What the 2026 AI listings
reveal about debt

Most of the artificial intelligence (AI) companies that came to market in 2026 owe their banks very little. Of the sixteen that listed shares or filed a prospectus in the United States, Hong Kong and mainland China this year, fourteen carried little bank or bond debt on their latest balance sheets, and the two large borrowers, SpaceX and the Chinese memory maker CXMT, held more cash than debt.¹ The burden sits elsewhere. Anthropic, the maker of the Claude models, has no drawn borrowings but has named $637 billion of commitments to buy computing capacity, 4.8 times its cash.² Nscale, a builder of AI data centers, has signed about $9.0 billion of loan and lease facilities since February, and the two listed Chinese developers of large language models, Z.AI and MiniMax, issued zero-coupon convertible bonds that fall due in 2027.³

The listings came during a sharp turn in Chinese technology shares. The STAR 50 index of Shanghai’s technology board rose 64% in the first half of 2026 and then fell 43% from its June 30 peak to October 27, erasing the gain, while the Nasdaq Composite stood only 5.2% below its October record.⁴

This report reads the filings to answer three questions: how much these companies owe and in what form, what their operating measures show, and whether they can repay. It also sets out who sold Chinese AI shares and why, tests five common claims about AI finance, and lists the signposts we will follow. It does not value any company or recommend any security (see sidebar, “About the research”).

About the research

• Company filings. SEC registration statements, prospectuses and quarterly reports; Hong Kong prospectuses and 2026 interim reports; mainland prospectuses and 2026 half-year reports.
• Market data. China Securities Index, FRED, and the Yahoo Finance and Tencent quote feeds.
• Reporting. Reuters, Bloomberg, the Financial Times, Caixin and other outlets, as cited.

Net cash is cash and short-term investments less borrowings and leases; free cash flow is operating cash flow less capital spending. Balance sheets are at June 30, 2026 (SpaceX: September 30) and prices at October 27 and 28. Currencies are converted at FRED rates of October 23, 2026. OpenAI, which has not published accounts, is described from press reports.

Sixteen companies, about $108 billion raised

The thirteen companies that priced their offerings raised about $108 billion of gross proceeds, $86.25 billion of it in the June offering of SpaceX, which now includes the AI developer xAI (Exhibit 1).⁵ The eleven Chinese listings raised about $16 billion between them. Four Chinese chip and model companies listed in Hong Kong in the first ten days of January; the two largest mainland deals, CXMT and the robot maker Unitree, came after the STAR 50 had peaked. Anthropic filed its registration statement publicly on October 26, Nscale on September 18, and OpenAI has submitted only a confidential draft.⁶ Kunlunxin, Baidu’s chip unit, and the model developer StepFun have filed or are reported to be preparing to file in Hong Kong but have not listed.⁷

The question for lenders, suppliers and investors is whether these companies can meet what they owe. The filings answer it more directly than share prices do, because they set out cash, debt, leases, purchase commitments, customer concentration and the terms of each instrument. Several of the companies are also each other’s customers, suppliers and shareholders, so a shortfall at one could pass to others. Companies that listed before 2026, such as CoreWeave, Moore Threads and MetaX, appear here as context only.

China’s AI shares gave back their 2026 gains in four months

A rally built on chips and models

The rally ran on expectations for domestic chips and models. The STAR 50 rose 64.3% in the first half of 2026.⁸ On June 30 the AI chip designer Cambricon became the first company on the STAR Market valued at more than RMB1 trillion,⁹ and electronics overtook banking as China’s largest listed sector, worth RMB27.6 trillion against RMB14.6 trillion.¹⁰ Goldman Sachs estimates that AI-related stocks account for 42% of market value and 57% of turnover across China’s onshore and Hong Kong markets. At its peak the STAR 50 traded at nearly 250 times trailing earnings; by late October the multiple had fallen to about 100, against about 30 for the Nasdaq 100.¹¹

The reversal came in two waves (Exhibit 2). In the week to July 17 the STAR 50 fell 16.9% as leveraged traders sold, the fastest unwind of borrowed positions since 2015 to 2016. State-backed buyers then put a record RMB13.8 billion into one STAR 50 exchange-traded fund on July 20, and the index rose 10.7% the next day (Exhibit 3).¹² The second wave began in mid-October. US consumer prices rose 3.7% in the year to September, markets priced further rate rises, and a global sell-off in technology shares followed; South Korea’s Kospi fell 7.57% on October 15.¹³ The CSI 300 posted a record ten-week losing streak and entered a bear market on October 22; Bloomberg estimates that Chinese stocks lost $2.4 trillion of value from their May peak.¹⁴ On October 27 the STAR 50 fell 4.8% and the CSI Semiconductor index 5.8%, and the Shanghai Composite closed at its lowest level since August 2025.¹⁵

Individual stocks fell further. Robot maker Unitree’s shares closed on October 15 at 52% below their first-day close. Memory maker CXMT fell more than 25% in a month, and the optical-module makers Innolight and Eoptolink fell about 60% from their June and July highs.¹⁶ Reuters estimates that about $480 billion of value was wiped off the STAR 50 alone.

Who sold, and who bought

Domestic traders who had borrowed to buy were the largest sellers. Margin financing and securities lending fell from RMB3.03 trillion on June 25 to RMB2.46 trillion on October 22, a drop of RMB577 billion or 19%, led by hardware and semiconductor stocks.¹⁷ Mainland investors who buy Hong Kong shares through the Stock Connect sold a net HK$17.0 billion in the week to October 23.¹⁸ Foreign active funds in Asia sold over three months while remaining overweight China, according to HSBC research cited by Reuters. Wealthy individuals had their own reason to raise cash: a 20% tax on income from offshore trusts, collected from October 22.¹⁹

The state was on the other side, after selling earlier in the year. Central Huijin, the main vehicle of the “national team” of state-backed investors, sold about $231 billion of exchange-traded funds (ETFs) in the first half to cool the rally and has bought back about $43 billion since July, according to Bloomberg Intelligence, which estimates that about $188 billion of buying capacity remains.²⁰ Broad-based ETFs took in RMB153.4 billion in the week of October 19 to 23, the largest weekly inflow of 2026,²¹ and regulators asked insurers to report their holdings and not to lock in losses.²² Equity ETFs drew more than RMB310 billion of net subscriptions in October to the 26th, even as margin financing fell by more than RMB140 billion.²³

Why investors sold

No single event caused the sell-off. Reuters reporting points to high valuations, doubts about profits, tax bills and the threat of US curbs on Chinese data center equipment. US interest rates mattered. The Federal Reserve raised its target range to 3.75% to 4.00% in September, its first rise in three years,²⁴ and the US ten-year Treasury yield reached 5.31% on October 5.²⁵ So did US restrictions. A Senate bill introduced in September names the optical-module makers InnoLight and Eoptolink, although it covers only federal national-security systems and allows five years to comply.²⁶ On July 28 the US Federal Communications Commission added foreign-made advanced robots to its Covered List, which bars authorization of new models.²⁷ Analysts cited by The Guardian expect no broad stimulus from Beijing before the 21st Party Congress in 2027.²⁸

How this sell-off compares with 2015, 2007 and 2000

Measured by the technology board alone, the fall matches the first leg of China’s 2015 crash: the Shanghai Composite fell 43.3% between June 12 and August 26, 2015, the same as the STAR 50’s fall from its 2026 peak (Exhibit 4). The broad market has fallen much less, 20.0% for the CSI 300 and 14.0% for the Shanghai Composite, against 72.0% for the Shanghai Composite in 2007 and 2008 and 77.9% for the Nasdaq from 2000 to 2002.²⁹

Three facts set 2026 apart from 2015 so far. Margin debt equals 2.73% of tradable market value, against 4.39% at the 2015 peak, and the average collateral ratio of 251% sits well above the 130% to 150% levels at which brokers must sell; the selling has been voluntary.³⁰ The yuan touched 6.6682 per dollar on October 26, its strongest level since June 2022.³¹ And on October 26 the People’s Bank of China (PBoC) created a facility that lets systemically important bank wealth-management subsidiaries and fund managers raise cash against central-bank bills if redemptions surge, to be used only in extreme bond-market stress.³² The International Monetary Fund (IMF) judges that advanced Asian markets with the highest sensitivity to AI would suffer most in a correction and that most emerging markets are less exposed.³³

US filers owe little to banks and much to their suppliers

The four US filings describe four ways of paying for computing power. Anthropic rents most of it under long contracts. SpaceX, through its AI unit, builds it with bond money and cash from its satellite business. Cerebras sells it, with a loan from its main customer. Nscale builds it with loans secured on each site. OpenAI, the other large US developer preparing to list, has not published financial statements.

Anthropic owes banks nothing and its suppliers hundreds of billions of dollars.

Anthropic: Fast growth, no drawn debt, $637 billion of commitments

Anthropic’s revenue reached $16.3 billion in the first half of 2026, 14 times the $1.2 billion of a year earlier, and $11.6 billion in the second quarter alone (Exhibit 5). Its costs fell as a share of revenue: computing and infrastructure absorbed 48% of revenue in the second quarter, against 177% a year earlier, and the company earned operating income of $974 million in the quarter, its first quarterly profit. Operating cash flow was $3.9 billion in the half and capital spending $0.5 billion. Customers paying more than $100,000 a year rose from more than 950 to more than 11,900 in the year to September, and net revenue retention, a measure of how much more existing customers spend, was about 900%. The five largest customers provide about 25% of revenue.³⁴ Reuters has reported third-quarter revenue of $16.8 billion, which would mean slower growth than in the second quarter; the S-1 does not include third-quarter figures.³⁵

One measure deserves attention. Almost half of first-half revenue, 48%, came through the cloud marketplaces of Amazon, Google and Microsoft, and Anthropic records those sales at their full value because it acts as the seller. The same three companies are Anthropic’s largest suppliers of computing capacity, and Amazon and Google are also its investors.

The balance sheet carries almost no conventional debt. Anthropic had $131.7 billion of cash and short-term investments on June 30 and had not drawn on its revolving credit facility, which it enlarged to $15 billion in September. The facility runs to 2031, is secured until the company holds investment-grade ratings from two agencies, and requires at least $5 billion of liquidity at each quarter end from December 2026. The $108.6 billion of convertible notes on the balance sheet is the fair value of $6.2 billion of non-interest-bearing notes held by Amazon, which convert into shares. The rising value of these notes and related instruments accounted for $72.4 billion of the $80.7 billion first-half net loss, a charge that uses no cash. Anthropic also has $77 billion of undrawn note facilities from Google, Amazon and Broadcom. The Broadcom notes, up to $42 billion, carry 9% interest paid in kind, rising to 13% if revenue tests are missed, and may be used only to pay rent on Broadcom equipment.³⁶

The commitments are where the obligation lies (Exhibit 6). Anthropic has agreed to spend at least $114.5 billion with Google to 2033, at least $110 billion with Amazon to 2036 and at least $44.9 billion with Microsoft to 2033; up to $120.5 billion with SpaceX’s AI unit into 2029; up to $44.6 billion with Nscale and $41.3 billion with Lambda; and $161.2 billion of rent over five years under equipment leases with Broadcom. These named contracts total $637 billion, 4.8 times cash and 13.8 times second-quarter revenue at an annual rate. The company also estimates $125 billion to $150 billion for further Google TPU capacity from 2028 to 2031 and more than $20 billion for AMD systems, neither of them in its commitments table, and it reports $417 billion of new obligations signed since the end of 2025.³⁷

Timing matters more than the totals. Two contracts alone, the Broadcom leases at a straight-line $32 billion a year and the SpaceX contract at its full rate of about $3.1 billion a month, would cost about $69 billion a year, 1.5 times second-quarter revenue at an annual rate; the S-1 does not yet give a payment schedule.³⁸ Some commitments are softer than others. The Nscale and Lambda fees accrue only when capacity is delivered, and the SpaceX contracts can be cancelled on 90 days’ notice after their initial terms end in January and March 2027. The Broadcom leases cannot be cancelled and carry cross-default clauses. Our reading is that Anthropic can meet its obligations for the next two years from cash, provided revenue keeps growing well ahead of the fixed payments; slower growth would bring forward the need for the note facilities or new capital. Investors have discussed a valuation of about $1.5 trillion for the listing, against $965 billion in the May private round, according to reporting cited by PitchBook.³⁹

SpaceX and xAI: Bond-funded building, carried by satellite cash flow

SpaceX, which merged with xAI and X before its June listing, replaced the AI unit’s expensive debt before going public. In March a $20 billion bridge loan repaid xAI’s 12.5% secured notes and floating-rate loans and X’s term loans, at a cost of $646 million in prepayment penalties. In June, $25 billion of senior unsecured notes with coupons of 5.35% to 6.65% repaid the bridge. At September 30, SpaceX had $38.0 billion of debt, including $12.9 billion recorded for sale-and-leaseback deals on AI equipment with a party related to a director, against $89.9 billion of cash and marketable securities. Only $2.9 billion falls due before the end of 2027.⁴⁰

The AI segment, now called SI, is still absorbing capital. It earned $7.6 billion of revenue and $431 million of operating income in the third quarter but lost $3.3 billion over nine months, and its third-quarter capital spending of $16.4 billion was 2.2 times its revenue. One SI customer provided 34.5% of group revenue in the quarter. The filing does not name it, but Anthropic’s S-1 discloses a contract with SpaceX’s AI unit that ramps to about $3.1 billion a month.⁴¹ Group operating income covered interest 4.1 times in the third quarter. The shares closed at $148.58 on October 27, 10% above the $135 offer price and 30% below their June 16 peak.⁴² SpaceX can service its debt from its satellite business and its listing cash; the AI unit’s returns depend on a few large customers.

Cerebras: A backlog that rests on one customer

Cerebras listed in May at $185 a share. Its first-half revenue was $374 million, its operating loss $492 million, and it spent $549 million on equipment. It holds $7.9 billion of cash and investments against a single loan: $918 million outstanding of about $1.0 billion that OpenAI lent in January to fund capacity for OpenAI itself, at 6% interest and repayable mainly through services. Its $25.4 billion of contracted future revenue, 34 times first-half revenue at an annual rate, is mostly a 750-megawatt contract with OpenAI. The filing names OpenAI, G42, the Mohamed bin Zayed University of Artificial Intelligence and Amazon Web Services among its significant customers; the largest provided 34% of second-quarter revenue.⁴³ The shares closed at $119.31 on October 27, 36% below the offer price.⁴⁴ Cerebras’s own debt is small; its credit risk is OpenAI’s ability to fund its commitments (Exhibit 7).

Nscale: The most leveraged model

Nscale, which builds and runs AI data centers in Norway, Portugal, Iceland and the United States, filed in September. First-half revenue was $140.6 million, against a net loss of $1.02 billion and negative adjusted EBITDA of $199 million. Net interest expense equaled 68% of revenue, mostly interest accrued on customer prepayments. On June 30 it had $1.5 billion of cash, $137 million of debt, $2.6 billion of lease liabilities and $6.5 billion of deferred revenue, most of it prepayments from customers. Since February it has signed about $9.0 billion of facilities, among them GPU loans in Texas and North Carolina at 2.375 percentage points over the secured overnight financing rate (SOFR), a Dell equipment-leasing program and a Macquarie loan in Iceland at 5.5 points over SOFR.⁴⁵

Its contracted backlog is $103.4 billion, and its largest customer provided 52% of first-half revenue. The auditors’ going-concern note records that reliance on uncommitted financing “raised substantial doubt” about the company’s ability to continue, a doubt management resolved by showing it could defer, reduce or cancel capital spending.⁴⁶ Of the US filers, Nscale is the one whose repayment depends most on new financing.

OpenAI: The largest developer, without public accounts

OpenAI submitted a confidential draft registration in June and has not set a date. Press reports give a partial picture: revenue of $5.7 billion and cash burn of $3.7 billion in the first quarter, according to The Information as reported by Reuters;⁴⁷ annualized revenue approaching $50 billion, according to the Financial Times;⁴⁸ and $278 billion of projected cash burn from 2026 to 2030, according to a company presentation seen by the FT.⁴⁹ OpenAI itself borrows little; Bank of America extended a $520 million credit line in July.⁵⁰ The leverage sits with its backers. SoftBank took a $40 billion bridge loan in March and sold $11.1 billion of bonds in September, at coupons of 8.625% to 9.75%, to fund its OpenAI stake.⁵¹ OpenAI is now seeking about $30 billion, with investors in the UAE reported among those leading the round.⁵²

Many of these companies are each other’s customers, suppliers and owners

The filings document a web of relationships in which the same companies buy from, sell to, lend to and own each other (Exhibit 8). Anthropic sits at the center of the largest. Amazon and Google hold or may buy its notes, sell it computing capacity under minimum-spend contracts and resell its models in their marketplaces. Broadcom leases it equipment and may lend it the money to pay the rent. AMD will buy up to $5 billion of its shares as AMD chips are deployed. Anthropic’s spending with SpaceX’s AI unit appears in SpaceX’s accounts as its largest AI customer, and its contract with Nscale depends on Nscale first raising the debt to build the capacity.⁵³

The pattern recurs elsewhere. OpenAI lent Cerebras $1.0 billion and accounts for most of its backlog. Tencent bought 74.9% of Enflame’s output in 2025, owns 17.95% of the company and prepaid RMB6.3 billion in the first half of 2026.⁵⁴ Alibaba owns 17.06% of MiniMax’s Class A shares, was a cornerstone investor in its listing and supplies its cloud computing under a cap of $300 million for 2026.⁵⁵

These arrangements are disclosed and accounted for under the usual rules; their effect shows up in risk. When a company’s investor, supplier and sales channel are the same firm, a slowdown at one reaches the others through several routes at once: lower purchases, unused capacity, weaker collateral and lower equity values.

China’s filers borrow cheaply and raise equity quickly

The Chinese filers fund themselves differently from their US peers. Their bank loans are cheap and mostly unsecured, high share prices have allowed large follow-on placements, and the two model developers have turned to zero-coupon convertible bonds.

The model developers: Z.AI and MiniMax

Z.AI, the developer formerly known as Knowledge Atlas or Zhipu, grew revenue fivefold to RMB954 million in the first half of 2026, but its gross margin fell from 50.0% to 26.4% as sales shifted to its application programming interface (API), now 86.5% of revenue. It used RMB3.4 billion in operations in the half, against RMB1.3 billion a year earlier, and its bank loans of RMB2.2 billion cost about 2.1% to 2.2% (Exhibit 9).⁵⁶ After June it raised about RMB60 billion: two share placings that netted HK$47.0 billion and a RMB20.1 billion zero-coupon convertible bond due September 16, 2027. The bond converts at HK$892.50, 56% above the October 28 close of HK$571.50, so at today’s price it would be repaid in cash; the placings cover it.⁵⁷ Z.AI’s shares stand 392% above their January offer price but 64% below the July placing price. The company has been on the US Entity List since January 2025.

MiniMax earned $116.6 million of revenue in the first half, 61% of it outside mainland China, and lost $358 million. Its cash outflow, including $188 million of its own equipment, ran at about $84 million a month, three times the $28.1 million the prospectus assumed for 2025.⁵⁸ It had $1.3 billion of cash and $134 million of bank loans at June 30. In July it sold shares for a net HK$9.4 billion and issued a HK$6.5 billion zero-coupon convertible bond due in July 2027 that converts at HK$335, 86% above the October 28 close. The Abu Dhabi Investment Authority was a cornerstone investor in its January listing.⁵⁹

The chip designers: Cheap loans, fast-growing working capital

The Hong Kong-listed chip designers carry little debt but need large amounts of working capital, because they buy wafers and components ahead of sales (Exhibit 10). Biren’s revenue rose to RMB1.24 billion in the first half from RMB59 million and it held RMB5.8 billion of net cash, but it used RMB2.0 billion in operations and had collected only 23.3% of RMB1.1 billion of receivables by the report date.⁶⁰ Iluvatar CoreX raised its bank borrowings from RMB1.0 billion to RMB2.7 billion in six months as inventories rose to RMB2.0 billion, and its first-half profit of RMB106 million rested on a RMB760 million unrealized gain on shares in SJ Semi.⁶¹ Axera, which designs chips for cameras and cars, has RMB1.1 billion of bank loans at rates of up to 5.9% and a RMB467 million liability that accrues 6% a year if a subsidiary does not complete a listing.⁶²

At their first-half rate of cash use, liquidity at June 30 covered six to sixteen months for most of these companies. Share placings after June lengthened that for all except Axera and Xizhi. Xizhi, which makes optical computing chips, needed no new money because it holds RMB2.9 billion against small cash use; Axera has not raised since its listing and its shares trade 33% below the offer price. Iluvatar CoreX also has banking facilities of more than RMB12 billion, over 80% of them unused.⁶³

The mainland listings: One large borrower, flush with cash

CXMT, China’s largest maker of DRAM memory chips, is the only large conventional borrower in the group. Its revenue rose to RMB150.3 billion in the first half of 2026 from RMB15.4 billion a year earlier, with a gross margin of 84.7%, as AI demand lifted memory prices, and net profit attributable to shareholders reached RMB77.6 billion. At June 30 it had RMB141.1 billion of borrowings and leases against RMB182.3 billion of cash and trading assets, a net cash position of RMB41.2 billion, before RMB66.3 billion of net listing proceeds arrived in July. At the end of 2025 its net debt had been 1.98 times earnings before interest, taxes, depreciation and amortization (EBITDA). Its long-term loans cost 1.60% to 3.63%, RMB82.0 billion of them fall due after five years, and it has RMB39.7 billion of undrawn facilities.⁶⁴ The prospectus itself warns that the rise in memory prices may not last, and in June the US Department of Defense added a CXMT subsidiary to its list of Chinese military companies.

Unitree has no bank debt and positive free cash flow, and it held RMB7.4 billion of cash after its listing. Its profit before one-off items fell 19% in the first half, and its US sales, 13.3% of 2025 revenue, face the FCC restriction.⁶⁵ Enflame, an AI chip designer, shows how customer money can stand in for debt. Its debt-to-assets ratio rose from 30.4% to 78.7% in the first half, largely because a Tencent entity prepaid RMB6.3 billion, which Enflame used to prepay its own suppliers RMB7.5 billion; RMB5.8 billion of listing proceeds arrived in September.⁶⁶ SJ Semi, a chip-packaging company, and DapuStor, a maker of data center storage drives, carry modest debt. DapuStor’s operating cash flow was negative despite a first-half profit of RMB1.33 billion, and both companies’ shares are about 61% below their May peaks.⁶⁷

The state is a shareholder across the group. The second phase of the China Integrated Circuit Industry Investment Fund owns 7.86% of CXMT and 3.89% of Enflame, social security and pension portfolios bought RMB7.5 billion of CXMT shares in its listing, and investors linked to local governments in Beijing, Shanghai and other cities hold stakes in Z.AI, Biren and Iluvatar CoreX.⁶⁸

Which companies can repay

We grouped the companies by the cash they hold, the cash they generate, and the size and timing of their largest fixed claims (Exhibit 11). The grouping is our judgment from the filings, not a credit rating. In order of repayment capacity:

1. Strong. SpaceX with xAI, CXMT, Unitree and SJ Semi hold more cash than debt or generate enough cash to service it, and their largest claims fall due after 2027.
2. Cash-rich, with large claims. Anthropic holds $130.6 billion of net cash and generates cash from operations, but its commitments are many times its revenue. Its position holds if growth stays well ahead of the fixed payments.
3. Adequate after raises. Z.AI, MiniMax, Biren, Xizhi, Cerebras and Enflame use cash, but placements, listing proceeds or customer prepayments cover their needs and their 2027 maturities. Cerebras and Enflame each depend on one customer.
4. Watch. Iluvatar CoreX, Axera, DapuStor and Nscale combine cash use with rising or short-term debt. Nscale’s auditors recorded a going-concern doubt that management resolved through its ability to defer spending.

Five claims about AI finance, tested against the filings

Commentary on AI finance often rests on five claims. The filings support some of them more than others:

1. AI hardware wears out in three to five years. The accounts partly agree. Anthropic and Cerebras depreciate equipment over three to five years, but the two companies with the most equipment debt, Nscale and SpaceX, use five to six years for GPUs and servers.⁶⁹ Longer accounting lives at the borrowers lower reported depreciation; if the equipment loses value faster, collateral would fall faster than book values suggest.
2. Debt outlasts the chips it pays for. The evidence is mixed. Nscale’s Texas GPU loan runs to February 2033 and its North Carolina loan to December 2031, at or beyond the equipment’s accounting life, while its Iceland loan amortizes over 46 months and Anthropic’s Broadcom leases run for five years.
3. The largest platforms pay from their own cash flow. Less so in 2026. SpaceX refinanced its AI unit’s debt with $25 billion of bonds, Meta announced a $95 billion share sale, SoftBank borrowed to fund its OpenAI stake, and Anthropic is taking $161 billion of equipment on finance leases.⁷⁰
4. Circular deals inflate revenue. The loops are real and disclosed. Revenue from them is recognized under normal accounting rules; the concern is concentration, because the same few counterparties appear on both sides of the ledger.
5. Cheaper models could strand the build-out. The IMF warns that low-cost open-source models such as Moonshot’s Kimi could squeeze margins along the AI value chain. Anthropic states that lower costs per task drive greater usage, citing a 40% cut with its latest model. Both can hold: falling prices help the developers that win volume and hurt owners of capacity sold by the hour.

What could change the picture

Three sets of forces will decide whether these companies meet their obligations over the next twelve months, against a calendar of dated tests (Exhibit 12).

Geopolitical. US restrictions on Chinese chips, robots and data center equipment, and limits on China’s access to US chips, shape costs on both sides. Reuters reported that ByteDance, Alibaba and Tencent ordered more than $35 billion of Nvidia chips but may receive only 5% to 10% of what they asked for.⁷¹ Oil near $101 a barrel, held up by the conflict in the region, feeds US inflation and interest rates.⁷²

International. Interest rates set the cost of the debt and the value of the equity that funds these companies. Most economists in CNBC’s survey expect a further Fed rise in December, and the hyperscalers’ capital spending, about $800 billion this year, sets demand for chips and capacity.⁷³

Domestic China. Beijing has shown it will support the market: the national team has about $188 billion of buying capacity, the PBoC has built a bond-market backstop, and a RMB60 billion state AI fund began investing in October. Tighter curbs on overseas investment take effect on December 1.⁷⁴

Our view rests on three assumptions: that US rates stay near current levels, that no new US measure cuts a major customer off, and that Chinese policy keeps supporting the market without broad stimulus. On those assumptions, our analysis suggests that the next twelve months will test most the companies that need new money: Nscale, Axera and Iluvatar CoreX among the filers, and suppliers whose backlogs rest on one developer. The larger tests come in 2027, when Anthropic’s SpaceX contracts become cancellable, Broadcom capacity and AMD deliveries ramp up, and the Z.AI and MiniMax convertibles mature.

---

The 2026 listings show an industry that has moved much of what it owes out of loans and into contracts. Anthropic owes banks nothing and its suppliers hundreds of billions of dollars. The Chinese developers owe little to banks and have used high share prices to raise more equity and convertible debt than they now need. CXMT, the one heavy borrower, has turned a memory boom into net cash.

For lenders, the useful measures are the ones the filings now disclose: commitments against cash and revenue, the share of revenue from a few counterparties, and the date of the first payment that cannot be cancelled. For companies and investors in the Gulf, where state-linked investors and customers appear in several of these deals, among them the Abu Dhabi Investment Authority in MiniMax’s listing and G42 as a Cerebras customer, and where data center projects worth more than $93 billion are active or planned across the GCC,⁷⁵ the same tests apply to every lease, offtake and capacity contract.

The sell-off in China has so far been a repricing of shares, absorbed by voluntary deleveraging and state buying. Whether it stays that way depends on revenue growth at a handful of developers, the path of US rates, and policy in Washington and Beijing. We will update this analysis as Anthropic prices its offering and the first 2027 payments approach.