CRWV

As of 2026-10-19 · partly unblinded · revised twice after review (section 2)

Partly unblinded: the 10-Q (share-price information masked) still showed the April 2026 private placement at $109.00, two H1 2026 placements at about $86 and $111, the $40 IPO price, and the convertibles' implied conversion price of about $116 (assumptions.md lists each). The model was pushed blind at b8bdc78 ($28.41), before we read the price, the Street's numbers or our research note 2026-10-19-neocloud-financing-vs-demand. Both revisions in section 2 came after the market pass, from reported figures, the reviewers' findings and our chief investment officer's (CIO's) ruling on the cost lead (how many months before a megawatt goes live its costs begin). They moved fair value to $35.26. assumptions.md is left as committed, and section 2 amends it.

We rate CoreWeave Avoid for new money at $63.33. Our central fair value is $35.26 a share, 0.557 times the price. That is well under the 0.85 line ($53.83) between Avoid and Hold. Only one reading of the costs, the six-month reading that was our round-1 central case, reaches that line, and only just. The model can read CoreWeave's 2025-26 costs two ways, and the choice moves fair value from -$9.51 to $53.33 (section 2). The Q3 2026 results, due 2026-11-09, will show which is closer (section 6).

Terms used below.

• Bear, central and bull cases. The central case is our best estimate and sets fair value. The bear is a severe but plausible bad outcome, and the bull is the matching good outcome.
• Two estimates of the exit price. The fade lets the company's returns on new capital drift toward its cost of capital. The outside view applies the multiple that comparable companies trade at. The exit price averages the two.
• Round 1 and round 2. The two revisions after the blind model (section 2). Round 1 answered two reviews; round 2 applies the CIO's ruling on a third reviewer's finding.
• Avoid. Fair value under 0.85 times the price.

1. A megawatt of capacity earns 10.7% on its equipment, just above its funding cost

CoreWeave rents GPU clusters to a few very large buyers on take-or-pay contracts of about five years (the buyer pays whether or not it uses the capacity; 98% of revenue is committed, 15-25% of contract value typically prepaid; 10-Q 2026-06-30). It leases the buildings and power, buys the servers mostly with secured debt, and earns the spread between the rent and the cost of the server, the building, the power and the debt.

Reported, Q2 2026: revenue $2.575B, adjusted EBITDA (earnings before interest, tax, depreciation and amortization, as the company adjusts it) $1.510B (58.6%), GAAP operating income -$49M, capex (capital spending) $9.35B on the company's definition, active power (capacity live) 1.5 GW, contracted power (signed, including capacity not yet live) 4.2 GW, backlog $104.2B (earnings release, presentation of 2026-09-17). Customers A, B and C were 36%, 26% and 10% of the quarter's revenue (10-Q).

A megawatt of active power brings about $10.7M of revenue a year on its first contract, costs about $27M to equip (in-service gross PP&E, property, plant and equipment, per active MW at 2026-06-30 is $27.1M, 10-Q note 5) and, in the central case, $2.65M a year to run once live, plus the same rate for the four months it is being built. On a five-year contract it repays its equipment with a good return: after $4.27M of depreciation and interest at 9% on 70% of its capex, a first-contract slot earns about 19% of revenue, a little under management's "mid-20s" stabilized contribution margin (presentation, 2026-09-17). The question is its second life: servers have a six-year book life (10-K), so a slot either re-rents an older GPU for its last year or is re-equipped at full cost. With renewals at 55% of the first rate after 0.15 years idle, the steady state across all vintages earns 10.7% on net PP&E (output.txt), just above the 9.65% WACC (weighted average cost of capital, the blended cost of CoreWeave's debt and equity).

2. Two revisions took fair value from $28.41 blind to $53.33, then to $35.26

Round 1 answered two reviews; round 2 applies the CIO's ruling on a third reviewer's finding about the cost calibration. Each change is in model.py; the bridge is printed at the end of output.txt.

Round 1 had step 3 at a six-month lead and $2.33M ($53.33). Round 2 changes only that lead, six months to four, with the cost refit: $53.33 to $35.26 (-$18.07). The bull case already used the six-month fit and is unchanged.

Calibration. The blind model's EBITDA ran 2.7-7.3 points above reported, depreciation 11% low and 2026 EBIT at +$1.36B against H1 GAAP operating income of -$193M. All three are now calibrated together to reported figures, in calibrate():

• Capex per MW: no change; it reconciles. The model's H1 2026 capex of $18.23B is 2.4% under gross additions to PP&E of $18.68B (10-Q note 5), which include the $2.54B rise in construction in progress that the company's $16.14B leaves out. In-service gross PP&E per active MW at 2026-06-30 is $27.1M against the model's $27M.
• Depreciation: a timing gap, now closed. With the power path in monthly steps and depreciation from one month before activation, the model's in-service gross PP&E at 2026-06-30 is $40.5B against $40.7B reported. Q2 2026 depreciation is $1.422B against $1.393B (+2%), Q1 $1.013B against $1.147B (-12%).
• Cash cost: part build-phase, by judgment. Cost that starts some months before activation fits reported cost (revenue less adjusted EBITDA) better than cost from activation, and the leases support it: Q2 2026 lease cost of $667M (10-Q) is about $2.1M a year per average active MW, while the $35.5B of leases not yet commenced, on the 2.7 GW contracted but not active, is about $1.1M per MW-year, so leases start before the power is live. But no filing discloses a lead, and the fit is a spike (output.txt):

The fit has two free parameters on four quarters, and only Q3 and Q4 2025 test it. Reported cost per average active MW runs $3.41-4.15M (Q2 2025 to Q2 2026); the model's Q2 2026 is $3.79M including the cost of power being built. The CIO ruled the central case at four months for those reasons; the bull case takes the six-month fit.

Model against reported, central case:

H1 2026 GAAP operating income is -$0.098B in the model against -$0.193B reported. The 2025 revenue misses (6-12% low) are the older contracts' rate, which no longer drives value.

The other reading: fixed overhead (25% of the odds). The third reviewer fitted the same quarters as a fixed overhead plus a higher cost per live MW from activation. Refit by least squares in fit_overhead(): $0.096B a quarter plus $3.455M per live MW-year, within 5.5% in each quarter (Q3 2025 +1.0%, Q4 2025 +0.4%, Q1 2026 -5.5%, Q2 2026 +3.6%). Four quarters cannot reject it. It gives -$9.51 (nil). It replaces round 1's persistent-cost world in drivers.json at the same 25% weight, drawn as the central lead with $3.254M per MW-year, which reproduces its value.

Does each gap persist to 2031? Each removed alone, central case:

The build-phase cost costs $1.06B of EBIT in 2026 and only $0.41B in 2031, because the build slows from 1.0 GW a year to 0.5 GW. The fixed-overhead reading takes $4.82B off 2031 EBIT, because its higher cost per live MW never goes away.

Steady EBITDA margin. The central steady state earns a 70.3% EBITDA margin against the best reported adjusted margin of 62.1% (Q2 2025; Q2 2026 58.6%). It is above 65% because revenue per active MW-year is the same as in Q2 2026 ($8.93M against $8.79M), but Q2 2026 cost of $3.79M per active MW-year carries four months of cost on power still being built while the fleet grows by half or more a year; at 5% growth that share nearly vanishes and cost is $2.65M. The fixed-overhead world would not get there.

Second-half 2026 capex. The blind model added each year's power in one lump at mid-year, so nothing was spent in H2 2026 for 2027's power: $7.9B. Built month by month, H2 2026 capex is $13.50B, $6.75B a quarter, against Q2's cash capex of $6.42B (market data cash flow) and gross additions of $11.69B in Q2 alone. Full-year 2026 is $31.7B on the gross-additions basis, under the 2026 guide of $35-39B in the earlier scan (not verified; the call transcript is not in our sources).

Outside view: why 8.4x and not the peers' 14.9x. The peers' 14.9x EV/EBIT is 19.4x NOPAT. At CRWV's 9.65% WACC and 5% growth, that multiple needs a 35% return on new capital for ever on year-six NOPAT (about 50% on year-five NOPAT, the basis the reviewer used). The same peers earned a median 5.3% on capital over 2016-2025 (market data, output.txt: AER 5.2%, AL 2.8%, GATX 4.6%, URI 11.3%, HRI 5.3%, R 6.1%). Their multiples rest on investment-grade funding and on aircraft, railcars and rental fleets that last far longer than six-year servers, not on returns CRWV can copy. With no durable moat, the outside view is a lessor whose new capital earns exactly its cost of capital: year-six NOPAT capitalised at the WACC, (1 - 23%) x 1.05 / 9.65% = 8.4x year-five EBIT. It is recorded in Forecast.changes with its reason and source. At the peers' 14.9x with every other revision, fair value would be $66.71. The fade alone gives $42.79 today and the outside view alone about $27.7.

Cash. Starting cash excludes the $1.38B of restricted cash, which secures the special-purpose entities' debt (-$1.63). The $5.0B kept as operating cash is not counted as surplus at exit, and the $0.221B of finance-lease liabilities is added to the exit claims (-$4.37 together).

Leases. Operating leases ($16.3B of liabilities at 2026-06-30) are an operating cost: their $500M a quarter of lease cost and $150M of variable cost sit in cost of revenue, so they are in the cash cost per MW and not in the bridge. Treating them as debt as well would count them twice. The $35.5B of leases not yet commenced enter the same way, as the power they cover is built and goes live, through the four-month build-phase cost and then the cost per active MW; at about $1.1M per MW-year they are inside the $2.65M. Finance leases ($221M) are debt in the exit bridge.

First contract term. The RPO schedule (remaining performance obligations, the contracted revenue still to be recognised; 10-Q: 41% of $103.7B within 24 months, 39% in months 25-48, 20% in months 49-78) was fitted with the Q2 2026 revenue run-rate as the running book and the rest of the backlog starting over the following months. Terms of 4.5 to 5.25 years fit each bucket within about a point, best 4.75 years (40%, 40%, 21%) and 5.0 years next (42%, 38%, 20%). Central stays at 5.0, and drivers.json brackets the term at 4.5-5.25 years. At 4.75 years fair value is $23.38 (-$11.88), at 4.5 years $11.46.

Cost of equity. Three builds (market.txt):

We keep the coupon-plus-premium build. The peer default gives 10.04%, only 0.4 points over the coupon on CRWV's own unsecured notes and 5 points under their trading yield. That is too low for the most junior claim in a company whose debt is about 60% of its capital, while the peers are mostly investment-grade. CRWV's own beta comes from 19 months of trading through an IPO and a 60% fall, so it is noise. The exit multiple moves with the WACC, so each line above moves both the discounting and the exit.

The July price claim. Our research note of 2026-10-19 says CoreWeave raised prices "about 25% in July 2026". The presentation filed with the 8-K of 2026-09-17 reads "~70% price increase across SKUs in July 2026". The 25% beside it on the slide is the "higher expected contribution margin on Q2-signed contracts". We cite the 70% (a management claim); the note is left unedited.

3. Five questions: funding, cost of equity, parent guarantee, asset life and Q3 tests

These are the five questions this report answers. DDTL is a delayed-draw term loan, and SOFR is the benchmark short-term borrowing rate.

How long each rate lasts, as modelled. Existing debt keeps its own effective rate (8.41% blended; 10-Q) until it matures: $4.4B in the rest of 2026, $6.2B in 2027, $4.4B in 2028, $2.4B in 2029, $3.2B in 2030 and $14.9B later. The senior notes are fixed (9.0-9.75% USD, 8.5% EUR) to 2030-2032; the DDTLs float at SOFR plus their spread and amortize as their contracts pay (DDTL 5.5 to 2031-09-01); the convertibles cost 2.875% in cash to 2033 plus the shares they may become. Maturities and new money are refinanced at the case's marginal rate and held there to 2031. A trading yield reaches CoreWeave only when it borrows: 15% on the 2031 notes costs nothing on notes already issued, and costs the most if it lasts through the 2027-2028 refinancings ($10.6B) and the new money ($15.8B by 2027). The cost of new money matters little to the cash flows; it matters through the fade's cost of capital (-$20.54 at the trading yield).

4. The central case reaches $54.9B of revenue in 2031; the bear is worth nothing and the bull $227.84

model.py builds the fleet month by month from the reported power history and a path to 6.0 GW by 2031 (contracted 4.2 GW arrives by 2029, then 1.8 GW more). Central revenue: $12.7B in 2026, $24.6B in 2027, $35.1B in 2028, $44.3B in 2029, $51.5B in 2030, $54.9B in 2031; EBIT $0.92B in 2026 rising to $11.68B in 2031 (a 21.3% margin); capex $21-32B a year; net debt peaks near $61B in 2029 (output.txt).

The bear is a coherent world of 4 GW by 2031, rents 7% lower, renewals at 40% after 0.35 years idle, a 4.5-year first term and the top two customers leaving, on the central four-month cost lead: the steady state earns 0.5%, net debt reaches about $74B, and the equity is worth nothing. The bull runs 8 GW by 2031, rents at $11.5M rising 1% a year, renewals at 75% with no idle time, a 7-year life and the six-month cost fit: the convertibles convert (exit price $447 against a $116 conversion price), and fair value is $227.84.

5. Under 20% of our scenarios reach the price, and the Street's numbers do not rescue it

At $63.33 the price needs one of $11.2M per MW-year on new slots, cash cost of $2.28M, renewals at 91%, a 7.9-year life or capex of $24.9M per MW (odds.json). Under 20% of 5,000 draws reach the price (P 12%): 15% in the base world, 3% in the fixed-overhead world, 82% in the bull-leaning world and under 1% in the bear-leaning one.

Do we charge the same risk twice? A review check asks this. We checked the three places credit risk could enter. The cash flows pay expected rates on new money (8.92%), not the 15% trading yield, so the credit market's fear is in the discount rate only (the 4.20% premium over the coupon), and the trading-yield cases are sensitivities, not the central case. The 8.4x outside view assumes new capital earns its cost, with no added haircut for leverage; the leverage sits once, in the exit bridge's net debt. The fixed-overhead reading is a separate world, not stacked on the central case. The input the market most plausibly reads differently is the cost lead: at the six-month fit fair value is $53.33, and the price needs cash cost of $2.28M, a little under that fit's $2.33M.

The simulated median is $2, 95% under the central $35.26, which a review check asks us to explain. A quarter of the draws come from the fixed-overhead world, where the median is nil. In the base world the median is $11, because five distributions lean against the central value on purpose: idle time (mean 0.25 years against 0.15), capex per MW (mean $27.3M), server life (mean 5.8 years), first term (mean 4.92 years) and cash cost (mean $2.74M). Each costs a few dollars, and the value is steep and floored at zero, so many base draws land at nil.

The stock is near its 52-week low ($59.51; high $141.79), under its 50-day ($87.03) and 200-day ($92.44) averages (market data quote). Our research note dates the fall to a credit move: the 2031 notes at 81.80, yielding 15.08% (Benzinga, 2026-10-16), and CRWV down 15.5% on 2026-10-14 (press, via the note).

The Street through our model (street.py, street.txt). A review check requires this when fair value is far from the Street target. The Street's average revenue (market data, 13-31 analysts) is close to ours in 2026-27 and 69% above it in 2030; its EBIT is far below ours in every year to 2029 and $9.83B against our $12.29B in 2030.

The Street's revenue at our margins is worth $56.30, still under the price, and needs $43.0B of new money by 2027. The Street's revenue and EBIT together are worth nothing in our model. Its $141 target is about 8.4 times its 2030 EPS of $16.87 (6 analysts), so it comes from a multiple on 2030 earnings at its scale, not from a cash-flow value of its own estimates (our inference). Even the lowest target, $74, is above the price.

Alpha's dated view (alpha rank CRWV --raw --json --allow-stale-model-cache, read in place under scripts/readonly-data-run.sh, 2026-10-19; cache status partial, integrity verdict CLEAN with 3 warnings that were not reviewed). Rank 437 of 704, score 38.1, Below Average: growth A (rank 16), expectations B (166), valuation D (508), quality E (533), momentum E (686). Alpha and this report point the same way, away from the stock; why is our inference and was not tested.

6. The Q3 2026 EBITDA margin settles which cost reading is right

• Q3 2026 adjusted EBITDA margin (2026-11-09, market data date, unconfirmed) picks the cost world: 64.5% central, 66.1% at the six-month fit, 59.1% fixed overhead. Above 65.5% moves us toward the six-month fit ($53.33); 62% or less toward fixed overhead (nil); in between holds the central case.
• Q3 revenue $3.59B central (bear $3.09B, bull $4.10B); under $3.3B or active power under 1.65 GW says the power path is slower than central.
• Q3 capex $6.75B on the model's basis (gross additions, including construction in progress); near $10B says the path or capex per MW is low.
• Revenue per MW-year above $11.5M for two quarters supports the bull; under $10.0M the bear.
• Any disclosed renewal of a 2023-24 contract under 40% of its original rate is the bear's renewal world; at or above the original rate is the bull's.
• A new unsecured note or convertible at 12% or more, or a secured facility over SOFR + 6%, is the bear rate mix.
• Microsoft and Alphabet capex guides on 2026-10-28: two of the four largest buyers cutting 2026-27 capex moves us toward the bear.

7. The other side's case: a weighted value of $74.59, above the price

The case-weighted value is $74.59, 18% above the price, with the bear floored at zero as limited liability requires (the method's unfloored weighting gives $57.38). A buyer at $63.33 is paying for an option: lose the whole stake in the bear (25% weight), collect about $228 in the bull (25%). The bull is not exotic. It needs rents to hold at the July 2026 level, older GPUs to renew at 75% (management claims H200 renewed at a premium and A100 into 2029), seven-year lives, 8 GW by 2031 (below the CEO's 8 GW by 2030), and the cost fit that matches the last three quarters best. Management also claims short-dated contracts at about $40M per MW of annualized revenue in Q3. And at the peers' 14.9x outside view, with every other revision kept, fair value is $66.71, above the price.

Why we still say Avoid: the central case, not the weighted value, is the most likely world, and it is 44% under the price. The weighted value rests on a quarter weight on a bull that needs every driver to break the same way at once, and it gives no weight to the fixed-overhead reading, which four quarters cannot reject and which is worth nothing. The disagreement is large, dated and cheap to settle: the Q3 margin in three weeks.

8. Sources

• SEC filings, fetched 2026-10-19 with the $EDGAR_IDENTITY User-Agent and masked: 10-K 2025; 10-Qs for 2025-09-30, 2026-03-31 and 2026-06-30; earnings releases Q3 2025 to Q2 2026; 8-Ks of 2026-06-18 (2032 notes), 2026-08-10 (DDTL 5.5), 2026-09-17 (presentation) and 2026-09-22 (2033 notes). Saved in data/filings/.
• Market data statements (data/), the exit anchor (data/exit-anchor.json) and peer statements (data/peers/), fetched 2026-10-19.
• Market pass, 2026-10-19: market data quote, price-target consensus and summary, analyst estimates and profile betas (data/market/); our research note of 2026-10-19 for the bond yield (Benzinga, 2026-10-16, press).
• Reviews of 2026-10-19 (two in round 1, a third on the cost calibration) and the CIO's ruling on the cost lead, relayed by the CIO.
• Model: model.py, output.txt; market tests: market.py, market.txt; the Street: street.py, street.txt; drivers and odds: drivers.json, odds.json.