The financialization of compute

Compute means the processing capacity that trains and runs artificial intelligence models. In practice it is a bundle of four things: chips (mostly graphics processing units, or GPUs, made by Nvidia and increasingly AMD and others), the memory and networking around them, a building with cooling, and a large, reliable supply of electricity. A company that owns this bundle can use it or rent it out by the GPU-hour, which is the price of one GPU for one hour.

Oil is priced per barrel and electricity per megawatt-hour. Compute is increasingly priced per GPU-hour for a named chip, for example an Nvidia H100 or B200. The chip model works like a grade of crude: buyers know roughly what they get. The units are not perfectly interchangeable (cluster quality, networking and location vary), which is why index providers try to normalise for performance. Silicon Data says its SiliconMark product is meant to address performance differences between clusters built from identical chips.

Large AI deals are now announced in power terms. OpenAI has announced plans to deploy at least 10 gigawatts (GW) of Nvidia systems and 6 GW of AMD GPUs, and in 2025 Anthropic committed to buy at least $30 billion of Microsoft Azure capacity including 1 GW of servers. One gigawatt is roughly the output of a large power plant, and it is the unit lenders and utilities now use because power is the hardest input to add quickly.

The International Energy Agency (IEA) estimated global data-centre electricity use at about 415 terawatt-hours (TWh) in 2024, roughly 1.5% of world electricity, and projected about 945 TWh in 2030 in its base case, just under 3%. The United States and China together account for nearly 80% of the growth to 2030, and US consumption is projected to rise by about 240 TWh (130%) over 2024 to 2030. The 2035 range runs from about 700 TWh to more than 1,700 TWh depending on how fast demand and efficiency evolve.

A commodity becomes financeable when supply cannot respond instantly to demand. New GPUs depend on a small number of manufacturers and on memory supply. Amazon cited higher memory costs when it raised its 2026 capex outlook to about $220 billion on its 30 July 2026 call, as summarised by datacentres.com. New data-centre campuses depend on land, grid connections and equipment lead times. The result is that contracts, not spot purchases, secure most capacity: customers sign multi-year commitments, and those commitments become the cash flows that lenders lend against.

Silicon Data chief executive Carmen Li said in August 2026 that A100 and H100 rental prices had risen about 20% since January and had been broadly flat for the previous 20 days, and that the forward curve sloped upward, meaning buyers paid a premium to lock in longer terms (reported by The Next Web, 13 August 2026). A rising rental price on a chip several years old is evidence that older GPUs still have use, which matters for the depreciation debate in chapter 04. It is one data point from one index provider, not a settled fact.

The last row matters. The proposed CME contracts settle in cash against an index, rather than delivering GPU time, partly because the product is not yet uniform enough for physical delivery. Silicon Data has said that normalising for performance could later allow physical delivery.

• GPU: a chip built for the parallel arithmetic that AI needs.
• Hyperscaler: one of the very large cloud providers, here Amazon, Microsoft, Alphabet (Google) and Meta, plus often Oracle.
• Neocloud: a newer cloud company focused on GPU capacity, such as CoreWeave, Nebius, Lambda or Crusoe.
• Capex: capital expenditure, money spent on long-lived assets such as chips and buildings.
• Take-or-pay: a contract in which the customer pays whether or not it uses the capacity.

The four largest US cloud spenders are Amazon, Microsoft, Alphabet and Meta. Their 2026 capex guidance, as compiled by datacentres.com from the companies' July 2026 releases and calls, was about $220 billion for Amazon, $195-205 billion for Alphabet, $130-145 billion for Meta and about $175 billion for Microsoft (later reported as about $170 billion after its 28 October 2026 results). The combined total is roughly $720-745 billion, but definitions differ: Microsoft counts finance leases, Meta adds finance-lease principal, and Amazon nets some incentives, so the sum is indicative only.

Free cash flow is operating cash flow minus capex: what is left after investing. A secondary summary of the September 2026 quarter results (useluminix.com, which says it did not re-check filings) shows the pattern below. Alphabet and Meta spent more than they generated from operations in the quarter, and Amazon's trailing twelve months free cash flow was reported as about negative $42 billion. Treat these as secondary figures to verify against the 10-Q filings.

Amazon is omitted from the chart because its quarterly operating cash flow was not available in the same source. It reported net cash capex of about $75.9 billion for the quarter. UBP estimated 2026 capex of about $820 billion against roughly $750 billion of operating cash flow for the group, as cited in the same compilation; that is an estimate, not company data.

J.P. Morgan Asset Management figures, as cited by useluminix.com, put gross bond issuance by five hyperscalers at $17 billion in 2024, $109 billion in 2025 and $194 billion in the first half of 2026. Different data providers give different totals because of currency and definition choices: Vanguard's series shows about $93 billion in 2025, and Capital Group counts $240.7 billion year to date through 31 August 2026 including $66 billion outside US dollars. Reuters, citing BNP Paribas, put the five-company total at about $220 billion through 10 August 2026.

• Alphabet: $20 billion of US dollar notes plus $31.8 billion of foreign-currency notes during 2026, per its 30 June 2026 10-Q (via MLQ).
• Meta: $25 billion of senior unsecured notes in six series maturing 2031 to 2066, May 2026 (via MLQ citing the filing).
• Amazon: sought $25 billion across eight tranches maturing 2029 to 2066 in July 2026 (Bloomberg via Investing.com); the target, not a confirmed total.
• Oracle: reported plan of about $20 billion of debt inside a $40 billion fiscal 2027 funding plan (FactSet); its February filing separately called for $45-50 billion of gross proceeds in calendar 2026 using debt and equity.

Investors are asking for a little more to hold this debt. Reuters data (via MLQ) shows the median spread of hyperscaler bonds over comparable Treasuries rose from 30 to 40 basis points for 2 to 4 year bonds, from 50 to 60 for 5 to 7 year bonds and from 108.5 to 118 for 20-year-plus bonds between 2025 and 2026. A basis point is one hundredth of a percentage point. Reuters also found that 78 of 91 hyperscaler bonds issued in 2026 traded at higher yields on 28 July than at issue, with a median rise of about 22 basis points.

A special purpose vehicle (SPV) is a separate legal entity that owns an asset and borrows against it. Meta's Hyperion campus in Louisiana is the best-documented example. According to a 21 July 2026 industry write-up and a separate compilation, Blue Owl holds 80% of the joint venture (about $23.0 billion of equity) and Meta holds 20% ($5.8 billion). The vehicle issued about $27.3 billion of notes (Beignet Investor LLC, due 2049, rated A+, priced about 225 basis points over Treasuries). PIMCO anchored about $18 billion and BlackRock took more than $3 billion. Meta leases the campus and provides a payment guaranty and a residual value guarantee, a promise to cover outstanding debt at the end of each renewal term if the property is worth less. The lease obligation is reported to begin on 1 June 2029.

The structure keeps the debt off Meta's balance sheet while leaving real obligations behind: Moody's, in February 2026, put the undiscounted data-centre lease commitments of five companies at $969 billion at end-2025, with about $662 billion not yet commenced (secondary citation). A Financial Times analysis in December 2025 reportedly found more than $120 billion of debt and equity placed in SPVs for Meta, xAI, Oracle and CoreWeave.

Oracle is the large builder with the weakest credit among the group. A compilation of filings and agency actions (useluminix.com) reports that S&P cut Oracle to BBB- on 9 July 2026, one notch above sub-investment grade, with Moody's at Baa2 on negative outlook. Oracle's 31 August 2026 10-Q reportedly shows $125.0 billion of senior notes and borrowings, and about $288 billion of data-centre leases not yet on its balance sheet (up from $260 billion at 31 May). S&P forecast a fiscal 2027 free operating cash flow deficit of nearly $42 billion. S&P also said OpenAI accounts for about half of Oracle's roughly $638 billion of remaining performance obligations (contracted future revenue). These are secondary citations of agency and filing data.

Private credit, meaning loans by non-bank funds, plays a growing part. Morgan Stanley estimated in October 2025 that roughly $1.5 trillion of external financing would be needed, with up to $800 billion from private credit in asset-backed deals by 2028. The Bank for International Settlements counts more than $200 billion of private loans to AI-related companies, per the same compilation. These are estimates, and the sources define their categories differently.

Neoclouds are specialist companies that buy GPUs in bulk and rent them out. They have less cash flow history than the hyperscalers and need financing to buy chips before revenue arrives. The financing pattern is consistent: sign a multi-year take-or-pay contract with a creditworthy customer, then borrow against the GPUs and the contract together. Lenders look at the customer's credit more than at the neocloud's.

CoreWeave reported 2025 revenue of $5.1 billion and first quarter 2026 revenue of about $2.1 billion, up 112% year on year, with a net loss of $740 million and adjusted EBITDA of $1.2 billion (a 56% margin), according to a TIKR summary of its results published 12 June 2026. Depreciation and amortisation was $1.15 billion and net interest expense $536 million in that quarter. Press coverage of the second quarter of 2026 reports revenue of about $2.575 billion and interest expense of about $640 million (versus $267 million a year earlier). The adjusted operating margin reportedly fell from 17% to 1% between the first quarters of 2025 and 2026, showing how depreciation and interest absorb the gross profit.

Backlog is contracted future revenue not yet recognised. TIKR reports that about 36% was expected to convert within 24 months and about 75% within four years as of 31 March 2026. A backlog is not cash: it is a promise to pay, and it is only as good as the customer.

A TokenPost article dated 27 October 2026 reports CoreWeave debt of $35.6 billion and cash of $5.5 billion at 30 June 2026, plus about $4.2 billion of debt added afterwards. In the first quarter of 2026 CoreWeave closed an $8.5 billion investment-grade term loan and in June announced a $3.5 billion senior notes offering (TIKR). Capex guidance for 2026 was raised from $30-35 billion to $35-39 billion in the second quarter coverage. Some outlets quote far lower debt figures; those appear to be stale or use a different definition, and the filings should be checked.

• Take-or-pay contracts were reported at 98% of second-quarter revenue.
• In Q1 2026 one customer was 45% of revenue and a second was 20% (TIKR). In Q1 2025 one customer, widely believed to be Microsoft, was 72%.
• Second-quarter coverage reports that three customers made up 72% of revenue, including Microsoft and OpenAI.
• Full-year 2026 revenue guidance was reportedly raised to $12.4-13.2 billion.

Nebius announced on 17 July 2026 its first senior secured facility of about $775 million, backed by deployed GPU infrastructure and contracted cash flows from an investment-grade customer, priced at SOFR plus 2.50% and maturing 31 October 2030. SOFR is the benchmark short-term US dollar interest rate. The company said the facility covers more than 100% of the capex needed to deploy those GPUs, that it was significantly oversubscribed, and that Microsoft and Meta have more than $40 billion of additional contracted revenue with Nebius. MUFG was the structuring agent and bookrunner (Nebius press release, 17 July 2026). Nebius has described the structure as repeatable.

Lambda reportedly closed a $926 million senior secured term loan B, rated Baa2 by Moody's, backing GPU deployment for an investment-grade customer (StorageNewsletter, 22 September 2026), and another report of a roughly $1 billion fixed-rate loan tied to Microsoft contracts exists but comes from a single outlet. Lambda is reported to be preparing a listing, with a pre-IPO round of about $3 billion in discussion at a valuation near $12 billion; that is unconfirmed. Crusoe reportedly began with $200 million of GPU-collateralised borrowing in 2023 and later raised $9.6 billion of JPMorgan-led debt plus $5 billion of equity for its Abilene campus, with talks in July 2026 about a raise at a valuation near $30 billion (secondary reports).

A source abbreviated CCIR was reported to count about $63 billion of commitments secured by GPUs, and Nvidia's widely reported $500 billion financing initiative is described as a target with only about $3.5 billion of guarantees actually in force (compilation by useluminix.com, with the caveat that several figures are secondhand).

Depreciation spreads the cost of an asset over its estimated useful life. If a chip costs $10 billion and is written off over six years, the annual charge is about $1.67 billion; over three years it is about $3.33 billion. The estimate is a management judgment, and it moves reported profit directly.

• Amazon: in early 2025 it shortened the life of a subset of servers from six years to five, citing the pace of AI hardware, per its FY2024 10-K. The change reportedly cost about $700 million of operating income, with $920 million of accelerated depreciation (Dave Friedman analysis of the filing).
• Meta: in early 2025 it extended server lives to 5.5 years, which reportedly reduced depreciation expense by about $2.9 billion, roughly 4% of estimated pre-tax profit (secondary report; verify against the filing).
• Amazon in 2020 moved from three to four years, and by 2023 the largest cloud providers had converged on about six years for servers, according to analysts' summaries.
• Google reported increasing server life to six years (DCD report); Microsoft's current figure was not verified for this piece.

Investor Michael Burry argued publicly in November 2025 that real GPU lives are closer to two to three years, because Nvidia now releases a new generation roughly every year. His claim is that this understates depreciation and overstates profit by about $176 billion across 2026 to 2028. That is one critic's model, not a finding; a commentator who reviewed it noted its supporting arithmetic was not shown in full and that one of the illustrative figures it used was internally inconsistent.

Defenders say a GPU does not retire when a new chip arrives. It moves from training the newest models to serving existing models (inference) and to less demanding work, so its useful economic life can exceed its technical leading-edge life. The August 2026 observation that H100 and A100 rental prices were up about 20% since January is consistent with that argument, but a few months of rising rents on 2022 and 2023 chips does not prove six-year lives for 2025 and 2026 chips.

A loan secured by GPUs is only as safe as the chip's value at the end of the loan. If lenders assume a six-year life and the true life is three, the collateral is worth less than modelled when needed. This is why Meta's Hyperion structure includes a residual value guarantee: the lessee, not the lender, takes the risk that the asset is worth less at the end of a renewal term. Moody's data shows large lease commitments that have not yet begun, which means the residual value question is deferred, not removed.

Accounting treatment also changes the optics. A finance lease puts the asset and obligation on the balance sheet and counts in capex; an operating lease keeps most of it out of the capex line. A July 2026 compilation reports that Microsoft's capex guidance fell from about $190 billion to about $175 billion after a lease accounting change, with underlying plans unchanged, which shows that headline capex can move because of classification alone.

The residual value of old GPUs is the least-observed number in this market. There is no mature secondary market with transparent prices, and the compute price indices described in chapter 05 are rental prices, not sale prices. Until a liquid resale market exists, lenders and accountants are using assumptions.

A last point is that the cost of being wrong is asymmetric. If lives prove longer than assumed, profits are understated for a time and nothing breaks. If they prove shorter, companies must either take impairment charges (write-downs) or accelerate depreciation, and any lender relying on collateral value or a guarantee faces a loss. Amazon's 2025 change shows a company can shorten a life without a crisis; the test comes when the shortening is large and sudden.

A financial market needs a reference price. For compute, several firms now publish them, and exchanges are building contracts on top. The picture is early and contested: the same chip can show prices several times apart depending on what is measured.

A spot marketplace listing is a single provider's price for interruptible capacity. A median on-demand price is across 49 providers. A hyperscaler on-demand band is a premium service with reliability and support. Silicon Data's own neocloud and hyperscaler indices show about a threefold gap. The roughly eightfold spread across trackers therefore compares different products, not a failure of one tracker. Silicon Data's hyperscaler series was also nearly unchanged over the March to April 2026 window (39 of 50 daily readings unchanged, net decline of 0.27%), which suggests list prices are sticky.

CME Group and Silicon Data announced in May 2026 and detailed on 11 August 2026 two cash-settled futures: Silicon Data H100 Rental Index Futures and B200 Rental Index Futures, listed on NYMEX, targeted to start trading on 5 October 2026 pending regulatory review. Each contract represents 730 GPU-hours (about one month of one GPU), quoted in dollars and cents per GPU-hour with a $0.01 minimum move, worth $7.30 per tick (CME Group release, 11 August 2026 via search summary). Whether trading actually began on that date was not confirmed in the sources reviewed for this piece.

• A data-centre operator could sell futures to lock in future rental revenue.
• An AI developer could buy futures to cap future compute costs.
• A lender could use the contracts to hedge the rental price underlying a GPU-backed loan, if liquidity develops.
• Basis risk is the gap between the index and the actual contract a user holds; with trackers disagreeing, it is large.

ICE and the data firm Ornn were reported in May 2026 to be working on futures based on the Ornn Computing Power Price Index (OCPI), which tracks traded spot prices across H100, H200, B200 and RTX 5090 chips. Sources conflict on whether those contracts have launched, so the status is unverified. Reports also say Architect's AX exchange runs perpetual futures on GPU and memory rental rates, and Kalshi lists contracts referencing an Ornn index (secondary reports, not confirmed with the venues).

Ornn's index was reported to show H100 rates falling from $2.94 on 25 May 2026, and a later Kalshi-related report put H100 at $1.70, while Silicon Data's neocloud index showed $2.53. These are not the same measures and should not be read as a consistent time series.

Silicon Data's statement that the curve slopes upward means buyers pay more for longer commitments, which would be unusual for a commodity whose supply is expanding fast. It could reflect scarcity, or it could reflect immature price discovery with few participants. Forward curves on thinly traded new products are often driven by a handful of trades.

In circular financing, a company invests money in a customer who then spends it on that company's products or services. Revenue is real, but part of the demand is funded by the seller. The AI industry has several such loops, and they matter for lenders because they make demand look more independent than it is.

• Nvidia and OpenAI (September 2025): Nvidia said it intends to invest up to $100 billion in OpenAI as gigawatts come online, under a plan to deploy at least 10 GW of Nvidia systems. Reports describe the agreement as a letter of intent rather than a final contract. A later compilation says Nvidia holds a stake of about $30 billion in OpenAI and has provided a guarantee capped at $105 billion on an Ohio project that pays only if OpenAI defaults; verify against Nvidia's filings.
• AMD and OpenAI (October 2025): AMD issued OpenAI warrants for up to 160 million AMD shares, about 10% of the company, vesting in tranches as OpenAI deploys 6 GW of AMD Instinct GPUs; reports say the warrant price is $0.01 per share and that one tranche requires AMD's stock to reach $600.
• Microsoft, Nvidia and Anthropic (18 November 2025): Nvidia committed up to $10 billion and Microsoft up to $5 billion in Anthropic, and Anthropic committed to spend at least $30 billion on Microsoft Azure capacity, including up to 1 GW of Nvidia Grace Blackwell and Vera Rubin systems (Capital Brief).
• AMD and Anthropic: a report dated 22 July says AMD would invest up to $5 billion in Anthropic, with Anthropic committing to deploy 2 GW of AMD MI450 GPUs, with the first gigawatt in the first half of 2027; the year was not stated in the excerpt reviewed, so this is unverified.

Microsoft has invested more than $13 billion in OpenAI and holds about 27% of its for-profit arm, and OpenAI separately has a $38 billion cloud agreement with Amazon (Capital Brief, November 2025). OpenAI is also the largest customer behind Oracle's contracted backlog: S&P said OpenAI accounts for about half of Oracle's roughly $638 billion of remaining performance obligations. A lender to Oracle is therefore indirectly lending against OpenAI's ability to pay. The compilation describes this as a lease chain where OpenAI is the economic user, Oracle signs the lease and lenders underwrite Oracle.

The Financial Times, confirmed by Reuters according to the compilation, reported internal OpenAI projections of $278 billion of negative free cash flow from 2026 to 2030, and that a $122 billion fundraising in March 2026 was projected to run out by 2028. These are reported projections, not company filings. The point is structural: many of the largest contracts rest on customers who are themselves burning cash and depend on new capital.

If a supplier funds its customer's purchases, a downturn in the customer's funding hits the supplier's revenue and its investment at once. If a lender finances the data-centre that the customer will rent, a fall in the customer's funding hits the lease. The risk is correlation. Indexing these claims by name, as the bond market does, hides that several apparently different borrowers depend on the same few revenue sources.

Defenders reply that these arrangements are tied to physical deployment milestones (gigawatts actually switched on) and that the tranche structures release money only as capacity is delivered. That is true of the AMD warrants and the reported Anthropic structures. It limits the downside for the supplier, but it does not remove it, because the chips and buildings still have to be financed by someone.

Disclosure is uneven. Equity stakes and guarantees show up in filings when material, but many of the commitments listed above are announcements, letters of intent or conditional frameworks. A reader should separate binding contracts from stated intentions.

Who stands behind a buyer can be unclear too: Culper Research, a short seller that posted on 13 May 2026 that it is short Nvidia, says documents suggest Alibaba financed Megaspeed, a Singapore cloud buyer, through a shell company (Culper Research, 13 May 2026, report not opened; Benzinga, May 2026). I found no reply to that financing claim. Nvidia's reply that I found concerns diversion, not financing: it said diversion claims about Megaspeed were investigated and debunked (The Hill, 29 Oct 2026). I found no ruling by a court or regulator on either. The Nvidia page (/investing/nvidia) sets out both sides.

A data centre is real estate with unusually heavy electrical and cooling equipment. Its value depends on three things: the tenant, the lease length, and access to power. Because the lease can be a stable cash flow, the asset has been packaged into securities in the same way as warehouses and cell towers.

Asset-backed securities (ABS) are bonds backed by a pool of cash-flow-producing assets; commercial mortgage-backed securities (CMBS) are backed by commercial property loans. Impax Asset Management put combined data-centre ABS and CMBS issuance at about $26 billion in 2025, more than ten times 2020, and Kroll Bond Rating Agency counted $27 billion. Datacentres.com (1 September 2026) reported $38.4 billion issued from January to August 2026 against $24.1 billion in the same period of 2025 and projected the full year could pass $55 billion, which is a secondary projection. Morgan Stanley estimated ABS and CMBS could ultimately finance about $150 billion of this demand.

The Structured Finance Association (July 2026) and Barclays, as cited there, give different outstanding-issuance figures, and one market publication reports that the SEC has relieved data-centre ABS issuers of filing 15-g disclosure forms and waiting five business days before selling bonds, which speeds issuance. Switch, a data-centre operator, raised $768 million in a data-centre ABS deal (company release). Spreads have been reported as very tight, though investors are becoming more selective about tenant quality, diversification and operator expertise.

The securities rest on long leases to tenants such as cloud providers. That is why the credit of the tenant matters more than the building itself, and why a residual value guarantee like Meta's matters when the lease ends. For GPU-intensive buildings, the shell may last decades but the equipment inside has a much shorter life, so the real estate and the chip financing are different layers of risk.

The IEA projects that data centres will take just under 3% of global electricity by 2030 in its base case, with the US and China driving nearly 80% of the growth. Grid operators set rules for connecting such loads. The grid operator PJM, which covers a large part of the eastern United States, lists a 2028/2029 capacity auction and guidance on co-located load (data centres built beside power plants) on its site, but the clearing prices and the auction date could not be verified in the sources reviewed for this piece.

• Power purchase agreements (PPAs): long-term contracts to buy electricity at an agreed price from a specific generator. They fix cost and are often needed for financing.
• Interconnection: the process and queue to connect a new load or generator to the grid; delays can strand a finished building.
• Behind-the-meter supply: on-site generation that avoids the grid but needs fuel and permits.
• Capacity markets: auctions in which grid operators pay for the right to call on generation in the future; prices feed into utility bills and politics.

This piece did not verify specific PPA terms, utility rate cases or transformer lead times, and so does not quote numbers for them. What the verified data says is that electricity is a growing share of a limited system, which makes power access a source of both value and conflict.

Moody's reported $969 billion of undiscounted lease commitments across five companies at end-2025, of which about $662 billion had not yet begun. Oracle's reported $288 billion of leases not yet on its balance sheet at 31 August 2026 illustrates how much of the pipeline is a promise to occupy buildings still being built. A building that is leased but not yet delivered carries construction, power and delay risk that a stabilised property does not.

The bull case is that demand for AI is real, capacity is scarce, and contracts with strong customers make the debt safe. The bear case is that the financial structure has outgrown the evidence for the revenue behind it. Both rest on the same facts, which is why the questions below matter.

1. Concentration: a few customers (OpenAI, Anthropic, Microsoft, Meta) support a large share of neocloud and Oracle contract backlogs. CoreWeave reported three customers at about 72% of Q2 2026 revenue.
2. Obsolescence: if GPU lives are shorter than assumed, collateral and guarantees are worth less than modelled (chapter 04).
3. Counterparty links: suppliers fund customers, so a funding shortfall at a lab hits chip makers, clouds and lenders together (chapter 06).
4. Rates and credit appetite: spreads on hyperscaler bonds widened in 2026, 78 of 91 new bonds traded at higher yields than issue by late July, Oracle's bonds were reported near 85 cents on the dollar in October, and Hyperion notes were reported at a discount. Refinancing risk rises if appetite fades.
5. Power and delivery: leases that begin only when buildings and grid connections are ready can slip, delaying revenue while interest is paid.

Further signals: Blue Owl's retail investors reportedly supply about 40% of its assets under management, and withdrawal requests at its OTIC fund reached 39% of shares against a 5% quarterly cap (compilation, secondary). That shows that part of the private-credit funding comes from investors who can ask for money back.

Supporters note that the biggest spenders still produce large operating cash flow (Microsoft covered its quarterly capex from operations), that investment-grade customers sit behind the neocloud loans, that debt is roughly one third of capex rather than all of it (Goldman Sachs, cited by Reuters on 29 July 2026), and that rental prices for older chips have risen in 2026. They also point to milestone-based structures that release funds as capacity is delivered. None of these points can be tested until the revenue from AI products reaches the scale implied by the spending, and that evidence is still arriving.

The list below describes exposure, not a view on any security. It uses the facts gathered in this piece and flags where a company's role is only reported.

Private and unlisted names that matter to the structure but cannot be bought on a stock exchange include OpenAI and Anthropic (the main end customers), Lambda and Crusoe (neoclouds preparing for larger raises or listings), PIMCO and BlackRock (large buyers of the Hyperion notes), and data providers such as Silicon Data and Ornn. Utilities and power producers (not reviewed for this piece) are the other half of the exposure map.

• Instruments: investment-grade corporate bonds; 144A private-placement notes (Hyperion); senior secured term loans (Nebius, Lambda); notes and delayed-draw loans (CoreWeave); data-centre ABS and CMBS; lease obligations; cash-settled rental-index futures (CME, announced).
• Credit signals to track: Oracle's agency ratings and bond prices; CoreWeave's interest cost against gross profit; hyperscaler spreads; ABS spreads.

What would change the picture is observable: a large customer cutting a contract, a downgrade that forces index exclusion, a loan refinanced at a much higher spread, depreciation lives being shortened across the industry, or a futures market that gains real volume and shows a falling forward curve.