The Infrastructure Scramble

The tech industry is experiencing a fundamental shift in how it acquires and deploys computing resources. What used to be a straightforward path through Nvidia and cloud hyperscalers is fracturing into competing ecosystems, each trying to secure supply chains and build moats. At the same time, the ability to build globally is becoming constrained by regulation, geography, and the finite nature of critical resources. The winners won’t be those with the most chips, but those who can orchestrate the broadest ecosystem of supply, software, and compliance.

Deep Dive

Databricks Reaches $134B Without the Nvidia Dependency Trap

The data infrastructure layer is becoming the real battleground in AI adoption, and Databricks’ \(5 billion Series J raise at a \)134 billion valuation signals a critical shift in how enterprises think about AI economics. Databricks is not building chips or training models. It is building the plumbing that makes AI workloads actually deployable at scale, and its financials reveal something crucial: companies are willing to pay for software that reduces their dependence on expensive, scarce GPU inventory.

The company projects $4.1 billion in revenue for 2025, representing 55% growth year-over-year while operating at breakeven. That combination is almost unheard of in enterprise software and signals that the market has matured past the “build everything at any cost” phase. Databricks is profitable not because it is cutting costs, but because the underlying demand for data orchestration and cost-efficient AI deployment has become so acute that customers simply pay what it costs. The breakeven operating margin at scale suggests the unit economics are real, not subsidized.

This matters because it reframes the entire AI infrastructure narrative. While everyone focuses on the race between Nvidia and custom silicon like Google’s TPUs, the actual leverage point is the software layer that makes diverse hardware interoperable. Databricks runs on multiple cloud providers, works with Nvidia GPUs but also custom chips, and increasingly acts as the abstraction layer that lets enterprises avoid vendor lock-in. In a world where chip supply is constrained and geopolitics is fragmenting silicon availability, that abstraction is worth far more than another GPU.

Google’s TPUv7 Ironwood Signals the Real Chip War Isn’t About Performance

Google’s TPUv7 Ironwood represents something more important than another data point in the AI chip arms race. It represents Google finally building hardware that is not just competitive with Nvidia on narrow benchmarks, but actually deployable as a strategic alternative for enterprises willing to migrate workloads. The real signal is not that TPUs are catching up. It is that Google has learned to build chips in concert with software and now integrates them so tightly that the cost-per-inference is what matters, not raw TFLOPS.

Nvidia’s dominance was never really about absolute performance. It was about the ecosystem. CUDA became sticky because it locked in developers, frameworks, and entire research communities. Breaking that lock requires not just better silicon, but better software integration, better pricing, and better incentives to migrate. Google, by controlling both hardware and the infrastructure stack (Vertex AI, BigQuery, Dataflow), can offer something Nvidia cannot: a vertically integrated system where the hardware is optimized for specific software patterns, and the pricing reflects that optimization.

What makes this genuinely threatening to Nvidia is not competition on performance. It is the emergence of viable alternatives that make the AI infrastructure ecosystem less homogeneous. If enterprises have 70% of their workloads running on Nvidia and can shift 20% to TPU at lower cost, Nvidia’s pricing power erodes. The market becomes segmented, and in segmented markets, the leader loses disproportionate share. Google is not trying to out-Nvidia Nvidia. It is trying to make Nvidia optional, and Ironwood is the proof of concept.

US Startups Adopting Chinese Models Signals a Crack in Export Control Strategy

The emergence of US startups adopting Chinese open-source AI models represents a fundamental failure of export control policy to keep pace with technology distribution. What was supposed to be a moat around frontier AI is becoming a sieve. The reason is simple: models like Deepseek, which Chinese developers trained on vastly cheaper compute, are now good enough for broad applications and available on GitHub. No tariff, no export license, and no regulatory framework moves fast enough to stop this flow.

This is not about Chinese models being superior. Many still lag frontier US models in reasoning and instruction-following. But “good enough” is a powerful force in economics, and good enough at 80% of the cost is irresistible to bootstrapping startups. The geopolitical logic of export controls assumes scarcity creates leverage. But AI models are not scarce. They are infinitely copyable. Once they exist in open-source repositories, the only way to prevent adoption is not through controls on supply, but through controls on demand, which no democratic government can impose at scale.

The second-order effect is more interesting: this trend accelerates the shift toward open-source AI as the baseline and proprietary models as premium offerings. Startups that can build differentiated products on commodity models gain leverage against those trying to build differentiated models themselves. The cost structure of AI development shifts from “train bigger models faster” to “find better use cases for existing models.” That fundamentally changes who wins in the ecosystem. Smaller, more focused teams with deep domain expertise can now compete with scale players who have always relied on brute-force compute advantage.

Signal Shots

New York Algorithmic Pricing Transparency Law Sets Template for State Regulation — New York became the first US state to require retailers to disclose algorithmic pricing tied to personal data, with 10+ states considering similar legislation. This is the first major attempt to regulate dynamic pricing by mandating transparency rather than restricting the practice. Watch whether the law actually changes behavior or becomes a disclosure checkbox. If retailers simply comply with notification without changing pricing algorithms, the law’s real impact is far weaker than intended.

Coupang Data Leak Exposes 65% of South Korea’s Population — South Korean authorities are investigating a breach at e-commerce giant Coupang that exposed 33.7 million accounts in a country of 51.7 million people. The scale is functionally catastrophic for privacy. What matters next is whether Seoul imposes penalties large enough to change behavior or treats this as another incident to move past. Korean regulators have historically been aggressive, but the sheer scale of the breach may signal that even strong regulatory regimes struggle against determined attackers at massive platforms.

Chinese IP Arbitrage Shows the Fundamental Scarcity Driving Infrastructure Conflict — A Chinese entrepreneur has amassed 10+ million IPv4 addresses mostly from Africa, then leases them back to companies outside Africa, leaving African ISPs struggling to build capacity. This is infrastructure arbitrage in its purest form: scarcity in one market (IPv4 addresses in Africa) creates opportunity for those who can acquire and redeploy assets. It reveals that the real constraint in global connectivity is not technology, but the ability to control finite resources. IPv4 exhaustion is not hypothetical anymore; it is reshaping economics.

Anthropic’s Multi-Product Strategy Shows Enterprise AI Is About Workflow Integration — Anthropic reports that 60% of enterprise customers use multiple Claude products, accelerating since Claude Code launched. This signals that enterprise AI adoption is shifting from “try a chatbot” to “integrate AI across workflows,” which means customers evaluate platforms, not just models. The vendor who owns the most workflow integration points wins disproportionate stickiness and revenue per customer. This is how Anthropic might eventually compete with OpenAI despite OpenAI’s earlier dominance in the consumer and developer spaces.

Swiss Government Bans SaaS and Hyperscale Cloud for Sensitive Data — Switzerland’s data protection authority called on public bodies to avoid Microsoft 365, cloud services, and SaaS generally due to lack of end-to-end encryption. This is the first major regulatory regime to treat SaaS as fundamentally incompatible with government data protection standards. If other EU nations follow, it fragments the European market, potentially spawning European alternatives and reducing Nasdaq tech companies’ total addressable market in regulated sectors. The implication: US SaaS companies may have already peaked in European government penetration.

Airbus A320 Software Rollback Reveals Hidden Risk in Automation Supply Chains — Airlines worldwide rushed to roll back Airbus software that could cause autopilot to exceed structural limits. Corrupt data in an OTA update nearly broke a critical safety system. This is a high-stakes glimpse into a future where more critical infrastructure runs on software with shorter development cycles. As automation penetrates manufacturing, aviation, and infrastructure, one corrupt update or miscalibrated algorithm compounds across millions of units instantly.

Scanning the Wire

• David Sacks leverage play accelerating — Trump administration AI and crypto czar David Sacks is formulating policies that directly aid his own tech investments and those of Silicon Valley allies. (New York Times)

• OpenAI losing engagement to Google — Similarweb data shows Gemini users spend longer per visit than ChatGPT and Claude users, signaling potential shift in engagement patterns despite OpenAI’s scale. (Financial Times)

• Pat Gelsinger details Intel’s “decay” — Ex-Intel CEO discusses structural breakdown at Intel, calls Chips Act execution “hideous,” notes irony that successor Lip-Bu Tan is following his original strategy. (Financial Times)

• China’s central bank reaffirms crypto crackdown — Beijing called virtual currency activity illegal and flagged stablecoins for failing KYC and AML compliance rules, showing no softening on regulatory stance. (Reuters)

• Varda Space Manufacturing Moving Toward Profitability — Founder Will Bruey says the company has proven space manufacturing works and is now shifting focus from proving the concept to making it “boring” and operational. (TechCrunch)

• Meesho IPO Becomes India’s First Major E-Commerce Listing — SoftBank-backed Meesho raises $606 million in IPO, opening doors for Indian e-commerce exits after years of private market dominance by Flipkart and Amazon. (TechCrunch)

• Automakers Pursuing Rare-Earth-Free Motors — Car companies are actively developing electric motors that don’t rely on Chinese rare-earth magnets to reduce geopolitical exposure and supply chain risk. (New York Times)

• SmartTube APK Compromised With Malware — Official SmartTube APK was compromised with malware, warning users to upgrade and highlighting risks in sideloaded Android ecosystem. (AFTV News)

Outlier

AI Proves Erdos Problem #124 — Researchers used AI to solve an open combinatorial problem from the Erdos Problems list, marking the first time automated reasoning cracked a problem from the famous unsolved set. This signals that AI has moved beyond pattern matching into genuine mathematical creativity. The deeper pattern: AI systems are now capable of genuine problem-solving in domains where humans thought only human intuition mattered. That changes everything about where AI labor becomes economically viable next.

See you in the next edition. The infrastructure race is just getting started, and every player is scrambling to avoid being locked into someone else’s ecosystem.