Capital, Chips, and Compliance

The AI infrastructure race is revealing a fundamental tension: capital wants to move at software speed, but it's increasingly constrained by compliance timelines and geographic realities.

Cognition's billion-dollar raise at $492 million in annualized revenue demonstrates how quickly AI software companies can scale when they solve real developer productivity problems. But the infrastructure required to support this software layer is moving in the opposite direction. Nvidia's $150 billion commitment to Taiwan rather than US facilities signals that despite political pressure for reshoring, semiconductor economics still favor established clusters. Meanwhile, Snowflake's $6 billion AWS chip deal shows hyperscalers building alternatives to Nvidia's dominance, fragmenting what was becoming a dangerously concentrated supply chain.

The regulatory environment is tightening in parallel. The EU's €200 million fine against Temu and insider trading charges against a Google employee using proprietary search data mark a shift from warning letters to actual enforcement. Regulators are no longer just studying these markets.

The pattern: software valuations accelerate while hardware concentrates in geopolitically sensitive regions, even as compliance costs rise sharply. Second-order effect to watch is whether this divergence creates new choke points in the AI stack beyond just chips.

Deep Dive

Nvidia's Taiwan Commitment Exposes the Limits of Industrial Policy

The semiconductor industry cannot be reshored through political pressure alone, no matter how much capital Washington deploys. Nvidia's decision to invest $150 billion annually in Taiwan rather than accelerating US manufacturing reveals a hard truth: advanced packaging capabilities, supplier ecosystems, and engineering talent concentrations take decades to build, not presidential terms to relocate.

Huang's Taiwan bet directly contradicts Trump's AI Action Plan to make the US the global AI hub, yet Trump has little leverage to change the calculus. The president can impose tariffs and offer subsidies, but he cannot conjure TSMC's packaging expertise or the dense network of specialized suppliers within hours of Nvidia's facilities. The company started US chip production last year, projecting $500 billion in domestic AI infrastructure over four years, but crucially still ships those chips to Taiwan for advanced packaging. That dependency is not changing anytime soon.

The strategic risk for US policymakers is that aggressive reshoring demands may backfire by pushing companies to double down on existing clusters. Taiwan's semiconductor ecosystem delivers better economics and faster time to market than greenfield US facilities can match in the near term. For Nvidia, maintaining dominance in the Vera Rubin era requires accessing that ecosystem at scale, and $150 billion buys a lot of certainty.

The implication for founders: assume semiconductor supply chains will remain concentrated in East Asia for the next 5-10 years regardless of policy rhetoric. Design your infrastructure strategy around that reality rather than betting on rapid geographic shifts. For policymakers, the lesson is that industrial policy requires patient capital measured in decades, not election cycles. The alternative is watching companies like Nvidia optimize for business fundamentals over political preferences.

AI Coding Economics Shift from Model Makers to Application Layer

Cognition's $25 billion valuation on under $500 million in revenue challenges the assumption that foundation model makers would capture all the value in AI coding. The company doubled its valuation in eight months while giants like Anthropic, OpenAI, and Google all ship competing products, suggesting the market has room for specialized players despite intense competition from well-funded model makers.

The key signal is not the valuation multiple, which remains aggressive even by software standards, but rather the customer concentration in enterprises like Mercedes-Benz, NASA, and Goldman Sachs. These organizations are paying for Devin because it integrates into their specific workflows and security requirements in ways that general-purpose coding assistants cannot easily match. The 50% month-over-month usage growth for six consecutive months indicates actual product-market fit, not just pilot programs.

This creates a template for AI application companies: build vertical-specific tooling that solves workflow problems rather than competing on model quality. Foundation models are becoming commoditized infrastructure, but the integration layer, security controls, and enterprise-specific features remain valuable. Cognition's acquisition of Windsurf assets and subsequent $1 billion raise suggest consolidation in this layer is starting, with winners pulling away quickly.

For venture investors, the lesson is that application-layer AI companies can achieve venture-scale outcomes if they move fast enough to build defensible positions before model makers fully verticalize. For founders, the window to establish these positions is narrowing as model makers improve their own tooling, but it has not closed. The risk is that today's $25 billion valuations become tomorrow's write-downs if model makers successfully bundle comparable functionality, making the race to revenue and customer lock-in existential.

Signal Shots

Fusion Startup Economics Still Require Patient Capital : Thea Energy raised $100 million to build stellarator fusion reactors using software-tunable magnets, bringing total funding to $130 million with a 2034 commercial target. The pixel-inspired magnet approach could reduce manufacturing complexity compared to traditional stellarators, but the company still relies on 12 conventional large magnets for core plasma confinement. This positions Thea among better-funded fusion startups but does not fundamentally change the economics requiring billions and decade-plus timelines before commercial viability.

Political Patronage Shapes Defense Procurement : Dell secured a $9.7 billion Pentagon software contract months after founder Michael Dell pledged $6.25 billion to Trump-branded children's investment accounts. The five-year Microsoft licensing deal consolidates Pentagon software purchasing and claims $422 million in annual savings, but the timing raises questions about procurement independence when executives cultivate White House relationships. Watch whether other major government contracts follow similar patterns of political engagement preceding awards, potentially reshaping how tech companies approach federal sales.

European AI Independence Faces Scaling Reality : Mistral is accelerating superintelligence development and signing deals with Airbus and BMW to reduce European dependence on US AI providers. The company frames this as technological sovereignty, but CEO comments acknowledge Europe's fundamental challenge is matching the capital scale that US hyperscalers deploy. Watch whether European enterprise customers prove willing to accept potential capability gaps for supply chain independence, or if Mistral's growth remains constrained by the same funding disadvantages facing other regional AI challengers.

SpaceX Grounding Complicates IPO Timeline : The FAA ordered SpaceX to investigate why its upgraded Starship V3 booster failed during the May 22 test flight, grounding the program until the probe concludes. The timing pressures SpaceX's anticipated mid-June IPO, as Starship reliability is critical to the company's Starlink expansion economics. The V3's all-new Raptor engines and design changes were meant to improve reliability, but the booster experienced immediate engine failures after stage separation. Watch whether SpaceX can complete the investigation and return to flight before the IPO window closes.

Memory Shortage Pushes Consumer Hardware Costs Up 40 Percent : Valve raised Steam Deck prices by over $200, with the 1TB OLED model now costing $949 versus $649 previously, citing rising memory and storage costs. The increases match similar moves by Lenovo, Sony, and Nintendo as the global memory shortage filters through to consumer hardware. This marks a shift from intermittent stock shortages to sustained price increases that could reshape gaming hardware economics if memory costs remain elevated through 2027.

ByteDance Pursues Dual-Track CPU Strategy : ByteDance is developing custom CPUs on both Arm and RISC-V architectures to power its AI infrastructure as Intel and AMD raise data center processor prices 10-35 percent quarterly. The dual-track approach hedges between proven Arm designs and RISC-V's export control advantages, which Beijing explicitly endorses for strategic autonomy. This follows ByteDance's Qualcomm ASIC deal earlier this week, signaling a comprehensive chip diversification program. Watch whether SMIC's 7nm production capabilities can support leading-edge CPU designs or if foundry constraints limit ByteDance's independence ambitions.

Scanning the Wire

YouTube launches AI-powered custom video feeds : The platform now lets users describe what they want to watch and generates personalized feeds based on interests, moods, or topics, which can be pinned to the homepage for quick access. (The Verge)

Meta rolls out Facebook Plus subscriptions globally : After testing premium tiers earlier this year across Facebook, Instagram, and WhatsApp, Meta is now deploying paid subscriptions worldwide and beginning tests of Meta AI subscriptions as the company diversifies revenue beyond advertising. (The Verge)

SOND exits stealth with $7 million for sleep tech earbuds : Founded by Bose's former head of sleep, the startup launched Dreambuds, which capture 12 physiological signals and respond in real time to improve sleep quality through closed-loop audio intervention. (TechCrunch)

CrowdStrike and Google dismantle Glassworm botnet : The joint operation took down infrastructure cybercriminals used to inject malware into open source projects, targeting developers and companies that rely on compromised software packages. (TechCrunch)

London reclaims position as Europe's top tech hub : The city overtook Paris with startups raising $17.7 billion in 2025, ranking fourth globally behind Silicon Valley, New York, and Beijing, according to Dealroom data. (Reuters)

Websites can now fingerprint visitors through SSD activity : Researchers demonstrated that simple JavaScript can measure telltale solid-state drive patterns through browser timing attacks, creating a new vector for tracking users without cookies or traditional identifiers. (Ars Technica)

Amazon begins selling its AI shopping tools to retailers : The company signed Kate Spade as its first customer for the recommendation and personalization technology that powers Amazon's own e-commerce experience, creating a new enterprise software revenue stream. (CNBC)

ElevenLabs ships music model with mid-track genre switching : The new generative model allows users to regenerate specific sections of a song without affecting the rest of the composition, enabling more granular creative control than previous generation-based approaches. (TechCrunch)

Outlier

Beijing Hoards Its AI Talent : China is increasingly blocking its top AI researchers from leaving the country, treating technical expertise like a strategic resource rather than mobile labor. This mirrors how medieval guilds controlled craftsmen or how the Soviet Union restricted scientist emigration, but applied to machine learning PhDs. The signal is not about innovation capacity, which benefits from open exchange, but about Beijing's belief that AI advantage comes from concentration rather than circulation of talent. If other nations follow this logic, expect the global research community to fragment into regional clusters with limited movement between them, reversing decades of scientific mobility and potentially slowing overall progress as ideas stop crossing borders freely.

The medieval guild comparison feels apt: we're watching governments rediscover that you can fence in people more easily than ideas. The ideas just take longer routes and arrive in unexpected forms.