One AI boom, two bills. Only one hides the risk.

East/West AIpha | Edition 19 | Tuesday, October 6, 2026

Automate Work, Amplify Life.

If you own AI, you own a funding structure. In the US the risk sits in contracts and credit. In China it sits in the share market's mood. Of the two bills, in our view the debt one is more dangerous: it hides the risk until it comes due in full.

Free cash flow at Microsoft, Amazon, Alphabet, Meta and Oracle fell from $191 billion in 2025 to $14 billion in the first half of 2026, while the cash they raised from debt rose from $90 billion in all of 2025 to $163 billion in the first half alone (Rhodium Group, 17 Sep). The US labs lock in compute with hard contracts and keep their shares private.

In China, the labs sell shares. Zhipu (智谱), MiniMax (稀宇科技), DeepSeek (深度求索) and Moonshot (月之暗面) took in 179 billion yuan of venture money, IPO proceeds and placements in the first eight months of 2026, per Rhodium.

Same boom, two failure points. Equity pain is loud, daily and shared by everyone on the register. Debt pain is quiet, fixed and owed to someone specific. A share price can recover in a quarter. A missed coupon cannot.

The allocator's two-bill test

Start with the most detailed disclosure so far. Reuters reported (Reuters, 29 Sep) that it had reviewed Anthropic's confidential IPO prospectus (Anthropic, 1 Jun). According to Reuters, Anthropic expects to spend at least $518 billion over a decade with six infrastructure partners, and about 80% of that is non-cancelable or payable regardless of usage (Reuters, 29 Sep). Reuters reported the split as at least $111.1 billion with Google, $110 billion with Amazon and $31.4 billion with Microsoft, plus about $161.2 billion of Broadcom-related equipment leases. Until Anthropic files publicly, treat these as reported, not filed.

One line Reuters quoted from the filing explains the US model better than any chart: "If our actual spend falls short, we must pay Google the difference." That's take-or-pay: when demand slips, the buyer carries the gap.

Now look at the lenders. On 23 Sep, a joint venture including Blue Owl Capital affiliates sold $1.12 billion of 8.875% senior secured notes due 2031 to fund a 76 MW data center in Virginia (Bloomberg, 23 Sep; Kirkland & Ellis, 30 Sep). The building is fully leased to CoreWeave for 15 years, a contract worth about $2.94 billion, Bloomberg reported.

A fully leased building still had to pay close to 9%: the notes sold at 98.5 cents to yield about 9.25%, Bloomberg reported. S&P rated them BB-, below investment grade, and said, "Exposure to a single, speculative-grade tenant remains the key risk," according to Bloomberg. That's the market saying a long lease from one AI cloud tenant isn't a government bond, however clean the paperwork looks.

The credit market is marking it in public. About $18 billion of bank loans behind Project Jupiter, the New Mexico campus Oracle is leasing for OpenAI, were privately quoted at 89 to 91 cents on the dollar in September (Financial Times, 18 Sep). Jim Chanos, the short seller, has criticized the off-balance-sheet structures behind deals like Meta's Louisiana campus, and the Financial Times reported that Big Tech groups, among them Nvidia, Broadcom and Meta, have offered up to $300 billion of residual value guarantees in less than a year, backing AI data center and chip debt that mostly stays off their balance sheets (Financial Times, 20 Sep).

Then Oracle issued a force majeure notice on Jupiter; a source told Reuters it faces a one-year delay over power, while Oracle says it remains on schedule (Reuters, 24 Sep). Either way, the contract has already decided who bends first. The Chinese have a phrase for this: 寅吃卯粮 (yín chī mǎo liáng), eating next year's grain this year. It works until you run out of next years.

OpenAI aims to raise at least $30 billion at about $1.4 trillion before the new money, Bloomberg reported, as a bridge in place of an IPO (Bloomberg, 29 Sep). Its round that closed on 31 March raised $122 billion at an $852 billion post-money valuation (OpenAI, 31 Mar).

Anthropic's confidential prospectus, as reported by Reuters, shows why these labs are coming to market at all: reportedly $4.6 billion of 2025 revenue against a $42 billion net loss, including a roughly $34 billion non-cash charge, with a listing possible after the November midterms at above $2 trillion (Reuters, 28 Sep). Until then, US lab equity is priced in a room, once a round, by a few very large holders.

SoftBank paid its final $10 billion tranche into OpenAI on 1 Oct, bringing its total to $64.6 billion and its stake to about 13% (SoftBank Group, 1 Oct). A week earlier it priced $11.1 billion of foreign-currency senior notes, with dollar coupons of 8.625% to 9.750% (SoftBank Group, 24 Sep), money The Next Web linked to the final payment (The Next Web, 2 Oct). Equity at the lab, high-yield debt at the holder.

Suppliers sit in the stack too. AMD committed to buy up to $5 billion of Anthropic stock and to supply it more than $20 billion of compute, Reuters reported, citing the prospectus (Reuters, 29 Sep). If you own the supplier, you own a slice of its customers' contracts.

Rhodium tracked 13 Chinese firms, including unlisted Huawei, and expects their AI capex to double to 932 billion yuan ($139 billion) this year, around 15 to 20% of US levels. Alibaba (阿里巴巴), Tencent (腾讯) and Baidu (百度) went from 170 billion yuan of free cash flow in 2025 to minus 16 billion yuan in the first half of 2026, per Rhodium.

The fix is different. Venture money into the four labs came to 9 billion yuan in 2025, before IPOs and placements lifted the 2026 total to 179 billion yuan (not like for like). Zhipu raised about 4 billion yuan in its January Hong Kong IPO and 27 billion yuan in a July placement, per Rhodium.

Listed labs reset their cost of capital every trading day. Zhipu ran from its January IPO to an intraday peak near HK$2,980 in June, then gave back roughly 80% to about HK$610 by late September (Sina Finance, 30 Sep), the round trip we traced in Edition 16. Each leg down was a funding-channel event, not a model event: a lock-up expiry in July (down 28.5% in a day), a DeepSeek price cut in September (down 10.34%) and a data-handling controversy (down 12.4%) (Caijing, 17 Jul; Guancha, 10 Sep; CNR Finance via Eastmoney (央广财经), 23 Sep).

In mid-September, with the stock still sliding, Zhipu sold 21.97 million new H-shares at HK$714, about a 10% discount, alongside $3 billion of zero-coupon convertible bonds, and the shares fell 9.08% and 5.69% on 14 and 15 Sep (Tencent News (腾讯新闻), 17 Sep; Futu, 2 Oct). The price fell, the company sold shares anyway, and on a 16 Sep investor call it raised its year-end ARR guidance to $3 billion from $2.4 billion. In our view, that sequence shows how much the next raise depends on the share price. I'm not showing you those prices as a trade. I'm showing you a funding channel.

China has its own Chanos, on the equity side. Yang Dong (杨东) of Ningquan Asset (宁泉资产) wrote in July that his team had "severely underestimated how big the AI-infrastructure bubble could get" and that many hot AI stocks "could fall 80 to 90 percent" (our translation) (DoNews, 7 Jul). Some Chinese retail investors call stocks that move like this 妖股 (yāo gǔ), "demon stocks".

More listings are coming: Moonshot confidentially filed for a Hong Kong IPO of about $3 billion in September, with CICC (中金公司) among the banks (Reuters, 3 Sep).

Debt plays a role through loans and, at Zhipu, convertible bonds. ByteDance (字节跳动) reportedly secured a $29.6 billion offshore syndicated loan in September at SOFR plus 68 basis points (Quartz, 4 Sep). Tencent and Baidu's long-term yuan bonds cost around 2.6%, against an average effective rate of about 4.4% on the five big US firms' long-term bonds, per Rhodium. Lay the borrowing costs side by side: about 2.6% in Beijing, about 4.4% for Big Tech in New York, about 9.25% on a leased building in Virginia, and up to 9.75% on SoftBank's dollar notes. One boom, four borrowing costs.

A third pool of money fits neither pattern.

The large Asian pools are picking the hard layer of risk, power and long contracts, rather than buying "AI" as one thing.

Commandment Nine in our book The Ten Commandments of Investing is Don't Lose Money. It carries Warren Buffett's warning from his 2010 letter: "any series of positive numbers, however impressive the numbers may be, evaporates when multiplied by a single zero." He was writing about borrowing: "history tells us that leverage all too often produces zeroes."

There's a Chinese phrase for the discipline both systems are skipping: 量入为出 (liàng rù wéi chū), spend within your income. Both can work. They just put the risk in different hands.

So my position is a question, not a trade. Before you add AI exposure, ask who pays if demand falls short. The equity bill reprices itself. The debt bill does not care what year it is. That is why, of the two, I fear the fixed one.

Next Tuesday, Edition 20 studies the top value investors in tech: four from the East, four from the West and two from the rest of the world. Among the East names: Zhang Lei (张磊) of Hillhouse, Li Lu (李录) of Himalaya Capital, Duan Yongping (段永平), and Capital Today's Kathy Xu (徐新). I interviewed Kathy for our book.

Where we've been on this

A preview of Premium, which opens soon. Free in this issue.

Our horizon. Our horizon is five years or more. These cases are about long-term business value, not short-term price moves.

The allocator's two-bill test, applied

The case for buying. Contracted demand, in our view, is the strongest argument. Reuters reported at least $111.1 billion of Anthropic commitments with Google, $110 billion with Amazon and $31.4 billion with Microsoft. Take-or-pay puts the gap on the buyer. These firms also borrow cheaply: Rhodium puts the average effective rate on their long-term bonds at about 4.4%. A riskier borrower shows the contrast: BB- notes on a Virginia data center fully leased to CoreWeave sold to yield about 9.25%, Bloomberg reported. In our view, strong balance sheets, for most of the five, are a real cost advantage.

The case for holding. Part of the borrowing funds capacity already under contract, such as Anthropic's reported commitments. What has to happen next is customer revenue growing into the contracts. OpenAI is near $70 billion of annualized recurring revenue, Axios reported, via Reuters. The signposts: free cash flow recovering as new data centers come online, and lab revenue keeping pace with commitments. We do not judge here what share prices already assume. That is the question to ask before adding or trimming.

The case for selling. Free cash flow went from $191 billion for all of 2025 to $14 billion in the first half of 2026, while debt raised went from $90 billion in all of 2025 to $163 billion in the first half alone, per Rhodium. Rhodium says AI firms' revenues are failing to cover aggressive capex plans by wide margins. The biggest contracts sit with labs that are not yet profitable: Anthropic reportedly posted a $42 billion net loss in 2025, including a roughly $34 billion non-cash charge, on $4.6 billion of revenue, per Reuters. A contract is only as good as the buyer's ability to pay. The Financial Times reported up to $300 billion of residual value guarantees from tech groups including Nvidia, Broadcom and Meta in less than a year.

What would change our view.

The question we would ask. If AI demand slipped for two years, does your hyperscaler exposure rest on contracts customers cannot cancel, or on those customers being able to pay?

Our read: all five are borrowing to build, and four of them, Microsoft, Amazon, Alphabet and Oracle, also sell the long contracts. That is why we watch their free cash flow before their headlines.

You decide. These are our views as of 6 October 2026, not a recommendation to you. Weigh them against your own goals, horizon and risk.

Disclosure and terms. This section is general research and education, the same for every reader. It is not personalized investment advice and does not take into account your objectives, finances or needs. WorkOptional is not a registered investment adviser or broker dealer. The cases above are our views as of 6 October 2026. They may change without notice and may be wrong. Do your own research and consider speaking with a licensed adviser before you invest. Past performance is no guide to future results.

Holdings: At the time of writing, San Eng and entities he controls hold a small SpaceX position, held since 2021, and no other positions, public or private, in the companies named here, other than through broad index funds.

Trading policy: We do not trade securities discussed in an issue from the day we start writing it until three full trading days after it is sent, and we do not trade against a published view within 30 days of publishing it. Any change in our holdings in a company we have covered is disclosed in the next issue.

Not directed at anyone in a jurisdiction where receiving it would require us to hold a licence.

Forward this to one allocator who still thinks "AI exposure" is one thing. We grow by word of mouth, not by algorithm.

Got this from a friend? Subscribe free to East/West AIpha

Which bill do you hold more of, the fixed one or the floating one? Reply and tell me.

Premium opens soon: $39.99 a month, or $335.92 a year (USD, plus any applicable tax). Annual includes a physical copy of Book 1 (free US shipping) and a 1-hour session with San on AI strategy and adoption inside your organization. The session is not investment advice and does not discuss specific securities. Plans and contents may change before launch.

San Eng

Chairman, WorkOptional.ai

Co-author, The Ten Commandments of Investing, a WSJ and USA Today best seller

$500M+ deployed as an investor

DISCLOSURES

For information and education only. Not investment, legal or tax advice. Nothing here is a recommendation to you to buy, sell or hold any security. Share prices in the main body appear only to show how a funding channel works. The investment cases in the Premium preview are our general views, the same for every reader.

At the time of writing, San Eng and entities he controls hold a small SpaceX position, held since 2021, and no other positions, public or private, in the companies named here, other than through broad index funds.

Researched and checked with our AI assistants, then edited and approved by San.

Sources: