CoreWeave at ~$90: does the debt destroy the equity, or is the market underpricing compute scarcity?

I've been trying to value CoreWeave without falling into either of the two easy camps:

1. “AI compute demand is enormous, therefore bullish.”
2. “$35B+ of debt and negative earnings, therefore obvious short.”

I think both miss the interesting question.

CoreWeave may be an unusually leveraged bet on **the residual economic value of AI compute**.

Q2 numbers:

* Revenue: $2.575B, +112% YoY
* Revenue backlog: \~$104B
* Additional early-Q3 commitments not included in that backlog: >$25B
* Active power: 1.5 GW
* Q2 capex: \~$9.4B
* Q2 net loss: $626M
* Q2 interest expense: \~$640M
* Adjusted operating income: $128M
* FY26 revenue guidance: $12.4–13.2B
* FY26 capex guidance: $35–39B
* YE26 annualized revenue run-rate guidance: $18.5–19.5B

So yes: the balance sheet risk is enormous.

But here's what changed my thinking.

In August CoreWeave financed another $2.6B of infrastructure with debt maturing in roughly five years, while the underlying customer contracts average only around three years.

That means lenders are taking some renewal/re-leasing risk on the GPUs.

This is subtly different from simply lending against a long-term Meta/OpenAI contract.

If that financing model scales, the market is starting to treat GPU clusters as assets with financeable residual value.

That's potentially important.

Suppose an AI customer can generate $20M of economic value from a MW of compute while another customer generates $5M.

The first customer can rationally outbid the second.

If AI capability and monetization keep improving, compute becomes an auction.

CoreWeave's 1.5 GW of active capacity—and eventually its targeted 8+ GW—then isn't merely a pile of depreciating GPUs. It's an option on future compute scarcity.

The obvious problem is that creditors get paid before shareholders.

My rough 2031 scenarios from \~$90:

**Bear — 25%**

AI succeeds but compute commoditizes. Hyperscalers internalize capacity, GPU residual values fall quickly and CoreWeave gets stuck with mediocre economics.

\~$30B revenue, \~10% operating margin, \~$55B net debt.

Equity can plausibly be worth approximately zero.

**Base — 50%**

CoreWeave becomes a durable specialist AI infrastructure provider.

\~$65B revenue, 20% operating margin, \~$55B net debt, 18x operating income.

After allowing substantial dilution, I get roughly $255/share.

\~23% CAGR.

**Bull — 25%**

Inference explodes, AI agents and eventually robotics create persistent compute demand, financing gets cheaper, and CoreWeave captures software/orchestration economics above the hardware.

\~$120B revenue, 25% operating margin, \~$65B net debt, 22x operating income.

Even assuming \~750M diluted shares, I get roughly $790/share.

\~54% CAGR.

Probability-weighting those gives me expected terminal value around $325/share, approximately a 29% annualized return on expected terminal wealth.

Obviously the model is insanely sensitive to margin and debt assumptions. That's the point.

This is effectively leveraged equity on compute scarcity.

The key things I'm watching aren't GPU shipment headlines:

1. Does customer concentration continue falling?
2. Can adjusted operating margins get sustainably into the teens and eventually outrun interest expense?
3. Do GPU-backed financing spreads keep falling?
4. Are lenders increasingly willing to underwrite residual/re-leasing value?
5. What happens to revenue and operating profit per active MW as the fleet scales?

My current conclusion is that the expected value is attractive, but the permanent-loss branch is too real to size it like NVIDIA or another established AI-infrastructure compounder.

I see it as roughly a 0.5% starter rather than a core position.

The part I'm least confident about is terminal margin. If CoreWeave ultimately becomes a commodity infrastructure lessor, 20% operating margins may be fantasy. If its software layer + deployment speed + scarcity position lets it retain real economics, they may be conservative.

Curious how others are thinking about the residual value of the GPU fleet. Is there a good historical analogue here—or is this basically an entirely new infrastructure financing market?

*Disclosure: I write about these themes at* [K-Shaped](http://kshaped.substack.com)*; this post is adapted from a longer piece where I work through the thesis in more detail.*