Steven Fiorillo — CoreWeave Compute Gap Investment Thesis

Source: Coreweave: Why I have been buying shares into weakness, Steven Fiorillo, June 13, 2026.

The Framework: Four Steps From Hyperscaler RPO to CoreWeave Cash

Fiorillo's proof chain ties hyperscaler earnings to a single neo-cloud beneficiary:

Step	Claim	Evidence
1. Demand is real	Signed obligations, not pipeline guesses	MSFT $627B RPO; GOOG $462B backlog; AMZN $364B; ORCL $638B RPO
2. Supply is short	Buyers admit they cannot keep up	All four say demand runs ahead of capacity on earnings calls
3. CoreWeave refills the gap	Purpose-built AI cloud between models and silicon	~6 weeks powered-shell → live supercomputer vs 3–6 months for peers
4. CoreWeave monetizes	Take-or-pay contracts + mid-20s contribution margins	56% adjusted EBITDA margin even in heaviest build quarter

"When all the buyers say the same thing, I treat it as a structural shortage, not a one quarter blip."

Investment Thesis #1: Hyperscaler RPO Proves a Structural Compute Shortage

Argument: Earnings season closed with four hyperscalers reporting record contracted backlog — Microsoft $627B RPO (+99% YoY), Alphabet cloud backlog ~$462B (nearly doubled in one quarter), Amazon ~$364B (before a new $100B+ Anthropic deal), Oracle $638B RPO (+363%). Combined capex exceeds half a trillion dollars this year and could reach $1T+ in 2027 from these four alone — yet every management team says they remain compute constrained.

AI revenue is already real, not pilot: Microsoft AI at $37B run rate (+123% YoY); AWS AI past $15B run rate (triple-digit growth); Alphabet Gen-AI product revenue +800% YoY; Oracle OCI +93% YoY driven by AI workloads. Microsoft CFO expects to stay capacity constrained through at least 2026; Sundar Pichai said cloud revenue would be higher if they could meet demand — money left on the table for lack of compute.

Contrarian element: Bears treat AI capex as cyclical hype; Fiorillo treats four independent CFO/CEO admissions as proof of a multi-year structural gap.

Trigger: Hyperscaler capex guides rising again in 2027; continued RPO growth despite massive spend; capacity-constrained language persisting on earnings calls.

Names: Microsoft (MSFT), Alphabet (GOOG), Amazon (AMZN), Oracle (ORCL) as demand proof; CoreWeave (CRWV) as gap-filler.

Investment Thesis #2: CoreWeave Backlog Is Contracted Demand, Not Hope

Argument: CoreWeave's revenue backlog grew 284% YoY to ~$99.4B and nearly doubled in a single quarter. It signed >$40B of new commitments in Q1 alone — more backlog in one quarter than most AI clouds have added in their entire history. 10 separate customers each committed $1B+; every major model runs on CoreWeave — OpenAI, Anthropic, Meta. Enterprise is scaling: financial services ~$10B backlog, physical AI past $1B.

This is the same demand showing up in hyperscaler RPO — but sitting on CoreWeave's books as contracted obligations.

Contrarian element: Market focuses on net losses and dilution; Fiorillo focuses on signed backlog as the leading indicator that demand has a destination.

Trigger: Backlog continuing to grow quarter-over-quarter; new $1B+ customer announcements; diversification beyond AI-native credits (<30% of backlog from non-investment-grade AI natives).

Names: CoreWeave (CRWV), Meta (META), Anthropic (PRIVATE), OpenAI (PRIVATE).

Investment Thesis #3: Take-or-Pay Economics + Six-Week Speed = Cash Engine

Argument: CoreWeave is not generic cloud rental — it is a purpose-built AI stack (data centers, networking, storage, orchestration software) tuned for GPU workloads. >90% of reserved customers use 2+ products; 75% use 3+ — wallet share and stickiness. It turns a powered shell into revenue-generating capacity in ~6 weeks vs 3–6 months for most peers — faster monetization is the product.

Contracts are take-or-pay for 4–6 years: fixed price per GPU hour whether capacity is used or not — same methodology as energy infrastructure. Contribution margins underwritten in the mid-20s; adjusted EBITDA $1.2B at 56% margin in Q1 even while posting a net loss. Cost of debt down ~700 bps since 2024; $8.5B DDTL 4.0 was the first investment-grade-rated loan backed by HPC infrastructure, priced under 6%.

"If I contract a gigawatt, I pay for a gigawatt. Even if I use half a gigawatt, I love that model."

Contrarian element: Street penalizes reported margin dip during buildout; Fiorillo says dip is mechanical — lease/power/depreciation hits 1–2 months before revenue, normalizes by month three. Management called Q1 the trough and guided margins expanding every quarter to low double-digit adjusted operating margin by Q4.

Trigger: Contribution margins holding mid-20s on new deployments; adjusted operating margin reaching low double digits by Q4; continued debt refinancings below 6%.

Names: CoreWeave (CRWV).

Investment Thesis #4: Inference Mix and Pre-Sold 2027 Revenue = Durability

Argument: Revenue compounded $982M → $2.08B (+112% YoY) with uninterrupted quarterly growth: 982M → 1.2B → 1.36B → 1.57B → 2.08B. More than half of compute is now inference, not training — the recurring, usage-driven production side of AI rather than one-time training builds. New contracts average ~5 years; >75% of exit-2027 run-rate revenue (> $30B target) is already contracted before renewals.

Component and power inflation gets priced into contracts at signing with POs in hand — spiking memory costs should not silently erode margins if the pass-through mechanism holds.

Contrarian element: Training-heavy AI clouds look cyclical; inference majority + multi-year take-or-pay shifts the revenue mix toward production monetization.

Trigger: Inference share continuing to rise above 50%; exit-2027 run-rate guidance holding with >75% pre-contracted; pass-through holding margins through memory spikes.

Names: CoreWeave (CRWV), Micron (MU) and memory suppliers as industry cost pass-through context.

Investment Thesis #5: Nvidia Partnership and Neo-Cloud Scale Moat

Argument: CoreWeave sits closest to the chips — among first to market on each new Nvidia generation. Named Nvidia exemplar cloud for inference on GB200 NVL72; Nvidia qualified its software as reference architecture. $2B Nvidia equity investment plus partnership to build >5 GW of AI factories by 2030. Early chip access + full-stack orchestration (stripped virtualization, proprietary GPU health layer) delivers higher price per GPU hour with lower total cost to customer.

Execution scale: ~50 live data centers; no single provider >17% of active power; six-week speed-to-live when rivals need 3–6 months.

Contrarian element: Hyperscalers could self-build — but Fiorillo cites CoreWeave IR counter: no hyperscaler has ever shrunk a revenue-generating fleet; AI gear economics favor leasing. Hopper pricing rising supports management view that supply won't catch demand this decade.

Trigger: New Nvidia generation deployments on CoreWeave first; GB200 NVL72 inference ramp; >5 GW AI factory milestones with Nvidia.

Names: CoreWeave (CRWV), NVIDIA (NVDA).

Investment Thesis #6: Neo-Cloud Pair Trade — CoreWeave and Nebius

Argument: Fiorillo is adding CoreWeave into weakness despite dilution and capital intensity — if you believe in the AI revolution and compute constraint, "the NeoCloud should win." He also holds Nebius and is adding to both positions.

Contrarian element: Weakness/dilution fears create entry; Fiorillo sizes neo-clouds as the levered expression of hyperscaler demand they cannot satisfy internally.

Trigger: Nebius backlog/RPO growth mirroring neo-cloud demand theme; CoreWeave share price weakness on dilution without backlog deterioration.

Names: CoreWeave (CRWV), Nebius (NBIS).

The Ecosystem Map

Fiorillo's active positions: CoreWeave (adding into weakness), Nebius (adding)
Hyperscaler demand layer: MSFT, GOOG, AMZN, ORCL — record RPO/backlog, all capacity constrained
Model customers on CoreWeave: OpenAI, Anthropic, Meta — every major model
Enterprise verticals: Financial services (~$10B backlog), physical AI (>$1B backlog)
Nvidia stack: $2B equity stake, exemplar cloud status, >5 GW AI factories by 2030, reference architecture qualification
Capital structure: $8.5B investment-grade DDTL 4.0 under 6%; debt cost down ~700 bps since 2024
Value capture: Fiorillo places bets at the neo-cloud layer between hyperscaler demand and Nvidia silicon — not at the model or chip layer directly

Key Risks

Heavy losses and interest: Q1 net loss $740M; $536M net interest — leverage remains high until more capacity earns revenue
Rates moving against them: Cheaper debt is a weapon; rising rates reverse the ~700 bps improvement since 2024
Capital intensity: Q1 capex $6.8B; FY guidance $31–35B — land, power, chips laid out before revenue; cheap capital access must stay open
Customer concentration: Improving (<30% backlog from non-investment-grade AI natives) but handful of very large contracts still carry weight
Component and power inflation: Memory costs spiking industry-wide — pass-through at signing must keep holding
Hyperscaler self-build at renewal: Big customers could insource capacity; counter is leasing economics and no fleet shrink precedent
Commoditization: If supply catches demand, GPU pricing compresses — management argues not a this-decade problem; long-term watch item
Dilution: CoreWeave diluting shareholders to fund tremendous capital spend — Fiorillo accepts this as cost of the thesis

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	CoreWeave (CRWV)	$99.4B backlog (+284% YoY); 56% adj EBITDA margin; 6-week speed-to-live	Fiorillo adding into weakness; >$40B Q1 bookings
2	NVIDIA (NVDA)	$2B equity stake; exemplar cloud; >5 GW AI factories by 2030; early chip access	GB200 NVL72 inference reference architecture
2	Oracle (ORCL)	$638B RPO (+363%); OCI +93%; demand far exceeds supply	Validates structural shortage CoreWeave fills
2	Microsoft (MSFT)	$627B RPO (+99%); Azure +40%; capacity constrained through 2026	CFO admission — not "bull on X" narrative
2	Alphabet (GOOG)	$462B backlog; Cloud +63%; compute constrained	CEO: would generate more revenue if they could build
2	Amazon (AMZN)	$364B backlog; AWS +28% (15-quarter high); $150B run rate	AI revenue past $15B run rate, triple-digit growth
3	Nebius (NBIS)	Second neo-cloud position alongside CoreWeave	Fiorillo investor adding to both neo-clouds
3	Meta (META)	Named major model customer on CoreWeave platform	Part of diversified $1B+ customer base

Monitoring Checklist

Bottom Line

Four hyperscalers with >$2T combined contracted backlog and >$500B annual capex still cannot meet demand — Fiorillo treats this as structural shortage, not a one-quarter blip.
CoreWeave's $99.4B backlog (+284% YoY) is the contracted destination for that demand — signed in one quarter more than most AI clouds have booked in their history.
Take-or-pay contracts (4–6 years) and 56% adjusted EBITDA margins prove the model works even in the trough build quarter — the margin dip is lease/power timing, not broken pricing.
Six-week speed-to-live vs 3–6 months is a monetization moat — revenue starts sooner; inference now >50% makes the book increasingly production-recurring.
Fiorillo is buying CRWV into weakness despite dilution — paired with Nebius as the neo-cloud expression of a compute-constrained AI supercycle he believes is still accelerating.