The AI landlord in the news

Today, I'm digging into CoreWeave (CRWV).

CoreWeave buys the GPUs, leases the data centers, wires up the power and networking, and sells the computing time to AI labs and big tech companies on multi-year contracts. Microsoft, OpenAI, Meta, and Anthropic are all customers.

They’ve been all over the financial news this month because their bonds fell to about 80 cents on the dollar, with yields near 15%, and the stock dropped 30% in two weeks. investors are nervous about whether customers like Anthropic will be able to fulfill the sizable revenue commitments that they've made with CoreWeave and others.

The company started in 2017 as a crypto mining operation in New Jersey. Nine years later it has a signed backlog of about $104 billion.

A few things that stood out to me:

• Revenue went from $16M in 2022 to $7.6B over the last twelve months. Q2 2026 alone was $2.6B, up 112% from a year earlier, and management guided 2026 revenue to $12.4B to $13.2B.

• The backlog is roughly 14 years of trailing revenue, and it doesn't include another $25B-plus of commitments signed in early Q3. Meta alone has committed about $35B through 2032.

• They pay for all of it with debt secured by the GPUs and the customer contracts. CapEx over the last twelve months was $20.6B, or 2.7x revenue, and 2026 CapEx is guided to $35B to $39B.

• The bond market is nervous. Last week some of CoreWeave's unsecured bonds traded near 79 cents on the dollar, yields went above 15%, and the cost to insure its debt hit a record. The stock is down 30% since October 7 and hit a new low of $57.24 on Tuesday.

• Interest expense was $640M in Q2, about the same as the $626M net loss. Q3 interest is guided to $860M to $940M, against $200M to $260M of operating income.

• Gross margin after GPU depreciation is approximately 15%. That is ugly. (There appears to be a mismatch that is driving gross margin lower because they have to start depreciating the GPUs on day zero. The revenue gets delayed by 3 to 9 months. This is still a considerably lower gross margin than I would have expected to see.)

• It's my understanding that CoreWeave uses leverage on a data-center-by-data-center basis, with no corporate guarantee or cross-collateralization. Given their customer concentration, this is smart.

On Monday, Goldman upgraded the stock to Buy, arguing most of the debt is backed by GPUs and customer contracts, so borrowing costs should only rise 1 to 2 points. On Tuesday, Elon Musk, whose one-word "Yikes" tweet helped kick off the selloff, told SpaceX investors that talk of AI overcapacity is "ridiculous" and that every hyperscaler is begging to rent more compute. But he bonds barely moved. CoreWeave's 2032 notes were still around 80 cents today.

A few weeks ago I wrote about Powell Industries, which builds switchgear for the same data centers and gets its customers to pay up front. CoreWeave sits on the other end of that deal. It writes the checks for the chips, the buildings, and the power, then collects rent over five or six years. One is funded by its customers and one is funded by its lenders.

The key takeaway here for me is that a signed contract lets you borrow, but it doesn't set your interest rate. In a capital-heavy business, your cost of money is part of your product.

Tt's going to be fun to follow these AI-heavy companies over the next few years. I tend to be in the camp that views intelligence as a service as something that is going to have an insane amount of demand but I question the overall economics of those selling the tokens. Humans are quite good at driving the margins of anything that appears to be a commodity down to break-even or worse. And that would spell big trouble for the likes of CoreWeave.

With that, I'll see you tomorrow!

Nick

TL;DR

• CoreWeave rents Nvidia GPU capacity to AI labs and hyperscalers on multi-year, take-or-pay contracts, from data centers it mostly leases.

• TTM revenue is $7.6B and backlog is about $104B. Q2 2026 revenue grew 112% and FY2026 is guided to $12.4B to $13.2B.

• Adjusted EBITDA margin is 59%, but TTM operating margin is -3% because depreciation ate 53% of revenue and interest ate another 25%.

• Net debt including leases is about $45B against a ~$34B market cap. Bonds trade near 80 cents with junk ratings, and the stock is down 51% over the past year, versus Nebius (+65%) and Oracle (-52%).

• Operator lesson: a contract can be collateral, but the economics depend on what the money costs and how long the asset lasts.

The 30,000-Foot View

CoreWeave is a "neocloud," a cloud provider built only for AI workloads. Amazon, Microsoft, and Google run clouds that do everything. CoreWeave does one thing: large clusters of the newest Nvidia GPUs, networked together and tuned for training and running AI models.

The business model works like a landlord's. A customer like Meta signs a multi-year contract for a fixed amount of GPU capacity, usually paying whether they use it or not. CoreWeave takes that contract to lenders, borrows against it, buys the GPUs, and installs them in data center space it has leased with power already secured. Revenue starts when the capacity goes live, and the contract pays down the loan over its life.

What CoreWeave sells beyond the chips is speed and uptime. It has been first or near-first to bring several Nvidia generations online at scale, and its software (Mission Control, the fleet and node lifecycle controllers, and the Weights & Biases developer tools it acquired) is built to keep thousands of GPUs running a single job without failing.

Revenue mix: CoreWeave reports one segment, so the more useful split is by customer. In Q2 2026, the largest customer was 36% of revenue, the second 26%, and the third 10%. In FY2025, Microsoft alone was 67%.

Key Stats

• Market cap: ~$34.1B

• TTM revenue: $7.6B

• TTM gross margin: 67.4% (before depreciation; see below)

• Revenue backlog: ~$104B at June 30, 2026

• Active power: 1.5 GW, with ~4.2 GW contracted

• 1Y total return: -50.6%

• Employees: ~2,189

• Industry: AI cloud infrastructure

Company History

• 2017: Michael Intrator, Brian Venturo, and Brannin McBee start Atlantic Crypto, mining Ethereum with GPUs.

• 2019: Crypto prices crash. The founders buy up cheap GPUs, rename the company CoreWeave in December, and start renting the hardware out for rendering, visual effects, and machine learning.

• 2022: Ethereum moves off GPU mining for good. CoreWeave does $16M of revenue as a small GPU cloud, right before ChatGPT launches.

• 2023: Revenue jumps to $229M. Nvidia invests, Microsoft becomes the anchor customer, and CoreWeave raises one of the first large loans secured by GPUs.

• 2024: Revenue hits $1.9B, with Microsoft at 62%. Debt and leases climb past $10B.

• March 2025: IPO on Nasdaq at $40, below the original marketed range. OpenAI signs an $11.9B contract around the same time, later expanded to about $22.4B.

• 2025: Buys Weights & Biases, agrees to buy Core Scientific in an all-stock deal that Core Scientific shareholders vote down, signs a $14.2B Meta deal, and peaks at $183.58 a share in June. FY revenue reaches $5.1B.

• April 2026: Meta commits another $21B through December 2032. Anthropic signs a multi-year deal the next day.

• August 2026: Q2 revenue of $2.6B, backlog of $104B, a $1B strategic investment from Jane Street, Nasdaq-100 inclusion, and raised guidance: $12.4B to $13.2B of revenue and $35B to $39B of CapEx for 2026.

• October 2026: Unsecured bonds fall below 80 cents, credit default swaps hit records, Gartner says more than 80% of neoclouds may not survive independently past 2030, and the stock drops from $88 to $61 in a week. Goldman upgrades the stock to Buy on October 19, HSBC holds at Reduce, and the shares hit a low of $57.24 on October 20.

Show Me the Money

Standout financial features:

• Gross margin is real, but it leaves out the biggest cost. CoreWeave's cost of revenue doesn't include depreciation on the GPUs. TTM gross margin is 67.4%. Subtract the $4.0B of depreciation and amortization and it drops to about 15%. That's why a company with a 59% adjusted EBITDA margin has a negative operating margin.

• Operating margin went from 16.9% in 2024 to -0.9% in 2025 to -3.0% TTM. Revenue grew 4x in that stretch, but depreciation grew faster, because CoreWeave starts depreciating GPUs the moment they're installed and the revenue ramps after.

• Interest is the second wall. TTM interest expense was $1.9B, about 25% of revenue. Q3 guidance puts quarterly interest at roughly four times adjusted operating income.

• CapEx went from $2.9B in 2023 to $20.6B TTM. Net debt including leases went from $1.8B to $45.2B in two and a half years. Every dollar of new revenue starts as borrowed money spent on chips.

Financial Data

TTM runs through June 30, 2026. Gross profit excludes depreciation. Net debt includes lease obligations.

Stock Performance

CoreWeave IPO'd in March 2025, so there's no 5-year or 10-year history. Returns are through the October 21, 2026 close.

1-year head-to-head vs. peers:

CoreWeave is still up 55% from its $40 IPO price, but it's down 66% from the June 2025 peak. Oracle, the other company borrowing heavily to build AI capacity, got hit about as hard. Nebius, which has less debt relative to its size, is up over the same year. Investors are separating AI infrastructure companies by their balance sheets.

The N.O.O.B. Nine: Competitive Powers

The Nerd Out on Business Nine is made up of Hamliton Helmer's famous "7 Powers" of competitive advantage (Scale Economies, Network Economies, Counter-Positioning, Switching Costs, Branding, Cornered Resource, and Process Power) combined with two of my own (Data Flywheel and Distribution Advantage).

Average Score: 2.9/5 - a speed-and-capacity moat built on power, Nvidia access, and execution, held up by a balance sheet that has to keep borrowing.

Memorable Marketing

CoreWeave doesn't market to the public. Its marketing is aimed at a few dozen AI buyers and the lenders who fund the buildout, and the main tools are being first with new hardware and announcing big customers.

Notable tactics:

• First to new Nvidia hardware (ongoing): CoreWeave has repeatedly announced being among the first to offer new Nvidia systems at scale. For an AI lab, a few months' head start on the newest chip is the whole pitch.

• Customer announcements as marketing (2025 to 2026): The OpenAI, Meta, and Anthropic contracts were each announced publicly, and the stock moved on each. In April 2026, the Meta and Anthropic deals landed a day apart.

• Buying the developer audience (2025): Weights & Biases gave CoreWeave a tool that machine learning engineers already use every day, plus a direct line to the people who pick infrastructure.

• Third-party ratings and benchmarks: CoreWeave leans on independent cluster ratings and benchmark results to prove performance.

• Selling to lenders: Brannin McBee has talked openly about winning investment-grade ratings on chip-backed debt to borrow from insurers. Lenders are the second customer CoreWeave has to win.

• The staff memo (October 2026): As the bonds fell, Michael Intrator told employees the market was acting "irrationally" and pointed to the backlog: "the money is coming, and it covers the bills." A week later the company was still shipping product news, launching a Zero Trust AI security framework and Physical AI templates on October 21.

Tactical takeaways:

1. If your buyers are few and sophisticated, announce the customer wins. A logo is proof.

2. Buy the tool your users already love to get closer to the buying decision.

3. Being first with new capability is a marketing event.

4. In a capital-heavy business, your financing story needs the same care as your sales story.

AI Uses & Opportunities

Current exposure:

• CoreWeave is AI infrastructure. Nine of the ten largest AI model providers are customers, per company commentary.

• Managed inference passed $100M of ARR in Q2, with a target of at least $250M by year-end. AI products beyond raw GPU rental passed $400M of ARR.

• Management said pricing went up about 25% across products in July, and new Q2 contracts carry higher margins than recent ones.

• Competitors are public now too. SpaceX reported $7.6B of quarterly revenue from its AI compute segment (up 830%) and said payback on its compute is under a year at current prices.

Future opportunities:

• Inference over training: training runs are lumpy and concentrated. Running models for millions of users is steadier and spreads revenue across more customers.

• Software margin on top of hardware: every dollar of managed services, observability, and developer tooling earns more than renting a GPU by the hour.

• Smarter capacity allocation: using demand forecasting to keep idle GPUs near zero between contracts.

• Enterprise customers: Caterpillar, Bentley Systems, and Isomorphic Labs are early signs CoreWeave can sell beyond AI labs and cut concentration.

Bumps in the Road

• Funding risk. The plan needs $35B to $39B of CapEx this year. If bond buyers keep demanding ~15%, new contracts earn less, and some may not pencil out at all. Goldman thinks secured borrowing costs rise only 1 to 2 points. HSBC's $39 target says otherwise.

• Customer concentration. Microsoft was 67% of FY2025 revenue. Q2 was better, at 36%, 26%, and 10%, but a pullback by any one of the top three would hurt.

• Hardware that ages fast. Nvidia ships a new generation roughly every year. If GPUs lose value faster than the contracts that pay for them, the math breaks at renewal.

• Building on schedule. Revenue only starts when capacity goes live. Delays at leased data centers push revenue out while interest keeps running.

• Circular money. Nvidia is supplier, investor, and in some deals a capacity backstop. Investors are asking how much demand is end demand.

Your Swipe File

• If you're ever in a business that relies on customer contracts to fund CapEx, make sure you have a rock-star lawyer(s) looking over your contracts and, likely, a good finance team to do due diligence on your customers

• Judge capital-heavy businesses after depreciation. A 59% EBITDA margin and a -3% operating margin describe the same company. Know which number pays your loans.

• Match the length of the asset, the contract, and the loan. Trouble starts when any one of those runs shorter than the others. a mismatch like this is what got companies like WeWork in trouble. They signed long-term leases and had short-term rents. Any sort of mismatch like that is a recipe for disaster

• CoreWeave can't do much about customer concentration given the composition of the market, but it's still a huge risk. When I worked in private equity, anything more than, say, 20% customer concentration was a huge red flag.

• Get customers to fund the buildout when you can. Powell takes deposits and grew with no debt. CoreWeave takes loans and now has to refinance in a nervous credit market.