GPU Capacity Outside the Hyperscalers

Four advantages hold up under the published record. A fifth — "we have the chips you cannot get" — has largely stopped working.

1. Speed, which the buyer cannot buy from itself

The strongest evidence is behavioural: Microsoft has committed over sixty billion dollars to companies it could out-spend a hundred times over (Microsoft has committed over 60 billion dollars to neoclouds). It does that because Crusoe energised a 1.2 GW campus in about a year while a conventional operator takes several, and because Nscale can stand up 240 MW faster than an internal programme. See Hyperscalers rent because they cannot build fast enough and Crusoe's co-developed power and shell.

2. Accounting, not just engineering

Renting converts capex into opex spread over a contract term — Capex-to-opex arbitrage. For a hyperscaler under investor scrutiny for its capex line, that is a feature independent of whether it could build the thing.

3. Operational depth at the top of the market

The premium tier is not a commodity. SemiAnalysis rates eighty-five providers and only CoreWeave holds Platinum; the rating maps to real pricing power (Platinum rating carries a large pricing premium). Reliability on a ten-thousand-GPU training run is worth more than a cheaper rate card — Reliability is the moat, not price and ClusterMAX tiering.

4. Power, land and permits, held early

The scarce asset is an energised site. Crusoe runs generation behind the meter, Nscale holds Nordic hydro at low cents per kilowatt-hour, and the The bitcoin-miner conversion cohort owns interconnects nobody can now replicate inside a decade. See The binding constraint is power, not silicon and Behind-the-meter generation.

The threat to all four

None of these advantages survives if the customer builds it themselves at a quarter of the cost, which xAI has now demonstrated (xAI built Colossus 2 at a quarter of the industry cost per megawatt, A frontier lab should build, not rent).

The advantage that has gone

Allocation arbitrage. In 2023 and 2024, holding H100s was the whole business. That window closed as hyperscalers locked supply and buyers got choosier — Only five to seven independent GPU clouds survive. What remains is operations, power and price, which is a harder fight and a fairer one.

And a fifth, contested

Sovereignty. Sovereignty is demand that capital cannot replicate is the strongest version; Nscale is the test case and the counter-evidence is in its own note.

Read the other side at Where neoclouds struggle.

Five problems, roughly in order of how quickly they bite.

1. The cost of money, which is now the business

CoreWeave pays more in annual interest than it earned in revenue two years ago (CoreWeave's annualised interest cost exceeds its 2025 revenue), and its credit default swaps have traded at roughly even odds of default (CoreWeave's credit default swaps priced roughly even odds). Neocloud paper prices 275 to 325 basis points wide of investment grade (Neocloud debt prices 275 to 325 basis points over investment grade). The whole sector faces a concentrated maturity window — see What happens at the refinancing wall.

2. Concentration on both sides

One vendor sells the chips, guarantees the loan and takes a revenue share (Nvidia, Nvidia demand backstops and revenue share). A handful of buyers carry the backlog (Microsoft was two-thirds of CoreWeave's 2025 revenue). That is Two-sided concentration risk, and it is structural rather than a phase.

3. The customer is building the competitor

The July 2026 Meta Compute report removed roughly a sixth of the equity value of the listed neoclouds in a session (The Meta Compute report cut neocloud equities by a sixth in a day). See The customer is becoming the competitor and Meta.

4. Depreciation, which nobody can settle

Operators book four to six year lives (Operators depreciate GPUs over four, five and six years); the bear case says two to three (The 176 billion dollar understated depreciation estimate). The counter-evidence is unusually good (CoreWeave re-leased three-year-old H100s at 95 percent of price). This is the hinge — Residual value as the hinge, How long does a GPU actually earn.

5. Physics, which does not respond to capital

Transformers at 160 weeks, turbine slots gone to 2029, interconnection queues in years — Substation transformer lead times passed 160 weeks, Heavy-duty gas turbine slots are allocated into 2029, Only a third of 2026 US capacity is under construction. A contracted gigawatt is a claim, not an asset — Is a contracted gigawatt a real gigawatt.

6. Disintermediation

In many of the largest contracts the neocloud never meets the end user — the hyperscaler or the lab does. Who owns the customer relationship asks what happens to the margin when capacity stops being scarce and the relationship is the only thing left worth owning.

7. The supplier competing on margin

Nvidia's own marketplace routes buyers to whichever partner has capacity, which makes providers interchangeable by construction — The DGX Lepton marketplace, Nvidia is commoditising its own partners. Whether it is working is contested (The marketplace layer is failing).

8. Everyone else in the stack earns more

The landlords out-earn their tenants, and the networking and memory vendors out-earn both — The plumbing captures more than the platform, Broadcom's AI semiconductor revenue tripled to 16.7 billion dollars, The landlord layer. The operator holds obsolescence, utilisation and refinancing risk simultaneously; nobody else in the chain holds all three.

The quiet one

Margin. Bare-metal gross margins look healthy until depreciation lands (Bare-metal gross margins run 55 to 65 percent before depreciation) and thin badly afterwards (Post-depreciation gross margins run 14 to 16 percent), against a break-even that sits near seventy percent utilisation (The utilisation cliff). Read Where neoclouds win for the other side.

The word "neocloud" is applied to companies whose balance sheets have nothing in common. Sorting them is the first analytical move, because each type fails differently.

1. The AI hyperscaler. Owns or leases shells, buys GPUs with debt, sells multi-year contracted capacity at the premium tier. CoreWeave, Nebius. Failure mode: leverage and refinancing.

2. The energy-first integrator. Starts from power and builds the compute on top. Crusoe, and increasingly Nscale. Failure mode: construction and anchor-tenant risk, not chip cycles. Crusoe's own framing is that it is Compute is a utility asset, not a technology asset.

3. The converted miner. Owns energised industrial sites from a previous business and re-tenants them. IREN, TeraWulf, Cipher Digital, Core Scientific, Applied Digital, Hut 8 and Riot Platforms. Failure mode: they are landlords being valued as operators. See The bitcoin-miner conversion.

4. The inference platform. Sells tokens, not GPU-hours, and rents the underlying capacity from everyone in categories one to three. Fireworks AI, Together AI, Baseten, Modal, Groq. Failure mode: margin compression from above and below — see Renting capacity you do not own and Fireworks runs at roughly half the gross margin of software.

5. The broker and the long tail. Aggregates or resells other people's fleets. The long tail of GPU resellers. Failure mode: no floor under price.

6. The self-build and the sovereign. Two categories that are not neoclouds and compete with them anyway: labs building their own (xAI) and states funding national capacity (G42, HUMAIN, Yotta and the IndiaAI Mission, SoftBank). Failure mode: neither has to earn a commercial return, which is the point — Sovereign compute is industrial policy, not a market, Build versus rent for a frontier lab.

7. The landlord. Owns powered shell and sells it on ten to twenty year leases, taking tenant credit risk instead of hardware obsolescence risk — Equinix, Digital Realty, Aligned Data Centers, Vantage and the sponsor-backed platforms, Galaxy Digital. Failure mode: the tenant. See The landlord layer.

Why the distinction earns its keep

A reader comparing Fireworks AI at a 17.5 billion dollar valuation with CoreWeave at a similar enterprise value is comparing an asset-light demand aggregator with a leveraged infrastructure owner. One is exposed to token prices, the other to interest rates and transformer lead times. See Asset-light beats asset-heavy for the argument that this asymmetry is the point, and The four-tier market structure for how SemiAnalysis cuts the same cake.

Signals, not opinions. Each is observable and each would move the conclusions in Where neoclouds win or Where neoclouds struggle.

Reserved pricing turning down. The one-year contract rate is the cleanest read on whether supply has caught demand — Reserved H100 pricing rose 40 percent in five months. A sustained fall takes the scarcity rent out (Two-thirds of a GPU-hour price is demand premium, Scarcity rent).

A large re-lease at a discount. The bull case on residual value rests on one disclosed data point (CoreWeave re-leased three-year-old H100s at 95 percent of price). A public re-lease well below original pricing would settle How long does a GPU actually earn against the operators.

Meta Compute shipping a price list. Reported intent moved markets (The Meta Compute report cut neocloud equities by a sixth in a day); an actual product would reprice the whole category — The customer is becoming the competitor.

A failed refinancing. Not a wider spread, an actual failure. Watch the 2026 to 2028 maturities described in What happens at the refinancing wall and Neocloud debt prices 275 to 325 basis points over investment grade.

Nvidia withdrawing the backstop. The programme was already paused in part in August 2026 (Nvidia's compute-access programme is reported at 36 billion dollars). Withdrawal would expose which operators were financeable on their own credit — Circular financing is a permanent feature.

Enterprise revenue arriving. The bear case is that neoclouds stay wholesale suppliers to five buyers — Neoclouds are brokers unless they reach the enterprise. Disclosed enterprise revenue mix is the tell.

Turbines and transformers loosening. The physical constraint is the reason incumbency in Behind-the-meter generation is worth anything — Heavy-duty gas turbine slots are allocated into 2029.

A merchant capacity market forming. Today almost everything is contracted — Merchant versus contracted capacity. A liquid spot market for large clusters would change the risk profile of every operator in The four-tier market structure.

Mechanic. The operator borrows to buy GPUs and pledges the GPUs, and often the associated contracts, as collateral. The lender underwrites two things: the contracted cash flow, and what the hardware will be worth when the loan matures.

This is how CoreWeave, Nebius, Lambda, Crusoe and Applied Digital funded their fleets, and it is why tens of billions of predominantly GPU-collateralised loans come due between 2026 and 2028 in a narrow window among a small, highly correlated circle of lenders (What happens at the refinancing wall, Private credit funds).

What is unusually good about it

It let the sector scale far faster than equity would allow, and against an asset that — so far — has held value better than sceptics predicted (CoreWeave re-leased three-year-old H100s at 95 percent of price).

What is missing

The collateral and the cash flow are correlated. If demand softens, utilisation falls, rental rates fall, and the resale value of the collateral falls at the same time and for the same reason. There is no diversification inside the structure. That is the difference between this and a mortgage, and it is why the depreciation argument (The economic life of a GPU is two to three years) is really a credit argument.

Second-order details that move recoveries: whole servers resell better than stripped parts, and major-OEM servers resell around 18 percent above lesser-known builders. Lenders who underwrote generic hardware get a different recovery from lenders who underwrote Dell.

Two 2026 developments address the correlation directly: GPU residual value insurance, which moves tail risk to a third party, and The OpCo and ComputeCo split, which separates the hardware from the building so each is financed by a lender equipped to value it.

The better structure

Contract-level asset financing, where the borrowing is against an investment-grade customer's payment obligation rather than the chips. Nebius has shown it can cover more than 100 percent of the capex required to deploy the underlying GPUs. Expect the sector to migrate toward it.

Links: Residual value as the hinge, Duration mismatch, Neocloud debt prices 275 to 325 basis points over investment grade.

The revenue mechanism underneath almost every large number in this vault.

Mechanic. A customer commits to a fixed block of GPU capacity at a fixed rate for two to five years — increasingly five to seven with prepayments — and pays whether or not it uses the capacity. The operator borrows against that contracted cash flow to buy the GPUs. On-demand and spot rental is the residual market and carries full price risk.

What is unusually good about it

It converts a depreciating asset into something a lender will underwrite. It is what makes Contract-level asset financing possible at Nebius and what allows project-level EBITDA margins reported in the 80 percent range for fully contracted deployments. It also transfers utilisation risk to the customer, which is the single most important protection an operator has against The utilisation cliff.

Duration is lengthening, which matters: CoreWeave's contracts over 48 months went from 10 to 21 percent of its backlog mix in a year.

What is missing

The rate is fixed and the shortage is not. An operator that signed a three-year block in late 2025 watched reserved rates rise forty percent by March 2026 (Reserved H100 pricing rose 40 percent in five months) and captured none of it. That is the trade in Merchant versus contracted capacity, and it explains the appearance of Subletting contracted compute — tenants monetising the spread their landlord gave away.

Take-or-pay also does not survive counterparty stress. A contract is only as good as the buyer, which is why Hyperscaler credit wrappers exist and why Backlog is not revenue.

The published analogue

Power purchase agreements and merchant plant offtake. The structure, the language and the failure modes are all borrowed from energy project finance, which is the strongest support for Compute is a utility asset, not a technology asset.

Links: GPU-collateralised lending, Microsoft, Duration mismatch.

CoreWeave Q2 2026 results, reported 11 August 2026.

• Revenue backlog of approximately 104.2 billion dollars as at 30 June 2026, up 246 percent from 30.1 billion a year earlier and from 99.4 billion at the end of Q1.
• Excludes more than 25 billion of additional net new customer commitments signed in the early weeks of Q3, taking the reported figure toward 129 billion.
• Contracts exceeding 48 months grew from 10 to 21 percent of the mix.
• 1.5 gigawatts of active power.

Why it matters here

It is the largest contracted forward book in the sector and the single number most often quoted as evidence the demand is real. The duration shift is the more informative part: lengthening contracts reduce Duration mismatch and improve refinanceability.

The caution is in Backlog is not revenue — conversion depends on power arriving on schedule (Contracted power is not connected power) and on 31 to 35 billion dollars of 2026 capex being funded (GPU-collateralised lending).

Sources

CoreWeave Q2 2026 release · Investing.com on the slides

The sector reports contracted backlog as its headline metric — CoreWeave's backlog reached 104 billion dollars, over 40 billion at Nebius, 9.3 billion at Cipher Digital, 12.8 billion at TeraWulf, 4 billion at IREN. Backlog conversion requires three things that are not guaranteed.

Power on schedule. Most contracted capacity is not connected (Contracted power is not connected power). A contract that begins delivery in 2027 depends on transformers and interconnects that are not in hand (Substation transformer lead times passed 160 weeks).

Counterparty follow-through. Many commitments are stated as "up to" — Meta's Nebius agreement is 12 billion dedicated plus up to 15 billion of additional purchases (Meta committed up to 27 billion dollars to Nebius). The optional half is a call option on someone else's demand.

Capital to build against it. Backlog is a liability before it is an asset: CoreWeave's 104 billion backlog implies 31 to 35 billion of 2026 capex, funded with debt at around nine percent (CoreWeave borrowed at around 9 percent in high yield).

Assessment

The honest counter is that these are take-or-pay contracts with investment-grade counterparties (Take-or-pay capacity contracts), and that IREN's disclosure of 4 billion contracted with 1 billion already generating cash is exactly the transparency the criticism asks for. Backlog with duration lengthening — CoreWeave's contracts over 48 months went from 10 to 21 percent of the mix — is better information than backlog alone.

The useful discipline is a ratio: contracted revenue divided by connected megawatts, tracked over time. That converts a marketing number into an operating one, and it is the number this vault would most like to see disclosed (Is a contracted gigawatt a real gigawatt).

Links: Merchant versus contracted capacity, Duration mismatch, The 2026 state of play.

The landlord layer's own view, stated bluntly in 2026 data centre investment analysis:

The tidy "investment-grade hyperscaler tenant" story breaks down with the neoclouds. Their leases are underwritten on venture funding and the resale value of GPUs, not an Amazon-grade balance sheet, so a facility leased to a neocloud carries far more counterparty risk than its hyperscale equivalent, even when the headline rent looks identical.

Assessment

This is independent confirmation of Two-sided concentration risk from the party with capital at stake, and it matches the pricing: neocloud paper runs 275 to 325 basis points wide of investment grade (Neocloud debt prices 275 to 325 basis points over investment grade).

The phrase that does the work is "even when the headline rent looks identical". Two buildings at the same rent are not the same asset, which is why Hyperscaler credit wrappers exist and why Galaxy Digital's fifteen-year sole-tenant lease to CoreWeave is a different instrument from Cipher Digital's to Amazon Data Services (Sole-tenant build-to-suit leases).

The qualification

Landlords are not risk-free either. Digital Realty carries an unfunded pipeline against a flat dividend and heavy equity issuance, and a purpose-built AI shell has few alternative tenants. Some lenders explicitly decline to underwrite the tenant at all and instead underwrite the facility's intrinsic re-leasing value — a bet on power scarcity rather than on any operator (The binding constraint is power, not silicon).

Which side is safer is genuinely open (Who is the safer credit, the landlord or the operator).

Links: The landlord layer, Private credit funds, Backlog is not revenue.

Tens of billions of dollars of predominantly GPU-collateralised loans from CoreWeave, Nebius, Lambda, Crusoe and Applied Digital come due between 2026 and 2028 — within a narrow window and among a small, highly correlated circle of lenders.

What is known

Spreads are already wide (Neocloud debt prices 275 to 325 basis points over investment grade). CoreWeave repriced a leveraged loan 100 to 125 basis points wider in 2026 and lenders began demanding covenants (CoreWeave's credit default swaps priced roughly even odds). The ABS market that would normally take this paper out sits around 25 billion against needs approaching 300 billion (Private credit lending to AI passed 200 billion dollars).

What is not known

Whether the wall is a wall. Three things could dissolve it:

1. Migration to Contract-level asset financing, where the borrowing is against an investment-grade payment obligation rather than the chips. Nebius has shown this covers over 100 percent of deployment capex.
2. The The Nvidia financing consortium institutionalising GPU compute as an asset class, bringing genuinely new capital.
3. Continued shortage keeping utilisation and rates high enough that refinancing is routine (Reserved H100 pricing rose 40 percent in five months).

The likeliest outcome, and why it is not a headline

Not failure but repricing. Refinancing 300 basis points wider turns a 14 to 16 percent post-depreciation margin into nothing (Post-depreciation gross margins run 14 to 16 percent) without a single default — a slow squeeze rather than an event, and the mechanism behind Only five to seven independent GPU clouds survive.

Links: Duration mismatch, The capital cycle, Private credit funds.

The most broadly agreed proposition in the sector, and by 2026 close to consensus across the IEA, Goldman Sachs Research, EPRI, Wood Mackenzie, Uptime Institute and Lawrence Berkeley National Laboratory.

The case: chips shipped, and the projects still did not energise. Roughly 2,060 GW of generation and storage sat in US interconnection queues at the end of 2024 with typical waits stretching to five years (Two thousand gigawatts sit in the US interconnection queue). Substation transformer lead times went from about 140 weeks in 2023 to over 160 in 2026 (Substation transformer lead times passed 160 weeks). Heavy-duty gas turbine slots are allocated through 2029 (Heavy-duty gas turbine slots are allocated into 2029). Only about a third of the 12 GW of US capacity expected in 2026 was under construction (Only a third of 2026 US capacity is under construction).

One line captures it: electrical equipment is under ten percent of total data centre cost and one hundred percent of the bottleneck.

Memory is now documented well enough to stand as a peer constraint rather than a footnote: HBM sold out through 2026, packaging booked through 2027, server DRAM up 60 to 70 percent in a quarter (Memory as the second bottleneck, HBM capacity is sold out through 2026).

Labour is a fourth: roughly 340,000 data centre positions are projected to go unfilled, and the scarce trades are the same high-voltage and commissioning specialists every campus needs at once (Labour as a delivery constraint, 340,000 data centre positions are projected to go unfilled).

A fifth constraint belongs beside power, memory and labour: the hall itself. Only one operator in five is ready for the 50 to 70 kW racks already common, and 68 percent of pre-2015 facilities cannot host AI density at all (Average rack density rose from 16 to 27 kilowatts in a year, Density as an obsolescence clock).

Assessment

The consequences run through the whole vault. It is why Behind-the-meter generation is a moat, why the The bitcoin-miner conversion cohort has an asset nobody can replicate, why Contracted power is not connected power is the most misread disclosure in the sector, and why Compute is a utility asset, not a technology asset is a serious valuation argument rather than a pitch.

The qualification worth stating: memory is a competing bottleneck. HBM and CoWoS constraints, and a parabolic rise in DRAM and NAND contract pricing in early 2026, forced OEMs to reprice AI servers and contributed to the shortage independently of power. "Power, not silicon" is directionally right and slightly too clean — the honest version is that 2026 had at least three simultaneous constraints, and capital was not among them.

Links: Where neoclouds win, Crusoe, Scarcity rent.

The single most misread number in the sector. Operators disclose three different things and press coverage collapses them into one.

• Pipeline — sites under evaluation. Crusoe cites a pipeline exceeding 45 GW.
• Contracted — land with secured power rights. Crusoe reports 4.9 GW contracted; CoreWeave and Nebius each around 3.5 GW at mid-2026; Nebius targeting 2.5 GW contracted by end-2026.
• Connected or active — power delivered to a built facility. CoreWeave had 1.5 GW active at Q2 2026 and targets 1.7 GW by year end. Nebius expects 800 MW to 1 GW connected by end-2026, from 170 MW at the end of 2025.

Nebius therefore converts perhaps a third of contracted power into connected power within the year. That ratio, not the headline gigawatt figure, is the operating metric.

Why the gap persists

Not capital. Interconnection queues, transformer and switchgear lead times, and turbine allocation — see Substation transformer lead times passed 160 weeks, Heavy-duty gas turbine slots are allocated into 2029 and Two thousand gigawatts sit in the US interconnection queue. This is the mechanism behind The binding constraint is power, not silicon.

Why it matters here

A backlog is only deliverable if the power lands on schedule, which links this note directly to Backlog is not revenue and to Is a contracted gigawatt a real gigawatt. It also explains why Behind-the-meter generation commands a premium and why the The bitcoin-miner conversion cohort has an asset that cannot be bought at any price today.

Sources

CoreWeave Q2 2026 · DCD on Nebius · Trending Topics

Wood Mackenzie, reported May 2026, plus corroborating industry data.

• Substation transformer lead times stretched from roughly 140 weeks in 2023 to more than 160 weeks in 2026.
• Generator step-up transformer lead times of 144 weeks and large power transformer lead times of 128 weeks, with a 30 percent supply deficit projected through 2027.
• Switchgear timelines held closer to a year but remain elevated.
• Pre-2020 lead time on a high-power transformer was 24 to 30 months; other 2026 estimates put it at up to five years.
• US data centre capacity is expected to scale from roughly 24 GW to 100 GW between 2026 and 2030, and data centres account for roughly 68 percent of US load growth through 2030.

The line worth keeping: electrical equipment is under ten percent of total data centre cost and one hundred percent of the bottleneck.

Why it matters here

It is the load-bearing evidence for The binding constraint is power, not silicon and the reason Contracted power is not connected power is a real operating gap rather than conservatism.

It also explains the premium on Behind-the-meter generation and on operators like Crusoe that manufacture their own switchgear (Crusoe's co-developed power and shell).

Sources

Data Center Knowledge · Tech Fund

Manufacturer disclosures, 2025 and 2026.

• GE Vernova booked over 30 billion dollars in gas power orders during 2025; its Q1 2026 earnings reported heavy-duty production slots allocated through 2029.
• Mitsubishi Power moved its heavy-duty H-class fleet to selective order acceptance.
• Siemens Energy reported double-digit growth in heavy-duty orders and a multi-year delivery backlog.
• GE has roughly 100 GW on contract across about 90 distinct customers in 24 countries, with about 10 GW of capacity remaining across 2029 and 2030 combined, and is emphasising capital efficiency rather than adding manufacturing capacity.

The structural deficit is confirmed across all three major heavy-duty manufacturers. Turbines are now slot-allocated rather than catalogue-procured.

Why it matters here

Because behind-the-meter generation is the main escape route from the interconnection queue (Two thousand gigawatts sit in the US interconnection queue), and this is the queue behind that escape route (Behind-the-meter generation).

It also dates the constraint. Turbine slots allocated to 2029 mean the physical shortage underwriting Scarcity rent has a scheduled, visible duration — which is the most important input to The capital cycle in this sector.

The workaround emerging is fuel cells: Oracle's Project Jupiter pivoted to a fully islanded Bloom microgrid with no turbines at all.

Sources

Savrn brief · Tech Fund

The frame. Power is the constraint everyone names (The binding constraint is power, not silicon). Memory is the constraint that binds independently, on its own schedule, and it repriced the sector in 2026 without reference to grid queues.

The mechanics. HBM consumes roughly three times the wafer capacity of DDR5 for the same gigabyte count, so shifting production to it removes conventional DRAM supply. Capacity is sold out through 2026 across all three suppliers (HBM capacity is sold out through 2026); packaging is booked through 2027 (CoWoS packaging is booked through 2027); server DRAM rose 60 to 70 percent in one quarter (Server DRAM prices rose 60 to 70 percent in one quarter). Data centres are projected to take up to 70 percent of global memory supply in 2026.

What the frame explains

Why AI server prices rose in 2026 well beyond the increase in memory content, tightening rental supply regardless of GPU or power availability (The OEM layer). Why NVIDIA cut gaming GPU production 30 to 40 percent. And why operators who signed fixed-rate contracts before the repricing captured a windfall they did not underwrite (Take-or-pay capacity contracts).

It also complicates the depreciation argument: an accelerator whose bill of materials has risen is not obviously cheaper to replace than to re-lease (Residual value as the hinge).

Known weakness

Memory cycles turn. DRAM has never been sustainably sold out, and if hyperscaler capex growth normalises, HBM pricing weakens before capacity slack appears — a faster correction than anything in the power supply chain (The capital cycle).

Links: The memory and foundry layer, Scarcity rent, Density as an obsolescence clock.

SemiAnalysis segments GPU supply into four tiers, and the segmentation predicts behaviour better than revenue rankings do.

The gap between tiers is priced. Hyperscaler list prices run three to six times the lowest neocloud listings for the same silicon, which is why the mid-market exists at all. Within the neocloud tiers, ClusterMAX tiering does the sorting: the tier a provider lands in maps onto a pricing premium (Platinum rating carries a large pricing premium).

The tier boundaries are moving in both directions. Groq moved from custom inference silicon into being an Nvidia cloud partner. The The bitcoin-miner conversion cohort entered from underneath, owning power rather than software. And the hyperscalers are becoming customers of tier two rather than purely competitors — Hyperscalers rent because they cannot build fast enough.

What the table hides is that tier two's largest customers are tier one. That is not a competitive market structure; it is a supply chain, and it is the reason Two-sided concentration risk matters more here than in ordinary cloud.

The table also omits the layer underneath: whoever owns the building and the interconnect (The landlord layer).

A fifth category sits outside the table entirely: state-funded and self-built capacity that competes for the same GPUs, turbines and customers without competing for the same returns (The global map, xAI).

And a routing layer now sits above all of it — The DGX Lepton marketplace — which if it works makes tier membership matter less than it does today (Who wins if the marketplace layer works).

The count is large and shrinking. SemiAnalysis reviewed eighty-five providers in detail against a market view of over two hundred; the consensus expectation is that Only five to seven independent GPU clouds survive.

Sources

SemiAnalysis ClusterMAX 2.0 · ModulEdge

What it is — the reference case for the whole category: a crypto miner that became the only Platinum-rated GPU cloud. Model — leveraged infrastructure owner, premium tier. Tension — the best operator in the sector also carries its heaviest balance sheet.

Founded 2017, IPO'd March 2025 at roughly a 23 billion dollar valuation, added to the Nasdaq-100 in 2026. Q2 2026 revenue of 2.575 billion dollars, up 112 percent, with 1.5 GW of active power and guidance of 12.4 to 13.2 billion for the full year against capex of 31 to 35 billion.

What it does better than anyone

It is the only provider in SemiAnalysis's Platinum tier, two ratings in a row, scoring on security, health checks, networking and orchestration rather than price — ClusterMAX tiering, Platinum rating carries a large pricing premium. It charges accordingly (What a GPU-hour costs in 2026) and wins the workloads where a failed training run costs more than the rate card (Reliability is the moat, not price).

The balance sheet

Roughly 35 billion dollars of debt at end-June 2026. Net interest expense of 640 million in the quarter, annualising above its entire 2025 revenue (CoreWeave's annualised interest cost exceeds its 2025 revenue). Net loss of 626 million. Its CDS has priced roughly even odds of five-year default (CoreWeave's credit default swaps priced roughly even odds) and it repriced a 2.6 billion leveraged loan wider shortly before Q2 earnings. It has borrowed in high yield at around nine percent (CoreWeave borrowed at around 9 percent in high yield) and raised over ten billion in unsecured and convertible debt including a first Eurobond.

Customers

Microsoft was about 67 percent of 2025 revenue (Microsoft was two-thirds of CoreWeave's 2025 revenue). The backlog has diversified toward OpenAI at around 22.4 billion and Meta at up to 35.2 billion, reaching roughly 104 billion (CoreWeave's backlog reached 104 billion dollars) — but see Backlog is not revenue and Two-sided concentration risk.

What it tried and failed to do

The 9 billion dollar all-stock acquisition of Core Scientific, which would have added about 1.3 GW, was voted down by Core Scientific shareholders in October 2025. It now pursues power organically and through partnerships including Applied Digital and Galaxy. That failure is the clearest evidence for Vertical integration is the only durable moat being a live strategic question rather than marketing.

Its funders are now also its suppliers and landlords: Nvidia put in a further $2bn against a 5 GW-by-2030 target, Blue Owl committed $4bn to a Pennsylvania joint venture, and Galaxy Digital delivers 800 MW on a fifteen-year lease (Nvidia invested a further 2 billion dollars in CoreWeave).

Links: Take-or-pay capacity contracts, GPU-collateralised lending, Nebius, The landlord layer.

What it is — an energy company that builds data centres, sold as a cloud. Model — vertically integrated, power first. Tension — the valuation assumes a fourfold revenue jump on construction schedules.

Founded 2018 by Chase Lochmiller and Cully Cavness to burn flared gas for compute, divested bitcoin mining to NYDIG, and is now the developer of the flagship Stargate campus at Abilene, Texas for OpenAI and Oracle. Revenue of 276 million in 2024, 998 million in 2025, guided to roughly 2 billion in 2026 (Crusoe grew from 276 million to a projected 2 billion dollars). Series E of 1.375 billion at over 10 billion in October 2025; reported in mid-2026 to be raising around 3 billion at roughly 30 billion.

The actual differentiator

Not GPUs. Crusoe co-develops power and shell rather than treating them as sequential — Crusoe's co-developed power and shell — and manufactures its own switchgear and modular cooling to route around supply bottlenecks that stop everyone else (Substation transformer lead times passed 160 weeks). Abilene went from groundbreaking to operational in under a year, running ten simple-cycle turbines supplying roughly 360 MW behind the meter (Behind-the-meter generation).

Scale

4.9 GW contracted against a pipeline above 45 GW (Contracted power is not connected power). Abilene reaches roughly 2.1 GW including a 900 MW campus announced for Microsoft in March 2026. Project Jade in Laramie County, Wyoming was approved at 2.7 GW initially with plans contemplating 10 GW — nearly three times the current power demand of the state.

How to read it

Crusoe's own argument is that it should not be valued as a neocloud at all but as a contracted power-and-compute utility — Compute is a utility asset, not a technology asset. The evidence broadly supports the claim: its binding constraint is turbines and permits, not chip cycles, which makes it less exposed to Residual value as the hinge than CoreWeave or Lambda. It is more exposed to construction risk and to any single anchor tenant walking.

Crusoe is also the only provider in this vault with AMD MI300X and MI355X on a public rate card, which makes it the visible test of Does a second source change the price.

Links: The four-tier market structure, Nscale, Private credit funds.

Sources

Crusoe Series E · Sacra · AI Grid Insider

What it is — the European-domiciled listed challenger, spun out of Yandex's non-Russian operations and relisted on Nasdaq in 2024. Model — contracted capacity for hyperscalers, funding the AI-native business. Tension — the fastest margin inflection in the sector, on the least operating history.

Q2 2026 revenue of 582.3 million dollars, up 454 percent; AI cloud ARR around 3 billion. Contracted revenue above 40 billion from investment-grade counterparties. Targeting 2.5 GW contracted and 800 MW to 1 GW connected by end-2026, from 170 MW at end-2025 (Contracted power is not connected power).

The contracts

Microsoft at up to 19.4 billion dollars over five years signed September 2025. Meta at 3 billion, then extended in March 2026 to 12 billion of dedicated capacity plus up to 15 billion of additional purchases — up to roughly 27 billion in total (Meta committed up to 27 billion dollars to Nebius), on one of the first large-scale Vera Rubin deployments. Nvidia holds an investment.

The financing move worth copying

Nebius has converted an operational contract into asset-level financing covering more than 100 percent of the capex required to deploy the underlying GPUs (Contract-level asset financing). That is a materially better structure than lending against the chips themselves (GPU-collateralised lending), because it borrows against an investment-grade cash flow rather than a depreciating asset. Its stated intent to repeat it across other deployments is the most important financing development in the sector this year.

The stated strategy

Co-founder Roman Chernin has said the hyperscaler deals are signed with margin, not revenue, in mind — the large contracts finance the rest of the business, whose customers are AI natives like Mistral, Cursor and Black Forest Labs. That is a direct answer to Neoclouds are brokers unless they reach the enterprise.

Nebius depreciates GPUs over four years, the shortest schedule among the giants (Operators depreciate GPUs over four, five and six years) — a conservative choice that makes its reported margins harder to dismiss (Nebius's AI segment margin swung 38 points in a year).

Links: CoreWeave, The customer is becoming the competitor, Take-or-pay capacity contracts.

Sources

Nebius 6-K, Meta agreement · Nebius 6-K/A, asset financing

What it is — an inference platform, not a capacity owner. Included here because it is routinely grouped with the neoclouds and is not one. Model — asset-light, rents from more than twenty suppliers. Tension — a software multiple on a business with hardware in cost of goods sold.

Founded by ex-PyTorch engineers. Series D of 1.505 billion dollars at 17.5 billion in July 2026, led by Atreides, Index and TCV with Nvidia participating — up from a 4 billion valuation nine months earlier. ARR passed 1 billion, roughly five times year-on-year, serving over 40 trillion tokens a day for more than 10,000 companies including Cursor, Notion, Uber, DoorDash and Shopify.

Why it belongs in a capacity graph

Because it consumes capacity without building it. Fireworks draws compute from more than twenty suppliers (Fireworks draws capacity from more than 20 suppliers) and has renewed and expanded contracts with IREN among others. It is the demand side of the buildout described in Renting capacity you do not own, and the clearest live test of Asset-light beats asset-heavy.

The margin question

Gross margin sits around 50 percent, well below the 70-plus percent typical of software, because GPU cost sits in cost of goods sold (Fireworks runs at roughly half the gross margin of software). Investors have been explicit that this is the risk: competitors that own compute, or labs and hyperscalers that do not lease it, do not carry that line. One quoted investor put it as VCs doing a revenue multiple and assuming the margin does not matter.

What it actually sells

Not raw GPUs. Fine-tuning and customisation depth — over 95 percent of its tokens come from models specialised on customer data — plus a disaggregated inference engine and day-zero support for new open-weight releases. Its nearest competitors are Together AI and Baseten, with Modal adjacent. The strategic bet is that open-weight models reach parity and enterprises want to own their intelligence rather than rent it from a frontier lab.

Links: Five business models under one label, The GPU cascade into inference, Groq.

Sources

Sacra · Qz · Newcomer

Supplier, shareholder, lender, guarantor, revenue-share participant and, through GroqCloud licensing and cloud partnerships, occasionally a competitor. No single entity has more influence over which neoclouds exist.

The four roles

Seller. Every operator in The four-tier market structure buys from it. Investor. Equity in CoreWeave disclosed at IPO, in Nebius, Nscale, Fireworks AI, and a right to invest up to 2.1 billion dollars in IREN. Guarantor. Up to 860 million dollars of Nscale's lease obligations (Nvidia guaranteed 860 million dollars of Nscale's leases); utilisation floors and rent-back commitments to smaller operators (Nvidia demand backstops and revenue share). Tenant. Roughly 1.5 billion dollars of arrangements renting capacity back from Lambda (Sale and leaseback with the chip vendor).

The 2026 programme, and its wobble

In July 2026 Nvidia introduced a compute-access financing programme reported at roughly 36 billion dollars, backstopping neocloud GPU purchases in exchange for a share of rental revenue (Nvidia's compute-access programme is reported at 36 billion dollars). By late August, the Wall Street Journal reported that some revenue-share deals were paused after partner pushback over a clause giving Nvidia approval rights over which end-customers a neocloud could serve, plus internal antitrust concern. Nvidia's public line was that the model remains in place and continues to evolve.

Separately, Nvidia signed memoranda with a consortium of Private credit funds — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — to move the capital commitment off its own balance sheet (The Nvidia financing consortium). Jensen Huang's framing was that the demand is real and the capital is not Nvidia revenue.

The fifth role: broker

The DGX Lepton marketplace routes buyers to whichever partner has capacity, at that partner's price. Its stated effect is lower cost per unit of performance for end users; its structural effect is to make partners interchangeable (Nvidia is commoditising its own partners). Whether it is succeeding is disputed (The marketplace layer is failing).

How to read it

Nvidia's position is the strongest argument both for and against the sector. For: it has better information than anyone about demand and is putting capital behind it. Against: a vendor that finances its own customers' purchases and takes a share of their revenue is getting paid twice on the same silicon while standing behind the loan, which is the structure examined in Vendor financing, historically and argued in Circular financing is a permanent feature.

Links: Two-sided concentration risk, Custom accelerators are captive and cannot be rented.

Sources

CNBC · Capacity

• Microsoft accounted for approximately 67 percent of CoreWeave's FY2025 revenue.
• CoreWeave's SEC filings indicate no single customer accounts for more than 35 percent of the backlog — a disclosure that confirms diversification while still implying heavy reliance on a handful of accounts.
• The backlog has since diversified toward OpenAI at around 22.4 billion dollars and Meta at up to 35.2 billion.

Why it matters here

It is the canonical instance of Two-sided concentration risk and the empirical core of Neoclouds are brokers unless they reach the enterprise. A supplier deriving two-thirds of revenue from one customer is a contract manufacturer with a cloud interface.

The trajectory is the encouraging part: the shift from a single customer to three very large ones is real diversification in credit terms, even if all three remain hyperscale AI buyers exposed to the same demand cycle. That correlation is why the The Meta Compute report cut neocloud equities by a sixth in a day hit CoreWeave and Nebius simultaneously.

Sources

Sacra · Tech Insider

The mechanic. A neocloud has one dominant supplier and a handful of dominant customers, and the same firms frequently appear on both sides.

Supply side. Nvidia sells the chips, holds equity, guarantees leases, provides utilisation floors, takes revenue share and — through the paused end-customer approval clause — sought a say in who the cloud may serve (Nvidia's compute-access programme is reported at 36 billion dollars).

Demand side. Microsoft was about two-thirds of CoreWeave's 2025 revenue (Microsoft was two-thirds of CoreWeave's 2025 revenue) and is most of Nscale's and IREN's flagship contracts. Meta is up to 35.2 billion at CoreWeave and 27 billion at Nebius.

Both sides at once. Nvidia rents capacity from Lambda and sells cloud services through IREN. Google backstops Fluidstack's debt rather than buying its capacity. Meta is a customer building a competitor (The customer is becoming the competitor).

Why this is a distinct risk, not just two risks

Because the correlations are hostile. The event that reduces customer demand — an AI capex digestion — simultaneously reduces the value of the collateral (GPU-collateralised lending), the willingness of the vendor to keep backstopping, and the credit of the guarantor. A firm exposed to one supplier and five customers in one end-market has no diversification anywhere in its structure.

The market has priced episodes of this precisely: a 25 percent swing in Hut 8 and Riot Platforms on the identity of a lease counterparty, and a double-digit fall across the sector on a Bloomberg report about Meta.

The systemic version

One circle of institutions now owns the buildings, lends against the GPUs and backstops the chips: the consortium that bought Aligned Data Centers for about $40bn (Aligned sold for about 40 billion dollars to an AI infrastructure consortium) is substantially the same set Nvidia recruited for compute financing (The Nvidia financing consortium, The sponsor-backed platform roll-up).

The mitigations that exist

Contract duration (Take-or-pay capacity contracts), investment-grade counterparties (Contract-level asset financing), and diversifying into the inference layer — IREN selling to Together AI and Fireworks AI is a small but real example.

Links: Where neoclouds struggle, Duration mismatch, Vendor financing, historically.

The load-bearing bull argument for the whole sector, and the one with the best evidence behind it.

Microsoft has committed over sixty billion dollars to neocloud partnerships (Microsoft has committed over 60 billion dollars to neoclouds) despite effectively unlimited capital. Meta has committed roughly forty-eight billion across CoreWeave and Nebius while guiding to 115 to 145 billion of its own capex. When companies of that size buy capacity from two-year-old startups, the constraint is evidently not money.

Assessment

The mechanism is documented rather than inferred. Building 240 MW from grid connection to operation takes a traditional operator several years; Nscale and Crusoe compress that, Crusoe to under twelve months at Abilene (Crusoe's co-developed power and shell). The queue is the constraint (Only a third of 2026 US capacity is under construction).

There is a second, less discussed motive: renting converts capex into opex (Capex-to-opex arbitrage), which matters to a company whose capital spending is under investor scrutiny. That motive is durable in a way the physical one is not — it survives even if construction speeds up.

Where it weakens

The argument implies a temporary condition. If the constraint is speed, then capacity leases are a bridge, and the neoclouds' terminal value depends on either locking in long duration before the bridge closes — hence contracts running to 2031 and 2032 — or reaching customers hyperscalers do not serve (Neoclouds are brokers unless they reach the enterprise). Meta Compute is what the other side of the bridge looks like (The customer is becoming the competitor).

Links: Take-or-pay capacity contracts, Where neoclouds win, Backlog is not revenue.

On 1 July 2026, Meta was reported to be building Meta Compute to sell surplus AI capacity. CoreWeave fell about 14 percent and Nebius about 17 in a session (The Meta Compute report cut neocloud equities by a sixth in a day). The company that had committed roughly 48 billion dollars to those two providers may sell against them.

The pattern is not isolated. SpaceX has marketed Colossus capacity. Google backstops Fluidstack rather than buying capacity outright. Nvidia rents capacity back from Lambda and sells cloud services directly through IREN. Every large counterparty in this graph occupies more than one side of the trade.

Assessment

The structural driver is indivisibility. Hyperscale capacity arrives in large blocks timed to demand forecasts, so surplus is a normal state rather than a strategic choice. That means the supply of competitive capacity is a function of other people's forecasting errors, which is a bad thing to have as your marginal competitor because it prices at whatever recovers something rather than at a required return.

The counter, from analysts at the time: the AI infrastructure market can support a fourth source of bare-metal capacity but not a fourth full-stack cloud, and raw compute resale does not touch what CoreWeave sells at the Platinum tier (Reliability is the moat, not price). Meta has no cloud operations organisation, no enterprise sales motion for infrastructure, and a 2026 in which its own capacity was reportedly constrained.

The honest summary: this is a margin risk rather than an existence risk, and it arrives precisely when neoclouds most need margin (Post-depreciation gross margins run 14 to 16 percent).

Links: Two-sided concentration risk, Where neoclouds struggle, Merchant versus contracted capacity.

In most of this sector's largest contracts, the neocloud does not meet the end user.

Nscale supplies GPU infrastructure that Microsoft uses to serve Azure customers — the neocloud provides the hardware layer and the hyperscaler provides the customer relationship. Cipher Digital provides turnkey space and power to AWS. Fluidstack holds the relationship while Google holds the credit and TeraWulf holds the site. Fireworks AI holds 10,000 customer relationships and rents the compute underneath from twenty suppliers.

Why the question is not academic

Because whoever owns the relationship captures the margin when capacity stops being scarce. In a shortage, the scarce layer wins regardless of position (Scarcity rent). Afterwards, value accrues to whoever the customer thinks they are buying from.

The Anthropic pattern is the sharpest version: labs leasing power and shell directly from converted miners, bringing their own hardware and orchestration. That reduces the neocloud's role to nothing at all.

The three answers currently in play

1. Own the customer. Nebius's stated strategy — hyperscaler contracts finance an AI-native customer base of its own.
2. Own the asset. Crusoe and the converted miners: never mind the relationship, own the thing nobody can build (Vertical integration is the only durable moat).
3. Own the software. The inference layer's bet (Can an inference platform survive without owning compute).

Each is defensible. What is not defensible is the middle — renting colocation to resell GPUs to a hyperscaler, which is Neoclouds are brokers unless they reach the enterprise with no way out.

Links: Five business models under one label, Two-sided concentration risk, Where neoclouds struggle.