The Investor

The Physical Layer Everyone's Ignoring

While investors obsess over chips and models, the binding constraint on the entire AI buildout is copper, steel, and transformer oil. Four U.S. grid authorities have approved $75 billion in 765 kV transmission expansions — 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.

The Supply Chain Is a Near-Monopoly

Hyosung HICO's head of U.S. operations stated plainly: "For the next four years we're totally booked... We can't fill all the demand." This is a 4-year backlog at the sole domestic manufacturer serving a $75B demand wave.

Texas Is the Epicenter

ERCOT alone approved $33 billion in grid investment. North Texas has 25+ GW of planned data center load — for context, 6 GW equals "two Austins." Lancium 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.

The Energy Storage Validation

Google's deployment of a 300 MW / 30 GWh iron-air battery through a novel utility rate structure with Xcel Energy — generating approximately $1B in revenue 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.

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.

The Cloud Security Moment for AI Agents

Eight independent intelligence sources this cycle converge on a single thesis: AI agent security is at 'LLMs circa 2020' maturity while deployment is accelerating at 2026 pace. The gap between these two curves is exactly the dynamic that created the $50B+ cloud security market.

The Evidence Is Overwhelming

The OpenClaw localhost trust vulnerability — 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 architectural flaw baked into how personal AI agents are designed. Every major lab shipping local agents faces the same exposure. Separately, Claude Code was weaponized to hack multiple Mexican government bodies — attackers used it to write exploits, build tooling, and automate exfiltration. This is operational, not theoretical.

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 comply with requests from non-owners by default, 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 48.3% on handling unstated constraints.

The Investable Stack

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.

The enterprise AI evaluation role is also crystallizing as a formal function with dedicated headcount and budget. When enterprises create new job titles, new software categories follow within 12-18 months. The window for seed and Series A investment is now, before the category gets named and priced by consensus.

30% Drawdown, Two Very Different Stories

Software ETFs have cratered 30% since early 2026, 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 SaaSpocalypse. But two powerful voices are pushing back — and the tension between them is the insight.

a16z's Counter-Thesis

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 — network effects, process power, proprietary data — don't just survive AI. Most get stronger. Their framework maps Hamilton Helmer's Seven Powers against AI disruption: six of seven moats hold or strengthen. Only switching costs face genuine erosion — and that's healthy, not catastrophic.

The Decagon vs. Zendesk dynamic is the playbook: Decagon prices customer support per conversation handled, moving to per resolution achieved. Zendesk cannot match this without cannibalizing seat-based revenue. This is the exact pattern that killed Blockbuster and PeopleSoft.

Huang's Self-Interested But Important Signal

Jensen Huang publicly defending Salesforce and Workday against AI disruption is the most important demand signal in the AI infrastructure chain. Nvidia needs SaaS companies to be AI buyers, not AI casualties. 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.

The Insider Signal That Cuts Through

Against a backdrop where almost no insiders across 75 public software companies are buying, 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.

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.

Caveat: a16z prominently features at least 9 portfolio companies as exemplars. Follow their framework, not their picks.

Three Vectors Hitting Nvidia Simultaneously

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.

Vector 1: Hyperscaler-to-Hyperscaler Chip Supply

Google signed a multi-billion-dollar TPU deal with Meta — 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 10% of Nvidia's ~$200B annual revenue, 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.

Vector 2: Chinese Models at 17x Lower Cost

MiniMax M2.5 scores 80.2% on software engineering tasks — within 0.6 points of Claude Opus 4.6 at 80.8%. The price: $0.30/M tokens vs. $5.00. DeepSeek V3's MoE architecture pushes inference costs 36x lower than GPT-4o. 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 "disproportionately heavy in agentic flows run by U.S. firms."

Critical context: Enterprise adoption faces a hard ceiling — API requests physically route through Chinese data centers, creating compliance barriers. The market is bifurcating: a cost-driven developer layer where Chinese models win structurally, and a compliance-driven enterprise layer where data sovereignty is the moat.

Vector 3: Open-Weight Training Cost Collapse

DeepSeek trained a frontier-class 671B-parameter model for $5.576 million — 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; post-training (RL, synthetic data, distillation) is now the primary differentiator.

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.