The Investor

The Numbers That Force the Conversation

CoreWeave printed Wednesday and the tape answered in the only way it knows how, down 6.6% intraday and another 8.7% after the bell. Revenue doubled to two billion dollars, capex ran to seven point seven billion (five and a half times last year), and the quarter burned four point seven billion in cash. The balance sheet carries twenty-four point eight billion of debt against three billion of cash, roughly two-thirds of which arrived via Nvidia's equity check. Jensen Huang, unhelpfully for the bulls, said the quiet part on a microphone: "If we didn't help CoreWeave exist, they would not exist."

That sentence is the trade. Or rather, the more interesting version of the trade, because vendor financing dressed as strategic investment is precisely what Cisco was doing with the CLECs in 1999, and this is probably wrong but the ratio is hard to argue with — CoreWeave is spending three times revenue on capex while Alphabet runs twenty-five to thirty-three percent and Amazon fifteen to twenty. Those are not degrees of the same business.

The Broadcom Signal Nobody Is Discussing

The less-discussed data point in the same cycle: OpenAI's eighteen-billion-dollar custom Broadcom chip deal is stuck because Microsoft will not commit to forty percent of production offtake, and OpenAI is reportedly shopping the remainder. If the apex consumer of AI infrastructure cannot frictionlessly close custom-silicon financing, the assumption that capital is infinite at the top of the stack is already failing, quietly, in a place the headlines have not reached.

Meanwhile xAI is renting compute TO Anthropic, which is competitors as each other's customers, which is compute behaving more fungibly than the dedicated-neocloud thesis required it to.

When the category leader only survives because its supplier is also its investor, every tail-end neocloud is un-bankable without the same arrangement — and Nvidia cannot write that check for all of them.

What This Means for Your Book

The comp set just got repriced against solvency math rather than growth multiples, and private names with similar capital structures and less patient lenders are now quoted off a different curve whether their cap tables have noticed or not. Backlog moved from sixty-seven to one hundred billion in a quarter; the concentration in Meta and Anthropic multi-year deals means one anchor softening cascades.

The alpha is not in shorting CoreWeave. It is in three adjacent positions: AI observability and FinOps names that capture spend regardless of who pours the concrete; non-Nvidia silicon paths that gain optionality the moment the Nvidia-as-bank model visibly strains; and the discipline to pass on any new neocloud deal whose survival requires Nvidia to participate. The first two are what you are doing with the capital. The third is what you are not.

The Closed Loop

Amazon and Alphabet booked ~$53B in private-market investment gains in Q1, with Amazon alone marking its Anthropic stake up by $15.6B. Hyperscaler 'Other Income' now runs at more than 33% of net income, against a historical norm of 5-10%. These are auditor-blessed, 10-Q-filed numbers sitting on Mag7 balance sheets. They are also, to borrow a useful phrase, gains on positions the buyers themselves helped price through the capital they committed to the companies being marked up.

The reflexivity is now quantifiable. Roughly 60% of combined quarterly income at Alphabet and Amazon is paper mark-ups on AI stakes, which means a single down-round in a top-five private AI name hits the Mag7 EPS that public portfolios benchmark against. The risk decomposes two ways. Marks that look validated may be self-referential, and any correction flows straight into the comps LPs watch daily. The first is probably wrong in the median case and very right in the tail.

The $4B Existence Proof

Core Automation — incorporated late March 2026, founded by ex-OpenAI researcher Jerry Tworek — went from $0 to $1B in weeks, and from $1B to $4B six weeks later. No product. No revenue. Nvidia anchored the seed at $100M, traditional VCs treated the Nvidia check as pre-validated demand, and a 4x markup round materialized before most IC committees could schedule a follow-up.

This is not an anomaly. It is the structure now:

FPV's Pegah Ebrahimi is on record that fast-follows "only work in boom times". When disciplined early-stage GPs say that out loud, the cycle is not early. The two-part round — lead investors anchoring at lower prices while the headline valuation rides on trailing tranches — creates information asymmetry that benefits the issuer and distorts every downstream comp.

Entry points for an outside investor are now either pre-Nvidia or post-$10B. The middle has been removed. That is not a complaint — it is a description of where capital is not going.

The Opportunity Cost Nobody Prices

Every dollar marked up inside a hyperscaler's venture book is a dollar those firms are not returning, not spending on their own infrastructure, and not using to buy the companies outright. The optionality has been spent. $15.6B of paper gain on Anthropic is $15.6B Amazon is not deploying elsewhere. That is a feature if Anthropic is the right horse. It is a cost if it is not, and the AI app layer keeps misbehaving for anyone who modeled it like SaaS.

For fund positioning, foundation-model primary rounds above $2B are no longer an institutional VC game. The asymmetric returns, or rather the more interesting version of them, live in pre-formation researcher access before Nvidia prices the deal, application-layer bets with real revenue at 15-25x, and structured secondaries on leaders with preference protections. The froth is the signal. It tells you where not to deploy, and by subtraction, where to look.

The Datadog Disclosure Changes the Comp Set

Olivier Pomel handed the public markets the cleanest AI monetization disclosure they have yet received, which is that 20% of Datadog customers using AI alongside the product drive 80% of ARR. The stock ran 30% in a single session, its largest move in more than six years. Atlassian had done a version of the same thing the prior week, up 25% on AI search pulling cross-sell through. The debate about whether AI-attach drives incremental spend is no longer a debate. It is disclosed, audited, and being repriced live.

The inverse printed on the same tape with equal conviction. HubSpot beat revenue by $20M and was marked down 20% because the guide implied AI is not yet lifting spend. Cloudflare cut 20% of headcount (1,100 people), pointed to 600% AI usage growth as the rationale, and gave back 18% after-hours. The tape read AI-justified cuts at a growth company as demand weakness rather than productivity leverage.

The Sorting Mechanism

The single-day dispersion across SaaS tells you exactly how AI exposure is being priced today:

The winners share one feature: their products observe or orchestrate AI workloads, which is to say they get paid regardless of which model wins. HubSpot's AI is competing with something the customer already pays for. The distinction used to be subtle. It is now a 50-point spread in one session.

The AI-RIF narrative just broke. Until last quarter, announcing a layoff and calling it 'AI efficiency' was worth a multiple. This week it was worth negative eighteen points.

What This Means for the Portfolio

Cloudflare is the more dangerous print for private marks. Every growth-stage portco CEO planning to dress their next RIF in an 'AI-first operating model' deserves a quiet briefing this week, because the public market is now reading AI-justified headcount cuts as demand weakness, not margin expansion. The narrative flipped. Private marks have not caught up yet, and this is probably wrong if hyperscaler capex tells a different story next quarter, but the direction is clear.

ServiceNow's 5x customer spend uplift from AI agents, disclosed in the same cycle, is the counter-example of what actually works. Agent revenue showing up as net-new budget rather than cannibalized seat revenue is the argument that earns a multiple premium. Any workflow tool with a credible agent overlay and provable spend expansion deserves a rerating. Anything framing AI as cost reduction deserves the opposite.

The Displacement Window

HubSpot's stumble opens a 12-18 month displacement window in CRM and marketing tools, or rather, the more interesting version: a window for any challenger who can show AI usage mapping to incremental customer spend before HubSpot recovers the story. Build the list this week, not next quarter.