App Economy Insights — Neocloud Economics (Aug 2026)

One-line take: The premise is a supply shortage — "there is more demand for compute than the biggest cloud companies can supply" — and the issue's point is that the interesting differences between neoclouds are financial, not technical: "capacity must be funded and built months before it can generate revenue, driving negative free cash flow. Debt, leases, depreciation, and customer concentration matter almost as much as growth." CRWV is the purest expression and the most levered: revenue +112% to $2.6B with 98% of it from committed contracts (OpenAI, Microsoft, Meta), backlog $104B, +246%, plus another $25B signed in early Q3 — but gross margin fell 8 points to 66%, the net loss was $626M with $640M of interest expense on GPU-collateralized debt, and Q2 capex hit $9.4B — more than 3x quarterly revenue on the way to $35–39B for 2026 and >1.85 GW of active power. NBIS is the same business with better unit economics: revenue +454% to $582M, gross margin +6 points to 77%, $236M of adjusted EBITDA (41%), four new deals averaging >$1B of contract value at $20–25M per megawatt (short-term capacity $40–50M/MW), payback compressed to ~22 months from two-to-three years, and >$9B of customer prepayments covering 50–60% of 2026 capex — the customer, not the bond market, funds the build. CBRS is the odd one out, a wafer-scale chipmaker turning into an inference cloud: revenue +74% to $180M with cloud +281% to $126M passing hardware (−23% to $54M) for the first time, a $477M GAAP operating loss that is mostly post-IPO stock comp against a $34M core loss, core gross margin 41% (dented by renting back systems it had already sold), FY26 core revenue guided up to $880–890M, >600 MW live or contracted through 2027 and $25.4B of RPO with OpenAI a major customer. The closing question is competitive, not financial: with META scaling custom silicon and gigawatts of GPUs and SPCX already selling access to its Colossus cluster, are neoclouds "essential infrastructure partners or temporary relief valves once mega-cap AI capacity fully comes online"? Author disclosure: long AMZN, MSFT, GOOG, NVDA, AMD, TSM and ASML in the App Economy Portfolio — a disclosure, not a set of views argued in this issue.

1. Stocks & names mentioned

Ticker	Name	Research	View	What's said	Source
NBIS	Nebius Group	QT · SA · STK · FA	Positive	"Nebius Finds Pricing Power" — the one section whose Bottom Line leans. Revenue +454% Y/Y to $582M, of which AI Cloud is $575M (98%); gross margin +6 points to 77% and adjusted EBITDA $236M at a 41% margin. The $176M operating loss is smaller than the $260M of D&A now landing on the income statement as "billions of dollars of new infrastructure started hitting" it. Unusual origin: Nebius is the international remnant of the 2024 Yandex breakup — the Nasdaq-listed Dutch holding company sold the Russian operations for $5.4B, rebranded, and put co-founder Arkady Volozh back in charge, then used the engineering talent and capital base to build a purpose-built AI cloud rather than rebuild an internet conglomerate. The unit economics are the story: four new deals averaged more than $1B of contract value at $20–25M per megawatt, with shorter-term capacity fetching $40–50M/MW; management now estimates Q2 contracts repay their capex and operating costs in about 22 months, down from a two-to-three-year range. The build is enormous — Q2 capex $5.7B, almost 10x quarterly revenue, against $20–25B for the year — but customers are financing it: more than $9B of prepayments in 2026, covering roughly 50–60% of the associated capex. Bottom Line: "rising prices, faster paybacks, and customer prepayments are making each new megawatt more attractive. That capital efficiency may ultimately matter more than its 454% revenue growth."	read ↗
CRWV	CoreWeave	QT · SA · STK · FA	Neutral	"CoreWeave is Sold Out" — the purest version of the neocloud model: buy NVIDIA GPUs, install them, rent the compute to OpenAI, Microsoft and Meta. Revenue +112% Y/Y to $2.6B with committed contracts generating 98% of it and on-demand just 2%; gross margin contracted 8 points to 66% as data-center rent, power and ramp costs grew faster than revenue. A $49M operating loss became a $626M net loss, "heavily weighed down by $640 million in interest expense tied directly to its GPU-collateralized debt facilities" — the financing cost, not operations, is what makes the loss. And the P&L understates the spend: Q2 capex $9.4B, more than three times quarterly revenue. What's improving: backlog reached $104B (+246% Y/Y) before another $25B of commitments in early Q3, so near-term capacity is "effectively sold out"; adjusted operating margin rose sequentially 1% → 5%, new Q2 contracts should carry contribution margins 5–10 points above recent deals, and that is before July's roughly 25% price increase. Mix is broadening too — storage, CPU, networking and software now exceed $400M of ARR, and managed inference went from $1M to more than $100M of booked ARR in a single quarter. FY26 capex outlook raised to $35–39B chasing >1.85 GW of active power by year-end. Bottom Line: the $104B backlog "sounds almost absurd next to a targeted 2026 exit revenue run rate of $19 billion," but conversion is capped by physical capacity — the case "comes down to how quickly it can turn power and GPUs into revenue without letting financing costs overwhelm the margin gains."	read ↗
CBRS	Cerebras Systems	QT · SA · STK · FA	Neutral	"Cerebras Moves to the Cloud" — the odd one out, because instead of buying NVIDIA GPUs it designs its own wafer-scale processor and monetizes it either by selling systems or renting compute through Cerebras Cloud. The mix flipped this quarter: revenue +74% Y/Y to $180M, with Cloud & Other Services +281% to $126M while hardware fell 23% to $54M — core cloud revenue "nearly quadrupled to $128 million and surpassed hardware for the first time." The headline loss needs context: a $477M GAAP operating loss against a core operating loss of just $34M, the gap being substantial stock-based compensation triggered by May's IPO plus customer-warrant charges and pass-through items — an overhang that "should normalize over upcoming quarters as initial post-IPO equity grants settle." Same split on margin: 14% reported vs 41% core (up ~9 points Y/Y but down from 46.5% in Q1), partly because Cerebras is temporarily paying to rent back systems it previously sold in order to meet cloud demand. Guidance improved — FY26 core revenue raised to $880–890M with higher gross- and operating-margin expectations — and capacity is the binding constraint: more than 600 MW live or contracted through 2027, with core gross margin expected to bottom in Q3 before new capacity reduces the need for expensive rented capacity. Demand is far ahead of revenue: RPO of $25.4B with OpenAI still a major customer. Bottom Line: it is "evolving from a chip seller into a fast-inference cloud… the real test is converting its enormous backlog into revenue while rebuilding margins as new capacity comes online."	read ↗
META	Meta Platforms	QT · SA · STK · FA	Neutral	Both a customer and the closing risk. Meta is named among the enterprises whose long-term commitments fill CRWV's 98%-committed revenue base — and then reappears in "What to Watch" as the reason that base may not be permanent: "the same compute shortage is pushing tech giants to build their own capacity. Meta is scaling custom silicon and gigawatts of GPUs." The open question the issue leaves for every neocloud investor: whether they "remain essential infrastructure partners or temporary relief valves once mega-cap AI capacity fully comes online." Referenced; not a stance call.	read ↗
MSFT	Microsoft	QT · SA · STK · FA	Neutral	Named with OpenAI and Meta as the customers CoreWeave rents compute to — i.e. one of the counterparties behind the 98% committed-contract revenue and the $104B backlog, which is also the customer-concentration risk the issue flags ("customer concentration matter[s] almost as much as growth"). Azure also appears on the other side of the trade, as one of the "broad software ecosystems" neoclouds deliberately do not try to replicate. Referenced; not a stance call. (Also a disclosed author holding.)	read ↗
NVDA	NVIDIA	QT · SA · STK · FA	Neutral	The dividing line between the three models. CRWV is defined as the "purest version of the neocloud model" precisely because "it buys NVIDIA GPUs, installs them in data centers, and rents the compute" — and those GPUs are the collateral behind the debt costing $640M of quarterly interest. CBRS is defined by the opposite choice: "instead of buying NVIDIA GPUs, it designs its own wafer-scale processor." Referenced as the ecosystem's supply anchor; not a stance call. (Also a disclosed author holding.)	read ↗
SPCX	SpaceX	QT · SA · STK · FA	Neutral	One line, but it is the competitive one: "SpaceX has already started selling access to its Colossus cluster" — a non-cloud owner of large-scale compute monetizing spare capacity, which is exactly the "temporary relief valve" scenario for dedicated neoclouds. Named alongside Meta as evidence that the compute shortage is pulling non-cloud giants into supplying compute themselves. Referenced; not a stance call.	read ↗
OpenAI	OpenAI (private)	—	Neutral	The demand side of two of the three stories. Named first among the customers CRWV rents compute to (part of the 98% committed-contract base), and again at Cerebras, where "OpenAI remains a major customer" inside $25.4B of remaining performance obligations — a backlog the issue notes "requires substantially more infrastructure" to convert. Referenced; not a stance call.	read ↗

"View" here is referenced — App Economy Insights is financial-analysis journalism, not buy/sell calls (BUY/SELL/HOLD ratings are shared only with App Economy Portfolio members). Following the source's convention, rows are Neutral by default, with Positive/Negative used only where the section's own closing "Bottom Line" clearly leans. NBIS is the single Positive row because its Bottom Line lands on an improvement, not a condition — "rising prices, faster paybacks, and customer prepayments are making each new megawatt more attractive." CRWV and CBRS stay Neutral: both Bottom Lines end on an unresolved test ("without letting financing costs overwhelm the margin gains"; "converting its enormous backlog into revenue while rebuilding margins"). Research: QT Qualtrim · SA Seeking Alpha · STK Stock Analysis. The "Source" links open the newsletter (no per-name timestamps — it's a written post). Author disclosure: long AMZN, MSFT, GOOG, NVDA, AMD, TSM and ASML in the App Economy Portfolio. MSFT and NVDA are rowed on body substance (a named CoreWeave customer; the GPU supplier defining the CRWV/CBRS split), not on the disclosure — AMZN, GOOG, AMD, TSM and ASML are disclosure-only here and get no rows. Named but not given rows: AWS, Azure and Google Cloud (one clause on the "broad software ecosystems" neoclouds do not replicate — Microsoft is rowed for the separate customer mention), Yandex (Nebius's 2024 predecessor; its Russian operations were sold for $5.4B and are not the listed entity), Colossus (a compute cluster, not a security), Fiscal.ai (the chart source), and Berkshire, Tencent, Cisco, Sea Limited, Nu and Adyen (a one-line teaser for Saturday's PRO edition, with no analysis in this issue).

2. Talking points

The setup: a capacity gap creates a business model

"AI has created a massive capacity gap. There is more demand for compute than the biggest cloud companies can supply." That gap is what neoclouds exist to fill.
The definition is deliberately narrow: specialized providers "built around AI infrastructure," focused on securing raw power, specialized data-center capacity, and accelerator clusters — explicitly not "replicating the broad software ecosystems of AWS, Azure, or Google Cloud."
Three models reported in the same week, which is what makes the comparison clean: CRWV rents an NVIDIA GPU fleet, NBIS built an AI cloud from scratch on international data-center infrastructure, CBRS designs its own chip and is pivoting into hosted inference.

Why the economics are "counterintuitive"

The structural point the whole issue rests on: "capacity must be funded and built months before it can generate revenue, driving negative free cash flow."
Consequence for analysis: "debt, leases, depreciation, and customer concentration matter almost as much as growth" — a growth-rate-first read of any of these three will mislead.

CoreWeave is sold out — and 98% pre-sold

Committed contracts generated 98% of Q2 revenue; on-demand usage was just 2%. This is a contracted-utility revenue profile, not a cloud-consumption one.
Revenue +112% Y/Y to $2.6B, but gross margin −8 points to 66% as "data-center rent, power, and other ramp costs grew even faster" — growth bought at a margin cost while capacity ramps.
Customers named: OpenAI, Microsoft, Meta — the same names that make the backlog credible also make it concentrated.

The loss is a financing line, not an operating one (CRWV)

Operating loss $49M; net loss $626M. The bridge is $640M of interest expense "tied directly to its GPU-collateralized debt facilities."
That is the single most important number in the section: the operating business roughly breaks even, and the capital structure produces the loss. Any margin improvement has to outrun a rising interest bill.

The income statement hides most of the spending (CRWV)

Q2 capex $9.4B — more than three times quarterly revenue. "The income statement captures only part of the spending."
FY26 capex outlook raised to $35–39B, chasing more than 1.85 GW of active power by year-end. Power, not chips, is the unit of ambition.

Backlog vs the ability to serve it (CRWV)

Backlog $104B, up 246% Y/Y, plus another $25B signed in early Q3 — "near-term capacity is effectively sold out."
Set against a targeted 2026 exit revenue run rate of $19B, the backlog "sounds almost absurd" — the constraint is physical: "conversion is constrained by physical capacity."
That reframes the metric: backlog growth is no longer the bull signal; energized megawatts per quarter is.

Where CoreWeave's margin is supposed to come from

Adjusted operating margin improved sequentially 1% → 5%, and new Q2 contracts are expected to carry contribution margins 5–10 points above recent deals.
Crucially, that is before July's roughly 25% price increase — pricing power showing up on the same shortage that constrains supply.
Mix helps too: storage, CPU, networking and software now exceed $400M of ARR, and managed inference went from $1M to more than $100M of booked ARR in a single quarter — the fastest-moving line in the issue.

Nebius's unusual origin story

Not a startup: Nebius emerged from the 2024 breakup of Yandex. The Nasdaq-listed Dutch holding company sold the Russian operations for $5.4B and kept the international businesses.
It retained the listing, rebranded as Nebius Group, and put Yandex co-founder Arkady Volozh back in charge.
The pivot decision matters: rather than rebuild an internet conglomerate, it redeployed "engineering talent, cloud expertise, and capital base" into a purpose-built AI cloud — it started this race already funded.

Nebius's Q2: high growth, and unusually high margin for a builder

Revenue +454% Y/Y to $582M, with AI Cloud at $575M — 98% of revenue, earned by renting GPU capacity through its own platform.
Gross margin +6 points to 77% (versus CoreWeave's 66% and falling) and $236M of adjusted EBITDA at a 41% margin.
The $176M operating loss sits beneath $260M of D&A — i.e. the loss is depreciation of newly energized infrastructure, not cash burn from operations.

Pricing per megawatt is the metric to steal

Four new AI cloud deals averaged more than $1B of contract value and $20–25M per megawatt.
Shorter-term capacity is fetching $40–50M per MW — roughly double, which is the clearest evidence in the issue that scarcity is being priced, and that duration is now a discount rather than a premium.

Payback compresses to ~22 months

Management estimates Q2 contracts will repay their associated capex and operating costs in about 22 months, "down from the previous two-to-three-year range."
Against multi-year contracts and hardware with a longer useful life, a sub-two-year payback changes what the capex means: less a bet, more an amortization schedule.

The customer becomes the lender (NBIS)

Q2 capex $5.7B, almost 10x quarterly revenue, with $20–25B still expected for the year.
But Nebius expects more than $9B of customer prepayments in 2026, covering roughly 50–60% of the associated capex — vendor financing in reverse, and the structural contrast with CoreWeave's GPU-collateralized debt and $640M interest bill.
Bottom Line: "capital efficiency may ultimately matter more than its 454% revenue growth."

Cerebras: the mix flips from hardware to cloud

Revenue +74% Y/Y to $180M: Cloud & Other Services +281% to $126M while hardware fell 23% to $54M.
On a core basis cloud revenue "nearly quadrupled to $128 million and surpassed hardware for the first time" — a chipmaker becoming a rental business, with the recurring revenue and the capital intensity that implies.

Reading a first post-IPO income statement (CBRS)

GAAP operating loss $477M versus a core operating loss of $34M; the difference is dominated by stock-based compensation triggered by May's IPO, plus customer-warrant charges and pass-through items.
Margin shows the same split — 14% reported vs 41% core — and the overhang "should normalize over upcoming quarters as initial post-IPO equity grants settle." A newly public company's first prints need the core-vs-GAAP reconciliation before any trend is drawn.

Renting back what you already sold (CBRS)

Core gross margin of 41% was up ~9 points Y/Y but down from 46.5% in Q1 — partly because Cerebras is "temporarily paying to rent back systems it previously sold" to meet cloud demand.
Management expects core gross margin to bottom in Q3, before new capacity comes online and reduces the need for expensive rented capacity. A dated, falsifiable margin trough is a useful thing to diarise.

Capacity is the bottleneck everywhere (CBRS)

More than 600 MW of data-center capacity live or contracted through 2027; FY26 core revenue guidance raised to $880–890M with better margin expectations.
RPO of $25.4B with OpenAI still a major customer — but "converting that backlog requires substantially more infrastructure." The same backlog-vs-megawatts tension as CoreWeave, at a tenth the revenue.

What to watch: partners or relief valves?

The shared tension: "demand is outpacing available compute, but serving it requires enormous amounts of capital before revenue arrives."
The competitive threat comes from the customers: META "is scaling custom silicon and gigawatts of GPUs," and SPCX "has already started selling access to its Colossus cluster."
The question left open: are neoclouds "essential infrastructure partners or temporary relief valves once mega-cap AI capacity fully comes online"?
The closing formulation is the thesis of the issue: neoclouds must "turn contracted demand into energized infrastructure, improve returns as utilization rises, and fund the next wave of expansion without letting debt or dilution overwhelm the economics. The demand is locked in. Capital efficiency will separate the winners."

3. In plain English

A jargon-free summary of the read behind each name. (Plain-language companion to the table above; renders on each ticker's consolidated page.)

NBIS — Nebius Group Positive

Nebius rents out AI computing power, the same basic business as CoreWeave. What makes this quarter interesting is not that revenue grew five-fold to $582 million — it is how the growth is being paid for.

Building an AI data centre is like building a toll road: you spend billions up front and collect fees for years afterwards. The question is always who lends you the money for the gap. CoreWeave borrows it against its chips and pays heavy interest. Nebius is increasingly getting its customers to pay in advance — it expects more than $9 billion of prepayments this year, covering 50–60% of what it spends on new capacity. Money from a customer costs no interest and cannot be margin-called.

The price it charges is also going up, and the newsletter gives the metric to watch: dollars per megawatt of power sold. Nebius's four newest deals — each averaging over $1 billion — were struck at $20–25 million per megawatt, and shorter contracts are going for $40–50 million per megawatt. In other words, customers pay a large premium for flexibility, which is what a genuine shortage looks like.

Put those together and the "payback period" — how long a new site takes to earn back everything it cost to build and run — has fallen to about 22 months, from two-to-three years previously. That is fast for infrastructure this expensive.

It is also an unusual company. Nebius is what was left of Russia's Yandex after the 2024 breakup: the Amsterdam-based, Nasdaq-listed parent sold the Russian business for $5.4 billion, kept the international operations and the cash, and re-pointed the whole thing at AI under founder Arkady Volozh. It started this race already funded.

The caution: it is still spending roughly ten times its quarterly revenue on construction, and its reported operating loss ($176M) exists because the new equipment is being written down ($260M of depreciation). Everything depends on the contracts being real and the customers staying solvent. Analysis, not a recommendation.

CRWV — CoreWeave Neutral

CoreWeave does one thing: it buys NVIDIA chips, plugs them into data centres, and rents the computing power to companies like OpenAI, Microsoft and Meta. It is the cleanest way to own the AI compute shortage — and the most financially exposed.

The demand side is extraordinary. 98% of its revenue this quarter came from contracts signed in advance, and the order book — work already sold but not yet delivered — reached $104 billion, up 246%, with another $25 billion signed weeks later. Revenue more than doubled to $2.6 billion. Management says near-term capacity is effectively sold out, and it pushed prices up about 25% in July.

The problem is the bill. Building capacity costs money years before customers pay for it: CoreWeave spent $9.4 billion on construction and chips in a single quarter — more than three times its revenue — and plans $35–39 billion for the year. To fund that it borrows against the GPUs themselves, and this quarter the interest on those loans was $640 million. That single line is why a business with a $49 million operating loss reported a $626 million net loss. Operationally it is close to break-even; financially it is deep in the red.

Two things pull in the right direction: profit per contract is improving (adjusted operating margin went from 1% to 5%, with new deals expected to be 5–10 points better), and the mix is broadening beyond raw chip rental — storage, networking and software passed $400 million of annual recurring revenue, and managed inference went from $1 million to over $100 million of booked recurring revenue in one quarter.

Why Neutral: the enormous backlog cannot be converted faster than CoreWeave can physically energize data centres — it is targeting more than 1.85 gigawatts of live power by year-end — and every month of delay is another month of interest. The whole case is a race between rising margins and a rising interest bill. Analysis, not a recommendation.

CBRS — Cerebras Systems Neutral

Cerebras is the outlier of the three. Rather than buying NVIDIA's chips, it makes its own — a single processor the size of an entire silicon wafer, designed to answer AI queries unusually fast. Historically it sold those machines outright. Now it is renting them out itself, through Cerebras Cloud.

That switch happened this quarter: revenue rose 74% to $180 million, but the composition flipped — cloud rental revenue jumped 281% to $126 million and overtook hardware sales (which fell 23% to $54 million) for the first time. Renting is a better long-term business, but it means Cerebras now has to fund the data centres itself, the same trap the other two are in.

Its reported numbers look alarming and mostly are not. Cerebras listed on the stock market in May, which triggered a huge one-off charge for employee share awards. Strip that out and the $477 million reported operating loss becomes a $34 million core loss; reported gross margin of 14% becomes 41%. That accounting distortion should fade over the next few quarters.

One real margin problem is unusual and worth understanding: Cerebras is paying rent to use machines it had already sold to other people, because demand for its cloud arrived before its own capacity did. That is expensive, and it is why core margin slipped from 46.5% last quarter. Management expects the low point in the third quarter, with new capacity — over 600 megawatts live or contracted through 2027 — relieving it after that.

The prize is a $25.4 billion pile of contracted-but-undelivered work, with OpenAI a major customer. The catch is the same as everywhere else in this issue: turning that into revenue requires building far more infrastructure first. Analysis, not a recommendation.

Key points & figures extracted from the App Economy Insights newsletter (article text in transcript.txt) for personal study. Not investment advice. © App Economy Insights for source material.