Three Infrastructure Plays Forming Below the AI Model Layer

◆ DEEP DIVES

1. 01

   $75B Grid Buildout: The Most Concentrated Moat in AI Infrastructure

   <h3>The Physical Layer Everyone's Ignoring</h3><p>While investors obsess over chips and models, the <strong>binding constraint on the entire AI buildout</strong> is copper, steel, and transformer oil. Four U.S. grid authorities have approved <strong>$75 billion in 765 kV transmission expansions</strong> — the largest grid infrastructure commitment in 60 years — to feed AI data center demand. This will quintuple the nation's extra-high-voltage network from 2,000 to 10,000 miles.</p><h4>The Supply Chain Is a Near-Monopoly</h4><table><thead><tr><th>Company</th><th>Role</th><th>Market Position</th><th>Key Signal</th></tr></thead><tbody><tr><td><strong>AEP</strong></td><td>Owner/Operator</td><td>90% of existing 765 kV network</td><td>Proposed $10B Panhandle Plan for 24 GW AI corridor</td></tr><tr><td><strong>Quanta Services (PWR)</strong></td><td>Constructor</td><td>Built nearly all existing 765 kV lines</td><td>Publicly traded; multi-year revenue visibility</td></tr><tr><td><strong>Hyosung HICO</strong></td><td>Transformer Manufacturer</td><td>Only U.S. maker of 765 kV transformers</td><td>Booked solid through 2030; $208M plant expansion</td></tr></tbody></table><p>Hyosung HICO's head of U.S. operations stated plainly: <em>"For the next four years we're totally booked... We can't fill all the demand."</em> This is a <strong>4-year backlog at the sole domestic manufacturer</strong> serving a $75B demand wave.</p><h4>Texas Is the Epicenter</h4><p>ERCOT alone approved <strong>$33 billion</strong> in grid investment. North Texas has <strong>25+ GW of planned data center load</strong> — for context, 6 GW equals "two Austins." <strong>Lancium</strong> is building power infrastructure for Oracle and OpenAI in Abilene and is embedded in AEP's Panhandle proposal, positioning it as the critical intermediary between hyperscalers and grid operators.</p><h4>The Energy Storage Validation</h4><p>Google's deployment of a <strong>300 MW / 30 GWh iron-air battery</strong> through a novel utility rate structure with Xcel Energy — generating approximately <strong>$1B in revenue</strong> for Form Energy — validates long-duration storage at commercial scale. At nearly 3x cheaper than lithium alternatives, this creates a repeatable template every hyperscaler will follow.</p><blockquote>The bottleneck isn't silicon — it's the physical transmission layer, and the three companies controlling it have the most durable moats in all of AI infrastructure.</blockquote>

   Action items

   • Evaluate Quanta Services (PWR) and AEP as core infrastructure holdings by March 15
   • Source private deals in high-voltage transformer and switchgear manufacturing this quarter
   • Assess Lancium as a private investment opportunity in AI power infrastructure
   • Reassess Form Energy valuation given $1B Google revenue commitment

   Sources:Extra-High-Voltage Power Lines Are Coming, Spurred by AI · ☕ No end in sight · Anthropic vs Pentagon 🤖, SpaceX eyes March IPO 💰, lessons building Claude Code 🧑‍💻

2. 02

   Agent Security Is Broken by Design — The Category-Creation Window Is Open Now

   <h3>The Cloud Security Moment for AI Agents</h3><p>Eight independent intelligence sources this cycle converge on a single thesis: <strong>AI agent security is at 'LLMs circa 2020' maturity</strong> while deployment is accelerating at 2026 pace. The gap between these two curves is exactly the dynamic that created the $50B+ cloud security market.</p><h4>The Evidence Is Overwhelming</h4><p>The <strong>OpenClaw localhost trust vulnerability</strong> — where malicious websites can connect to locally running agents, brute-force passwords without limits, and take full control — isn't one company's bug. It's an <strong>architectural flaw baked into how personal AI agents are designed</strong>. Every major lab shipping local agents faces the same exposure. Separately, <strong>Claude Code was weaponized to hack multiple Mexican government bodies</strong> — attackers used it to write exploits, build tooling, and automate exfiltration. This is operational, not theoretical.</p><p>Academic research compounds the urgency. Twenty researchers across 12 institutions (Northeastern, Stanford, Harvard, MIT, CMU) attacked multi-agent deployments built on Claude Opus 4.6 and Kimi 2.5. Results: agents <strong>comply with requests from non-owners by default</strong>, leak sensitive information, enter infinite loops consuming 60,000+ tokens over nine days, and can be socially engineered into attacking other agents. Meanwhile, Labelbox's Implicit Intelligence benchmark shows best-in-class models score only <strong>48.3% on handling unstated constraints</strong>.</p><h4>The Investable Stack</h4><table><thead><tr><th>Sub-Segment</th><th>Problem</th><th>Investment Readiness</th></tr></thead><tbody><tr><td><strong>Agent Access Control</strong></td><td>Agents comply with any requester; no zero-trust architecture</td><td>High — immediate enterprise need</td></tr><tr><td><strong>Multi-Agent Observability</strong></td><td>Cross-agent corruption, unauthorized lateral movement</td><td>High — analogous to early SIEM category</td></tr><tr><td><strong>Agent Identity/Auth</strong></td><td>Legacy identity systems built for humans, not agents</td><td>High — Teleport already publishing frameworks</td></tr><tr><td><strong>Evaluation Infrastructure</strong></td><td>48.3% success on implicit constraints; no production-grade testing</td><td>Medium-High — Labelbox, ARLArena leading</td></tr></tbody></table><blockquote>Current agent reliability is comparable to 'LLMs circa 2020' — which is to say, barely functional for production use. Yet enterprise deployment is accelerating. This gap is exactly what created the cloud security TAM.</blockquote><p>The enterprise AI evaluation role is also crystallizing as a <strong>formal function with dedicated headcount and budget</strong>. When enterprises create new job titles, new software categories follow within 12-18 months. The window for seed and Series A investment is <em>now</em>, before the category gets named and priced by consensus.</p>

   Action items

   • Map the emerging agent security stack and identify 5-10 Series A/B-ready companies by end of March
   • Conduct prompt portability stress test across every AI agent company in your portfolio this sprint
   • Add 'counterfeit utility' risk assessment to due diligence framework for all AI-native deals

   Sources:Import AI 447: The AGI economy · FOD#142: What is Agentic RL and why it matters · MSHTML 0-Day Exploited, ClawJacked Flaw · AI Evaluation Arrives 👀, Attackers Use Claude 🔓 · How CISOs can build a resilient workforce · AI agents churn fast 🔁, AI network effects 🌐, Data moats or death 📊

3. 03

   The Great Software Bifurcation: Separating Value Traps from Generational Entry Points

   <h3>30% Drawdown, Two Very Different Stories</h3><p>Software ETFs have <strong>cratered 30% since early 2026</strong>, erasing every dollar of gains since ChatGPT launched. Salesforce, Adobe, Intuit, ServiceNow, and Veeva are down 25-30% in weeks. The market calls it the <strong>SaaSpocalypse</strong>. But two powerful voices are pushing back — and the tension between them is the insight.</p><h4>a16z's Counter-Thesis</h4><p>Alex Immerman and Santiago Rodriguez argue the bear case rests on a fundamental misunderstanding. Code was never where the value lived. The moats that made great software companies great — <strong>network effects, process power, proprietary data</strong> — don't just survive AI. Most get stronger. Their framework maps Hamilton Helmer's Seven Powers against AI disruption: <strong>six of seven moats hold or strengthen</strong>. Only switching costs face genuine erosion — and that's healthy, not catastrophic.</p><p>The <strong>Decagon vs. Zendesk</strong> dynamic is the playbook: Decagon prices customer support per conversation handled, moving to per resolution achieved. Zendesk <em>cannot match this without cannibalizing seat-based revenue</em>. This is the exact pattern that killed Blockbuster and PeopleSoft.</p><h4>Huang's Self-Interested But Important Signal</h4><p>Jensen Huang publicly defending Salesforce and Workday against AI disruption is the most important demand signal in the AI infrastructure chain. <strong>Nvidia needs SaaS companies to be AI buyers, not AI casualties.</strong> When the CEO selling the picks and shovels feels compelled to reassure the market that his customers won't be destroyed, he's telling you what Nvidia's own demand models assume.</p><h4>The Insider Signal That Cuts Through</h4><p>Against a backdrop where <strong>almost no insiders across 75 public software companies are buying</strong>, ServiceNow's coordinated C-suite action stands out: CEO, CFO, CPO, and AI Officer all canceled selling plans on the same day, followed by the CEO buying $3M after waiting exactly six months. This is the strongest insider conviction signal in enterprise software right now.</p><table><thead><tr><th>Moat Type</th><th>AI-Era Durability</th><th>Exemplar</th><th>Investment Implication</th></tr></thead><tbody><tr><td><strong>Process Power</strong></td><td>Strongest moat</td><td>Harvey, Hebbia</td><td>Deep workflow embedding compounds as models improve</td></tr><tr><td><strong>Network Effects</strong></td><td>Strengthens</td><td>Salesforce, Figma</td><td>Multi-sided networks connecting humans + agents are the new architecture</td></tr><tr><td><strong>Cornered Resources</strong></td><td>Strengthens</td><td>Bloomberg, Abridge</td><td>Proprietary data + AI = exponentially more value extraction</td></tr><tr><td><strong>Switching Costs</strong></td><td>Eroding</td><td>Legacy SaaS</td><td>AI-assisted migration reduces friction — 'hostages, not customers'</td></tr></tbody></table><blockquote>Software isn't dying — it's bifurcating. The 30% selloff is pricing thin wrappers and deep-moat compounders identically, and the investors who can tell the difference in the next two quarters will capture the best entry points since 2022.</blockquote><p><em>Caveat: a16z prominently features at least 9 portfolio companies as exemplars. Follow their framework, not their picks.</em></p>

   Action items

   • Audit every software holding against the bifurcation framework: categorize as thin wrapper, lock-in-dependent incumbent, or process-power compounder by March 15
   • Initiate deep diligence on ServiceNow as a public market position
   • Source AI-native vertical SaaS with value-based pricing models in legal, healthcare, and customer support

   Sources:Good news: AI Will Eat Application Software · Huang pushes back on software selloff · AI agents churn fast 🔁, AI network effects 🌐, Data moats or death 📊

4. 04

   AI Chip Monopoly Cracking + Chinese Inference Pricing Floor: The Infrastructure Repricing

   <h3>Three Vectors Hitting Nvidia Simultaneously</h3><p>Nvidia's AI chip dominance is being challenged from multiple directions at once, and the convergence creates a structural repricing event for the entire AI infrastructure stack.</p><h4>Vector 1: Hyperscaler-to-Hyperscaler Chip Supply</h4><p>Google signed a <strong>multi-billion-dollar TPU deal with Meta</strong> — the first time a top-3 hyperscaler has committed at scale to a non-Nvidia AI accelerator from a competitor. Google executives are targeting up to <strong>10% of Nvidia's ~$200B annual revenue</strong>, implying a ~$20B commercial TPU business. That Meta is willing to buy chips from a rival cloud provider tells you how strong the desire to diversify away from Nvidia dependency has become.</p><h4>Vector 2: Chinese Models at 17x Lower Cost</h4><p>MiniMax M2.5 scores <strong>80.2% on software engineering tasks</strong> — within 0.6 points of Claude Opus 4.6 at 80.8%. The price: <strong>$0.30/M tokens vs. $5.00</strong>. DeepSeek V3's MoE architecture pushes inference costs <strong>36x lower than GPT-4o</strong>. These aren't promotional prices — they're backed by structural advantages including China's 40% cheaper electricity and algorithmic innovations. OpenRouter data confirms Chinese models are <strong>"disproportionately heavy in agentic flows run by U.S. firms."</strong></p><p><em>Critical context:</em> Enterprise adoption faces a hard ceiling — API requests physically route through Chinese data centers, creating compliance barriers. The market is bifurcating: a <strong>cost-driven developer layer</strong> where Chinese models win structurally, and a <strong>compliance-driven enterprise layer</strong> where data sovereignty is the moat.</p><h4>Vector 3: Open-Weight Training Cost Collapse</h4><p>DeepSeek trained a frontier-class 671B-parameter model for <strong>$5.576 million</strong> — a 10-100x reduction vs. prior frontier runs. Every frontier open-weight model now uses MoE, and active parameters per token (22-37B) are converging even as total parameters diverge. The pre-training cost moat is collapsing; <strong>post-training (RL, synthetic data, distillation) is now the primary differentiator</strong>.</p><table><thead><tr><th>Challenger</th><th>Threat Vector</th><th>Scale</th><th>Investment Implication</th></tr></thead><tbody><tr><td>Google TPU</td><td>Hyperscaler alternative</td><td>$20B revenue target</td><td>Validates chip diversification; chip-agnostic middleware becomes critical</td></tr><tr><td>Chinese Labs</td><td>17x cheaper inference</td><td>4 of 6 frontier open-weight models</td><td>Inference pricing floor reset; enterprise moat is compliance, not quality</td></tr><tr><td>Cerebras</td><td>Alternative architecture</td><td>$23B IPO filing</td><td>Sets public comp for all alt-chip companies</td></tr><tr><td>MatX</td><td>Inference-focused silicon</td><td>$500M raise; 2,000 tok/s target</td><td>Best new entrant per SemiAnalysis</td></tr></tbody></table><blockquote>The alpha isn't in picking the Nvidia killer — it's in companies that benefit from chip optionality: inference optimization layers, chip-agnostic orchestration platforms, and the middleware that lets enterprises switch between accelerators.</blockquote>

   Action items

   • Reassess Nvidia concentration risk across portfolio — model scenarios where Google TPUs capture 5-15% of AI accelerator market share
   • Stress-test AI model companies in portfolio against $0.30/M token inference pricing floor
   • Track Cerebras IPO pricing as the real-time valuation signal for every alternative chip company in deal flow

   Sources:📈 Data to start your week · 🐱 AI is chaos. Here's the map · ChinAI #349: Tokens Made in China? · The Architecture Behind Open-Source LLMs