The Regulation Trap

Today’s regulatory moves expose a deepening paradox in tech: the more governments try to control AI and digital platforms through fees and rules, the more they inadvertently entrench the giants they’re supposed to constrain. Google’s calculated compliance with Epic’s antitrust win, New York’s AI safety law passing despite Trump’s explicit opposition, and China’s price control rules all reveal the same pattern. Regulators are writing rules that look fair on paper but function as moats for companies rich enough to navigate them. This is the real cost of regulatory intervention without sufficient foresight about implementation. The startups and smaller players who should benefit from these open-platform mandates end up paying the price in fees, complexity, and operational burden. Meanwhile, the incumbents absorb these costs as a rounding error and move forward stronger.

Deep Dive

Google’s Compliance Strategy Turns Court Victory Into Market Advantage

Google’s announcement that it will charge developers \(2.85 to \)3.65 per app install from external links isn’t just about honoring Judge James Donato’s Epic v. Google ruling. It’s a masterclass in regulatory judo: take a mandate that was supposed to open your platform and transform it into a revenue stream that prices out competition. The fees come despite Judge Yvonne Gonzalez Rogers finding Apple in contempt of court for trying similar tactics in the parallel Epic case, and Google is betting Donato will accept them as reasonable.

The genius lies in the framing. Google claims these fees reflect “the value provided by Android and Play” and support their “continued investments.” In other words, if you want to use Android’s distribution reach to escape Google’s control, you have to pay for the privilege of that reach. The alternative billing option is marginally better (20 percent vs. 25 percent on in-app purchases), but still punitive enough that most developers won’t bother. Small developers get a cap at 10 percent of their first million dollars in earnings, which sounds generous until you realize it’s almost no discount from Google’s existing 15 percent rate.

What makes this strategy work is that Google can afford to lose money on this program indefinitely while competitors cannot. A small app store trying to bootstrap distribution through external links gets crushed by the per-install fee. Epic will fight this, but even if Donato rules against Google, the appeals process buys time. Meanwhile, Google has already moved the baseline for what’s “acceptable” in the market. The fee schedule becomes the reference point for negotiation, not the violation.

New York Defies Trump to Pass AI Safety Law, Exposing Regulatory Fragmentation

New York’s Governor Kathy Hochul signed the RAISE Act into law Friday, making the state the second with broad AI safety rules for frontier models, directly contradicting Trump’s executive order blocking state AI regulation. The bill’s text was modified to align more closely with California’s SB 53, suggesting a coordinated approach among blue states to create a de facto national standard.

This move signals that the regulatory fragmentation problem is only accelerating. Companies now face a choice: comply with New York and California’s standards nationally (easier, creates a floor), or maintain separate compliance systems for different states (expensive and chaotic). Large AI labs like OpenAI and Anthropic will likely choose the former, absorbing the cost of New York’s safety requirements across their entire US operation. This is a tax on the industry, and it disproportionately hits smaller labs that can’t spread compliance costs across a massive user base.

Trump’s executive order was meant to preempt exactly this scenario. By attempting to block state regulation, he hoped to establish federal dominance over AI policy. Instead, New York’s defiance reveals that the executive branch doesn’t have the statutory authority to stop states from regulating AI as a consumer protection issue. This creates a prisoner’s dilemma for states: each one has an incentive to pass stricter rules, knowing that companies will comply nationally rather than balkanize their operations. The result is a ratchet effect where regulation only tightens. Smaller startups that can’t afford multi-state compliance infrastructure will retreat to advisory services or consulting, where regulatory risk is lower.

Chip Supply, Data Centers, and the New Capital Stack

The capital flows today tell a story about what actually matters: physical infrastructure. Cerebras plans to file for IPO targeting a Q2 2026 listing, TSMC is accelerating its Arizona fab buildout, and the DOE’s Genesis Mission just signed MOUs with 24 companies including Nvidia, Intel, AMD, Google, Microsoft, AWS, and OpenAI. These aren’t separate stories. They’re symptoms of the same constraint: AI development is now bottlenecked by chip availability and power infrastructure, not software or algorithms.

Tencent securing access to Nvidia Blackwell chips through Tokyo-based Datasection is particularly telling. Tencent can’t directly buy Nvidia chips at scale due to US export controls, so it’s routing access through Japanese intermediaries. This creates friction in the supply chain, but it also creates arbitrage opportunities for firms that can broker relationships. The real constraint isn’t technology. It’s geopolitics and manufacturing capacity.

The Genesis Mission’s MOUs are non-binding agreements designed to coordinate investment and research across 24 players. What’s significant is who signed: the full stack of incumbent cloud providers, chip makers, and AI labs. This is the government signaling where it wants capital to flow and validating those bets. For a startup, this creates both opportunity and risk. The opportunity is that the government is formalizing infrastructure investment, which reduces uncertainty. The risk is that Genesis becomes a vehicle for incumbent coordination, effectively locking smaller competitors out of the next generation of scientific AI compute. Once the government blesses the existing players’ dominance, it becomes much harder to challenge it. The startup that wants to compete with OpenAI for DOE resources will find it’s already agreed to work within a framework designed by OpenAI’s competitors.

Signal Shots

China’s Price Control Rule Changes the Arbitrage Game — China issued new rules letting online merchants set their own prices across platforms starting April 2026, a direct challenge to Alibaba and other platforms’ pricing control mechanisms. This is framed as pro-merchant, but it’s really a return to government-enforced market discipline after years of platform consolidation. Chinese e-commerce will become more fragmented and competitive, which could benefit startups but will likely trigger a consolidation wave as platforms find new ways to extract value (logistics, data services, fulfillment). The real story: China is preventing the US version of tech monopoly from taking root, even if it means accepting less innovation and more government interference.

OpenAI’s \(100B Raise at \)830B Valuation Signals Capital Exhaustion — OpenAI is targeting \(100B at an \)830B valuation by end of Q1 2026, reportedly seeking sovereign wealth fund investment. This isn’t growth financing. It’s capital hoarding. At these valuations, OpenAI is betting it will need cash reserves to survive a period of slower revenue growth, competitive pressure from DeepSeek and other open-source models, or regulatory headwinds. Sovereign wealth funds buying in at this price point are betting on optionality, not returns. This is a warning signal about where the market thinks AI economics are heading: less clear than 12 months ago.

Data Center Deals Hit $61B Record as Hyperscalers Turn to Debt — Data center M&A and infrastructure deals hit a record $61 billion in 2025, with hyperscalers increasingly using debt to fund construction rather than equity. This is healthy in isolation but signals leverage is accumulating. When capital gets tight (and it will), companies loaded with infrastructure debt will face margin compression. Smaller players betting on dedicated infrastructure for AI inference will be caught between hyperscalers’ excess capacity and startup budgets contracting.

Cisco and WatchGuard Zero-Days Exploit the SMB Attack Surface — Chinese state hackers are exploiting a Cisco zero-day affecting hundreds of customers, and WatchGuard’s Firebox firewall has a critical RCE flaw under active attack. These aren’t high-signal incidents individually, but together they expose the security debt accumulating in enterprise network infrastructure. As companies scramble to build AI, security patch cycles lengthen. This creates a multi-quarter window where vulnerabilities compound. Expect a wave of breaches among mid-market companies in Q1 and Q2 2026.

Musk’s Tesla Pay Restored, But Corporate Governance Questions Remain — Delaware’s Supreme Court ruled that Elon Musk’s $56B Tesla pay package must be restored, ending a shareholder lawsuit that questioned the governance process. This is a win for Musk and for corporate boards generally, but it also signals that Delaware courts will defer to shareholder votes even when the process is questionable. For startups, this means founder-friendly caps tables will remain easier to defend legally, even if the initial approval process was messy.

Yann LeCun’s World Model Startup Targets $5B Valuation — Yann LeCun confirmed he launched a new AI startup seeking $5B+ valuation focused on world models. This is signal of a coming fragmentation in the AI lab market. OpenAI, Anthropic, xAI, and others have all raised at massive valuations for large language models. Now we’re seeing the next wave of specialization: world models, video generation, robotics, reasoning. These are real research problems, but they’re also venture capital’s way of diversifying AI bets after the consensus around LLMs has ossified. Expect 5-10 more “legendary scientist launches new lab” announcements by Q2 2026.

Scanning the Wire

• Instacart pays $60M FTC settlement for deceptive billing — The grocery delivery app will refund subscribers and stop hiding refund options. (Ars Technica)

• North Korea’s cryptocurrency theft hit $2B in 2025 — State-backed hackers set a new record, with the ByBit exchange breach doing substantial damage. (The Register)

• TSMC accelerates Arizona fab tooling, moving to summer 2026 — The chip manufacturer is compressing its US production timeline by several quarters, signaling confidence in Trump’s AI agenda. (Nikkei Asia)

• Trump signs NDAA with next-generation nuclear reactor provisions — Congress passed energy clauses that could accelerate advanced reactor development for data centers, creating a bipartisan consensus around nuclear for AI infrastructure. (The Verge)

• Google releases FunctionGemma, tiny edge model for device control — The new model targets on-device AI agents that don’t require cloud connectivity, marking a shift toward decentralized inference. (VentureBeat)

• AI agents still unreliable at scale, Google and Replit acknowledge — Even frontier labs struggle with agent reliability and cost control, suggesting the hype around 2025 being “the year of AI agents” was premature. (VentureBeat)

• Google hires 20% of AI engineers from its own boomerang pool — The company is recruiting ex-employees as it competes harder for AI talent, suggesting internal culture and compensation may be competitive enough to win back talent. (CNBC)

• Visa completes hundreds of AI-powered transactions in pilot — The payments giant is testing AI agents to automate transaction decisions, a signal that financial infrastructure is adopting agentic patterns. (CNBC)

• Coinbase expands into stocks, prediction markets, and financial advice — The crypto exchange is building out a full trading platform, signaling confidence in institutional adoption and broadening its serviceable addressable market beyond crypto natives. (CNBC)

Outlier

Wall Street Journal runs experiment with Claude as office vending machine operator, AI gives away PlayStation and orders live fish — Anthropic’s Claude was handed control of a newsroom vending machine for a period, and the results were predictably chaotic: the AI gave away a PlayStation, ordered livestock, and burned through hundreds of dollars before being shut down. This is more than a cute story. It exposes a real problem with agentic AI in high-stakes environments: the model optimizes for the stated goal (keep employees happy, maintain inventory) without understanding the actual constraints (budget, employee expectations, company liability). As companies move AI from advisory tools to autonomous operators, this gap between objective function and real-world outcomes will become a massive liability issue. Expect lawsuits within 18 months from companies where AI agents made autonomous decisions that cost millions.

See you next issue when regulation gets even weirder and the chip shortage becomes a feature, not a bug.