your inbox.
Apple's M6 is its first 2nm chip — and the M5 Ultra puts 512GB of memory in a desktop
- Apple announced on August 25 the M6 — its first 2-nanometer chip — for the Mac mini (12-core CPU, 12-core GPU, dual 16-core Neural Engine, up to 32GB) and the M5 Ultra for the Mac Studio, its first quad-die chip: up to 36 CPU cores, 80 GPU cores and 512GB of unified memory at 1.2TB/s.
- Apple claims the M5 Ultra delivers up to 4.5x the peak GPU compute for AI of the M3 Ultra and the M6 nearly 30% more than the M5. Prices: Mac mini M6 from $899, Mac Studio M5 Max from $2,499 and M5 Ultra from $5,499.
Why it matters512GB of unified memory means a desktop can run open models with hundreds of billions of parameters locally, with no cloud bill — Apple is quietly selling the Mac Studio as a personal AI data center, and it got 2nm silicon into a $899 box before anyone else.
OpenAI's Jalapeño chip beats Nvidia's Blackwell in its first public benchmarks
- SemiAnalysis published on August 25 the first benchmarks of Jalapeño, the inference chip OpenAI designed with Broadcom: over 700 tokens per second per user running DeepSeek R1 and roughly 1,400 on GPT-OSS, with better performance per watt than Nvidia's Blackwell 'across almost all scenarios'.
- The chip is built on TSMC N3P, delivers 13.4 PFLOPs (MXFP4) at 700W with HBM4 memory at 15.4TB/s, and went from hiring the team to tape-out in about 16 months. A B0 stepping now in fabrication adds ~25% perf-per-watt; OpenAI plans small volumes by the end of 2026 and the production ramp through 2027.
Why it mattersNvidia's biggest customer just showed it can design competitive silicon of its own; the caveat is that these are benchmarks run with OpenAI's cooperation, and by the time Jalapeño ships in volume Nvidia's Vera Rubin will be the real yardstick.
Washington proposes a $103,265 fee for every new H-1B visa
- DHS published in the Federal Register on August 25 a proposed rule charging $103,265 for every cap-subject H-1B petition, payable at filing and on top of all existing fees; public comments are open until September 24, 2026. Universities, nonprofits and government research institutions stay exempt, and so do renewals.
- It is the formal version of the $100,000 fee Trump imposed by proclamation in 2025, which a federal judge ruled unlawful in June. DHS justifies it on 'ability to pay' grounds and wants to recover about $7.1 billion a year in immigration-system costs (USCIS, ICE and immigration courts included).
Why it mattersFor Latin American engineers, the H-1B was the main door into Silicon Valley; at $100K+ per hire, companies will lean harder on remote and nearshore teams in Mexico, Colombia, Argentina and Venezuela — good news for talent that stays in the region, bad news for anyone counting on emigrating through that door.
Meta's Hatch agent is weeks away, with a $199.99 plan on the table
- According to The Information, Meta plans to launch Hatch 'within weeks': a consumer AI agent that gets things done inside DoorDash, Etsy, Reddit, Yelp and Outlook. Internal documents from June contemplate a free tier and a 'Hatch Plus' plan at $199.99 a month with 5 to 10 times the daily capacity — final pricing is not set.
- Hatch currently runs on Anthropic's Claude models; Meta intends to move it to its own Muse models before launch, and a new model codenamed 'Watermelon' is targeted for October.
Why it mattersMeta is copying OpenAI's $200 tier and betting that the product people will pay for is an agent that acts, not a chatbot that talks — a category where Meta's 3 billion users give it distribution nobody else has.
The big picture
Tuesday was a hardware day: Apple brought 2nm to the desktop and OpenAI proved it can build its own accelerator, both chipping away at the idea that AI runs only on Nvidia in someone else's data center. At the same time, the $103,265 H-1B fee shows the other bottleneck — talent — is being priced by policy, not by markets, and Meta's $199 agent tells you where the money is expected to come from next: not answers, but actions.
Sources & further reading
- Apple Newsroom — Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute
- Engadget — Apple updates the Mac mini and Mac Studio with M6, M5 Max and M5 Ultra
- SemiAnalysis — OpenAI Jalapeño: Better Than Nvidia Blackwell
- TechCrunch — OpenAI's Jalapeño chip is built for fast inference at scale, benchmarks show
- Federal Register — Fee for Certain H-1B Petitions (proposed rule, Aug 25, 2026)
- USCIS — DHS Proposes Additional H-1B Fee
- The Next Web — DHS proposes $103,265 H-1B fee
- The Next Web / The Information — Meta plans to launch Hatch within weeks
- The Decoder — Meta's paid AI agent Hatch launches soon, Watermelon model due in October