Below we discuss CoreWeave 2Q26 Earnings and Conference Call Analysis. We are more analytical than deterministic, we are not providing our future estimates in this pieces, but if interested, subscribe for more!
1. The Summary
This was the quarter management had been promising for several calls: operating margin is improving sequentially (op income -$144m → -$49m) on record revenue and a record power add — but the inflection is happening underneath an interest expense line that is now $640m, and the entire investment case rests on whether a $35–39bn capex year converts into what management just guided to!
2. The Numbers
| Metric | Q2 2026 | Change |
| Revenue | $2.6bn | +112% y/y, +24% q/q |
| Revenue backlog | $104.2bn | +246% y/y + $25bn beginning Q3 |
| Operating Income | $(49)m | $19m in Q2 2025 |
| Interest expense | $640m | +140% |
| Net loss | $(626)m | $(290)m in Q2 2025 |
| Capex | $9,4 bn ($6,4 bn cash) | — |
| Cash + equivalents | >$5.5bn | — |
CoreWeave delivered $2,5 bi in revenues in 2Q26, a 112% growth YoY.
However CoreWeave is in the middle of ramping up new data centres, which press operating margin. Depreciation expenses grew 149% YoY and Cash Cost and Expenses grew 135% YoY. The GAAP margin was negative 2%, but we noted an important sequential improvement that should keep stepping up next quarters!
Against our numbers, Revenue came just 5% below; depreciation came in 12% above and cash costs came 8% below.
NonOperating Expenses, despite increasing, is coming better than we expected, in fact management highlighted in the call the reduction in avg cost of debt they were able to deliver so far.
CoreWeave Cash Flow
Cash Flow from Operations came in lower than we expected, absent client prepayments and cash advances that helped last quarters.
CoreWeave continue to be FCF negative in the short term and leverage is still high, but company fast execution and high expected growth should be the levers to improve future earnings and cash profile!
Operating footprint
- Active power: 1.5 GW, ~500 MW added in the quarter — “more than any quarter in our history,” more than tripling y/y, and more than any single neo-cloud operates in total (third-party estimates, per Intrator).
- Over 300 MW came on in June alone — “June itself larger than any full quarter in our history.”
- Contracted power: 3.7 GW at quarter-end → 4.2 GW as of the call, plus >1.5 GW of powered land, existing-site expansion options and executed LOIs. Agrawal framed this as “close to about 6 gigawatts already” against the ≥8 GW by 2030 target.
- >1 GW contracted outside the US, including 360 MW in Indonesia (first APAC entry), coming online in ~18 months.
- First several self-builds underway; first self-build site expected online later this year.
Software / non-GPU revenue lines — this is where the margin story actually lives:
- Storage, CPU, networking and software: >$400m ARR as of Q2, described as margin-accretive.
- Managed inference: $1m → >$100m booked ARR in the few months since launch; exit-2026 target ≥$250m ARR. Growth “constrained only by our near-term capacity.”
- 7 new AI platform capabilities shipped; 1 billion model training runs tracked on the platform.
3. Guidance
| Q3 2026 | FY 2026 | |
| Revenue | $3.45–3.60bn | $12.4–13.2bn (raised) |
| Adj. operating income | $200–260m | $960m–1.15bn (raised) |
| Adj. operating margin | — | “low teens” in Q4 |
| Interest expense | $860–940m | — |
| CapEx | $11.5–13.5bn | $35–39bn (raised) |
| Year-end active power | — | >1.85 GW (raised from >1.7 GW) |
| Exit annualized run-rate revenue | — | $18.5–19.5bn (raised) |
Arithmetic:
- Q3 revenue guide implies +33% to +38% sequential growth — a step up from Q2’s +24%, consistent with the June power add landing in Q3.
- Q1 ($21m) + Q2 ($128m) + Q3 ($200–260m) = $349–409m. Against the FY guide of $960m–1.15bn, that leaves ~$551–741m of adjusted operating income in Q4 alone — roughly 2x the first three quarters combined.
- Q4 revenue implied at ~$4.25–4.9bn. $551–741m on that base is ~12–15% — which does reconcile cleanly with the stated “low teens.” The guide is internally consistent, but it is a very large single-quarter step.
- FY capex of $35–39bn is roughly 3x FY revenue of $12.4–13.2bn. Q3 capex guide alone ($11.5–13.5bn) exceeds the $6.9bn cash balance.
4. Management’s narrative: the four pillars
Intrator organised prepared remarks around four claims.
(i) Demand is broadening, not just deepening.
The argument: AI has left the frontier-lab ghetto and is entering “software, industrial systems, financial markets, enterprise workflows and national security missions.” Evidence offered was customer names rather than segment disclosure:
- Industrial / physical AI — Caterpillar, deploying NVIDIA Vera Rubin, using CoreWeave as a “data factory” for autonomous construction equipment.
- Life sciences — Isomorphic Labs (new).
- Financial services — Flow Traders and IMC, systematic trading firms, added to an existing roster.
- Public sector — Leidos, via CoreWeave Federal, for defense/national security/intelligence.
- AI-native + enterprise change-makers — Decart (Oasis 3, “first API-accessible world model for physical AI”), IBM (RL, agent tool use, model evaluation), and Nissan and ZF via Monolith field engineers.
- Inference production traffic — Grammarly and You.com.
- CoreWeave Omni — first deal signed in “the past few weeks,” begins to scale in 2027; interest from sovereign, enterprise and cloud customers.
Two pricing claims sit inside this pillar and matter more than the logos: “Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs, while pricing for prior generation SKUs is at or above where it was years ago,” and “Our near-term capacity remains effectively sold out.”
(ii) The “continuous loop” thesis.
The strategic argument of the call. Management’s framing: enterprises no longer treat a model as a deliverable (train → deploy → done). Training, inference, evaluation and improvement now form one continuous loop, so compute stops being a front-loaded one-time cost and becomes an ongoing requirement that grows with every application in production.
This is the intellectual justification for both the demand curve and for owning the tooling layer — Weights & Biases for monitoring, Mission Control for orchestration/observability, serverless RL and sandboxes for experimentation, and ARIA (AI Research and Iteration Agent) to analyse thousands of eval runs and recommend the next experiment.
“AI development services, which carry higher margins, are also being adopted by a broader set of customers than our core cloud… the developers building AI applications today are the AI cloud infrastructure customers of tomorrow.”
That is a land-and-expand funnel claim, and the >$400m services ARR plus the $100m inference ARR are the first datapoints against it.
(iii) Architecture and TCO.
First cloud provider to bring up and validate NVIDIA Vera Rubin NVL72, attributed to software-defined liquid cooling and rack management. New MLPerf records for training and inference on Grace Blackwell, and lowest cost per token in internal tests.
Third-party validation cited: Signal65 (TCO up to 47% lower than the average hyperscaler) and Gartner (Visionary, 2026 Magic Quadrant for Cloud AI Infrastructure, July).
Leaderboard leadership on cost-per-token and time-to-first-token claimed via Artificial Analysis for Kimi 2.6, K2.7 Code, GLM-5.2 and MiniMax M3.
(iv) Power and supply chain as the real moat.
Beyond the gigawatt figures, the mechanical argument Intrator made is the most analytically useful thing in the prepared remarks:
“Each new deployment is landing against a much larger installed base than it was even 1 quarter ago. As that base grows, each new build becomes a smaller part of the whole, while contracted revenue from existing deployments remains in place. This is how we are transforming scale into operating leverage.”
That is a denominator argument: ramp costs are roughly proportional to incremental capacity, revenue is proportional to installed capacity, so margin expands mechanically as long as growth decelerates relative to base — which also means the margin expansion is partly a function of growth slowing, not of unit economics improving.
Supply-chain de-risking was represented by the long-term Solidigm agreement (memory/storage) on top of NVIDIA, OEM and ODM relationships.
5. The economics: what Agrawal actually laid out
This was the most substantive section of the call and deserves separating from the narrative.
Pricing. The Q2 margin improvement came before the July pricing action of approximately +25% across SKUs, taken “in response to the current demand environment and the increasing ROI our customers are observing… as they shift to inference.” Component price increases are also being passed through. So none of the July repricing is in the Q2 print — it flows into H2, which is part of why the Q4 step is as large as it is.
Contribution margin. Contracts signed in Q2 carry expected contribution margins 5–10 percentage points above those added in recent quarters.
The cash-flow shape. Agrawal was explicit about the J-curve: a typical 5-year contract’s economics “do not arrive evenly.” Cost is front-loaded as capex, funded by asset-level debt, customer prepayments and corporate capital. Once the cluster is delivered, revenue is contracted, predictable and highly cash generative. The deployment fully repays the asset-level debt used to fund it and generates additional free cash flow — so at the end of the initial term the cluster is unlevered, and recontracting is pure incremental return.
The recontracting option. This is the newest strand of the story and the one that dominated Q&A. Claims made:
- CoreWeave remains “largely sold out of prior generations of NVIDIA GPUs in addition to the current SKUs.”
- Upside is being realised across the Ampere and Hopper fleet.
- The headline proof point: an A100 contract recently signed extending into 2029 — for a SKU introduced in 2020, i.e. roughly nine years of contracted life.
- The hedge: “we have built a business whose economics do not rely on recontracting after initial customer term.”
Capital structure. ~$18bn raised in Q2 across debt, convertibles and equity; >$32bn secured to date.
Firsts in the quarter: inaugural Eurobond; first delayed draw term loan backed by HPC infrastructure issued in public markets; and a second publicly syndicated term loan — DDTL 5.5, closing announced the day before the call — which is the first to include shorter-duration customer contracts.
Notably, it priced “during one of the most dislocated weeks for credit this year,” was completed at full size anyway, and spreads have retraced since. Weighted average cost of debt is down ~300bps over the past year, worth ~$1.1bn of annualized interest savings on the end-Q2 debt load.
The strategic point Agrawal made about DDTL 5.5 is easy to miss: it “unlocks our ability to serve a critical part of the enterprise market at scale while also allowing us to accelerate the ramp of our managed inference platform and grow our exposure to shorter-dated contracts that typically come at a higher ASP and margins.”
Backlog quality. Of the $104.2bn backlog, >50% is attached to contracts where customer delivery has commenced, expected to reach >2/3 by year-end. Excluded from that figure: >$25bn of net new commitments signed in the early weeks of Q3, which would take the pro-forma figure to roughly $129bn.
6. Other Highlights
Agrawal: “the capacity that’s coming up for renewal… is a very limited part of our fleet.” ASPs on older generations are “higher or at levels that we’ve seen about a year ago.” He then redirected to managed inference as the alternative home for maturing fleet, reiterating the ~$250m exit ARR. If the renewal pool is small, the A100-to-2029 datapoint is a proof of concept, not a material 2026–27 earnings driver.
Intrator on price: “many of our clients are monetizing their products. And so they are more aggressive about coming in and willing to pay us higher margins because they need access to the compute that will allow them to be successful.” >> This is the more important admission on margin expansion driven by customers’ own monetisation success.
Agrawal on competition: “Even with this increased competition, we are seeing demand, pricing and margin all expanding.”
Intrator: Vera Rubin is “seeing the margin expansion right from the start,” demand is “enormous,” pricing power is strong, and — the key attribution — “when we were talking about that 5% to 10% margin step function that we’re seeing, a lot of that is coming in, in the Vera Rubin SKU.” >> this materially narrows the interpretation of the 5–10pp contribution-margin improvement. It is substantially a new-SKU phenomenon rather than a fleet-wide repricing, which makes it dependent on continued Vera Rubin allocation and on the premium holding as Vera Rubin capacity proliferates across the industry.
7. Tensions and things to carry forward
The interest expense problem is now the story. Q2 interest expense of $640m is 5x adjusted operating income of $128m. The Q3 guide makes it worse in absolute terms: $860–940m of interest against $200–260m of adjusted operating income. Even taking the FY adjusted operating income guide at the top end ($1.15bn), Q3 interest alone nearly matches the full year’s operating profit. The ~300bps reduction in weighted average cost of debt (~$1.1bn of annualized savings) is real and impressive, but it is being applied to a debt balance that is growing faster than the rate is falling. Rate improvement is losing the race against balance growth — that is the cleanest single frame for the credit risk here.
Capex is ~3x revenue and cash is thin relative to the next quarter’s spend. $35–39bn FY capex against $12.4–13.2bn FY revenue; $11.5–13.5bn Q3 capex against a $6.9bn cash balance. The model is structurally dependent on continuous capital-markets access. That dependency was tested this quarter — DDTL 5.5 priced in “one of the most dislocated weeks for credit this year,” was taken at full size, and spreads retraced. That is a genuine positive datapoint on market access, and it is also a reminder of what the sensitivity is.
Near-term catalyst named on the call: the Fully Connected developer conference in September, with leadership and customers presenting on production AI.
Thank you,
CoreValue Research Team
