CoreWeave has just printed the best quarter of its short public life: revenue up 112%, a $104bn backlog, and adjusted operating income back to $128m after a first-quarter trough. It also paid its lenders $640m in a quarter when it earned nothing at the operating line. This is not an argument about whether the business works. It is an argument about what is left for the shareholder after $51.6bn of debt and lease obligations have been served. Our band is $90–135 against $103.81.
The lenders earn multiples of what the shareholders do. Interest expense of $640m against an operating loss of $49m, and a third-quarter guide of $860–940m of interest against $200–260m of adjusted operating income. Even in the quarter management calls an inflection, the capital structure takes everything the operations produce and more.
Roughly 64% of this year’s operating profit is scheduled for the last three months. The company’s own guidance implies about $676m of adjusted operating income on ~$4.6bn of fourth-quarter revenue — a 14–15% margin, against 5% just delivered. That is a hockey stick, and it is the number to hold management to in November.
This valuation is a discount-rate argument, not a business argument. On our base-case cash flows, a 9% cost of capital produces $113 a share and 11% produces $36. With enterprise value near $86bn and net debt near $30bn, the equity is a thin residual that behaves like a long-dated call option on the enterprise.
CoreWeave reported after the close on 11 August. The $103.81 used throughout this report is the after-hours print that followed, 14.9% above the $90.32 regular-session close. Our normal convention is to price against the last completed session, and we are departing from it here for one stated reason: the closing price predates every fact this report is about, and pricing a post-earnings report against a pre-earnings quote would print a 24% “undervalued” gap that exists only until the opening bell. The trade-off is that an after-hours print is a thinner, less reliable price than a close. Both numbers are on this page so you can use either. Against the $90.32 close the same band reads undervalued; against $103.81 it reads fairly valued, and every multiple below scales directly with whichever you pick.
With that stated, the business argument is largely settled and it is not where the disagreement lives. CoreWeave is the largest pure-play AI cloud in the world. Revenue grew 112% to $2,575m, the backlog reached $104bn at quarter-end with more than $25bn signed since, contracted power stands at 4.2 GW, and the guidance went up on revenue, on profit and on capital expenditure. Companies in trouble do not raise capital expenditure by $4bn.
The argument is about the claim structure sitting between that business and a shareholder. Financial debt of $35.1bn and lease liabilities of $16.5bn stand against $5.0bn of book equity. The bar below is drawn to scale.
Three numbers follow from that picture and they define the whole investment case. Interest expense is 24.9% of revenue and is guided to $860–940m in the third quarter — more than three quarters of the entire full-year adjusted operating income guide of $960m to $1.15bn. Depreciation is 54% of revenue, so the celebrated 59% adjusted EBITDA margin becomes a 5% adjusted operating margin almost entirely through one line. And first-half free cash flow was roughly minus $10.5bn: $3.66bn of operating cash flow against $14.12bn of cash capital expenditure. Cash and restricted cash of $6.9bn is less than one quarter of guided capex. This company does not fund itself. The capital markets fund it, continuously, and the model only works while that conveyor is moving.
| Measure | Value | Note |
|---|---|---|
| Reference price, 11 Aug 2026 after hours | $103.81 | The price used throughout. The regular-session close was $90.32; Q2 was released after it |
| Market capitalisation | ~$56.7bn | On roughly 545.6m shares, up 75.8% year over year |
| Enterprise value | ~$86bn | Adding $35.1bn of debt, less $5.5bn of unrestricted cash. Roughly $103bn including lease liabilities |
| Book value per share | ~$9.20 | $5.0bn of equity on $77.1bn of assets. The equity is 6.5% of the balance sheet |
| Debt / annualised adj. EBITDA | 5.8x | 8.5x including lease liabilities. Fitch rates the issuer BB− with a positive outlook |
| Current ratio | 0.46 | $6.2bn of recourse debt is current |
| 52-week range | $60.55–153.20 | 32% below the high, 72% above the low, on a stock listed in March 2025 at $40 |
| Short interest | 11.8% | Of shares outstanding; 20.7% of the float |
| Implied / realised volatility | ~92% / ~95% | Position sizing matters more than entry price on a security that moves like this |
| Consensus target, 43 analysts | ~$138 | 29 Buy ratings; range $61 to $155 |
Derived inputs. Enterprise value, net debt, book value per share, the leverage multiples and the interest and depreciation ratios are ours, computed from the Q2 2026 release and the condensed balance sheet at 30 June 2026. Net debt of $29.6bn is financial debt of $35.1bn less $5.5bn of unrestricted cash; $1.4bn of restricted cash is excluded. Lease liabilities are excluded from net debt in the model because rent already sits inside the EBITDA margin we discount — adding both would charge for the same obligation twice.
Terminal value is 83% of enterprise value, and net debt is a third of it. Both numbers are the risk in this model.
This is an unlevered discounted cash flow on the enterprise, with net debt of $29.6bn deducted to reach the equity. The base year is the midpoint of company guidance, $12.8bn of 2026 revenue. Growth is expressed as a five-year compound rate to 2031, and the terminal margin is the adjusted EBITDA margin the fleet earns once it is built.
Move the discount rate first. It does more work here than every operating assumption combined, because the equity sits behind $51.6bn of obligations: a 200 basis point change in the cost of capital is worth more than doubling or halving the terminal margin. That is not a modelling artefact. It is what leverage does.
Reading the inputs. The widget is a constant-rate five-year form, and CoreWeave’s real cash flows are nothing like constant: capital expenditure runs at roughly 290% of revenue this year and falls toward 33% by 2031. The capex input is therefore a constant-rate equivalent, deliberately set well above the terminal maintenance rate of about 35% so that the five-year stream absorbs a build the widget cannot draw. The full, fading path is printed in Section 03 and it is what the published band is built on. Cash tax is held at 10% for loss carryforwards, net debt at $29.6bn, and dilution is a slider rather than a constant because the share count rose 75.8% over the trailing year. The calibration is anchored at a 9% discount rate, where the widget and the full-path model agree within a few dollars; away from 9% the widget is the more generous of the two, and where they disagree the full-path table governs.
The number above is not a fact about CoreWeave. It is what our assumptions produce, and on this name the single most consequential assumption is not in the sliders at all: it is the six-year depreciation life the accounts use for GPUs. Shorten it to four years and depreciation rises by roughly half, replacement capital expenditure rises with it, and the terminal cash flow this model capitalises largely disappears. That is why the capex and margin inputs are exposed and why we say to move them. Change the inputs until they are yours rather than ours and use the number that falls out.
Same model, different inputs. Click any card to load it. The spread is extreme — the bull is ten times the bear — and on a business where the equity is a thin residual on a very large asset base, that is the honest shape of the distribution rather than a failure of analysis. Small changes in enterprise value swing the equity violently in both directions.
Weighting those 30 / 45 / 25 at a common 9.5% discount rate and 3% terminal growth gives $117. That weighted case cannot be loaded as a preset — the model supports three — so it is stated here instead. Note that it sits above the base case rather than below it, because the distribution is skewed right: the bull adds more than the bear subtracts. On a levered equity that asymmetry is real, and it is also exactly what a long-dated call option looks like.
The widget in Section 02 holds capital expenditure at a constant share of revenue. The real path does not, so here it is in full. This is the base case, in billions, and it is the version the published band comes from.
| Base case, $bn | 2026E | 2027E | 2028E | 2029E | 2030E | 2031E |
|---|---|---|---|---|---|---|
| Revenue | 12.8 | 26.0 | 38.0 | 46.0 | 52.0 | 57.0 |
| Adjusted EBITDA | 7.3 | 15.1 | 21.7 | 26.2 | 29.6 | 32.5 |
| Capital expenditure | (37.0) | (30.0) | (26.0) | (22.0) | (20.0) | (19.0) |
| Cash tax | — | — | — | (1.0) | (1.8) | (2.5) |
| Free cash flow to the firm | (29.7) | (14.9) | (4.3) | 3.2 | 7.8 | 11.0 |
| Capex as % of revenue | 289% | 115% | 68% | 48% | 38% | 33% |
Free cash flow does not turn positive until 2029 on these assumptions. Everything before that is funded by lenders and, at the margin, by new equity. One convention matters more than it looks: we discount from the 30 June 2026 balance-sheet date and count only the cash flow still ahead, so the 2026 column enters the model at roughly minus $19.2bn rather than minus $29.7bn. The first half’s $10.5bn of burn has already happened and is inside the $29.6bn of net debt we subtract; charging it again would double-count it, and would understate fair value by roughly $19 a share.
Base-case cash flows throughout, on 546m shares and $29.6bn of net debt. Read this table before anything else on the page. Moving the discount rate from 9% to 11% takes the equity from $124 to $53. Another 100 basis points roughly halves it again. That is not a modelling artefact; it is what happens when a $30bn net debt claim sits in front of an $86bn enterprise.
| WACC ↓ / Terminal g → | 2.0% | 2.5% | 3.0% | 3.5% | 4.0% |
|---|---|---|---|---|---|
| 8.5% | $110 | $128 | $149 | $175 | $206 |
| 9.0% | $91 | $106 | $124 | $144 | $169 |
| 9.5% | $74 | $87 | $102 | $119 | $139 |
| 10.0% | $60 | $71 | $83 | $98 | $114 |
| 11.0% | $36 | $44 | $53 | $64 | $76 |
| 12.0% | $17 | $23 | $30 | $38 | $47 |
| 13.0% | $2 | $7 | $12 | $19 | $25 |
More useful than arguing about what the shares are worth is inverting the model, because the answer then depends on no view of the company at all. It simply states what has to be true for a buyer at this price to earn a normal return.
Hold the base-case cash flows and solve for the discount rate: the price implies roughly 9.45%. That is close to the cost of capital of the enterprise, and it is defensible — the take-or-pay structure and the investment-grade rating on the asset-backed paper genuinely lower the risk of the assets. It is much harder to defend as the required return on a residual equity claim sitting behind $51.6bn of debt and lease obligations, in a stock with 95% realised volatility. Now do it the other way: hold an 11% discount rate — a plausible cost of equity for a claim like this — and the price requires revenue to compound at about 43% a year to roughly $76bn by 2031. That is not the base case. It is the bull case, and at today’s price it is what you are underwriting unless you accept the lender’s discount rate as your own.
| Method | Assumption | Per share |
|---|---|---|
| DCF, full path | Base case at a 9% WACC and 3% terminal growth — roughly the market’s implied rate | $124 |
| Same at 10% | $83 | |
| Same at 11% | $53 | |
| DCF, constant-rate form | The Section 02 widget on its base preset, 9% WACC | $113 |
| Exit multiple on 2028E EBITDA | 8x adjusted EBITDA of $21.7bn less ~$69bn of 2028 net debt, discounted back at 14% on ~600m shares | $127 |
| Same, at 9x | $153 | |
| Probability-weighted | 30 / 45 / 25 at a common 9.5% discount rate | $117 |
| Street | 29 Buy ratings of 43 analysts | $61–155 |
At $103.81 the shares sit inside that band, a little below the point estimate. We have deliberately set the band below the methods that assume the market’s own discount rate. The band spans a full-path cost of capital of about 9.8% at the bottom and 8.8% at the top; the point estimate is roughly 9.3%. That is a judgement rather than arithmetic and it deserves stating plainly: an A3-rated delayed-draw facility is a claim on a bankruptcy-remote vehicle holding a specific contract with a specific offtaker, and its pricing says very little about what the common stock behind it should yield. The honest way to state the disagreement is not that the model is wrong, but that today’s price asks you to discount the equity at the enterprise’s cost of capital. If you accept that, the shares are cheap. If you think a residual claim behind fifty billion dollars of obligations deserves a higher hurdle, they are not.
| Metric | CRWV | Nebius | IREN |
|---|---|---|---|
| Market capitalisation | ~$56.7bn | ~$46bn | ~$14bn |
| Enterprise value | ~$86bn | ~$47bn | ~$13–15bn |
| 2026E revenue | $12.4–13.2bn | $3.0–3.4bn | ~$0.75bn TTM |
| Latest quarterly revenue growth | +112% | +684% | ~flat |
| Exit-2026 ARR target | $18.5–19.5bn | $7–9bn | >$4bn |
| EV / 2026E revenue | 6.7x | ~14x | ~18x |
| EV / exit-2026 ARR | 4.5x | ~5.9x | ~3.5x |
| Adjusted EBITDA margin | 59% | ~30–40% target | ~41% |
| Free cash flow, trailing | −$13.7bn | negative | negative |
| Debt / equity | 10.3x | modest | ~1.5x |
| Short interest | 11.8% | high | 26% |
On the only multiple that normalises for where each company is in its build — enterprise value against exit-2026 annualised revenue — CoreWeave trades at 4.5x against Nebius near 5.9x, while being roughly two and a half times larger, further along the margin curve and the only one of the three with an investment-grade-rated financing platform. IREN is cheaper still at about 3.5x, which is the market correctly charging for a messier profit and loss account and a mining legacy. Against its own history the stock has already de-rated hard: 6.7x 2026 revenue against 12.2x on 2025 revenue, with the shares down roughly 30% over twelve months while revenue more than doubled. Against the hyperscalers it is expensive on quality-adjusted terms, because Microsoft’s cloud earns similar gross economics without depending on leverage, a depreciation estimate and renewal pricing.
Peer market capitalisations, enterprise values and multiples are derived from published share counts and prices at 11–12 August 2026 and are indicative rather than a like-for-like screen. Nebius reports second-quarter results on 12 August 2026, so its figures may move the day this publishes. CoreWeave’s own figures are computed from the Q2 2026 release.
| Q2 2026 | Result | Y/Y | Comment |
|---|---|---|---|
| Revenue | $2,575m | +112% | Up 24% sequentially; beat the $2.56bn consensus. Fifth consecutive revenue beat since listing |
| Cost of revenue | $879m | +181% | Growing faster than revenue as newer, costlier sites come online |
| Technology & infrastructure | $1,507m | +125% | Includes depreciation and amortisation of $1,393m |
| Operating income | −$49m | n/m | Against $19m of operating income a year ago. On a GAAP basis the business did not earn anything |
| Interest expense, net | $640m | +140% | 24.9% of revenue, and rising faster than revenue. The core bear metric |
| Net loss | −$626m | n/m | Improved from −$740m in Q1. Diluted loss per share $1.14 |
| Adjusted EPS | −$1.03 | — | Beat the −$1.20 consensus — only the second earnings beat in six quarters |
| Adjusted EBITDA | $1,510m | +100% | A 59% margin, up 3 points sequentially but still 3 points below Q2 2025 |
| Adjusted operating income | $128m | −36% | Up sixfold from $21m in Q1. This is the line the inflection story rests on, and it moved |
| Operating cash flow | $679m | n/m | Against −$251m a year ago; collections improved materially |
| Cash capital expenditure | $6,422m | +162% | $14.1bn across the first half, against $3.66bn of operating cash flow |
| Revenue backlog | $104bn | +246% | Up $4.6bn sequentially, and excludes more than $25bn signed in early Q3 |
| Active power | 1.5 GW | — | Roughly 500 MW added in the quarter, about 300 MW of it in June and therefore barely monetised |
Full-year revenue went to $12.4–13.2bn, full-year adjusted operating income to $960m–$1.15bn, and exit annualised revenue to $18.5–19.5bn. Capital expenditure went up too, from $31–35bn to $35–39bn, and third-quarter interest expense is guided to $860–940m. The company is choosing to grow faster and to pay more for it, which is a defensible choice with a $104bn backlog behind it and an unforgiving one if demand blinks.
The arithmetic is simple and nobody put it to management. Full-year midpoint revenue of $12.8bn less the $4.65bn already delivered in the first half less the $3.53bn third-quarter midpoint leaves roughly $4.62bn for the fourth quarter. Full-year adjusted operating income of $1.055bn at the midpoint, less $149m delivered and $230m guided, leaves roughly $676m. That is a 14 to 15% adjusted operating margin in a quarter that begins in seven weeks, against 5% just delivered and about 6.5% guided for the third. It holds at both ends of the guidance range. Roughly 64% of the year’s operating profit is scheduled to arrive in the final three months. The 300 MW that landed in June is the reason to think it can happen. It is still a hockey stick, and if it slips a quarter the guide breaks.
Management was confident and occasionally impatient. Michael Intrator’s rhetorical mode is to zoom out; the useful work on this call was done by CFO Nitin Agrawal, who clarified the exit-ARR raise and walked through the June-weighted capacity additions. The genuine positive surprise was pricing: newly signed deals are carrying contribution margins 5 to 10 percentage points above recent quarters, attributed to platform quality and to Vera Rubin. In a market everyone describes as commoditising, prices going up is real evidence rather than spin, and if it is durable it is the most important disclosure on the call. Managed inference went from $1m to more than $100m of booked annualised revenue in a single quarter, with a $250m exit target.
The negative surprises were capital expenditure up $4bn at the midpoint, interest expense guided up 41% sequentially, and an adjusted EBITDA margin still three points below last year. Analysts pressed on renewals as the 2023–24 fleet rolls off — management’s answer was that only a limited portion of the fleet is affected, that A100s from 2020 are contracted out to 2029, and that realised prices are at or above year-ago levels — and on whether data-centre moratoria threaten the 3 GW and 8 GW targets. Management framed regulatory risk as affecting where it builds rather than whether. That is true so far, and it is also exactly what you would expect management to say.
It buys NVIDIA accelerators — Hopper, Blackwell GB200 and GB300, and now Vera Rubin — installs them in leased or partner-operated data centres, wraps them in its own orchestration stack and sells the capacity under multi-year take-or-pay contracts to AI labs, hyperscalers and enterprises. Take-or-pay is the whole financial trick: the customer pays for reserved capacity whether or not it uses it, which converts a semiconductor purchase into a financeable, bond-like revenue stream. That is what lets a company with $5bn of book equity borrow $35bn.
Founded in 2017 as Atlantic Crypto, an Ethereum mining operation, it pivoted to AI cloud and listed in March 2025 at $40 a share. Revenue went from $229m in 2023 to $1.9bn in 2024, $5.1bn in 2025 and $12.4–13.2bn guided for 2026. Committed compute has historically been about 96% of revenue; managed inference, the platform layer (SUNK, Mission Control, Tensorizer, Weights & Biases) and the new cross-cloud Interconnect are small but strategically load-bearing, because they are what stop the business being a pure landlord.
Power, not silicon, is the binding constraint, and contracted power is the real asset. CoreWeave holds roughly 4.2 GW contracted as at 11 August plus about 1.5 GW of powered-land options, against 1.5 GW active today, a year-end target above 1.85 GW, more than 3 GW by end-2027 and 8 GW by 2030. Two structural pressures run against all of that. Customers are becoming competitors: Meta has explored selling cloud capacity, SpaceX has begun selling excess compute, and Microsoft — historically the largest customer at 67% of 2025 revenue — keeps building its own. And community opposition and moratoria are a live constraint on where the next gigawatt lands.
Concentration is improving fast but remains the second-largest risk on the page. Microsoft was 67% of 2025 revenue and the top two customers were 65% of first-quarter 2026 revenue. Meta has committed roughly $35bn in total with deliveries starting in 2027, OpenAI about $22.4bn and roughly a third of the forward book, Anthropic an undisclosed multi-year agreement and Jane Street $6bn, alongside Bentley, Caterpillar, Grammarly, Isomorphic Labs, Databricks, Hudson River Trading, Leidos and Flow Traders. Falling concentration among customers funded by the same capital cycle is not diversification in the way the word usually means. It is correlation wearing a different name, and it is why the OpenAI funding question in Section 09 matters more than its size in the revenue mix suggests.
| Dimension | Score | Assessment |
|---|---|---|
| Scale | 7/10 | 1.5 GW active, 4.2 GW contracted, more than 43 data centres. Scale buys NVIDIA allocation priority and access to the syndicated loan market. The strongest single pillar |
| Cost advantage | 6/10 | Purpose-built density, high utilisation, and critically a lower cost of capital than any other neocloud through SPV-financed, investment-grade-rated GPU paper. Undermined by the fact that NVIDIA sets everyone’s input cost |
| Brand | 6/10 | Nine of the ten leading foundation-model providers are customers; Gartner named it a Visionary in the 2026 Cloud AI Infrastructure quadrant. Brand matters when a lab is betting a training run on you |
| IP / technology | 6/10 | SUNK, Mission Control, Tensorizer, LOTA, ARIA, and the first Vera Rubin NVL72 bring-up. Real engineering, but a lead measured in quarters rather than years |
| Switching costs | 5/10 | High inside a contract through take-or-pay, tooling and data gravity. Low at renewal, which is when it counts — and CoreWeave’s own Zero Egress Migration cuts both ways |
| Distribution | 5/10 | Direct enterprise sales plus a Microsoft sub-leasing channel, with a first chief revenue officer appointed in 2026. Building, not built |
| Customer loyalty | 5/10 | Renewals reported at roughly 95% of original pricing on rebooked H100s, and 2020-vintage A100s contracted through 2029. But Microsoft declined the next contract without cancelling the last, and Meta is building its own |
| Network effects | 3/10 | Weak. Compute is not a network good. Modest ecosystem effects through SUNK, Interconnect and Weights & Biases telemetry |
| Regulatory barriers | 3/10 | Permitting queues and interconnect backlogs favour whoever already holds power contracts — an accidental moat. It is also a risk, in the form of moratoria |
| Data advantage | 3/10 | None of consequence. Operational telemetry only; customer workloads are not CoreWeave’s asset |
The three genuine advantages are contracted power at scale, NVIDIA allocation priority, and a financing template — six delayed-draw SPV facilities in under two years, now investment-grade rated — that no other neocloud has replicated. The threats are customers in-sourcing, the renewal cliff on 2023–24 vintage hardware, and a rate or credit-spread shock that removes the capital-cost advantage overnight. Direction of travel is mixed: expanding on scale, capital access and software, shrinking on competitive intensity. The durable moat here is a balance sheet and a power book, not a technology — which is an uncomfortable thing for a technology multiple to rest on, and it is the reason the discount rate in Section 03 does so much work.
We ran a deliberately hostile read of the release, the balance sheet and the insider filings. Severity is scored one to ten, where one is immaterial and ten is severe. Most of what follows is a consequence of the business model rather than evidence of bad faith, and the quality of disclosure is adequate. The quantity of risk is not, for most portfolios.
| Concern | Sev. | Evidence, and why it matters |
|---|---|---|
| Leverage | 9/10 | $35.1bn of financial debt plus $16.5bn of lease liabilities against $5.0bn of book equity. Debt to annualised adjusted EBITDA of 5.8x, or 8.5x including leases. Current ratio 0.46, with $6.2bn of recourse debt current. The equity is a thin residual claim on a very large asset base, and small changes in asset value swing it violently |
| Free cash burn and refinancing dependence | 9/10 | First-half free cash flow of roughly −$10.5bn, trailing twelve months near −$13.7bn, and a full-year figure that will approach −$28bn on current guidance. $16.7bn of debt raised and $3.0bn of equity placed in the first half alone. Solvency depends on capital markets staying open continuously. A closed window is an existential event here, not a bad quarter |
| Interest expense against operating profit | 8/10 | $640m of interest against a $49m operating loss, with third-quarter interest guided to $860–940m against $200–260m of adjusted operating income. Even in the inflection quarter the lenders earn multiples of what the shareholders do, and GAAP profitability is years away |
| Depreciation elasticity | 8/10 | Six-year straight-line on GPUs, extended from four years in 2023. Peers including Nebius and Lambda use shorter lives. A change in accounting estimate — entirely at management’s discretion — could erase reported profitability without a single customer leaving. At the current run-rate a move to four years adds roughly $2.8bn a year of depreciation, several times the entire adjusted operating income guide |
| Customer concentration | 7/10 | Microsoft roughly 67% of 2025 revenue; the top two 65% of Q1 2026; OpenAI about a third of the multi-year contracted book and not investment grade. Improving quickly, but one renegotiation still resets the model, and OpenAI’s own funding is a tail risk nobody at CoreWeave controls |
| Adjusted metrics flatter | 6/10 | Adjusted EBITDA of $1.51bn excludes depreciation of $1.39bn and interest of $640m — the cost of the asset and the cost of the money, which are the two largest genuine costs of a financed-hardware business. A 59% EBITDA margin is close to meaningless here. Use adjusted operating income, or better, free cash flow |
| Dilution | 6/10 | Shares outstanding up 75.8% year over year to roughly 545.6m, $2.98bn of private placements in the first half, about $165m a quarter of stock compensation, and convertibles issued in Q2 partly hedged by $492m of capped calls. Per-share value creation is running well behind enterprise value creation |
| Circular ecosystem | 5/10 | NVIDIA is supplier, roughly 6–7% shareholder, customer and backstop, having agreed to take up to $6.3bn of unsold capacity through April 2032 and placed $2bn of stock in January 2026 at $87.20. Jane Street invested $1bn and committed $6bn; OpenAI invested $350m. Related-party revenue is real revenue, but demand signals from inside a closed loop are lower-quality evidence of end demand |
| Backlog definition | 5/10 | “Revenue backlog” is remaining performance obligations plus other amounts management estimates will be recognised, subject to delivery and availability. It is not RPO. It is partly a management estimate and partly delivery-contingent, and should be discounted accordingly |
| Persistent insider selling | 5/10 | The chief executive sold roughly 985,000 shares across three tranches in July and early August 2026, at about $83, $81 and $92. Venturo entities, McBee, the CFO and a director are all net sellers. All under 10b5-1 plans adopted in November 2025, so this says nothing about the quarter — but it is a metronomic supply of stock and an alignment question |
| Governance | 5/10 | Multi-class A/B/C share structure, a founder-chairman who is also chief executive, and short average board tenure. Offsetting that, Glenn Hutchins, Meg Whitman and Karen Boone are genuinely strong directors. Limited shareholder ability to force a change of course if capital allocation goes wrong |
| Litigation | 4/10 | Masaitis v. CoreWeave, D.N.J. 2:26-cv-00355, class period 28 March 2025 to 15 December 2025, alleging concealment of reliance on a single third-party data-centre supplier and the resulting delivery shortfalls. Lead-plaintiff motions were filed in March 2026 and it is ongoing. Financially survivable; relevant to whether disclosure is complete |
| Receivables | 2/10 | Listed for completeness because it improved. Receivables of $2.54bn are about 90 days of sales, down sharply from roughly 185 days at 31 December 2025, and deferred revenue of $9.7bn means customers are prepaying. Watch for reversal rather than treat it as a concern today |
There is no accounting-fraud pattern here and we are not implying one. What there is: extreme leverage, single-vendor and single-customer dependence, and one discretionary accounting estimate — useful life — doing an enormous amount of work in the reported numbers. If you own this, the two things to watch are the depreciation policy footnote and the maturity ladder, in that order. Neither will appear in a headline.
| Area | Score | Assessment |
|---|---|---|
| CEO track recordMichael Intrator | 4/5 | Co-founder and chief executive since 2017, previously co-founder of a natural-gas hedge fund. Took a crypto-mining shell to a $12bn revenue run-rate in under four years. Non-technical, which shows in the emphasis — and the emphasis has been correct, because this business is won on power contracts and financing rather than on engineering |
| Insider ownership | 4/5 | Very high. Intrator holds roughly 10% directly plus substantial trust and derivative holdings, and Venturo entities hold more than 10m underlying shares. Alignment of exposure is not in doubt |
| Board quality | 4/5 | Genuinely strong for a company this young: Glenn Hutchins, Meg Whitman, Karen Boone. Offset by multi-class voting and a combined chair and chief executive |
| CFO credibilityNitin Agrawal | 3.5/5 | In post since February 2024, previously vice president of finance at Google Cloud. Clear and precise on calls. The financing innovation — six SPV delayed-draw facilities, an investment-grade rating on GPU-backed paper, a first public syndication — is a genuine achievement rather than a presentational one |
| Capital allocation | 3.5/5 | Disciplined in one important way: substantially all capital expenditure is tied to already-signed contracts, and the SPV structure matches debt to specific offtake. Aggressive in every other way, with the capex plan raised twice this year |
| Communication | 3.5/5 | Confident and articulate, occasionally dismissive of legitimate near-term questions. The Q4 margin bridge went unexamined on this call, which is a question the company should have been made to answer |
| Guidance accuracy | 3/5 | Revenue beaten in all five quarters since listing, by about 7% on average; earnings missed in four of the last five. The Q3 2025 delivery shortfall traced to a third-party developer. Read: they forecast demand well and cost and timing poorly, which is exactly the wrong way round for a company on this much leverage |
| Transparency | 3/5 | Adequate, not exemplary. “Revenue backlog” mixes remaining performance obligations with management estimates, the headline metrics lean heavily on adjusted EBITDA in a business where depreciation and interest are the main costs, and customer names are disclosed selectively |
| Acquisitions | 3/5 | Weights & Biases was a smart, on-thesis purchase. The $9bn all-stock Core Scientific acquisition was terminated in October 2025 when that company’s own shareholders voted it down — arguably a bullet dodged, but it was management’s plan |
| Compensation | 3/5 | The chief executive package is not egregious for the size of the company, but roughly $660m of annualised stock compensation is heavy against about $1.06bn of adjusted operating income |
| Insider selling | 2.5/5 | Persistent and large: roughly 985,000 shares sold by the chief executive across three tranches in July and August 2026 alone. Pre-arranged under plans adopted in November 2025, which removes the information-timing concern but not the message |
| Dilution management | 2/5 | Shares up 75.8% year over year, plus $2.98bn of placements and convertibles. Necessary given the capital expenditure plan, but shareholders pay for this growth twice — once in dilution and once in interest |
| Buybacks and capital returns | n/a | None, and none would be appropriate. Every dollar is needed for the build |
Do they act like long-term owners? On strategy, yes, and it is not a close call: the power book, the financing template and the software investment were all long-horizon decisions taken ahead of demand, and they have been right so far. On the share count, less so. Between 76% annual dilution and a metronomic insider sell programme, per-share outcomes rank below enterprise scale in the revealed preference ordering. That is a defensible choice in a land-grab, and it is worth knowing that it is the choice being made.
| Catalyst | Timing | What it tests |
|---|---|---|
| Nebius Q2 results | 12 Aug | A strong annualised-revenue bridge validates the whole neocloud complex; a miss de-rates all of it. Reported the day this publishes |
| NVIDIA Q2 FY27 results | Late Aug | A data-centre beat lifts the complex. Any hint of order digestion goes straight through to CoreWeave’s renewal assumptions |
| Q3 2026 results | Early Nov | The highest-information event on the calendar. Adjusted operating income at or above $260m and a credible bridge to a 14–15% Q4 margin. Any slippage in that hockey stick breaks the full-year guide |
| OpenAI funding and compute commitments | Through 2027 | The largest single tail risk. Roughly a third of the forward book sits with a counterparty whose ability to fund future compute was publicly questioned in April 2026. New funding secures it; distress does not |
| Vera Rubin first revenue | Q3 2026 | Whether the claimed margin expansion “from day one” shows up in the numbers, or the deployment slips |
| Disclosure of the >$25bn of early-Q3 commitments | Q3 2026 | Whether the counterparties are investment-grade enterprises or more AI-lab credit. This is the concentration question in its purest form |
| Further debt raises | Ongoing | Tighter spreads are validation and directly raise fair value through the discount rate. A pulled or sweetened deal, as with DDTL 5.5, is the opposite signal |
| Possible Fitch upgrade from BB− | H2 2026 | Cheaper unsecured capital at the parent, where it matters most. An outlook cut to stable would be a material negative |
| Fed policy path | Sep / Dec | Roughly half the stack floats over SOFR, so 100 basis points is worth $175–350m of pre-tax income. It also moves the discount rate that decides this valuation |
| Year-end power above 1.85 GW | Dec 2026 | Whether the third-party developer delays that caused the 2025 shortfall are behind the company |
| Managed inference at $250m ARR | Dec 2026 | Validates the off-contract cascade, which is management’s substantive answer to the depreciation bears |
| Q4 and FY2026 results, 2027 guidance | Late Feb 2027 | Exit annualised revenue at or above $19bn confirmed, and a 2027 guide at or above $26bn. Below that, the base case in Section 03 needs rebuilding |
| Microsoft renewals and roll-offs | Through 2027 | Renewal at 90% or more of original pricing settles the depreciation-life argument better than any disclosure would |
| Continued 10b5-1 insider sales | Ongoing | Low information content, steady supply of stock |
No dividend and no buyback is plausible in this window, and index-inclusion flow has already happened — the shares joined the Nasdaq-100 in June 2026. S&P 500 eligibility requires GAAP profitability and is not in view.
BULLStart with what actually happened last night. Revenue up 112%, a $104bn backlog with $25bn more already signed this quarter, adjusted operating income up sixfold sequentially, and guidance raised on revenue, on profit and on capital expenditure. Companies in trouble do not raise capex by $4bn. And the pricing comment is the tell: new deals carrying contribution margins five to ten points above recent quarters. In a supposedly commoditising market, prices are going up.
BEARPrices are going up because supply is short, and the industry is spending hundreds of billions a year to fix exactly that. You are extrapolating scarcity rents. Meanwhile look at the actual profit and loss account: operating income was minus $49m and interest expense was $640m. This company earned nothing and paid its lenders $640m for the privilege.
BULLThat is what building infrastructure ahead of demand looks like. Nobody valued Equinix on GAAP earnings in 2002. The right lens is adjusted EBITDA of $1.51bn at a 59% margin, against take-or-pay revenue contracted eight to ten years forward.
BEARAdjusted EBITDA is precisely the wrong lens here, and you know it. It excludes depreciation, which is 54% of revenue, and interest, which is 25%. Those are not accounting abstractions. They are the cost of the asset and the cost of the money, which is the entire business. Strip out the acronym and CoreWeave converted $2.575bn of revenue into $128m of adjusted operating profit. Five percent. Vertiv earns more than that selling cooling equipment.
BULLQ2 was the second quarter after the trough, with 300 of the 500 new megawatts landing in June and therefore barely monetised. Q3 guides to $200–260m and the full year to $960m–$1.15bn. The operating leverage is arriving on schedule.
BEARThen do that arithmetic in public. Full-year midpoint, less the first half, less the Q3 midpoint, leaves roughly $676m of adjusted operating income in Q4 on about $4.6bn of revenue. A 14 to 15% margin in a quarter that starts in seven weeks, against 5% just delivered. Two-thirds of the year’s profit is scheduled for the last three months. If that slips one quarter the guide breaks and the multiple goes with it.
BULLIt slips or it does not, and the June-weighted capacity is the reason to think it does not. Meanwhile the balance-sheet fear is stale. DDTL 4.0 was rated A3 — investment grade on GPU-backed paper, which had never been done — at SOFR plus 225 against SOFR plus 400 a year earlier. DDTL 5.5 closed oversubscribed. The market is lowering their cost of capital while you tell me they cannot fund themselves.
BEARDDTL 5.5 had to be sweetened before it cleared. And the rating attaches to a bankruptcy-remote vehicle holding one contract with one offtaker. It says nothing about the equity, which sits behind $51.6bn of debt and lease obligations against $5.0bn of book value. Free cash flow was minus $10.5bn in six months and cash on hand is less than one quarter of guided capex. That is not a balance sheet, it is a conveyor belt, and it only works while the belt is moving.
BULLIt is moving. Deferred revenue is $9.7bn and growing, so customers are prepaying. Days sales outstanding fell from about 185 to 90. Nine of the ten leading model providers are customers. NVIDIA has backstopped $6.3bn of unsold capacity through 2032.
BEARNVIDIA is the supplier, a 6% shareholder, a customer and the backstop. Jane Street invested a billion and then committed six. OpenAI invested $350m and is a third of the forward book. When the demand validation comes from inside the building, it is weaker evidence than it looks.
BULLEvery one of those is a real, signed, arm’s-length contract, and concentration is falling fast. Microsoft was 67% of 2025 revenue; the top two were 65% in Q1 with Meta, Anthropic, Jane Street, Caterpillar and Bentley layering in behind.
BEARFalling concentration where the replacement buyers are AI labs funded by the same capital cycle is not diversification. It is correlation wearing a different name.
BULLThen price it. At $103.81 you pay 4.5 times enterprise value to exit-2026 annualised revenue. Nebius is near 5.9 times with a quarter of the scale. The stock is down 30% over twelve months while revenue more than doubled. The de-rating has already happened.
BEARAnd the same cash flows at an 11% discount rate are worth $53. You only reach today’s price at about 9.45%, which is what you assume when you have already decided it works.
CHAIRThe bull wins the quarter and the bear wins the balance sheet, and neither of those is a rhetorical draw. On demand, pricing and execution the evidence is genuinely strong: five consecutive revenue beats, 4.2 GW contracted, an inference product that went from $1m to $100m of annualised revenue in three months, and — most persuasive — a falling cost of debt awarded by lenders who have done more diligence on the collateral than either side of this table. On the capital structure the bear’s point is unanswerable: the valuation is a function of the discount rate, and the discount rate is a function of whether you believe a six-year GPU life and an uninterrupted financing conveyor. Both are assumptions. The two sides are not disagreeing about facts. They are disagreeing about whose cost of capital applies to the residual claim. Our band spans that disagreement rather than pretending to resolve it, and the price sits inside it.
Three things, and nobody in the debate above knows any of them. Renewal pricing on the 2023–24 vintage fleet, which alone decides whether six-year depreciation is conservative or fictional. Whether OpenAI can fund its commitments. And whether the fourth quarter’s implied 15% adjusted operating margin arrives on time. Anyone quoting a single-point price target on this name — including the sell side at $138 — is expressing a view on the discount rate at least as much as on the business.
One. The customer-concentration footnote in the Q2 Form 10-Q — is the top customer now below 50%? Two. The debt maturity ladder and the floating-to-fixed split, which decide how much of Section 03’s discount-rate risk is actually hedged. Three. The counterparties behind the more than $25bn of early-third-quarter commitments. Four. The Q4 margin bridge in management’s own words on the November call. Five. Any change to the useful-life estimate in the accounting policy note, which would move reported profitability more than any operating development on this list.
AI companies need enormous amounts of specialised computing power, and the equipment is so expensive that very few of them can afford to own it. CoreWeave buys that equipment — vast quantities of NVIDIA chips, the ones everybody is fighting over — puts it in buildings with a great deal of electricity and cooling, and rents it out by the year. It is a landlord for AI computing, and its tenants are OpenAI, Meta, Microsoft, Anthropic and a growing list of ordinary large companies.
The rental contracts run three to five years and are “take or pay”, meaning the tenant pays whether or not it uses the capacity. That is a good arrangement, and it is why banks will lend against these contracts. The catch is the scale of the spending. To earn about $13bn of revenue this year, CoreWeave will spend $35–39bn buying and installing equipment. Roughly three dollars out for every dollar of annual rent coming in, and that money is borrowed.
Last quarter the company paid $640m in interest and made a small loss at the operating line. The lenders earned more than the business did. That is not a scandal — it is what the early years of an infrastructure build look like — but it means the shareholder is last in a very long queue. CoreWeave owes about $35bn, and closer to $52bn once you count the long-term leases on its buildings, against roughly $5bn of book value.
Right: the AI build-out continues, CoreWeave keeps filling buildings, and around 2029 the spending finally slows while the rent keeps arriving. At that point a very large amount of cash starts falling out of the business. That is the bull case and it is not fantasy — the contracts are real and the customers are the biggest names in technology. Wrong: three things, roughly in order. The debt, because the company survives by borrowing more, constantly, and if lenders get nervous this gets ugly quickly. The chips might not last as long as the accounts assume — six years is the assumption, and plenty of informed people think four to five is more realistic, which would mean today’s reported profit is partly an illusion. And the customers are building their own: Microsoft was two-thirds of revenue last year and now builds its own capacity.
One closing thought, and it is the most useful habit in this whole report. The stock has traded between $60.55 and $153.20 in the past year and moves around 95% a year in volatility terms. On a company where the equity is a thin slice on top of fifty billion dollars of obligations, position size is the entire decision, not entry price. If a 60% fall in a bad quarter for AI sentiment would be intolerable to you, the answer is not to wait for a better price. It is to own less of it, or none.
| Status | What it covers |
|---|---|
| Verified from CoreWeave | All Q2 2026, Q1 2026 and Q2 2025 income statement, balance sheet, cash flow and non-GAAP figures; the 30 June 2026 balance sheet including debt, leases, cash, deferred revenue and equity; revenue backlog, active and contracted power, and data-centre count; third-quarter and full-year guidance including revenue, adjusted operating income, exit annualised revenue, capital expenditure and interest expense; management and analyst commentary from the 11 August call, including the contribution-margin improvement on new contracts, the June-weighted capacity additions and the renewal commentary; the financing terms of DDTL 4.0 and the closing of DDTL 5.5; Form 4 insider filings from June to August 2026. |
| Our estimate or judgement | Enterprise value, net debt, book value per share, leverage multiples and every ratio in Section 01; the entire cash flow path in Section 03, the sensitivity grid, both reverse tests, the exit-multiple calculation and the 2028 net debt figure inside it; the $90–135 band and the $112 point estimate, including the explicit decision to set the band below the market’s own implied discount rate; the constant-rate calibration of the Section 02 widget and its capex equivalent; all scores in Sections 06, 07 and 08; and the reading of customer diversification as correlated rather than genuine. |
| Still open | Q2 2026 figures are drawn from the earnings release, the accompanying financial statements and the earnings call, and will be reconciled against the Form 10-Q when it files. Two capital expenditure measures circulate for the quarter — cash purchases of property and equipment of $6.42bn, and a broader company measure of about $9.4bn including non-cash additions. We use the cash figure throughout and will reconcile the difference at the 10-Q. Peer figures for Nebius and IREN are derived from published share counts and prices rather than company-reported multiples. The 4.2 GW contracted power figure is as at 11 August and post-dates the 3.7 GW reported at quarter-end. |
Method. Unlevered discounted cash flow, methodology version 1.2. Cash flows are discounted from the 30 June 2026 balance-sheet date, counting only the cash flow still ahead, against net debt of $29.6bn measured at that date; lease liabilities are excluded from net debt because rent is already charged inside the EBITDA margin being discounted. Share counts are 545.6m outstanding today, with terminal counts of 540m to 620m across the scenarios. The reference price used throughout is the $103.81 after-hours print of 11 August 2026, following the $90.32 regular-session close, and every multiple in this report scales directly with it.