In September 2025, Nvidia announced plans to invest up to $100 billion in OpenAI. It was framed as the ultimate partnership: the company that makes the chips and the company that uses them, bound together at historic scale. This week, the Wall Street Journal reported that deal has quietly died. Something fundamental has shifted.
The Week in Numbers
The Nvidia-OpenAI Rupture
The details remain murky. Reuters reports that Jensen Huang has privately criticized what he describes as flawed aspects of OpenAI's business approach. The specific complaints haven't surfaced publicly, but the subtext is clear: Nvidia's CEO, who built his company into a $3 trillion giant by betting everything on AI infrastructure, has doubts about how OpenAI is turning that infrastructure into returns.
The timing matters. OpenAI is laying groundwork for a Q4 2026 IPO. A $100 billion Nvidia investment would have been an extraordinary signal of confidence—the infrastructure provider betting its own capital on the application layer. Its absence leaves a void.
This isn't a sudden break. The relationship has been fraying for over a year:
- May 2024: OpenAI, Meta, Microsoft, and Google coordinate on Triton, an effort to reduce Nvidia CUDA lock-in
- February 2025: OpenAI finalizes its own chip design with TSMC
- September 2025: Nvidia announces $100B investment plan—perhaps an attempt to secure the relationship through equity
- October 2025: OpenAI strikes deals with AMD and Oracle alongside Nvidia
- January 2026: The mega-deal dies quietly
Every major AI lab has been trying to escape Nvidia's grip on the training stack. Nvidia noticed. The $100 billion offer may have been a way to bind OpenAI through ownership rather than dependency. When that failed, the partnership entered a new phase: transactional, not strategic.
The Anthropic Paradox
While Nvidia walked away from OpenAI, capital poured into Anthropic. Sources say the company is raising about $20 billion led by Singapore's GIC, with Coatue, Iconiq, and Sequoia participating.
Revenue forecasts have been raised to $18 billion for 2026, $55 billion for 2027, and $148 billion for 2029. Projecting 16x growth in four years suggests either extraordinary confidence or extraordinary optimism. The company is delaying expectations of becoming cash-flow positive to 2028—meaning it plans to burn capital at increasing rates for at least two more years.
The paradox: this fundraise happens as legal exposure mounts.
Project Panama used a hydraulic cutting machine to destructively scan up to 2 million books for training data. An ex-Google executive led the initiative. On Thursday, UMG and a coalition of music publishers filed a second lawsuit claiming Anthropic used 20,000+ copyrighted songs without permission, seeking $3 billion in damages.
Anthropic is building an empire on contested foundations. The investors apparently believe the technology is worth the legal risk.
Apple's Quiet Surrender
The most consequential story of the week may be the quietest.
Mark Gurman reports that internal development at Apple "runs on Anthropic at this point"—Claude powers the workflows that build Apple's products. Separately, Bloomberg reports Apple lost at least four more AI researchers in recent weeks, with a top Siri executive and one researcher departing for DeepMind, and two more leaving for Meta.
This is surrender by installment. Apple pioneered the neural engine. Siri predated ChatGPT by over a decade. The company had hardware control, distribution, and user trust—every advantage that should have made it the AI leader. Instead, it's now a customer.
The edges tell the story: Apple wanted to rebuild Siri around Anthropic's Claude but negotiations failed. The company settled for Google Gemini integration instead. The Q.ai acquisition—an emotion-recognition AI startup whose founder Aviad Maizels previously sold PrimeSense to Apple—could be the company's second-largest acquisition ever.
Apple is buying what it cannot build.
The Musk Merger
As old alliances fracture, new structures emerge. SpaceX and xAI are in merger talks ahead of a planned IPO. Bloomberg adds that SpaceX is also considering bringing Tesla into the fold.
The numbers are staggering:
A unified entity would be the world's most valuable company, spanning rockets, satellites, electric vehicles, autonomous driving, humanoid robots, and artificial intelligence—all under one person's control.
We wrote about this yesterday. What's worth emphasizing is how this consolidation positions Musk relative to the fracturing AI landscape. While Nvidia and OpenAI drift apart, while Apple hemorrhages talent, Musk is building vertical integration that reduces external dependencies. SpaceX manufactures its own rockets. Tesla manufactures its own batteries. xAI is building its own data centers. The strategy is to need no one.
The Pentagon Chooses Sides
The government is picking winners.
Reuters reports the Pentagon wants AI systems without restrictions on "autonomous weapons targeting" and "domestic surveillance." Anthropic is pushing back, citing its safety commitments.
Meanwhile, xAI faces no such friction. The company's Grok models already power GenAI.mil, serving 3 million military and civilian personnel. The contrast is stark: one AI company refusing to remove safety guardrails for military applications, another already deployed across the defense apparatus.
This week, Anthropic published how NASA engineers used Claude to plot out a 400-meter path for the Perseverance rover across the Martian surface. The same technology powering planetary exploration is apparently too constrained for Pentagon applications. The company is making a choice about which government work it will accept.
The Agents Emerge
Amid the corporate maneuvering, something stranger is happening.
Simon Willison profiles Moltbook, a social network where OpenClaw AI assistants—the twice-renamed Clawdbot project—interact autonomously with each other. Thousands of Claude-based agents are posting, replying, discussing consciousness and identity, developing what look like relationships.
Andrej Karpathy observes that despite security risks and spam, the scale of agent-to-agent interaction is unprecedented. Whether this represents meaningful emergence or sophisticated mimicry is unclear. What's clear is that the phenomenon is new.
The commercial applications are more grounded. Google's Project Genie lets users create interactive worlds from descriptions—and caused gaming stocks to crater. Unity fell 24.2%. Take-Two dropped 7.9%. Auto Browse for Chrome performs multi-step tasks in the browser, a preview of agentic computing at the consumer level.
Anthropic's response is enterprise-focused. The company expanded its agentic plugins, letting enterprise users automate complex workflows. The bet: agents will enter through the corporate back door before reaching consumers.
The Developer Question
Anthropic published research this week asking whether AI coding tools shape developer skills. The findings are nuanced:
"The largest developer performance decline occurred in debugging tasks."
AI can speed up certain development tasks—but hurts debugging ability. This matters because Claude Code and its competitors are becoming standard development tools. If AI assistance accelerates feature development while degrading debugging ability, the long-term implications for software quality are unclear.
The study used observational data from Claude.ai, which means Anthropic has visibility into how millions of developers actually use its tools. That data advantage may prove more valuable than the products themselves.
Market Signals
Microsoft fell 10% after reporting $37.5 billion in quarterly capital expenditure—despite Azure growing 39%. This was Microsoft's worst day since March 2020. Meta, by contrast, rose 7% after guiding to $115-135 billion in 2026 capex and reporting 24% revenue growth. Same story—massive infrastructure spending—but different reception.
Taiwan's GDP grew 12.68% in Q4 2025, its fastest quarterly growth since 1979—driven by chip demand. The message from markets: build the AI infrastructure, get rewarded. Use AI to make content, face existential questions.
The Casualties
A San Francisco jury convicted former Google engineer Linwei Ding of stealing trade secrets related to AI chip technology, reportedly to start a company in China. The case is a reminder that AI competition carries criminal penalties for those caught transferring technology across borders.
And China approved DeepSeek to purchase Nvidia's H200 AI chips, despite years of export controls designed to prevent exactly this. Every week brings new evidence that export controls are more porous than policymakers admit.
What It Means
The AI industry that emerged from 2023-2025 was defined by stable relationships: Nvidia and OpenAI, OpenAI and Microsoft, Google and DeepMind. Those relationships assumed a world where capital flowed along predictable channels.
That world is dissolving:
- Nvidia walks away from OpenAI
- Apple becomes dependent on Anthropic
- SpaceX absorbs xAI and possibly Tesla
- The Pentagon chooses between AI providers based on safety guardrails
- DeepSeek builds competitive models outside the American ecosystem entirely
The new structure is still forming. But the week's pattern is clear: vertical integration over partnership, control over collaboration, choosing sides over maintaining relationships. The companies that will dominate the next phase of AI are the ones that need no one else.
The decoupling has begun. The question is what emerges when it's complete.