The Splinter Effect
The Splinter Effect
The year-end data is clarifying something investors have been sensing for months: the AI market isn’t consolidating around a winner. It’s fragmenting into fundamentally different businesses with incompatible economics. On one side, you have the infrastructure play (Nvidia licensing Groq’s tech for \(20 billion). On the other, the agent builders (startups flooding the market with shopping assistants). And in between, the autonomous systems (Tesla's 30 cars versus Waymo's 200). The billion-dollar question isn't who wins anymore. It's whether these different AI markets can coexist or if the capital flowing in (\)200 billion across startups in 2025, up from 34% of global funding to 50%) eventually has to pick winners. The answer will determine which of this year’s 50 new AI billionaires still have meaningful wealth in 2027.
Deep Dive
Amazon’s Agent Defense: The Moat You Didn’t Know It Had
Amazon’s move to block AI shopping agents from accessing its platform while simultaneously building its own tools isn’t about blocking competitors. It’s about preserving the most valuable thing in e-commerce: customer behavior data and the friction-free transaction. Third-party shopping agents represent a fundamental threat not because they’re good at shopping but because they’re good at routing purchases away from Amazon’s recommendation engine and into a normalized, commoditized comparison layer.
What makes this defensible is that Amazon controls the infrastructure. When a third-party agent queries Amazon’s catalog, it’s making requests through APIs that Amazon can monitor, throttle, or redirect. The company can afford to block agents because it doesn’t need agent traffic to drive volume. What it needs is to own the moment when a customer decides to buy. By building its own agent, Amazon isn’t trying to beat startups at agent technology. It’s trying to own the translation layer between “I want a widget” and “I’m buying from Amazon.”
The broader signal: platformers with first-party data advantage will win the agent wars, but not by building better agents. They’ll win by controlling the conditional logic that sits between the user query and the transaction. This is why Alibaaba and ByteDance are investing in agents in China at the same time Amazon is in the U.S. The technology isn’t where the moat is. The moat is the ability to make agents default to your marketplace. For startups building agents, this means the real TAM isn’t e-commerce transactions. It’s verticals where no single platform has locked-in behavior yet, or where customers haven’t yet formed strong defaults. Expect the next wave of agent funding to focus on healthcare, travel, and professional services, not retail.
Nvidia’s Groq Licensing Is a Concession Disguised as Expansion
Nvidia paying Groq $20 billion to license its inference technology, with cash flowing directly to founders and investors like BlackRock and Tiger Global, should be read as a strategic admission. Nvidia doesn’t license technology from smaller chip companies. Nvidia acquires them or crushes them. The fact that it’s licensing from Groq signals that custom silicon optimized for specific workloads (in Groq’s case, inference) has become too differentiated to ignore, and that Nvidia’s architectural advantages in training don’t fully translate to serving pre-trained models at scale.
What Nvidia is really buying is credibility in the inference market and the ability to tell enterprise customers that it’s hedging its own bet on generic GPU dominance. It’s also buying out of a competitive threat that was starting to get real. Groq raised $1.8 billion in venture funding on the promise that it could do inference cheaper and faster than Nvidia. Instead of letting that narrative mature, Nvidia is absorbing it. The move also changes the incentive structure for other inference-focused chip startups. If you’re competing with Groq, you now know that the outcome isn’t market share. It’s a licensing deal or an acquisition by one of the big three (Nvidia, AMD, Intel).
The second-order effect is subtle but important: this legitimizes the idea that AI infrastructure is splintering into specialized layers. Nvidia owns training. Groq owns inference optimization. Someone else will own deployment. This is actually bad for startups building horizontal AI tools because it means the platform layer is getting more fragmented, not more consolidated. Founders should watch where the next round of infrastructure investment goes. If it goes to the systems that can abstract across this fragmentation (like model serving platforms), that’s where the value will sit.
The Vibe Coding Reckoning Is Coming for AI-First Startups
Cursor’s CEO Michael Truell warning about “vibe coding” and “shaky foundations” isn’t just a technical concern. It’s a signal that the market is entering a phase where the velocity advantage of AI-assisted development is colliding with the complexity ceiling of real systems. For the past year, startups have been moving at unprecedented speed because they could. AI coding assistants lower the friction of writing code. But lowering friction and lowering technical debt are different things.
What Truell is really saying is that startups built on the assumption of “move fast and refactor later” are about to hit a wall where refactoring becomes impossible without rewriting core systems from scratch. This happens in every technology cycle, but it’s accelerated in the AI era because the speed-to-market advantage dissipates once code accumulates. A startup that shipped a scheduling algorithm in two weeks using Cursor might spend six months untangling it when it needs to scale to millions of users. The startups that win in the next phase will be the ones that learned to write foundation code in 2024 and 2025. The ones that will struggle are the ones that optimized for speed at the expense of architecture.
This matters structurally because it means the next wave of VC funding will likely be more skeptical of velocity metrics as a proxy for quality. Founders will need to show they’ve thought about technical debt, not just shipped quickly. And for tool makers like Cursor, the opportunity isn’t in making coding faster. It’s in making it possible to code at speed without sacrificing maintainability. That’s a much harder problem to sell, but it’s the one that matters after the honeymoon phase ends.
Signal Shots
Tesla’s Robotaxi Ambition Meets Reality — Tesla has deployed an estimated 30 robotaxis with safety drivers in Austin while Waymo operates 200 driverless vehicles in the same market. The gap isn’t just scale; it’s operational readiness. Tesla is still in the data collection phase while Waymo is in the revenue phase. What to watch: whether Tesla’s full self-driving data can compress the timeline to driverless operation, or whether the two-year head start Waymo has is insurmountable.
AI Funding Reaches Escape Velocity — 50-plus new AI billionaires emerged in 2025 with investors pouring over $200 billion into AI startups, representing 50% of all global venture funding. The concentration is real, but the diversification across founders is signal that the market isn’t winner-take-all. What to watch: whether wealth retention tracks performance. Many of these billionaires are on paper; real returns come when exits happen.
Crypto Dealmaking Breaks Records on Policy Tailwind — M&A in crypto hit $8.6 billion across 267 deals in 2025, up from \(2.17 billion in 2024, with 11 crypto IPOs raising \)14.6 billion globally. The Trump administration’s crypto-friendly stance has unfrozen capital that sat on the sidelines for years. What to watch: whether this momentum sustains if policy shifts, or whether it was purely momentum-driven and the real money goes elsewhere.
China’s State-Backed Venture Strategy Signals Long Game on Hard Tech — China launched three state VC funds with $7.1 billion each targeting early-stage “hard technology” startups. This is patient capital playing a 10-year game on semiconductors, batteries, and advanced materials. What to watch: whether U.S. policy responses (export controls, visa restrictions) can meaningfully slow this, or whether state backing makes timelines irrelevant for Beijing.
Regulatory Blowback Against Content Moderation Framing Escalates — A U.S. judge temporarily blocked Trump administration detention of CCDH CEO Imran Ahmed over claims of promoting censorship, signaling judicial skepticism of the government’s framing. This echoes broader European resistance to characterizing content monitoring as censorship. What to watch: whether this judicial skepticism holds as political pressure increases, or whether the framing shifts to national security rather than free speech.
Scanning the Wire
Humanoid Robot Reality Check — Robotics startups have raised billions on the promise of humanoid workers, but insiders admit their robots can only perform simple tasks and timelines remain speculative. The hype-to-capability gap is widening.
Apple’s Regulatory Chess in Brazil — Apple settled with Brazilian regulators to allow alternative app stores at a 5% fee, creating a template for how U.S. tech giants manage regional compliance without wholesale platform changes.
Data Centers Ascend from Infrastructure to Strategy — The shift from viewing data centers as backend utilities to core competitive assets reflects the reality that AI workloads have inverted the importance hierarchy. Compute location now matters.
Abbott Glucose Monitor Failures Highlight Hardware Liability Risk — Seven diabetes patients died due to undisclosed bugs in Abbott’s Freestyle Libre monitors, exposing the regulatory and legal risks when connected medical devices fail silently.
Waymo’s Gemini Integration Signals In-Car AI Consolidation — Waymo is testing Google’s Gemini as an in-car assistant, suggesting the autonomous vehicle layer will converge with the large language model stack rather than compete with it.
Texas Age-Verification Law Gets Blocked, but the Pattern Holds — Apple paused App Store compliance with Texas age-assurance law after court intervention, but the regulatory pressure will persist across states. Expect fragmented compliance rather than nationwide solutions.
Ellison Father-Son Bid for Warner Bros. Shows Consolidation Still Moves — Larry and David Ellison’s $108 billion hostile bid for Warner Bros. Discovery demonstrates that even established media giants remain acquisition targets for tech capital with scale aspirations.
Gmail Address Changes Finally Coming — Google is rolling out the ability to change your Gmail address, resolving a 20-year UX friction point. This seems minor until you realize how many identity systems are built on email immutability.
Outlier
The Last Frontier of Surveillance Arbitrage — While tech companies fight regulators over content moderation terminology, a Spanish entrepreneur’s identification of a computer virus author decades ago led to Google opening a major cybersecurity hub in Málaga. The signal: geographic arbitrage on cybersecurity talent is the new tech race that nobody’s talking about. Madrid, Lisbon, and other European tech hubs are becoming low-cost strongholds for security research precisely because they’re outside the U.S. regulatory spotlight, but close enough to influence global standards. Watch for more big tech investments in secondary European cities disguised as “regional expansion” but really about building expertise moats outside of D.C. and Brussels’ direct gaze.
See you in the next one. Or don’t. The robots will probably send it for you by then.