CoreWeave (Trans): Margins inflect; interest expense rising in tandem

Dolphin Research's Trans of CoreWeave FY26 Q2 earnings call

I. Core financial highlights recap

1. Full-year guidance raised across the board

Revenue: $12.4bn–$13.2bn.Adj. OP: $960mn–$1.15bn. Year-end run-rate revenue: $18.5bn–$19.5bn.

Year-end power in operation: lifted from 1.7GW+ to 1.85GW+.Management emphasized stronger delivery cadence behind the raise.

2. Q3 guidance

Revenue: $3.45bn–$3.6bn.Adj. OP: $200mn–$260mn, with margins continuing to expand QoQ and reaching low-teens in Q4.

Interest expense: $860mn–$940mn, reflecting larger debt balances to support accelerated deliveries.Capex: $11.5bn–$13.5bn.

3. Full-year capex raised to $35bn–$39bn, driven by higher expected capacity deliveries to customers this year and several recent major wins.Q2 capex was $9.4bn, slightly above the prior guide range; CIP rose from $9.6bn to $11.9bn, with significant late-quarter power-in-place to be transferred to PP&E in early Q3.

4. Key quarterly metrics

Aggregate: revenue of $2.6bn (+112% YoY, +24% QoQ); adj. EBITDA $1.5bn (vs. $753mn a year ago), doubling YoY, with a 59% EBITDA margin.Adj. OP $128mn (vs. $21mn in Q1 and $200mn a year ago), OPM 5%, well above the guide ceiling.

Losses and opex: net loss $626mn (vs. $290mn net loss a year ago); adj. net loss $567mn (vs. $130mn a year ago).Interest expense $640mn (vs. $267mn a year ago); opex $2.6bn, including $165mn in SBC.

Backlog: revenue backlog of $104bn (+46% YoY), excluding >$25bn net new customer commitments signed in early Q3.Over 50% of the existing backlog is under delivery; by year-end, expected to exceed two-thirds.

Taxes: despite posting a net loss, income tax was accrued due to a valuation allowance against net deferred tax assets.Absent material discrete items or changes, the 2026 tax rate should remain broadly stable.

5. Capital structure and liquidity

Raised approx. $18bn in Q2 across debt, converts and equity, including the first EUR bond and the first publicly syndicated, HPC infrastructure-collateralized DDTL.Cumulatively secured over $32bn in debt and equity capital.

Weighted avg. debt cost fell by nearly 300bps over the past year, implying ~$1.1bn annualized interest savings on Q2-end debt balances.As of Jun 30, cash, cash equivalents, restricted cash and marketable securities totaled over $6.9bn.

II. Earnings call details

2.1 Management highlights

1. Demand and pricing

Demand is diffusing across industries, geographies, workloads and GPU generations.AI is no longer confined to frontier model labs; it is embedding into software, industrial systems, financial markets, enterprise workflows and national security missions.

Near-term capacity is effectively sold out, with contracting terms increasingly favorable to the company.Pricing and margins for Blackwell and Vera Rubin SKUs reached record highs, while prior-gen SKUs price at or above levels from several years ago.

Prices were raised in Jul, with an Avg. ~25% uplift across SKUs, supported by the current demand backdrop and higher ROI as customers shift to inference.Component cost inflation is being passed downstream.

Contribution margins on Q2 new customer contracts are expected to be 5–10pts higher than in recent quarters, and the Q2 margin uplift occurred before the Jul price increases.This underscores structural pricing power.

2. Customer expansion and vertical traction

Industrial: partnered with Caterpillar to deploy the NVIDIA Vera Rubin platform, supporting industrial-grade physical AI training and inference to enhance domain-specific models for autonomous construction equipment.Life sciences: added Isomorphic Labs as a customer, with the vertical becoming a key growth driver.

Financials: onboarded Flow Traders and IMC, enabling next-gen AI model development and deployment for systematic trading.Public sector: working with Leidos to advance CoreWeave Federal, delivering secure AI for defense, national security and intelligence missions.

Others: Descartes is developing Oasis 3, a first API-accessible physical AI world model, on the platform; IBM uses it for RL, agent tool invocation and model evaluation.Through Monolith's on-site engineers, the company directly serves Nissan and ZF.

CoreWeave Omni is seeing strong interest from sovereign, enterprise and cloud customers.The first deal was signed in recent weeks, with scale-up starting in 2027.

3. AI platform and developer tools

Core view: training, inference, evaluation and improvement form a continuous loop.Compute has shifted from an upfront spend early in the model lifecycle to ongoing demand that grows with each production app and iteration.

The platform spans cloud infra, managed inference, developer tools and agent solutions, plus orchestration and observability via Mission Control.Clients can deploy, monitor with Weights & Biases, evaluate, tune via serverless RL or sandboxes, and return to production after validating quality, performance and cost.

CoreWeave ARIA (AI Research & Iteration Agent) analyzes thousands of eval runs and provides minute-level insights, recommending the next experiment.In Q2 the company launched seven new AI platform capabilities and delivered several industry firsts.

Modeling training runs tracked on the platform surpassed 1bn this week.AI development services carry higher GP than core cloud and reach a broader customer base, serving as a natural expansion path: today's AI app developers are tomorrow's AI cloud infra customers.

4. Managed Inference

Within a few months, booked ARR grew from ~$1mn to over $100mn.Managed inference ARR is expected to reach at least $250mn by end-2026.

The only constraint on growth today is near-term capacity.Token monetization is enabled via both serverless and dedicated deployments, giving customers flexibility in consumption.

On rankings such as artificial analysis, open-source models (Kimi 2.6, K2.7 Code, GLM 5.2, MiniMax M3) lead on single-token cost and first-token latency.Customers like Grammarly and u.com have moved from trials to production traffic, running AI coding agents, proprietary fine-tuned models and open-weight models.

5. Performance, TCO and third-party validation

In Q2, CoreWeave became the first cloud to bring up and validate NVIDIA Vera Rubin NVL72, leveraging in-house capabilities in software-defined liquid cooling and rack management.On NVIDIA Grace Blackwell, the company set MLPerf records for open-source model training and inference.

Internal tests show the lowest inference single-token cost.Per Signal65, CoreWeave's estimated TCO is up to 47% lower than the Avg. hyperscaler.

In Jul, Gartner named CoreWeave a 'Visionary' in the 2026 Magic Quadrant for cloud AI infrastructure.This recognition reinforces competitive positioning.

6. Power and capacity

Power in operation reached 1.5GW at Q2-end, with ~500MW added in the quarter, the largest quarterly increase in history and over 3x YoY.In Jun alone, >300MW was added, exceeding any prior full quarter.

Signed power reached 3.7GW in Q2, with ~500MW added since quarter-end, now at 4.2GW.These figures exclude electrified land, expansion options at existing sites and executed LOIs for >1.5GW of potential power.

First self-built projects are underway, with the initial site expected online later this year.Electrified land underpins deeper vertical integration, enabling greater operating control and better long-term margins.

Outside the U.S., over 1GW has been signed.The company recently entered APAC, with 360MW in Indonesia slated to begin coming online in ~18 months; Intl markets are expected to be a major growth driver.

The company sees good visibility to at least 8GW by 2030.Management believes demand will significantly outstrip supply for years.

7. Supply chain and capital structure

Long-term relationships with NVIDIA and OEM/ODM partners underpin capacity delivery, while a long-term agreement with Solidigm helps de-risk critical component supply.The second publicly syndicated term loan was the first to include short-duration customer contracts; pricing occurred during one of the most volatile weeks of the year, yet the deal was fully sized, and spreads have since tightened.

This financing expands the ability to serve enterprise at scale and accelerates the managed inference ramp, increasing exposure to short-duration contracts.Short-duration deals typically carry higher ASP and margins.

8. Contract cash flow cadence and re-contracting value (CFO framework)

Unit economics on a typical 5-year contract remain strong and are improving, though returns are not evenly distributed over time.Costs are front-loaded as capex and funded by debt, customer prepayments and corporate capital; once a cluster is delivered, contract revenue ramps and becomes predictable and highly cash generative.

This cadence is embedded in pre-signing margin models.A single deployment can fully amortize asset-level debt used for capex and generate substantial incremental FCF.

At initial contract expiry, clusters are unlevered, allowing the company to re-contract or redeploy to the market.Each resale or renewal is incremental to the returns already realized in the initial term.

Economics do not rely on post-initial-term re-contracting, yet current observations point to longer useful lives and higher pricing, implying notable upside.Ampere and Hopper fleets already show this: a recently signed A100 contract extends to 2029 at attractive pricing, even though the SKU launched in 2020.

Legacy clusters are installed, electrified, production-grade and running at scale, with customer ROI validated.In a market with constrained new capacity and rising costs, they are scarce assets.

High-margin adjacencies across storage, CPU, networking and software reached >$400mn ARR in Q2 and are expected to expand rapidly.These will continue to thicken overall margins.

2.2 Q&A

Q: On renewals post initial term, what are typical customer intentions on tenor? How much installed equipment faces renewals in the next few years and how big is the opportunity?

A: The market is validating that older-gen infrastructure still has significant value for many AI use cases.NVIDIA's frontier solutions matter for the most demanding customers and cutting-edge use cases, but many workloads in the same ecosystem can run on later-stage, older SKUs.

We can sell a 2020-architecture GPU at full pricing into a contract expiring in 2029, offering a direct reference point for off-contract prospects of this fleet.Capacity coming up for renewal is a small share of the total fleet, and older-gen ASPs remain at or above levels from a year ago.

An interesting angle is routing these fleets into managed inference after expiry, a business evolving rapidly and expected to keep growing fast, with year-end ARR around $250mn.This provides another monetization path.

Q: Beyond existing agreements, do you need broader LT supply agreements to lock in future capacity?

A: Our mandate is to ensure we can deliver what customers need, which requires tightly managing a complex supply chain spanning land, power, shells, GPUs, networking and memory, all pressured by AI's growth.Over the past years we have built LT, deep relationships with ODMs, OEMs, NVIDIA and memory suppliers, and we systematically plan for the infra, components and capital needed to deliver at acceptable price and quality.

This is embedded in daily operations: continuously managing these relationships and securing everything required to deliver NVIDIA-based infrastructure.Importantly, value delivered by CoreWeave Cloud is expanding faster than input cost inflation in the supply chain, resulting in margin expansion: contribution margins on last quarter's new contracts were 5–10pts higher than recent quarters.

Q: Early learnings from managed inference? How do you allocate capacity between it and traditional take-or-pay contracts?

A: We manage the product portfolio holistically.In recent years, we focused on scale via LT take-or-pay contracts; as we enter the hyperscale phase, we are broadening the lineup with higher-margin products, software and CPU capabilities essential for customer success.

Managed inference scaled from ~$1mn to $100mn in a single quarter, offering a great opportunity to deliver frontier compute and to repurpose off-contract GPUs for long-tail value extraction.The market is deep, and with silicon control we have native advantages; we expect long-term success.

Additionally, yesterday's completion of DDTL 5.5 shows strong capital markets interest and support for underwriting CoreWeave's short-duration contracts, a clear tailwind for this segment.This will help expand the addressable market.

Q: The 5–10pt margin uplift on recent contracts: what are the drivers? How much from shorter-duration contracts vs. competitive differentiation? And how is broader market pricing?

A: It's multi-factor and hard to decompose precisely.First, infrastructure delivered via CoreWeave Cloud is more valuable to customers than alternatives: platform quality, infra reliability, security and TCO bring customers back and drive expansions on our cloud, as they recognize our delivery enhances the value of the same hardware.

Second, many customers are monetizing their products, making them proactive and willing to pay higher margins to secure the compute they need to succeed.This is happening across infra, but is particularly pronounced in our ecosystem: premium products are being priced at a premium by compute buyers.

Q: With multiple upward pricing drivers (higher delivered value, component cost pass-through, re-contracting), should year-end ARR still be viewed at $18bn–$19bn?

A: We have raised the year-end ARR in guidance to $18.5bn–$19.5bn for 2026.This is already embedded in the outlook.

Q: With tighter DC regulation and local pushback, how confident are you in exceeding 3GW of power in operation by end of next year and the 8GW roadmap by 2030?

A: Some communities have enacted moratoriums, but these don't affect demand, only siting.Engagement must be highly collaborative with host communities, grounded in transparency and working with local governments, utilities and policymakers to ensure projects integrate into the communities we enter.

Practically, we fund grid upgrades ourselves to avoid passing costs into rate base, create substantial construction jobs and sustain long-term DC jobs, while completed DCs contribute to the tax base.These factors are critical to market entry and community engagement, and this infrastructure is essential for U.S. AI leadership.

None of the figures guiding today's progress and outlook are impacted by current regulatory resistance.We are confident, will keep expanding and engaging, and operate DCs at industry-best standards, expecting to be held to that bar.

By the numbers: signed power stands at 4.2GW today, plus ~1.5GW from electrified land options and executed LOIs, approaching 6GW.Given we are only mid-2026, the path to >8GW of power in operation by end-2030 is tracking well.

Q: As inference rises in importance, how must the fleet evolve? Inference needs more CPU and storage; can current DCs support it?

A: We've discussed this for several quarters.We don't build 'for training' or 'for inference' separately; we build AI infrastructure with all components to serve the full AI loop, cycling between training and inference through multiple iterations required for customers.

As a result, deployed infrastructure will smoothly tilt toward inference over time.This is by design.

Q: With renewed focus on edge AI after Meta's news, how do you view the evolution among edge, small clouds, neocloud and hyperscalers?

A: A key underappreciated point is CoreWeave's position at the epicenter of massive information flows: hyperscalers use us, labs use us, and enterprises are scaling on the platform, feeding insights into how the future will look and informing our positioning and capacity allocation.Some workloads will run at the edge, others have looser latency and protection needs; our cloud is designed to serve both effectively, and we will continue to calibrate based on customer demand for more edge vs. more scale tolerant of latency.

We are seeing both types of workloads and are confident that infra scale and the ability to schedule across the two will be a durable competitive advantage.On competition: even with rising competition, demand, pricing and margins are all expanding simultaneously, signaling product strength and growth within an already large TAM.

Q: You added 500MW of power in operation this quarter, yet QoQ revenue growth was roughly in line with last quarter. How should we think about the ramp-in of capacity, and when does the 300MW added in Jun contribute at steady state? Any early monetization signals vs. prior gen?

A: Of the ~500MW added in Q2, ~300MW was concentrated in Jun, a single-month addition larger than any prior quarterly increase in CoreWeave's history.This capacity was back-end loaded within Q2, so contributions begin to show in Q3 and Q4.

On monetization, focus on Vera Rubin.This generation is driving margin expansion from the outset: demand for Vera Rubin is immense, and our pricing power on delivering it via CoreWeave Cloud is a strong signal.

A large portion of the 5–10pt margin uplift comes from Vera Rubin SKUs, which we expect to be a very successful generation.This should sustain improved unit economics.

Q: As GPU fleets roll off, how do you choose between re-contracting, spot allocation and routing into managed inference? How far ahead do you decide?

A: We are still learning.Inference is extremely dynamic and expanding so quickly that we are pushing to keep up on new infra build cadence; off-contract infrastructure gives us flexibility to scale managed inference and to size the true breadth of the opportunity.

Some retired capacity will be re-contracted, subject to attractive economics, including tenor and considerations for long-term stability and obligations supporting ongoing build-out.We also recognize near-term opportunities to sell on shorter tenors and extract additional margin, given a structural supply-demand imbalance that has persisted for years and will continue for the foreseeable future.

Q: DDTL 5.5 finances short-duration contracts, and the political climate around DCs suggests the biggest players benefit most from pricing tailwinds. How does this financing change your target contract tenors? Will you lean more into margin extraction on short-duration deals?

A: Exactly.DDTL 5 enables CoreWeave to fill the curve across the tenor structure we deem most profitable: selling both LT and short-term contracts to extract incremental margin; we have been aggressive here and are the first to bring 5.5 to market.

Shorter-term contracts also matter because enterprise customers prefer sub-5-year tenors, reflecting shorter planning cycles.By securing financing that supports such contracts under 5.5, we can diversify tenors and back more deals, opening a new market of customers wanting two- to three-year capacity, which was harder to address before—now it accelerates as we can offer capacity on their purchasing timelines.

<End of text>

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure