This group rents out artificial-intelligence computing power. The product is access to clusters of NVIDIA graphics chips (GPUs — the processors that train and run AI models), housed in data centers with the power, cooling, storage and networking needed to use them. The customers are the companies building AI: the big technology firms (Meta, Microsoft, OpenAI, Amazon, Google), AI labs, and enterprises that cannot get their own chip allocations from NVIDIA or build clusters fast enough. These "neoclouds" (a shorthand for newer cloud providers specialized in renting GPU compute, as distinct from the broad general-purpose clouds) exist because demand for compute has outrun what the customers can build for themselves — so even firms spending $100B+ a year of their own (Meta, Microsoft) still rent capacity from this group. The six names here split into two business shapes: operators that buy GPUs and rent the compute (CoreWeave, Nebius, IREN), and landlords that supply powered buildings and let a tenant install the chips (Core Scientific, Applied Digital, TeraWulf). Several of the landlords are former Bitcoin miners that already controlled cheap power and land and pivoted to AI.
On the facts available, contracted demand is running well ahead of what the group can currently deliver: backlogs across these names grew on the order of +280% to +325% year-over-year, and active capacity is a small fraction of contracted capacity (CoreWeave ~1 GW active vs >3.5 GW contracted; Nebius ~170 MW active vs >2 GW contracted). What limits supply is no longer chips alone but powered, grid-connected data-center capacity — and the capital to build it. In money terms, the operators trade at roughly 5x to 80x+ this year's revenue and most of the group does not yet generate owner cash (cash left for the owner after all costs and reinvestment): the market value here is set against contracted future revenue and a buildout that has not happened yet, not against today's earnings. Demand vs supply, and what is paid for the gap, is laid out below.
Source: company SEC filings Q1-2026 (CoreWeave 10-Q, Nebius 6-K); 500-stocks Data Centers scan; competitor disclosures per reports-v3 ticker pages (CRWV, NBIS, IREN, CORZ, APLD, WULF).
The product is rented GPU compute, sold by the GPU-hour or by the megawatt (MW) of power filled with chips. A customer that wants to train or run an AI model needs thousands of GPUs wired together with fast networking. Building that yourself takes a chip allocation from NVIDIA, land, a power connection from the grid, cooling, and 2-4 years. Renting it from a neocloud means you skip all of that and pay for what you use.
The companies earn cash in two main ways:
The economics are an arbitrage spread (the gap between the price charged and the cost to deliver): the rental price per GPU-hour, minus the wholesale cost of the GPU spread over its life (depreciation — the accounting cost of an asset wearing out / aging), minus interest on the debt used to buy the GPUs, minus power and operating cost. The two business shapes capture different slices of that spread:
Source: reports-v3 CRWV, NBIS, IREN, CORZ, APLD, WULF summaries; 500-stocks Data Centers scan (Neoclouds subsection).
The cleanest demand signal is the contracted backlog — what customers have already signed and (mostly) cannot walk away from. The accounting term is RPO (remaining performance obligations — signed future revenue not yet earned). Across this group the backlogs are very large and grew explosively in the last year:
| Company | Contracted backlog | As of | Notes |
|---|---|---|---|
| CoreWeave (CRWV) | $99.4B | Q1 2026 | contracted +~284% YoY; up ~$33B in one quarter |
| Nebius (NBIS) | $33.6B | Mar 31 2026 | contracted +57% in one quarter; ~71% lands 2028+ |
| TeraWulf (WULF) | $12.8B | Q1 2026 | contracted 522 MW; Fluidstack + Core42, Google-backstopped |
| Applied Digital (APLD) | ~$11B | 2026 | contracted CoreWeave 400 MW / 15-yr lease at Ellendale ND |
| Core Scientific (CORZ) | ~$10.2B | 2026 | contracted CoreWeave ~590 MW / 12-yr lease |
| IREN (IREN) | ~$3.1B ARR | Q3 FY26 | contracted ARR = annual recurring revenue run-rate; Microsoft + NVIDIA; targeting $3.7B ARR by YE 2026 |
These contracts are largely non-cancelable take-or-pay: at Nebius, anchor customers Meta and Microsoft must pay "irrespective of actual utilization," and the only escape is one-sided in the company's favor (a customer can drop a tranche only if the provider fails to deliver on time). So for the group the risk is mostly execution (can they build it?), not demand (will the customer walk?). The buyers are concentrated and very large: Meta is ~40% of CoreWeave's backlog and OpenAI ~25%; CoreWeave is essentially the only HPC (high-performance computing — large-scale GPU/AI workloads) customer at both Applied Digital and Core Scientific; Microsoft and NVIDIA anchor IREN; Fluidstack (with a Google payment backstop) and Core42 anchor TeraWulf.
Beyond the signed book, the forward industry demand picture (these are forecasts / industry estimates, mark as such):
Reasoning from the premise that AGI is arriving, compute is the binding input to further AI progress — more compute makes better models, which create more demand for compute — so the demand picture is reinforcing rather than saturating. That is the lens; the figures above are the facts, with the industry-wide ones marked approximate. The reader weighs them.
✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:
Source: reports-v3 ticker pages; SEC filings (CoreWeave 10-Q RPO $99.4B; Nebius 6-K RPO $33.6B; Oracle 10-Q RPO ~$550B); 500-stocks Data Centers scan; hyperscaler/industry figures from general knowledge, not live-verified.
Supply is measured in energized power — megawatts of data-center capacity actually filled with chips and earning revenue — versus contracted power (land + grid connection secured, but not yet built). Across the group, active capacity is a small fraction of what is contracted, which is the supply story in one number:
| Company | Active / energized | Contracted / pipeline | Model |
|---|---|---|---|
| CoreWeave | >1.0 GW | >3.5 GW (target ~8 GW by 2030) | Operator; ~600k GPUs est., ~32 data centers |
| Nebius | ~170 MW | >2 GW | Operator; owns greenfield (newly built from scratch) site in Finland, plus US/EU sites |
| Core Scientific | ~350 MW (~590 by early 2027) | ~4.5 GW gross / ~3+ GW leasable | Landlord; CoreWeave-anchored |
| TeraWulf | ~60 MW HPC (+ ~195 MW mining) | ~1,150 MW controlled | Landlord; Lake Mariner NY flagship |
| IREN | ~110 MW AI (ramping) | ~5+ GW pipeline (~480 MW lit by YE26) | Operator; owns land + GPUs (vertically integrated) |
| Applied Digital | ~100 MW (ELN-02 building) | ~900 MW contracted + ~1 GW marketed | Landlord; CoreWeave 400 MW anchor |
The real bottlenecks, in order of how binding they now are:
Who controls supply: CoreWeave is the largest pure-play, with roughly an order of magnitude more active power than Nebius and the deepest contracted pipeline. The landlords (CORZ, APLD, WULF) control physical sites but largely on behalf of one tenant. IREN and Nebius own their stack (land, buildings and, for IREN, the GPUs too). No single company holds a dominant share of the whole market on the facts here; the structural fact is that all of them are supply-constrained, not demand-constrained, today.
Source: reports-v3 ticker pages (active/contracted power, GPU counts, capex guides); 500-stocks Data Centers scan; SEC filings; $/MW build cost and transformer lead times are external industry estimates, not live-verified.
Putting the two sides together: the product is short (more demanded than can be delivered), not long. Contracted future demand far exceeds current delivery ability, and the price and lead-time signals all point to shortage rather than glut.
| Evidence | Reading | What it points to |
|---|---|---|
| Backlog vs run-rate | CoreWeave $99.4B backlog vs ~$8.3B annualized revenue; Nebius $33.6B vs ~$1.6B | ~12x-20x more contracted than currently delivered |
| Active vs contracted power | CoreWeave >1 GW active vs >3.5 GW contracted; Nebius ~170 MW vs >2 GW | Large buildout still to come — supply lags signed demand |
| Backlog growth | +~284% (CoreWeave), +~325% (Oracle) YoY | Demand accelerating faster than capacity |
| Rental price trend (new contracts) | Rising; capacity "largely sold out" on current-gen at CoreWeave | Shortage — sellers have pricing power on new business |
| Utilization / sold-out | Operators report waitlists and sold-out current-gen capacity | Tight market |
One important crosscurrent that complicates "short forever": older chips deflate fast. The H100 (last generation) on-demand price compressed from roughly $8/GPU-hour at peak scarcity toward ~$2-6 depending on segment, and AWS reportedly cut its H100 list price by roughly ~40%+ in mid-2025 (pricing figures are industry estimates, not live-verified). So the shortage is concentrated in the newest generation (Blackwell GB200, then Vera Rubin); the prior generation softens as it rolls off. The fixed long-term contracts cover the operators through roughly 2027-2028, with the main risk window being 2027-2030 contract renewals.
When could it flip to oversupply? On the facts here, the binding question is whether power buildout catches up to chip supply. As long as energized, grid-connected capacity is the bottleneck, the product stays short. A flip to oversupply is a 2027-2028+ question — it would require power and construction to catch up while demand growth slows, which, reasoning from AGI arriving (compute demand reinforcing), is not the current trajectory but is not impossible. Facts only.
Source: reports-v3 ticker pages; CoreWeave/Nebius/Oracle SEC filings; H100 pricing history from industry sources cited in the CRWV report (silicondata, intuitionlabs), not live-verified.
The six names, what each makes, how concentrated it is on this one product, rough size, and its position. (Prices and market caps are as of each report's date in spring-to-June 2026 and move daily; treat as approximate.)
| Company | What it makes | Exposure to this product | Rough size | Position / structure |
|---|---|---|---|---|
| CoreWeave (CRWV) | Buys NVIDIA GPUs, rents compute under long-term reserved contracts | Pure-play (~100% of revenue) | ~$60-73B mkt cap; ~$94B EV | Largest pure-play; deepest backlog ($99.4B); first to GB200 GA (general availability); NVIDIA-backed; heavy debt (~$33B) |
| Nebius (NBIS) | AI cloud (rented GPU compute); small edtech + autonomous side bets | Pure-play (~98% of revenue) | ~$67B basic / ~$94B diluted equity | Owns greenfield sites; $33.6B backlog; Meta + Microsoft anchors; foreign filer; founder control |
| IREN (IREN) | Owns land, power, data centers AND GPUs (vertically integrated); BTC mining residual | Mixed → pivoting (AI ~23% of rev, rising fast) | ~$17-22B mkt cap | Only one owning both real estate and GPUs; ~46% operating margin; Microsoft + NVIDIA contracts; ~5 GW pipeline |
| Applied Digital (APLD) | AI data-center landlord (cloud arm spun off as ChronoScale May 2026) | Landlord pure-play | ~$10-12B mkt cap | CoreWeave 400 MW / $11B / 15-yr anchor; A3 investment-grade-rated lease SPV (special-purpose entity that holds one lease) — a credit difference vs CORZ |
| TeraWulf (WULF) | AI-hosting landlord at Lake Mariner NY; BTC mining shrinking | Landlord, pivoting (HPC overtaking mining) | ~$11.5B mkt cap; ~$14B EV | 522 MW / $12.8B backlog; Fluidstack lease with a Google payment backstop; Core42 anchor; ~85% NOI margin |
| Core Scientific (CORZ) | Post-bankruptcy data-center landlord; BTC mining residual | Landlord, single-tenant heavy | ~$7.3B mkt cap; ~$11-12B EV | CoreWeave ~590 MW / $10.2B / 12-yr lease (~100% of HPC P&L one customer); CRWV merger failed Oct 2025; ~$5B debt |
The structural pattern: the operators capture the GPU rental economics (and its risk); the landlords capture fixed rent (and give up the upside). Three of the landlords (CORZ, APLD, partly WULF) are economically downstream of CoreWeave — if CoreWeave's GPU-rental spread holds, they collect their rent; if it doesn't, they face counterparty risk (the risk the paying customer can't or won't pay) on a single tenant. This is the bridge to individual deep-dives.
Source: reports-v3 ticker pages (market caps, contracts, structure); 500-stocks Data Centers scan. Market caps as of report dates spring-June 2026.
In plain money terms, here is roughly what the market is paying for each name, expressed two ways an owner can feel: dollars of market value per dollar of this-year revenue, and dollars of enterprise value per contracted megawatt. Enterprise value (EV — equity value plus net debt; what you'd effectively pay to own the whole business and settle its debt) matters here because several of these names carry a lot of debt. (The per-MW figures are back-solved from disclosed contracts and EV, so they are approximate.)
| Company | Price / this-yr revenue | EV per contracted MW est. | Owner-cash shape |
|---|---|---|---|
| CoreWeave | ~5-6x ($94B EV ÷ $12-13B 2026 rev guide) | ~$25-30M / MW (>3.5 GW) | High-capex; ~56-60% adj EBITDA margin (EBITDA = earnings before interest, tax, depreciation, amortization — a rough cash-from-operations proxy) but heavy debt + capex → net cash burn while building |
| Nebius | ~2.0x basic / ~2.8x diluted vs $33.6B RPO; ~42x basic vs $1.6B run-rate | ~$34M basic / ~$47M diluted per contracted MW | High-capex; operations burn cash before customer prepayments; reported "profit" is largely a paper markup on an equity stake, not cash |
| IREN | ~5-6x ARR ($23.6B EV ÷ ~$3.7B target ARR) | low single-digit $M / MW on full pipeline | High-capex (owns GPUs); ~46% operating margin; ~$1.25B distributable cash at steady state per the report's build (a forecast) |
| Applied Digital | ~$1.76M/MW base rent; ~77% NOI margin | ~$8-10M / MW (energized) | Landlord, more capital-light per dollar of rent; bond-like fixed rent |
| TeraWulf | ~80x+ EV / trailing revenue est. (revenue pre-ramp) | ~$12M / MW controlled (~1,150 MW) | Landlord; ~85% NOI margin once built; still EBITDA-negative today |
| Core Scientific | ~$120-150/kW/month fixed rent (back-solved) | ~$18-20M/MW energized; ~$3-4M/MW on full pipeline | Landlord; bond-like fixed rent + CPI escalator (rent that rises with inflation); ~$5B debt vs ~$350M run-rate |
The money-in/money-out shape of the group:
So in one breath: across the group the market is pricing in several times this year's revenue (operators) up to ~80x (pre-ramp landlords), and roughly $8M to $47M of enterprise value per contracted megawatt, for businesses that are mostly still building and mostly not yet generating owner cash. The price set today is against the contracted future and the buildout, not today's earnings. Neutral arithmetic.
Source: reports-v3 ticker pages (revenue guides, EV, RPO, per-MW math, margins); per-MW and per-kW figures back-solved from disclosed contracts; market values as of report dates spring-June 2026.
Where a company-level deep-dive is most information-dense, factually:
This is a pointer to where information is densest, not a recommendation.
Source: reports-v3 ticker pages; 500-stocks Data Centers scan.
What was used:
Hard facts (grounded in filings / ticker reports): the contracted backlogs (CRWV $99.4B, NBIS $33.6B, WULF $12.8B, APLD ~$11B, CORZ ~$10.2B, IREN ~$3.1B ARR); active vs contracted power for each; the contract structures (take-or-pay, single-tenant, Google backstop, A3 SPV); CoreWeave's $12-13B 2026 revenue and $31-35B capex guides; Oracle's ~$550B RPO. Company prices and market caps are point-in-time (spring-June 2026 report dates) and move daily.
Approximate / not live-verified (general knowledge, cutoff ~early 2026, web retrieval was unavailable): the ~$230B→high-$300Bs hyperscaler capex trajectory, the "100%+/yr" and "low-hundreds of billions by 2030" neocloud market sizing, the $35-45M/MW build cost, the ~600k CoreWeave GPU estimate, the transformer lead-time figure, and the GPU rental-price levels/history. These are directional context, tagged est. throughout and flagged in the unverified note in the Demand section. Where exact points were given in the draft they have been widened to ranges.
No buy/sell recommendation, price target, or valuation verdict is expressed. This page presents facts and neutral arithmetic so the reader can judge the demand-vs-supply gap and what is paid for exposure to it.
Source: as listed above. Company figures from primary SEC filings via the reports-v3 ticker pages; industry figures from general knowledge, not live-verified.