This website uses cookies

Read our Privacy policy and Terms of use for more information.

Recaply Tech

Good morning, tech enthusiasts. Nvidia and Microsoft are rebuilding the Windows PC around AI agents that run on the device, not in the cloud. The first RTX Spark laptops and desktops arrive this fall.

Meanwhile, Berkeley instructors say heavy AI use is helping push failure rates in intro CS classes far above past years. Who's actually ready for the agent era?

In today's recap:

  • Nvidia and Microsoft rebuild Windows PCs around AI agents

  • Broadcom slips as Marvell rides Huang's trillion-dollar praise

  • Berkeley CS failure rates spike as AI use grows

NVIDIA & MICROSOFT

Nvidia and Microsoft rebuild Windows around agents

Nvidia and Microsoft rebuild Windows around agents

Source: Nvidia

Recaply: Nvidia and Microsoft just unveiled RTX Spark, a superchip for a new class of Windows PCs built to run AI agents locally. Microsoft is also reworking Windows itself so those agents can run securely on the device.

Key details:

  • Windows gets new security primitives for agent identity, containment and policy. Nvidia's OpenShell runtime adds user controls over what agents can do, and Hermes Agent and OpenClaw are adopting both.

  • RTX Spark delivers up to 1 petaflop of AI compute with up to 128GB of unified memory. Nvidia says it can run 120B-parameter models with up to 1 million tokens of context locally.

  • Microsoft tuned the Windows scheduler for all 20 CPU cores and raised the memory limit for the GPU, Windows chief Pavan Davuluri wrote. Older x86 apps run through Prism emulation.

  • Nvidia announced RTX Spark on May 31, and Microsoft covered agent features at Build on June 2 and 3. RTX Spark laptops and desktops arrive this fall from six PC makers.

Why it matters: Nvidia says agents haven't gone mainstream because they can't yet run safely and privately on people's main PCs. RTX Spark is the two companies' answer: local models, OS-level guardrails and a GPU with room for agent workloads. Microsoft CEO Satya Nadella says the goal is to bring "unmetered intelligence" to every home and desk. Neither company has announced prices yet, so the real test comes this fall.

BROADCOM & MARVELL

Broadcom slips as Marvell rides Huang's praise

Broadcom slips as Marvell rides Huang's praise

Source: Yahoo Finance

Recaply: Broadcom shares just sank after its AI chip forecast missed Wall Street estimates. Rival custom chipmaker Marvell went the other way, surging 32% after Jensen Huang called it the next trillion-dollar company.

Key details:

  • Both companies design custom chips for AI data centers. Broadcom's customers include Google, Meta, Anthropic and OpenAI, while Marvell's networking chips link thousands of AI processors together.

  • Broadcom projected $16B in third-quarter AI chip sales, below the $17.2B analysts expected. Its AI chip revenue still grew 143% from a year earlier last quarter.

  • Huang made the remark onstage with Marvell CEO Matt Murphy in Taipei, CNBC reported. The 32.52% jump was Marvell's biggest one-day gain ever, lifting its value to just over $250B.

  • Marvell's shares jumped on June 2, after Huang's appearance at Computex Week. Broadcom reported results on June 3, and its stock fell in after-hours and premarket trading.

Why it matters: Broadcom's results weren't weak. Revenue of $22.19B beat estimates, and it guided $29.4B for the current quarter. But it didn't raise its full-year AI chip forecast, and CFRA's Angelo Zino says the bar was "really high" going in. Hyperscalers are expected to spend about $650B on AI this year. In that market, a soft AI forecast gets punished, and a nod from Huang gets rewarded.

UC BERKELEY

Berkeley CS failure rates spike amid AI use

Berkeley CS failure rates spike amid AI use

Source: The Daily Californian

Recaply: Failing grades in several UC Berkeley computer science classes just soared this spring. Instructors blame heavy reliance on AI, weak math preparation and too few teaching assistants.

Key details:

  • Teaching professor Dan Garcia says the "primary driver" is a "vast increase in academic dishonesty" from students using models like Claude, ChatGPT and Gemini. Nearly 30 CS 10 students were caught cheating on take-home exams.

  • Some 35.3% of CS 10 students and 10.6% of CS 61A students got F's this spring. Neither class topped 10% in spring 2024 or spring 2025.

  • Garcia is among 1,300 UC faculty who said they're being forced to teach "middle school" math, the New York Post reported. They want SAT and ACT scores back for STEM admissions.

  • Berkeley's student paper, The Daily Californian, published the grade data on June 3, drawing on the Berkeleytime grade site. Garcia plans to tell future classes about spring 2026 on day one.

Why it matters: Berkeley's own guidelines expect only 7% of students in lower-division courses to get D's or F's. CS 10's failure rate is five times that. Associate teaching professor Gireeja Ranade saw a 16.8% F rate in her upper-division optimization class, against a 5% norm. Both professors want students to work through hard problems, and Ranade says the age of AI calls for teaching "more, not less."

NEWS

Beyond the big stories in tech today

  • Monterey Park residents voted on June 2 to permanently ban data centers, the first US city to do so by ballot. Early results showed 86.3% of more than 7,000 counted votes in favor, and the ban stays until voters end it.

  • Apple's incoming CEO John Ternus signed off on dropping Vision Pro successors to focus on smart glasses, analyst Ming-Chi Kuo says. Kuo expects AI glasses in 2027, while Bloomberg's Mark Gurman says a Vision Pro 2 is still in testing.

  • Uber is cutting 23% of its people team, which includes HR and recruiting staff, CNBC reported. Uber says the cuts are well under 1% of its 34,000 employees and didn't result from AI.

  • Microsoft pushed developers at Build to make native Windows 11 apps with WinUI rather than web-based frameworks like Electron or React Native. It's dropping the "3" from WinUI 3 and says it has no plans for a new framework.

  • YouTube is flooded with pirated audiobooks narrated by AI, from Harry Potter to John Grisham, a New York Times investigation found. One pirated copy of Grisham's latest thriller has drawn over 80,000 views, The Conversation notes.

  • Samsung's Galaxy Z Fold 8 Ultra appeared in the Bluetooth SIG database, confirming the "Ultra" name for one of two Fold 8 models. The shorter, wider model may simply be called Galaxy Z Fold 8, 9to5Google reports.

  • Colorado Gov. Jared Polis vetoed House Bill 26-1210, which would have banned surveillance pricing based on personal data. Polis said its broad sweep would punish lower prices as well as higher ones, but the Center for Democracy & Technology called it a good bill.

  • Microsoft introduced MAI-Code-1-Flash, a fast coding model its AI team trained from scratch without distilling third-party models. It's rolling out to individual GitHub Copilot users in Visual Studio Code.

  • VS Code for the web had a bug that let attackers steal a GitHub token with a single click, a security researcher disclosed. The token could read and write every repo a user can access, including private ones. Microsoft merged fixes the next day.

  • GitHub launched the Copilot app, a desktop app for running and reviewing several coding agents in parallel from one view. It works with existing Copilot Pro, Pro+, Business and Enterprise plans.

  • BYD plans to start producing sulfide-based all-solid-state batteries in limited batches in 2027, ramping up to mass production by 2030. China's SAIC is also aiming for all-solid-state EV batteries in 2027.

  • Micron crossed a $1T market value on May 26, after its stock rose roughly ten-fold in a year, Reuters reported. A push from Nvidia three years ago moved it into custom memory chips for AI.

  • SpaceX won a 100% property tax abatement for its proposed Terafab chip plant after Grimes County, Texas, commissioners voted 4-1. Commissioner David Tullos voted no, saying SpaceX hadn't been transparent with the county.

  • Kasra Rahjerdi spent $1,500 testing whether AI models could hack a vulnerable app he built. GPT 5.5 found the hidden flag in 7 of 10 runs, Claude Opus 4.8 in 2, and both Gemini models failed every run.

🧡 Enjoyed this issue?

🤝 Recommend our newsletter or leave a feedback.

How'd you like today's newsletter?

Your feedback helps me create better emails for you!

Login or Subscribe to participate

Cheers, Jason

Connect on LinkedIn, & Twitter.

Reply

Avatar

or to participate