The AI legal wars just got a landmark ruling, Cursor dropped a model that rewrites the economics of coding agents, and the power grid quietly became the most contested infrastructure in tech. Here's what matters today.
Musk's $100B OpenAI Lawsuit Dismissed — And Why It Actually Matters
A California jury unanimously threw out Elon Musk's blockbuster lawsuit against OpenAI, Sam Altman, Greg Brockman, and Microsoft. The case — valued north of $100B — collapsed on a statute of limitations defense. Musk had alleged that Altman and Brockman "stole a charity" by converting OpenAI's nonprofit structure to a for-profit. OpenAI's lawyers built their defense around a simple timeline: Musk knew about the restructure for years before suing in 2023, long after California's two-year window closed. Musk called it a "calendar technicality" and vowed to appeal. The Microsoft claim was also dismissed.
Here's everything you need to know:
- The jury took three weeks to reach a unanimous verdict
- OpenAI argued Musk backed the for-profit conversion himself before launching xAI in 2023
- The ruling clears a major legal cloud hanging over OpenAI's path to going public
- OpenAI can now sell itself as investable AI infrastructure without a governance lawsuit attached
- Musk's appeal could take years and face a high bar given the statute of limitations facts
- Microsoft was dismissed as a co-defendant; its relationship with OpenAI remains intact
- The case was widely seen as a proxy fight over AI governance and Musk's stake in the industry's direction
The so-what is blunt: OpenAI just removed the biggest asterisk on its balance sheet. For founders and builders watching the AI infrastructure landscape, a legally clean OpenAI is a more formidable competitor — and a more reliable trading partner. This wasn't just a lawsuit. It was a shot across the bow of the industry's restructuring. The bow is still standing.
Cursor's New Model Approaches Frontier — At 10% Of The Cost
Cursor dropped Composer 2.5, an upgraded in-house coding model built on Moonshot's Kimi K2.5, and the benchmarks tell a story that should make every coding agent maker nervous. The model approaches Anthropic's Opus 4.7 and OpenAI's GPT-5.5 across top development benchmarks while delivering roughly a 10% improvement over its predecessor. The striking number: average CursorBench task costs under $1 with Composer 2.5. Opus 4.7 and GPT-5.5 at equivalent score levels run up to $11 per task.
Here's everything you need to know:
- Composer 2.5 was partially trained on Colossus 2
- Cursor is currently training a larger SpaceXAI model with 10x more compute
- The model approaches frontier-level performance on development benchmarks
- Average cost per CursorBench task: under $1 (vs. up to $11 for comparable Opus 4.7/GPT-5.5 scores)
- Built on Moonshot's Kimi K2.5 base
- The economics upends the assumption that frontier-quality coding requires frontier-level spend
The implication isn't subtle: if you are building a coding agent product and your cost per task is 10x higher than the new baseline, you have a pricing problem that will compound. Cursor isn't just competing on accuracy anymore — it's competing on unit economics. The model is a direct signal that the next phase of the coding agent race is won on margin, not just capability.
Anthropic Acquires Stainless — Owning The Developer Platform Stack
Anthropic acquired Stainless, the startup behind its official SDKs and MCP server tooling. The team previously built Claude's core developer libraries. The acquisition strengthens Claude's SDK, CLI, and MCP tooling at a moment when agent connectivity is becoming the defining platform battleground.
Here's everything you need to know:
- Stainless built the official SDKs used by thousands of developers interfacing with Claude
- MCP (Model Context Protocol) tooling is central to how AI agents connect to external tools and data
- The acquisition keeps Anthropic's developer stack in-house as competition for SDK dominance intensifies
- Agent connectivity is increasingly a differentiating factor in enterprise AI deployments
For founders building on Claude: expect tighter integration, better tooling, and a more opinionated platform. For builders watching the SDK wars: this is Anthropic signaling it won't cede the developer experience to third parties. The race to own the agent-to-toolchain layer just got a new bet.
xAI Launches Grok Build — Entering the Coding Agent Race
xAI launched Grok Build, an early-beta coding agent that puts Grok directly into the developer terminal. The move positions xAI as a direct competitor to Cursor, GitHub Copilot, and OpenAI's Codex.
Here's everything you need to know:
- Grok Build is currently in early beta
- Targets the developer terminal as the primary interface
- Competes directly with Cursor, GitHub Copilot, and OpenAI's Codex
- xAI is now a player in the coding agent market, not just the model race
The terminal is where developers live. xAI is betting that Grok's personality and capabilities can earn a seat in that workflow. Whether it can displace established players depends on how fast the beta ships quality and how Grok handles the boring stuff — code completion, context windows, API integration — that determines whether a coding tool actually ships.
NextEra Buys Dominion for $67B — Buying The Power Socket Closest To AI
NextEra Energy closed the largest US utility merger in history, paying $67 billion for Dominion Energy on May 18, 2026. The strategic logic is straightforward: Dominion sits on top of Northern Virginia's Data Center Alley, holding 51 gigawatts of signed data center demand — roughly fifty nuclear plants worth. NextEra isn't buying poles and wires. It's buying the power socket closest to AI.
Here's everything you need to know:
- NextEra is acquiring Dominion Energy for $67B — the largest US utility merger on record
- Dominion holds 51 GW of signed data center demand in Northern Virginia
- Virginia's data center load is projected to grow 121% by 2045
- PJM wholesale electricity prices spiked 76% in a recent report using the word "irreversible"
- Sixty Virginia data centers tripped offline last winter, dumping 1,500 MW of surplus in seconds
- "Large load tariffs" are quietly shifting transmission buildout costs onto hyperscalers
- Estimates suggest residential customers may absorb $700 billion of AI grid upgrades through their bills
The hidden invoice for AI's $400 billion capex cycle isn't on the balance sheet — it's on monthly utility bills. Every builder relying on cloud infrastructure should understand that power pricing is now an AI cost driver, not a background constant. This merger is the clearest signal yet that whoever controls electrons controls the pace of AI scaling.
Meta & Anduril Win A $159M Army Contract For Battlefield AR
The US Army and defense startup Anduril have built a battlefield AR prototype — built in partnership with Meta — that fuses AI, drones, and targeting into soldiers' line of sight. Anduril holds a $159M Army prototyping contract. Soldiers use eye-tracking and voice commands to cue drone or artillery strikes from the headset. Anduril is also self-funding EagleEye, an integrated helmet-and-headset combo with Meta; production expected after 2028. EagleEye plugs into Anduril's Lattice command-and-control platform — the same software the Army awarded a $20B integration contract in March.
Here's everything you need to know:
- Anduril holds a $159M Army prototyping contract for AR glasses built with Meta
- Soldiers use eye-tracking and voice commands to cue drone or artillery strikes from the headset
- EagleEye self-funded by Anduril, production expected after 2028
- Lattice platform won a $20B Army integration contract in March
- Microsoft's own $22B battlefield AR contract was cancelled after system failures
Microsoft's failure left the Army slot open. Whoever delivers a viable headset embeds their platform — and AI stack — into Pentagon procurement for years. The defense AI market isn't theoretical anymore. It's a procurement pipeline with real contracts and real compute requirements.
ArXiv Implements One-Year Ban For AI-Generated Scientific Slop
ArXiv, the world's dominant scientific preprint server, is enforcing a one-year ban for researchers who submit papers with hallucinated references or unchecked LLM output. The CS chair Thomas Dietterich stated that papers where authors failed to verify LLM output "can't be trusted." Red flags include hallucinated references and LLM prompts/responses left in submissions. After the one-year ban, all future submissions from the affected author must first be accepted by a peer-reviewed venue.
Here's everything you need to know:
- ArXiv CS chair Thomas Dietterich announced enforcement against AI-generated scientific content
- Triggering violations: hallucinated references, LLM prompts/responses left in submission
- One-year ban for violators, with future submissions requiring prior peer-review acceptance
- Moderators flag violations, section chairs confirm evidence, authors retain appeal rights
- ArXiv circulates cutting-edge research before peer review — AI hallucinations corrupt the record upstream
- Fabricated citations are already rising in biomedical literature
ArXiv's enforcement model is likely to be widely adopted. The preprint server sits at a critical choke point in scientific publishing — whatever gets through there spreads before peer review catches it. For builders working on scientific AI or research tooling: this is a preview of how AI output verification becomes a first-class requirement in high-trust domains.
China Switches On The World's First Offshore Wind-Powered Underwater Data Center
China began full commercial operation of the world's first offshore wind-powered underwater data center — a $226M facility sitting 30 feet beneath the East China Sea off Shanghai's Lingang Special Area. The 24 MW facility draws 95% of its electricity from wind generation and uses seawater for passive cooling without industrial chillers. It houses nearly 2,000 servers including GPU clusters from China Telecom and LinkWise for AI workloads. It achieves a PUE of 1.15 and claims electricity consumption down 22.8%, zero freshwater use, and land use reduced over 90%.
Here's everything you need to know:
- $226M facility, 30+ feet underwater in the East China Sea
- 24 MW with nearly 2,000 servers including GPU clusters for AI workloads
- Draws 95% of electricity from wind generation
- Passive seawater cooling — no industrial chillers
- PUE of 1.15, electricity consumption down 22.8%, zero freshwater use
- The unresolved problem: maintenance of servers 30 feet underwater is costly and complex
- Microsoft shelved Project Natick for the same reason
The maintenance tension is real. When a server fails underwater, the logistics are expensive and slow. China's showing that the concept works in controlled conditions. Whether it scales depends entirely on whether the economics hold when failure rates climb with density.
Odyssey Releases Real-Time Multimodal World Model And Multiplayer Variant
Odyssey released Starchild-1, the first real-time multimodal world model that generates synchronized audio and video on the fly while adapting to user inputs with no fixed generation length. Also dropped Agora-1, allowing up to 4 players to interact inside the same AI-generated world stream — demoed via a GoldenEye video game simulation where every pixel is produced live. Agora maintains shared game state across participants.
Here's everything you need to know:
- Starchild-1 generates synchronized audio and video in real time with no fixed output length
- Adapts to user inputs dynamically during generation
- Agora-1 enables up to 4 players in the same AI-generated world stream
- Demo used a GoldenEye simulation where every pixel was produced live
- Shared game state maintained across all participants
- Targets: multiplayer games, robotics, and agent training in simulations
The signal here is that AI-generated environments at interactive frame rates are real, not vaporware. For game developers and robotics engineers: this is early-stage but the trajectory is clear. AI-generated worlds you can step inside — with other people — are a product category now.
⚡ Quick Hits
-
OpenAI: Partnered with Dell to run Codex inside corporate data centers, connecting coding agents to enterprise private systems and internal data.
-
OpenAI: Partnered with Malta to offer free ChatGPT Plus to every citizen completing a national AI literacy course — the first country-wide deal of its kind.
-
Anthropic: Co-signed with Pope Leo XIV on AI ethics encyclical Magnifica Humanitas, launching May 25. Anthropic co-founder Christopher Olah sits on the lay panel alongside the Vatican's doctrine chief.
-
Amazon: Rolled out Alexa Podcasts in Alexa+, a NotebookLM-style feature that creates a two-host AI conversation on any topic.
-
Meta: Laying off up to 8,000 employees this week while no longer planning to hire for another 6,000 open roles, as part of its AI efficiency push.
-
Shopify: AI-referred orders grew nearly 13x year over year, signaling agentic commerce is becoming a real shopping channel.
-
Brookhaven National Lab: Building a $1.7B–$2.8B AI-native particle collider where machine learning is a design constraint from day one, not bolted on later. Switches on mid-2030s.
-
Nvidia: Said enterprise AI infrastructure demand is going "parabolic" as Dell pushes Rubin-based agentic workloads into production planning.
Techlook — AI & tech signal for founders and builders.