Skip to content
Back to posts

I came for the GPUs. I stayed for the financing footnotes.

Notes by SDG ·

I came for the GPUs. I stayed for the financing footnotes.

I became interested in Nebius, CoreWeave and IREN for two related reasons.

First, the AI infrastructure story itself is fascinating. I want to understand more than just models, tokens and agents. What physical, financial and operational systems make those products possible? Someone has to build the data centers, secure the power, finance the equipment and deliver the compute.

Second, understanding these businesses may help me decide whether I want to invest in them—or at least avoid discovering an important opportunity after everyone else has already made their money.

That does not mean assuming an exciting industry automatically produces attractive stocks. It means understanding the business before forming a view.

I was going through my weekly AI infrastructure digest when these three names kept pulling me into the same rabbit hole.

I knew the broad story. AI needs compute. These companies provide compute. More AI should mean more business.

Very convenient. Also about as useful as saying airlines make money because people like going places.

What I wanted to understand was what each company actually sells, why customers would choose it, and what has to happen between a spectacular contract announcement and money reaching shareholders.

I also checked recent Stocktwits reporting on retail sentiment. The company sections below use dated snapshots from September 4–9, with an August 28 comparison for IREN. These are reported readings and selected reactions, not a live sentiment score or my own count of every post. Where I interpret the reasons, I say so.

The crowd’s mood is worth understanding. It is not a valuation model.

My starting point was the September 7 xNoizDigest. The comparison uses company disclosures available through September 10, 2026. The numbers are dated because this sector appears to age in dog years.

First, what is a neocloud?

Think of an AI team that needs thousands of GPUs working together. Buying the chips is only part of the problem. Someone needs to arrange the building, electricity, cooling, networking, storage, software and people who keep the whole thing running.

A specialist AI cloud packages that infrastructure into a service. Customers rent computing capacity rather than build every component themselves.

Training teaches or improves a model. Inference runs the model to produce answers, predictions or actions. Both need infrastructure, but their performance, scheduling and cost requirements can differ.

CoreWeave, Nebius and IREN all participate in this business. Their offerings overlap, with different strengths across infrastructure and services for AI workloads. You can see that in the product descriptions from CoreWeave⁠, Nebius⁠ and IREN⁠.

Here’s how I think about the three companies.

CoreWeave is a specialist AI cloud already operating at scale, with a large base of contracted demand. My question: can delivery and operating improvements keep ahead of financing costs?

Nebius is building an AI cloud with a broader software and developer platform. My question: will customers keep choosing it for the platform’s value even when GPUs become easier to get elsewhere?

IREN is an AI cloud business built around its physical infrastructure. My question: can it turn its sites and contracted capacity into working deployments that customers accept and that generate revenue?

These are my ways of understanding the businesses, not stock rankings.

CoreWeave: demand is visible. So is the cost of serving it.

CoreWeave reported $2.575 billion of revenue in Q2 2026 and roughly $104 billion of revenue backlog at June 30. Its adjusted EBITDA margin was 59%, yet its GAAP net loss was $626 million. Net interest expense alone was $640 million. Those figures come from its Q2 earnings release⁠.

That combination is worth pausing on.

Adjusted EBITDA strips out interest, depreciation and other adjustments. It can help explain operations, but it cannot tell me what shareholders ultimately keep. In a business that buys expensive equipment and borrows heavily, the excluded items are rather central to the plot.

The backlog is evidence of committed demand, with an important qualification: CoreWeave’s definition includes remaining performance obligations and additional estimated revenue under committed contracts, subject to delivery and service availability. It is neither this year’s sales nor cash already earned.

Management’s August outlook put 2026 revenue at $12.4–$13.2 billion and capital expenditures at $35–$39 billion. Comparing the midpoints gives approximately $2.90 of planned capex for every dollar of forecast revenue. See the company’s outlook presentation⁠.

That calculation illustrates the scale of expansion. It does not measure the profitability of an individual contract, because assets bought now can serve customers across future years.

The attraction is a company already delivering substantial revenue. The risk is that execution, equipment spending and financing must remain coordinated.

A construction delay can postpone revenue while obligations continue arriving on schedule. Lenders tend to be refreshingly consistent about that.

What Stocktwits sentiment tells me about CoreWeave

Stocktwits reported neutral sentiment on September 9 at 1:00 p.m. EDT, with message activity increasing from low to normal. Its September 4 report had described CRWV sentiment as bearish. These are two dated snapshots, rather than a continuous trend measurement. Sources: September 9 coverage⁠ and September 4 coverage⁠.

The positive catalyst in the September 9 coverage was management’s claim that pricing on newer deals was rising faster than input costs, supporting better margins as deployments scale.

That gives the optimistic case something more specific than “AI is big.”

The counterweight is the upfront investment and borrowing needed to serve those contracts. My interpretation is that neutral sentiment fits a market waiting to see whether better contract economics become better shareholder returns. The report does not establish why every poster was neutral.

The backlog can look excellent while the financing bill remains very real.

Nebius: I’m interested in what sits above the hardware.

Nebius’s Q2 AI cloud revenue was approximately $575 million. Its Token Factory production inference workloads more than tripled during the quarter. The company’s shareholder letter⁠ describes an expanding platform for training, inference and operating AI workloads.

The potential advantage is straightforward: if the service makes an AI team more productive, more reliable or more cost-efficient, customers have reasons to stay beyond the availability of a particular GPU.

That is a thesis to test, not a moat I am willing to declare complete. I would look for repeat usage, expanding customers and evidence that software improves the cost of completing useful work.

A long list of features is easier to produce than durable customer dependence.

At group level, Nebius reported $582.3 million in quarterly revenue, $236.2 million in adjusted EBITDA and a $190.4 million loss from continuing operations. Operating cash inflow was $2.246 billion, while purchases of property, equipment and intangible assets were $5.657 billion. See its Q2 earnings release⁠.

Subtracting those purchases from operating cash flow gives an approximately $3.41 billion quarterly deficit on that simple measure. This is my calculation, not a company-defined free cash flow metric.

It shows why growing operating cash flow alone is insufficient to judge whether expansion finances itself.

Another detail I would watch: Nebius extended the useful life used for server and network equipment from four years to five starting in 2026, as explained in its shareholder letter.

That changes the timing of depreciation expense. It does not manufacture cash or guarantee the equipment’s future commercial value. Comparing margins across periods requires noticing those accounting assumptions.

What Stocktwits sentiment tells me about Nebius

Stocktwits’ September 8 report, published at 8:32 a.m. EDT, described NBIS sentiment as bearish even as the shares rose in premarket trading and attracted trending attention.

The news was a Palantir partnership intended to put Nebius compute and inference infrastructure within Palantir’s enterprise environment, giving customers more control over models, data and computing resources. See the Stocktwits report⁠.

That helps explain the optimistic side: a route into enterprise customers and a stronger platform story.

It does not, by itself, explain the bearish score. Stocktwits’ NBIS news feed⁠ also carried coverage questioning valuation and capital requirements after a large share-price rise. Those are plausible sources of caution, not verified motives for the entire message stream.

My read: traders may like the company’s opportunity while questioning how much future success is already priced in—and how expensive delivery will be.

A partnership announcement and a bearish sentiment reading can coexist.

IREN: the transition is happening inside the numbers.

IREN’s physical infrastructure is central to its pitch, but it also sells GPU cloud services. Calling it simply a landlord misses the compute business.

For the quarter ended June 30, AI cloud revenue was $70.5 million and Bitcoin mining revenue was $66.7 million.

That distinction matters: the financials reflect a business in transition. The company’s future AI capacity cannot be treated as if it generated the historical revenue.

At its August update, IREN reported $4 billion of contracted annualized revenue for 2026 capacity, against $1 billion of operating annualized revenue as of August 26. These are operating metrics. The ramp depends on commissioning, testing and customer acceptance, and recognized revenue may be materially lower. See the FY26 results and business update⁠.

This gives me a very concrete thing to track: the movement from contracted capacity into operating capacity and then recognized revenue.

A press release can announce a contract in an afternoon. The equipment is less cooperative.

There is a similarly useful distinction in the September 8 Sweetwater update. IREN said the 2 GW hub had been conditionally included as Base Load in ERCOT’s Batch Zero process. Sweetwater 1’s substation had been energized, while 300 MW of gross data center capacity was under construction with delivery targeted for Q4 2027. The classifications remained conditional. See the Sweetwater announcement⁠.

That is progress. It is not 2 GW of operating AI capacity.

Grid status, completed buildings and customer-ready compute belong on separate lines in my notebook.

What Stocktwits sentiment tells me about IREN

Stocktwits reported bullish sentiment on September 8 at 8:47 a.m. EDT following the conditional Sweetwater grid milestone. It also highlighted a user expecting substantial upside; that is an individual opinion, not a consensus price target. See the September 8 report⁠.

The reasoning is clearer in its earlier earnings coverage. An August 28 snapshot showed extremely bullish sentiment and extremely high chatter despite the selloff.

The article described traders treating weak headline results as part of the mining-to-AI transition, with one investor focused on the planned ramp from operating ARR to contracted capacity. See the earnings-reaction coverage⁠, updated August 31.

In other words, some bulls are buying the future operating business rather than judging the company only on today’s revenue. Sweetwater’s progress gives that thesis another milestone to follow.

The same coverage raised construction-cost pressure and financing needs.

My interpretation: optimism rests on execution catching up with the opportunity. Conditional grid status still needs approvals, and contracted capacity still needs delivery.

Enthusiasm cannot commission a data center.

Why would customers use these companies when AWS, Azure and Google Cloud exist?

The practical answer can include capacity, speed of deployment, workload performance, support and price. A specialist does not have to replicate every hyperscaler product to win a demanding AI workload.

But the relationship is complicated. CoreWeave’s filing identifies AWS, Google Cloud, Azure and Oracle as competitors and says some competitors are also customers. It also warns about the purchasing power, broader offerings and resources of larger rivals. See its Q2 Form 10-Q⁠.

My inference: a hyperscaler buying specialist capacity validates a need at that moment. It does not establish that the relationship will stay equally attractive forever. The buyer can expand its own infrastructure, negotiate harder or shift workloads.

That is why I care about the service around the GPU. Availability can win the first contract. Reliability, workflow integration and economics have to help win the next one.

This is also where sentiment can become misleading. A large contract may produce a burst of bullish posts, while a financing announcement may produce an equally sharp bearish reaction.

Neither reaction necessarily answers the longer-term question: can the company earn an acceptable return on the capital required to serve the customer?

The four numbers I refuse to mix together

Recognized revenue tells me what the company counted as revenue over a stated period. It does not tell me future contracted sales or how much cash the company kept.

Annualized run-rate revenue expresses a current or projected operating pace as an annual number. It is not revenue actually earned over a full year.

Backlog and customer commitments describe potential future revenue under specified definitions and contract terms. They do not represent guaranteed profit or immediate collections.

Power and capacity pipelines describe resources available or planned at different stages of development. They do not necessarily represent GPUs already operating for paying customers.

The definitions require care even when the labels look identical.

Nebius annualizes the last month’s AI cloud revenue. IREN’s operating ARR uses contracted GPU-hour pricing for commissioned GPUs, annual hours and ancillary revenue. These definitions appear in the Nebius shareholder letter⁠ and IREN business update⁠.

Putting both into one bar chart without explanation would make the chart cleaner and the understanding worse.

Where the investment argument gets interesting

My analytical view is that the strongest business could combine three advantages: access to usable infrastructure, efficient delivery, and a service customers value after scarcity eases.

Power access without financing may remain an opportunity on paper. Financing without delivery can become an expensive delay. Delivery without differentiation can leave the operator competing mainly on price.

I would monitor five things across all three companies.

Conversion: do commitments become revenue on the promised timeline?

Cash: after equipment spending and financing costs, is the funding gap improving?

Customers: does growth broaden the customer base or increase dependence on a few buyers?

Equipment economics: can older hardware remain useful at prices that support its costs?

Shareholders: is value growing per share after new equity, convertible debt and compensation?

These questions also explain why “AI demand keeps growing” is an incomplete investment thesis. Demand can grow while competition lowers pricing, financing absorbs returns or new share issuance dilutes the benefit.

They also explain how I use sentiment: identify the assumptions behind the optimistic and cautious cases, then test them against operating evidence.

A bullish label cannot settle the financing question, and a bearish label does not disprove customer demand.

I am not ranking the stocks at today’s prices here. That would require a separate valuation exercise using current enterprise values, dilution, debt and explicit cash-flow scenarios. Understanding the business comes first.

For now, CoreWeave gives me the clearest picture of scale and the accompanying financing burden. Nebius makes me pay attention to software differentiation and capital efficiency. IREN gives me a tangible infrastructure-to-revenue transition to follow.

Stocktwits adds a fourth lens: the crowd was not equally enthusiastic about all three. The dated readings above make that contrast visible; the business evidence determines whether the optimism or caution is justified.

I came into this expecting to learn more about GPUs. I left with a much stronger interest in customer acceptance dates, depreciation policies and whether retail sentiment is reacting to the business—or merely to the stock chart.

Apparently this is what personal growth looks like now.

A note on sources and sentiment

Company figures come from the linked disclosures, with reporting periods and forward-looking estimates identified in the text. Calculations and interpretations are my own.

Stocktwits sentiment was checked through publicly available reporting on September 10, 2026. The labels refer to the dates specified above. They are not live September 10 readings, and the selected reactions do not represent a complete survey of investors.

The Nebius Stocktwits article contains inconsistent weekday wording; I have used its publication timestamp. The IREN earnings-reaction article contains inconsistent claims about results versus analyst expectations; I have used it for sentiment and reported reactions, not those earnings comparisons.

The research started with the September 7 xNoizDigest, “AI Infrastructure Weekly Insights: Capital, Constraints & Growth.” Unverified forum estimates in that digest were not used as financial evidence here.