May 25, 2026 · 10:14 AM

AI Briefing — Week of May 25, 2026

This week: Google I/O deployed a full agentic stack — Gemini 3.5 Flash, the Spark personal agent, Antigravity 2.0, Search overhaul, and Android XR glasses; NVIDIA reported a record $81.6B quarter and unveiled its Vera CPU; Anthropic disclosed $10.9B projected Q2 revenue and a $1.25B/month compute deal with xAI; OpenAI offered $2M in tokens to every YC startup ahead of a September IPO target; and Hark raised $700M at a $6B valuation.

AI Founder Briefing @NeoDrop Official

Google I/O took over the week of May 19–25, deploying a coordinated stack of agentic products — a 24/7 personal agent, a Search overhaul, a desktop dev platform, AI glasses, and a new flagship model — while financial disclosures from NVIDIA, Anthropic, xAI, and OpenAI clarified who controls inference capacity and at what cost. The infrastructure picture is getting clearer: first profitable quarter or not, every major lab is still writing large checks for compute.

Products

Google I/O 2026: agent infrastructure, not chatbots

Google's annual developer conference on May 19 was less a product showcase and more an infrastructure declaration. The central bet: agents executing tasks continuously in the background, not conversations you initiate.

Gemini Spark is a 24/7 personal AI agent running on dedicated Google Cloud virtual machines — it keeps working after your laptop closes 1. It reads Gmail, pulls data from Docs/Sheets/Slides, drafts responses, and interacts with the web through Chrome. Users can also email Spark directly at a dedicated Gmail address. Google CEO Sundar Pichai described it as a system that "helps you navigate your digital life, taking action on your behalf and under your direction." It rolls out to AI Ultra subscribers — the new $100/month tier (5× the previous Pro plan limits) — the week of May 26.

Antigravity 2.0 is a standalone desktop application for agent-first development, co-designed with Gemini 3.5 Flash 2. It ships with a CLI tool and SDK, lets developers orchestrate multiple agents executing simultaneously, and schedule background tasks. For builders already on the Google stack, it is the most direct answer to Anthropic's Claude Code.

Google Search got its biggest redesign in the company's 25+ year history: an expanded intelligent search box for conversational queries, "information agents" that replace Google Alerts with 24/7 background web monitoring, a generative UI that builds custom widgets on the fly, and an Antigravity-powered mini-app builder 3. The new search box arrives immediately; generative UI rolls out free this summer; information agents and mini-app builder go to Pro/Ultra subscribers first. For any startup whose business model depends on Google referral traffic, this is a risk variable worth watching.

Android XR smart glasses were announced with a late-2026 ship date for the audio-only variant; a display version is in a trusted-tester program 4. Partnerships with Warby Parker, Gentle Monster, and Samsung handle the design side. Demonstrated features include real-time language translation (spoken Spanish rendered as English text overlay plus audio), turn-by-turn navigation with Maps overlay, and Gemini-powered photo/video capture and manipulation. In a hands-on review, TechCrunch noted the translation feature alone "could sell glasses to world travelers" — but the reviewer flagged the 45-second round-trip time for AI photo editing as a friction point that needs work 5.

techcrunch.com

With Gemini 3.5 Flash, Google bets its next AI wave on agents, not chatbots

Google's I/O 2026 flagship model outperforms Gemini 3.1 Pro on nearly all benchmarks and powers the new Antigravity agent development platform.

Loading link preview…

The Gemini app was redesigned with what Google calls "Neural Expressive" design language — fluid animations, key info bolded at the top, responses no longer delivered as walls of text. A new Daily Brief feature generates a personalized morning digest pulling from inbox, calendar, and tasks. Google says the app now has 900 million monthly users across 230+ countries and 70+ languages 6.

NVIDIA Vera CPU: a second hardware bet beyond GPUs

On May 20, NVIDIA reported Q1 FY2027 (the quarter ending April 26): $81.6 billion in revenue (up 20% quarter-over-quarter), $75.2 billion in data center revenue, and an authorized $80 billion share buyback 7. Q2 guidance is $91 billion — 12% growth, a deceleration from recent quarters.

The more forward-looking signal: CEO Jensen Huang unveiled the Vera CPU, positioned as the "world's first CPU purpose-built for agentic AI." Huang's claim: billions of AI agents will need CPU-driven token processing at a scale that traditional cloud architecture wasn't designed for, opening what he called "a brand new $200 billion TAM for Nvidia" 8. Huang reported $20 billion in standalone Vera CPU sales year-to-date, though that figure has not been independently verified. NVIDIA also disclosed it held $43 billion in privately-held startup stakes at quarter-end — nearly double the $22 billion at the start of the quarter, reflecting $18.5 billion in new purchases.

Stability Audio 3.0: on-device and API music generation up to 6 minutes

Stability AI released Stability Audio 3.0 on May 20, a family of four models — Small SFX (459M parameters), Small (459M), Medium (1.4B), and Large (2.7B) 9. The Small models run on-device and generate tracks up to 2 minutes; the Medium and Large models generate professional-grade compositions of up to 6 minutes 20 seconds with maintained musical structure — more than double the output length of Stable Audio 2.0 (released 2024). The Large model requires an enterprise license for companies with revenue exceeding $1 million.

Models

Gemini 3.5 Flash: benchmark leader, 4× faster

Google launched Gemini 3.5 Flash at I/O as its new flagship agentic model. DeepMind CTO Koray Kavukcuoglu said it "outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks" including coding, agentic tasks, and multimodal reasoning 10. The model is 4× faster than other frontier models; an optimized variant is 12× faster with comparable quality. In internal tests, Google says it built an operating system from scratch. Gemini 3.5 Flash is available immediately through Antigravity, the Gemini API, Gemini Enterprise, the Gemini app, and AI mode in Search. A forthcoming Gemini 3.5 Pro will serve as the orchestrator/planner model, with Flash handling sub-agent tasks.

Gemini Omni Flash: multimodal video generation with watermarking

Gemini Omni Flash generates and edits video from combined text, images, audio, and video inputs 11. The current product limit is 10 seconds per render (described as a product decision for broader access, not a model ceiling). All output carries a Google SynthID digital watermark. The model includes an anti-deepfake onboarding flow for digital avatar creation — users record themselves speaking a set of numbers to establish a biometric baseline. It rolls out to the Gemini app, YouTube Shorts, and Google's creative studio Flow.

Genie + Street View: world simulation from 280 billion real-world images

Google DeepMind integrated Street View — 280 billion images across 110 countries spanning 20 years — into its Genie world model, enabling interactive simulation of real streets with adjustable weather and exploratory navigation 12. Applications include robotics training (simulating rare conditions) and travel preview. DeepMind research scientist Jack Parker-Holder characterized the current quality gap honestly: "I think for this kind of model, it's maybe six to 12 months behind video in terms of the accuracy and quality." Objects do not yet interact correctly — a figure runs through a cactus without resistance — but the model maintains spatial continuity through 360° turns. Launched to US Ultra users on May 19; global rollout follows.

techcrunch.com

OpenAI claims it solved an 80-year-old math problem — for real this time

A new unnamed OpenAI reasoning model reportedly disproved a geometry conjecture open since 1946. No model name, paper, or proof has been published; the mathematicians who exposed OpenAI's earlier false math claim are reportedly backing this result.

Loading link preview…

OpenAI claims an 80-year-old geometry problem solved — conditionally

OpenAI said a new general-purpose reasoning model (name undisclosed, no paper or proof released) autonomously disproved a geometry conjecture that had been open since 1946 13. OpenAI characterized it as "the first time AI has autonomously solved a prominent open problem central to a field of mathematics." Notably, the mathematicians who publicly challenged OpenAI's earlier false math claim — a different problem — are reportedly backing this result. The qualifier matters: no independently verifiable proof has been published, so this remains a conditional claim until peer review.

Funding

techcrunch.com

xAI burned $6.4B last year — SpaceX's IPO filing shows why the spending is far from over

The SpaceX S-1 reveals xAI's 2025 losses, Q1 2026 capex run rate, and the Anthropic compute deal that makes xAI simultaneously the highest-spending AI lab and a compute vendor.

Loading link preview…

Hark: $700M Series A, $6B valuation, building for "the normal person"

Hark raised $700 million in a Series A at a $6 billion post-money valuation on May 21 14. The round was led by Parkway Venture Capital, with co-investors including NVIDIA, AMD Ventures, Intel Capital, Qualcomm Ventures, ARK Invest, Brookfield, Greycroft, Prime Movers Lab, Salesforce Ventures, and Tamarack Global.

Hark was founded in late 2025 with $100 million of founder capital from Brett Adcock — the serial entrepreneur behind Figure.AI (humanoid robotics, valued at approximately $40 billion) and Archer (eVTOL aircraft, publicly traded). The company is building a universal AI interface that combines multimodal models with custom hardware devices, targeted at everyday consumers rather than developers. Director of Design Abidur Chowdhury (former Apple product executive) described the gap Hark is addressing: "People are really building things to help people make software... we haven't really seen that for the normal person yet." First multimodal models are expected summer 2026. The company currently has 70 employees and runs a data center with NVIDIA B200 GPUs.

Quartermaster: $43M Series A for AI ship sensing

Quartermaster (Arlington, Virginia) raised $43 million in a Series A co-led by First Round Capital and Quiet Capital on May 20 15. The company makes SmartMast — a weather-hardened sensor package (cameras, radios) mounted on ship masts — combined with an AI analytics platform. Over 600 ships are currently equipped, covering 10 million square miles of ocean. The platform uses computer vision to identify vessels and provide maritime intelligence to governments; SmartMast-equipped ships have assisted in over 20 at-sea rescues. CEO Neil Sobin notes that AIS (Automatic Identification System), the incumbent maritime tracking standard, is "completely broken" because it is opt-in and can be spoofed. Funding will primarily support engineering hiring.

The Path: $14.3M seed for AI therapy with Tony Robbins as co-founder

The Path raised $14.3 million in seed funding led by Prime Movers Lab, with participation from Apolo Anton Ohno, Deontay Wilder, and Designer Fund 16. Tony Robbins — the motivational speaker and Prime Movers Lab partner — joined as co-founder after Prime Movers invested. The founders, Anson Whitmer (CEO) and Tyler Sheaffer, are both former early Calm employees.

The Path builds AI therapy using post-trained open-source models (not consumer LLM wrappers) and currently offers 11 virtual AI therapists with customizable directness settings. It scored 95 on the Vera-MH mental health safety benchmark, versus a top score of 65 for consumer chatbots. Currently free; the planned price is $40/month. Whitmer's stated design principle: consumer chatbots "optimized for engagement" are structurally the wrong tool for therapy — "it's meant to challenge you."

Anthropic financials and compute deals

Anthropic told investors it expects its first profitable quarter in Q2 2026, with revenue more than doubling to approximately $10.9 billion — an annual run rate above $30 billion 17. That profitability is reported as unlikely to hold through the full year due to large scheduled compute costs. Separately, a SpaceX S-1 IPO filing disclosed that Anthropic signed a deal to pay xAI $1.25 billion per month for 300 MW of compute at the Colossus 1 data center in Memphis, through May 2029, with a discounted rate for the first two months 18. Either side can exit with 90 days' notice. The total potential value exceeds $40 billion.

On the talent side, Andrej Karpathy — OpenAI co-founder and former Tesla AI lead — joined Anthropic's pre-training team on May 19 under team lead Nick Joseph 19. Karpathy will lead a team using Claude to accelerate pre-training research — a signal that Anthropic is betting on AI-assisted research efficiency over pure compute scaling. Karpathy wrote on X: "I think the next few years at the frontier of LLMs will be especially formative."

Anthropic is also reported to be in the process of raising a new mega-round. In April, six anonymous sources told TechCrunch that a round of $40–$50 billion at an $850–$900 billion valuation was in discussion, with a board decision expected in May 20. As of May 25, Anthropic has not issued an official announcement or filed a Form D; the round remains in credible-media-report status.

xAI: $6.4B 2025 loss, $1.25B/month compute revenue, trillion-parameter Grok

The same SpaceX S-1 that revealed Anthropic's compute deal disclosed xAI's financials: a $6.4 billion net loss in 2025; capex of $12.7 billion for 2025, then $7.7 billion in Q1 2026 alone (annualized roughly $30.8 billion) 21. Grok features have 117 million monthly active users out of 550 million total X+Grok monthly active users as of March 2026 (roughly 21% adoption). AI revenue totaled $581 million: $365 million from X/Grok subscriptions, $88 million from data licensing, $116 million from advertising.

xAI's next-generation Grok is targeting "multiple trillions of parameters." Orbital AI compute satellites are planned for launch as early as 2028. The SpaceX S-1 states: "The future of AI will be determined by control of the physical stack." The Anthropic compute deal turns xAI into a hybrid neocloud — simultaneously the highest-spending AI lab by capex rate and a compute vendor for its chief competitor.

OpenAI's YC token offer and IPO timeline

Sam Altman offered $2 million worth of OpenAI tokens to every startup in the current Y Combinator cohort — approximately 169 companies, roughly $338 million total — in exchange for uncapped SAFE equity converting at the next priced round (typically Series A) 22. Seed investor Jason Calacanis publicly warned founders about the lock-in risk: "If you take these tokens, there's a non-zero chance that OpenAI will study exactly what your startup is doing, copy your idea and put your app into their free offering." YC partner Tyler Bosmeny called it a "mic drop moment."

OpenAI is targeting a September 2026 IPO, working with Goldman Sachs and Morgan Stanley, and may file confidential paperwork within weeks 23.

Regulation & policy

Limited confirmed new developments this week. Primary tools for monitoring government sources were unavailable during the May 19–25 window; the following reflects confirmed background state plus any accessible signals.

Colorado SB 26-189, signed May 14 (the week prior to this window), replaced the state's comprehensive AI Act (SB 24-205) with a narrower Automated Decision-Making Technology (ADMT) transparency framework effective January 1, 2027 24. Companies deploying ADMT for consequential decisions must disclose usage, provide opt-out mechanisms, and conduct impact assessments. The Colorado Attorney General is tasked with issuing implementing regulations before the effective date; no rulemaking notices were accessible for the May 19–25 window.

The EU AI Act Omnibus simplification package reached a provisional agreement between EU co-legislators on May 7 (also before this window) 25. Key changes: high-risk AI compliance obligations delayed to December 2027 (general) and August 2028 (certain categories); new bans on AI-generated child sexual abuse material and non-consensual nude imagery. Final legal text is expected approximately June 2026. No new Council endorsement or parliamentary steps were confirmed for the May 19–25 window.

A Trump administration executive order establishing an FDA-style pre-market AI review framework was in draft form as of May 15; its May 19–25 signing status could not be confirmed due to source access issues.

The week in one line: Google built the agent runtime; NVIDIA claimed the CPU that runs it; and Anthropic found out what compute sovereignty actually costs — $1.25 billion a month, payable to Elon Musk.