05-22-Daily AI News Daily

Daily Summary

Nvidia's data center revenue hit $75.2B in a single quarter, doubling year-over-year—AI compute demand hasn't peaked yet.
Manus founders are planning to raise $1B to buy back their company from Meta; Tencent open-sourced a translation-specific small model ready to use.
Today's bottom line: compute power, Agents, and local deployment—three major tracks moving simultaneously. Worth diving into.

⚡ Quick Navigation

• 📰 Today’s AI News - Latest updates at a glance

💡 Tip: Want to experience the latest AI models mentioned in this article (Claude 4.5, GPT, Gemini 3 Pro) right away? No account? Grab one at Aivora —one minute to get started, hassle-free support.

Today’s AI News

👀 One-Liner

Nvidia’s data center revenue hit $75.2B in a single quarter—this is the number worth remembering today.

🔑 3 Keywords

#ComputePowerSurge #ManusEscape #OpenSourceTranslation

🔥 Top 10 Headlines

1. Nvidia Earnings Crush Expectations Across the Board

Data center segment: $75.2B in a single quarter. Last year at this time? Under $40B. Nearly doubled, and it beat analyst expectations of $73.4B.

This isn’t a one-product blowout—it’s the entire AI infrastructure demand exploding collectively. Nvidia’s Q1 total revenue hit $81.6B, up 85% year-over-year, with EPS of $1.87, also beating expectations. Q2 guidance came in hot too.

Stock popped after hours, then reversed and dropped 1.6%. This “sell the news” pattern has happened in 4 of Nvidia’s last 5 earnings calls. But don’t let the stock price fool you—the numbers themselves tell one story: AI compute demand hasn’t hit its ceiling yet.

2. Manus Founders Plan $1B Fundraise to Buy Back Company from Meta

The timeline on this story is wild. Meta announced a $2-3B acquisition of Manus in late 2025. In mid-April, China’s regulators demanded Meta cancel the deal. Now Manus founders—Xiao Hong, Ji Yichao, and Zhang Tao—are in talks to raise roughly $1B from external investors to buy the company back from Meta.

The new valuation reportedly needs to match Meta’s original $2B acquisition price. Translation: the founders are spending $2B to buy back their own company.

Manus originally went viral for “general-purpose AI Agents,” getting called the Chinese DeepSeek successor. If this buyback succeeds, it’ll restart as a fully independent Chinese AI company. Bloomberg reported it; Manus hasn’t responded yet.

3. Tencent Open-Sources Translation-Specific Model Covering 1.8B to 30B Parameters

General-purpose large models have always had a translation problem: the smaller the parameters, the worse the output. Tencent just open-sourced a complete translation-specific model suite covering 1.8B to 30B, called HY-MT2.

The key stat: small-scale models now match the translation quality of large-parameter general LLMs. What does that mean? Local deployment, cheap inference, translation running on edge devices—all suddenly viable.

For developers building multilingual products, this model suite is worth testing seriously—no more forced API calls to expensive large models just for translation quality. Models are already on HuggingFace, ready to use.

4. gstack: 23 Role Configurations for Claude Code Playing CEO, Designer, Engineer

Y Combinator president Garry Tan open-sourced his Claude Code setup, and it hit 100K stars in a single day.

The tool is called gstack—23 opinionated tools that simulate perspectives from CEO, designer, engineering manager, release manager, documentation engineer, and QA. Not just prompt templates, but a complete “AI team role-play” workflow.

For solo developers, the value is clear: write code solo, but get AI feedback from different functional angles on your product decisions. It’s like getting a small team’s multi-perspective input at the cost of one person. The 100K-star velocity shows this scratches a real itch.

5. Google Omni Flash Solves Math Problem Step-by-Step on a Blackboard

Prompt was one line: Solve 4(x-3)^2 = 25 on a blackboard, showing step-by-step solution.

Omni Flash generated a video of someone actually writing out the solution step-by-step on a blackboard. The visual output is clean, the logic is correct.

But the author’s take is sharp: Omni Flash is Google’s “Nano” tier in video, not the flagship. Wait for the Pro version before calling it production-ready. This demo is more of a capability preview—AI-generated teaching videos without needing real people on camera.

6. Tencent Launches Marvis, Zhipu Releases AutoClaw, Photoshop 27.7 Major Update

Three announcements bundled. Tencent rolled out Marvis, an OS-level AI assistant that can control phones and computers cross-device, with local privacy mode—sensitive operations need user confirmation before executing. That design detail matters; Tencent is thinking seriously about enterprise compliance.

Zhipu AI’s AutoClaw mobile app officially launched, supporting cloud and local execution modes, bringing AI Agent capabilities to phones.

Photoshop 27.7 is a substantial update—specific features weren’t detailed, but every major Photoshop version update reshapes creator workflows. Designers should pay attention.

7. Tesla FSD Supervised Version Takes Another Step Toward China

On May 21, Tesla officially published the list of countries where supervised FSD is available—China is on it. The news hit Weibo’s trending topics.

But reality check: Tesla China’s official channels haven’t made any announcement, and customer service still says “approval is in progress, we’ll push it as soon as it’s ready.” Same line as before.

Current status: China’s on the official list, but regulatory approval isn’t done. For Chinese Tesla owners, this is a “getting close, but not yet” signal. Musk’s attention is clearly elsewhere these days (politics, X), so FSD’s China rollout probably depends more on regulators than Tesla.

8. Practical Tips for Generating App Icons with AI

One of indie developers’ biggest headaches: finish the product, icon looks terrible, don’t want to pay a designer.

The core trick: find an icon style reference you like, then tell AI “look at these app icons’ style, generate a high-quality icon for ‘xxx app,’ strictly follow the exact same visual style.” Results are surprisingly good—the style consistency beats pure text descriptions by a mile.

The principle is simple: give AI a visual anchor instead of just words. You can use this today, no extra tools needed.

9. Google I/O 2026 Gemini Model Official Blog Screenshot

Someone questioned a claim about Gemini models; the author fired back with a screenshot from Google’s official blog—from blog.google/innovation-and-ai/sundar-pichai-io-2026/#gemini-models.

The value here isn’t the argument itself, but this: Google I/O 2026’s official blog has Sundar Pichai’s clear positioning and roadmap for the Gemini model family. If you want Google’s official narrative on their own models, go read the source instead of secondhand takes.

10. South Korea’s AI Memory Windfall: SK Hynix Employees Become “Tactical Nuclear Weapons” in Dating Markets

This isn’t pure AI news, but it’s the realest social snapshot of the AI wave.

Since 2026 started, Samsung and SK Hynix have been cashing in on the AI memory demand explosion. SK Hynix employees saw 25% monthly salary bumps, with average performance bonuses hitting 6.1M yuan. Result: Korean matchmaking companies now list “SK Hynix work uniform” as a dating market advantage. Engineers once called “engineering NPCs” are suddenly society’s most envied professionals.

Where does AI compute demand actually flow? Into memory chips, into engineers’ paychecks, into South Korea’s dating market. That supply chain is more direct than any earnings report.

📌 Worth Following

[Research] SVFSearch: Game-Domain Short Video Frame Search Multimodal Benchmark — Tests multimodal LLMs on retrieval reasoning across short video frames, with visual ambiguity in gaming scenarios far harder than general cases. Fills a real evaluation gap.

[Research] ArchSIBench: Testing Vision-Language Models on Architectural Spatial Intelligence — Core capability for robot navigation and 3D scene understanding. Existing spatial reasoning benchmarks only test basics; this targets architectural cognition specifically, valuable for embodied AI researchers.

[Research] New Solution to Portrait Generation’s Trilemma: Text Alignment, Realism, and Aesthetics — Text-to-image portraits have always traded off between “alignment/realism/aesthetics.” This paper proposes a Pareto optimization framework worth reading for image generation product teams.

😄 AI Fun

Anime Fan Uses AI to “Take Favorite Coser to Hot Pot”

Someone wrote an absurdly detailed prompt: asking AI to generate “a casual iPhone snapshot” of a Coser eating hot pot, face flushed from steam, a few strands of hair stuck to her neck, expression showing she just noticed the camera and threw up a peace sign.

The prompt was nearly 500 words long—down to “light blush on cheeks,” “outfit slightly shifted from sitting,” “don’t overdo perfection.”

The generated image actually looked pretty convincing.

The funniest part isn’t how good the image is, but this: the time spent writing the prompt probably exceeded the time it’d take to actually invite someone to hot pot. AI granted the wish, but the method was turning the dream into a requirements document first.

🔮 AI Trend Predictions

Translation-Specific Small Models Will Become Mainstream for Enterprise Local Deployment

• Timeline: Q3 2026
• Confidence: 75%
• Reasoning: Today’s news on Tencent’s HY-MT2 open-source proves 1.8B-30B specialized models match large-parameter general models on translation. This signals the path “specialized small models replacing general large model APIs” is validated for translation, and vertical-domain specialized models in healthcare, legal, and finance will replicate this pattern.

Manus Buyback Success Will Reposition It as Independent Chinese AI Agent Flagship

• Timeline: Q3 2026
• Confidence: 60%
• Reasoning: Today’s news on Manus founders’ $1B fundraise for buyback shows buyback talks are substantive, with valuation anchored at $2B. China’s regulators explicitly blocked foreign acquisition; founder team has strong independence intent; funding scale is basically viable. Post-buyback, Manus will likely restart as “Chinese autonomous AI Agent” and accelerate product iteration with fresh funding.

Nvidia Data Center Quarterly Revenue Breaks $100B

• Timeline: Q4 2026
• Confidence: 55%
• Reasoning: Today’s earnings show Nvidia Q1 data center revenue at $75.2B , nearly doubling year-over-year, with Q2 guidance continuing to beat. At current growth trajectory, hitting $100B in Q3-Q4 is plausible, but key variables are Blackwell Ultra shipment pace and whether China export restrictions tighten further.

AI-Generated Teaching Videos Enter Mainstream Content Platforms

• Timeline: Q3 2026
• Confidence: 65%
• Reasoning: Today’s demo of Google Omni Flash generating blackboard math videos shows AI can directly generate clear teaching videos—and this is just the Flash (lite) version. Pro version launch will push quality higher; platforms like Bilibili and YouTube will see an AI teaching content wave, compressing traditional live-action course production’s cost advantage fast.

❓ Related Questions

How Do I Access Claude Code?

Claude Code is Anthropic’s AI programming tool, currently requiring Claude Pro or Claude Team subscription. Mainland users face payment method restrictions and account registration barriers.

Today’s viral gstack is built on Claude Code configs and hit 100K stars—proof the demand is real.

Solution: Visit Aivora for ready-made accounts—instant delivery, hassle-free support.