The Investor

Record Capital, Record Concentration

Q1 2026 produced the largest venture quarter in history: $300 billion deployed into roughly 6,000 startups globally. But the headline is a distortion. Four AI mega-deals — OpenAI, Anthropic, xAI, and Waymo — absorbed $188 billion, or 65% of all capital. AI companies captured 80% of total VC, up from 55% a year ago. This isn't a rising tide lifting all boats. It's a tsunami concentrated in one harbor.

Strip out those four deals and the remaining ~$112 billion would still be a record in most prior years — the broad market is genuinely healthy. But the structural dynamics have shifted in ways that demand portfolio repositioning.

The Seed Market Just Bifurcated

Seed funding dollars rose 31% YoY, but seed deal count cratered 30%. Translation: average seed checks roughly doubled. Fewer companies are getting funded, but the ones that do are getting larger bets. The top deployers — Accel (16 deals), a16z (15), Lightspeed (14) — maintained aggressive pace, but the funnel narrowed dramatically for everyone below the top decile.

Pre-seed is now the entry point that seed was two years ago. Disciplined investors writing $250K-$1M checks can access the pricing dynamics that seed investors enjoyed in 2024.

Bits to Atoms: The $1B+ Cohort Reveals a Capital Rotation

Beyond the Big 4, the billion-dollar round club signals a critical shift. Cerebras and Rapidus (chips), Skild AI (robotics), Wayve (self-driving), and Shield AI (defense) all crossed the $1B threshold. This is the market pricing in what multiple sources confirm: the AI cycle has a physical layer. Unlike cloud and mobile — built almost entirely in software — AI requires factories, fabs, and fleets. Return profiles differ (longer duration, higher capex, deeper moats), but the competitive dynamics favor companies integrating atoms and bits.

McKinsey projects AI inference will surpass training as the dominant workload by 2030 at 35% CAGR, and the global semiconductor market is projected to double from $775B to $1.6T by 2030. The implication: the value chain is rotating from "build bigger models" to "run models everywhere, cheaply, at scale."

The Tension You Must Navigate

Here's the paradox multiple sources surface simultaneously: record deployment coexists with acknowledged bubble risk. Conference attendees acknowledge AI bubble dynamics while deploying capital at unprecedented speed. IPO uncertainty for SpaceX, OpenAI, and Anthropic compounds the tension. This is the classic late-cycle paradox — and it demands a barbell strategy: conviction bets on infrastructure and inference at one end, disciplined pre-seed entry points at the other.

Non-AI startups are in a capital desert — and that's the contrarian buy. When 80% of all VC flows to AI, companies with real revenue and defensible positions in cybersecurity, fintech infrastructure, and climate tech trade at significant discounts to intrinsic value.

The Capex Numbers Are Staggering — But the Physics Don't Cooperate

The AI infrastructure arms race reached a new phase this week with hard numbers attached. AWS committed $200B in 2026 capex — the largest single-year infrastructure spend ever disclosed by a cloud provider. Meta locked in $135B total, including $62B in third-party cloud commitments ($35B to CoreWeave through 2032, $27B to Nebius). OpenAI targets 30 GW of compute by 2030 — roughly the electricity consumption of the Netherlands.

But here's the constraint nobody's modeling: nearly half of U.S. data centers planned for 2026 face delay or cancellation due to power grid limits, equipment shortages, and local opposition. This isn't a temporary supply chain hiccup. Power constraints and local opposition are structural barriers that take years to resolve.

CoreWeave: The $87.8B Backlog That Should Keep You Up at Night

CoreWeave's numbers are simultaneously impressive and alarming. Revenue hit $5.13B in 2025 (2.7x YoY), but the company posted a $1.17B net loss. The headline $87.8B backlog implies ~17x backlog-to-revenue — but the composition is what matters:

If either anchor tenant renegotiates 20-30%, CoreWeave's unit economics — already negative — deteriorate while debt service obligations remain fixed. The $1.75B private debt raise rather than equity suggests even optimistic investors want downside protection.

Amazon's Custom Silicon Is the Underpriced Nvidia Threat

Andy Jassy disclosed that Amazon's custom chip business doubled from $10B+ to $20B+ annualized run rate in approximately two months. Graviton CPU adoption hit 98% among top 1,000 EC2 customers. Trainium 3 is nearly sold out before GA. Two unnamed customers tried to buy Amazon's entire Graviton supply for 2026.

More consequentially, Amazon is considering selling Trainium racks to third parties, which would open a direct competitive front against Nvidia in the merchant AI chip market. Every frontier AI lab — Anthropic, OpenAI, Meta, Google — is now pursuing custom silicon. This is bearish for Nvidia's margin sustainability over 3-5 years, though near-term demand remains overwhelming.

The investment thesis is splitting: infrastructure captures predictable, contract-backed revenue for the next decade, while the application layer can't figure out if its best customers are profitable. Allocate accordingly.

Software's First-Ever Discount to the Market

The iShares Software ETF (IGV) is down 21% year-to-date and approximately 30% from its September 2025 peak, erasing roughly $2 trillion in market capitalization. For the first time in the modern era, software stocks now trade at a discount to the S&P 500. The sector that commanded premium multiples for over a decade is now valued below the market average.

The cause isn't cyclical — it's structural. AI agents are breaking the link between headcount growth and SaaS revenue growth. If agents replace seats rather than complement them, the per-seat pricing model that powered predictable recurring revenue, 120%+ NRR, and 10-20x forward multiples faces existential challenge.

The Coding Agent Margin Crisis Validates the SaaS Thesis Break

Every major AI coding tool provider is scrambling to fix margins simultaneously — providing the clearest evidence yet that AI application-layer economics are fundamentally different from traditional SaaS:

The shift from request-based to token-based billing is telling: usage patterns are wildly variable, and the most engaged users are unprofitable under flat-fee models. This is a classic SaaS antipattern — the best customers destroy your margins.

The $100/Month Ceiling and What It Means

OpenAI explicitly price-matched Anthropic at $100/month, establishing a market clearing price for premium AI. The $200 Pro tier was delisted from the pricing page. For portfolio companies building on top of these models: your cost of intelligence is now benchmarked at $100/month per power user. Any pricing model that assumes customers pay significantly more for an AI wrapper is fighting gravity.

Anthropic's counter is architecturally superior: the advisor pattern routes cheap Haiku/Sonnet executors for routine work and escalates to Opus only for hard decisions, cutting effective inference cost by 12%+ while improving output quality. This is the SaaS margin playbook applied to AI inference — and it compounds over time.

Software's first-ever discount to the S&P 500 isn't a buying opportunity — it's the market repricing per-seat SaaS as a structurally impaired business model, and the private market marks haven't caught up yet.

The 80% Bypass Rate Nobody's Pricing In

The demand-side risk is equally concerning: 80% of white-collar workers bypass company AI tools, while LLM-referred traffic converts at 30-40%. Enterprise AI adoption is real in contract value but hollow in utilization. Any AI SaaS company reporting strong bookings but unable to demonstrate DAU/MAU ratios and feature adoption depth faces severe renewal risk.