MacrosLM Finance Knowledge
QUALITY OF EARNINGS  /  AI INFRA  /  CASE: CRWV
CASE STUDY · QUALITY OF EARNINGS · Q1 2026 10-Q · 12 MIN READ

Reading CoreWeave's earnings: where the GAAP loss meets normalized EBITDA

CoreWeave reported a $144M operating loss on $2.08B of revenue in Q1 2026. The GAAP number is hard to act on — it lives in a swirl of $7.7B of capex, $9.8B of lease liabilities, and $536M of quarterly interest. A QoE doesn't argue with GAAP. It builds the bridge from reported figures to a normalized run-rate a buyer, lender, or rating agency can underwrite against.

§ 01The bridge from operating loss to normalized EBITDA

Every QoE starts in the same place: reconcile GAAP to reported EBITDA, then layer normalizations. CoreWeave's bridge is unusual because depreciation alone is larger than the operating loss — a feature of GPU-heavy infrastructure where economic useful life is itself the debate.
CRWV Q1 2026 · QoE bridge ($M)
Line $M Notes
Revenue 2,078 10-Q income statement
Cost of revenue (1,484) incl. ~$680M GPU D&A
Operating expenses (738) SG&A + R&D
Operating loss (GAAP) (144) starting point
+ Depreciation & amortization 847 GPU + facility
+ Stock-based compensation 153 non-cash
Reported EBITDA 856 41.2% margin
QoE adjustments — typical provider ranges
− One-time IPO/listing costs (38) non-recurring
− Customer concentration premium (95) top-2 customers >60% of revenue
− GPU residual-value haircut (180) D&A understated vs economic life
− Power capacity ramp costs (48) pre-revenue MW build
+ Run-rate scaling (annualized new contracts) 62 illustrative
+ Tax shield from lease structure 48 recurring
Adjusted EBITDA · normalized 605 29.1% margin
Bridge in one number
$605M
Normalized

From $856M reported down to $605M after stripping non-recurring items, sizing concentration risk, and re-pricing depreciation against true GPU economic life. 29% margin, not 41%.

Reading the haircut

The $251M of net adjustments isn't an accounting argument — it's a view on durability. Concentration premium says: if the top 2 customers leave, revenue doesn't degrade linearly. GPU haircut says: 4-year straight-line books may understate true 3-year economic life given AI hardware refresh cycles.

CoreWeave snapshot

CRWV · $109.50MC $52.7B

Cash $2.24B · Capex $7.70B · Lease liab. $9.78B · Interest exp. $536M/Q

§ 02Where the EBITDA actually comes from

CoreWeave doesn't publish formal segments, but its disclosures and customer concentration commentary let us approximate the underlying mix. AI training compute dominates margin; inference and storage are dilutive.
AI training compute
~$1.45B

~70% of revenue · Adj. EBITDA margin ~38%. Long-term contracted GPU capacity, mostly hyperscaler-adjacent. The driver of every dollar of reported EBITDA.

AI inference
~$435M

~21% of revenue · Adj. EBITDA margin ~14%. Pay-as-you-go workloads, lower utilization, lower lock-in. Growing but dilutive to margin.

Storage & networking
~$195M

~9% of revenue · Adj. EBITDA margin ~22%. Attach to compute, capacity-priced. Useful for stickiness, not the EBITDA engine.

§ 03The depreciation argument, quantified

The single most consequential QoE input for CoreWeave is GPU economic life. The company books straight-line over a useful life that lies somewhere between 3 and 6 years — but the asset class has been around for less time than that, and the H100/H200/B100 transition is actively reshaping residual values.
Sensitivity · normalized EBITDA at varying GPU lives
GPU useful life D&A / yr Adj. EBITDA Margin
6 years (current book) 2,260 605 29.1%
5 years 2,712 492 23.7%
4 years (common QoE base) 3,390 323 15.6%
3 years (bear case) 4,520 40 1.9%
Reading the table

The choice of GPU life moves CoreWeave from a ~30% margin business to a ~2% margin business without changing a single contract, customer, or operating cost. This is why every buyer-side diligence stress-tests this assumption first.

The historical analogue isn't cloud — it's mainframe leasing in the 1980s, where 5-year vs 3-year lives created the gap between IBM Credit and the secondaries. Today the question is: does H100 hold residual value when B300 ships?

§ 04QoE adjustment magnitude vs peers

How much does CoreWeave's bridge actually take out of reported EBITDA, relative to comparable infrastructure businesses? Below: typical QoE adjustment as a percentage of reported EBITDA, observed across recent transactions and disclosures.
Adj. EBITDA / Reported EBITDA · indicative
Company Profile Rep. EBITDA Adj. / Rep. Driver
CoreWeave (CRWV) AI GPU infra · concentrated $3.4B '26E ~70% GPU life, concentration
Equinix (EQIX) Data center · diversified $4.2B ~94% long lives, low customer risk
Digital Realty (DLR) Data center · diversified $2.9B ~92% similar to EQIX
Iris Energy (IREN) GPU + bitcoin · transitional $190M ~62% asset mix, segment shift
Applied Digital (APLD) HPC + crypto hosting $45M ~48% single-customer concentration

§ 05What a buyer actually pays for

A QoE isn't a verdict — it's the input layer to a deal. Two debates set the multiple.
Debate 1 · GPU economic life

The 1.5x problem. At a 4-year GPU life, CoreWeave looks like a 16% margin business. At 6 years it's 29%. Buyers and sellers fight this one before they fight price. Lenders moonlight at the bottom of the range.

Debate 2 · Customer concentration

The 62% problem. With ~62% of revenue from two customers, every QoE applies a 10–20% haircut for concentration. The premium scales with contract length: 3-year customers get a haircut; 7-year customers get a discount.

SOURCE: CRWV Q1 2026 10-Q (period ended Mar 31, 2026); segment, sensitivity, and peer-comp figures illustrative of typical QoE methodology and provider-range observations. Not investment advice.MacrosLM · Finance Knowledge Series →