We FINALLY have some insight into a frontier lab's finances.
Anthropic’s revenue grew faster than its chips were delivered. It orders most of its compute twelve to eighteen months ahead, and the bill lands when the hardware does. Its compute suppliers have offered to lend it seventy-seven billion dollars to pay for more.
Creative accounting is a feature, not a bug, in AI infrastructure.
I'm Ben Baldieri. Every week, I break down what's moving in GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
Anthropic Made a $974m Operating Profit. Most of Its Compute Is Still on Order
The public S-1 shows how an AI lab turns a profit: revenue outruns hardware it ordered more than a year earlier.
Anthropic filed its S-1 publicly on 26 October for a Nasdaq listing under the ticker ANT. Revenue reached $11.6 billion in the three months to June, up from $787 million a year earlier, and the company made an operating profit of $974 million, its first. The net loss for the quarter was still $66.3 billion, mostly from revaluing convertible notes and other financing instruments as its valuation rose. Compute and infrastructure cost $5.6 billion, or 48% of revenue. In the first quarter of 2025 that ratio was 251%. Anthropic buys most of its compute under multi-year contracts 12 to 18 months before it becomes available, and books the cost when the hardware arrives, so the expense does not move with demand. The filing says margins will swing, "at times meaningfully", as each block of committed compute lands, and that buying more short-term compute, which costs more per unit, will squeeze them in the near term. Its suppliers are also its lenders. Google can buy as much as $20 billion of its convertible notes. Amazon can buy $15 billion, released in $5 billion steps as compute milestones are met. Broadcom's $42 billion funds a lease of chips and data centre equipment. Anthropic held $112.4 billion of cash and $19.3 billion of short-term investments at the end of June. Leases and purchase commitments signed since January carry about $417 billion of minimum payments, some running as long as 20 years. A default under some of those agreements could trigger cross-defaults on its credit facility and convertible notes.
Why this matters:
The profit is a timing gap. Revenue rose 2.4 times between the March and June quarters while compute cost rose 1.6 times, and the filing warns margins will move as each block of committed capacity arrives.
Vendor financing has reached the frontier lab. All $77 billion of note capacity is tied to three companies that sell Anthropic its compute. Broadcom's 8-K this week puts its maximum exposure at $43 billion if an unnamed customer defaults on leased AI racks.
The commitments outlast the customers. Most revenue is consumption-based and uncontracted. The compute obligations run as long as 20 years, and the risk factors say so.
SpaceX Will Lease $40bn of Vera Rubin GPUs From Apollo
What SpaceX calls the largest committed GPU financing went to a company that also launches rockets.
SpaceX entered a binding commitment letter on 24 October for a master finance lease with an entity managed by Apollo Global Management, which will buy GPU pods and related equipment and lease them to SpaceX. Each tranche runs five to five and a quarter years from delivery, starting in the first half of 2027. Principal payments total about $40 billion. Rent is paid quarterly in advance: one-twentieth of the financed cost, plus interest at a fixed rate that will be set at pricing, before the lease is signed. SpaceX expects to close by the end of the year. Apollo says the GPUs are NVIDIA's Vera Rubin, TechCrunch reports. SpaceX's AI segment, which includes Grok, the Cursor coding tool and cloud compute, made $7.6 billion of its $14.6 billion third-quarter revenue. Capital spending was nearly $20 billion in the quarter, up from $6.9 billion a year earlier, and SpaceX expects $35 billion to $40 billion in the fourth quarter alone. CFO Bret Johnsen described payback periods of under a year on AI infrastructure. Elon Musk wrote on X that the lease is primarily callable debt, and predicted that many of the more than 200 neoclouds will fail or consolidate.
Why this matters:
It is debt by another name. The rent repays one-twentieth of the cost every quarter, so the principal is gone by the end of the term, and Musk calls it callable debt.
Payback in under a year is SpaceX's own figure. On a five-year lease it implies the GPUs earn their cost back several times over, and no breakdown behind the claim has been published.
Neoclouds now compete for the same lenders as a borrower with Starlink and launch revenue behind it. The fixed rate has not been set yet, and it is the number every GPU borrower will want.
Xingyun Has 18.3bn Yuan of GPU Orders on 1.2bn Yuan of Equity
A Shenzhen-listed GPU lessor's filings show Chinese neoclouds borrow against customer leases, as CoreWeave does, with banks advancing 90% of the hardware.
Xingyun Technology, a Changsha company that emerged from court-supervised restructuring in December 2024, has disclosed about 18.32 billion yuan of compute and storage orders this year, it said in an investor-relations record on 25 October. Xingyun's third-quarter report shows 1.24 billion yuan of shareholders' equity against an accumulated deficit of 3.86 billion yuan. Liabilities reached 4.42 billion yuan, up from 439 million in January. Fixed assets rose from 1.7 million yuan to 2.13 billion. Xingyun says lenders advance at least 90% of equipment cost against its servers and five-year customer leases, which gives single projects equity leverage of more than 10 times. Xingyun times its monthly loan repayments to match customers' monthly rent. A Beijing subsidiary raised a 3.33 billion yuan syndicated loan in July, secured on servers and receivables. Customers appear only as codes, and Xingyun says the biggest train and serve large models. Customer V's 6.52 billion yuan leasing contract has been repriced up 79%, and VB's compute contract is up 231% after two increases. Xingyun puts the rises down to a supply shortage and to agent workloads that it says burn 4 to 15 times the tokens of a chat. It expects a nine-month net profit of 311 million yuan, excluding share-based pay. The controlling shareholder's 12.65% stake has been frozen by a Guangzhou court since July. China AI Dispatch first assembled the filings.
Why this matters:
The structure is the one CoreWeave uses. CoreWeave's $2.6 billion loan from August sits in a special-purpose vehicle that finances specific customer contracts, at SOFR plus 5.5% with a 1.35x debt service cover test. Xingyun's banks advance at least 90% of hardware cost, and it says single projects run above 10 times equity.
This week's American backstop came from the chip vendor. Broadcom will pay the gap if leased racks resell for less than 72.5% of what is still owed, while Xingyun says its banks look first at the end customer's credit and the length of the lease.
The hardware is the weak link. The VB contract commits Xingyun to export-control compliance and makes it pay 100% of the remaining contract value if seized equipment leaves the service down for more than ten days.
Moonshot Cut Kimi's Price by a Third and Now Ranks Fourth
Moonshot's open-weight model sits three points behind Claude Opus 5.5 on an independent index, at $10 per million output tokens.
Moonshot AI released Kimi K3.1 Pro on 23 October, a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active per token and a 1 million-token context. Its API costs $2 per million input tokens and $10 per million output tokens on OpenRouter, down from $3 and $15 for Kimi K3. Weights are due by 4 November, under a licence not yet announced. Artificial Analysis ranked it fourth on its Intelligence Index at 55, 11 points above Kimi K3 and three behind Claude Opus 5.5, and said it will be the leading open-weight model once the weights are published. On Moonshot's own tests it scored 65.7 on Terminal-Bench against 66.4 for Claude Opus 5.5, using its own harness and other labs' published scores. Arena placed it eighth in Code Arena WebDev, level with Claude Opus 5 (Max), at 70% lower cost, Arena says.
Why this matters:
A model three points behind Claude Opus 5.5 on an independent index sells output at $10 per million tokens. GPT-6 Astra lists at $50.
Chinese token prices are falling while Chinese GPU rents rise. Moonshot cut its price by a third in a year when Xingyun's customers agreed rent rises of 79% and 231%, which squeezes any lab that rents compute and sells tokens.
The closed labs' revenue sits under this price. Anthropic's S-1 shows most of its revenue is consumption-based and uncontracted, which is the revenue a cheaper model can take.
FuriosaAI Turned Down $800m From Meta. It Is Now Worth $2.9bn
Korea's inference chip maker has a card in mass production and a state fund writing almost half its cheque, a year and a half after saying no to Meta.
FuriosaAI closed a $620 million Series D at a $2.9 billion post-money valuation on 28 October, Bloomberg reports. Its 2025 round valued it at $735 million. Korea's National Growth Fund committed KRW 370 billion, about $280 million. The Export-Import Bank of Korea and Korea Development Bank also invested, alongside Infinitum, Jericho Capital and Cisco Investments. Its RNGD inference card entered mass production in January. Furiosa says it will make 20,000 this year and has secured the memory and fab capacity to more than double that in 2027. Each card draws 180W and is air-cooled. Eight go into a 4U, 3kW server with 384GB of HBM3. Furiosa claims RNGD matches NVIDIA's RTX PRO 6000 on throughput at the same card count and serves twice as many users per kW. Samsung SDS runs an NPU cloud service on it, and LG Group companies use it on-premises. The money goes to two things: more RNGD production, and a third-generation chip on a 2nm, multi-die design with HBM4/4E, developed with Broadcom. No revenue was disclosed. Furiosa rejected an offer of about $800 million from Meta in March 2025, Bloomberg reported.
Why this matters:
Korean state money wrote almost half the cheque. The National Growth Fund's $280 million is 45% of the round before counting Ex-Im Bank and KDB, so this is industrial policy as much as venture capital.
Furiosa's pitch is power, not volume. Chinese buyers alone have ordered more than 6 million of NVIDIA's RTX PRO 5500 cards (#130), and Furiosa's benchmark is against an RTX PRO 6000, not a B300. Every performance number is Furiosa's own.
Gen 3 is where the valuation gets tested. A 2nm chip with HBM4 for large clusters takes Furiosa from the enterprise rack into NVIDIA's core market, and puts it in the queue for the same memory everyone else wants.
Meta Is Raising $97.5bn in Equity After Its First Negative Free Cash Flow in a Decade
Two of the buyers sell Meta its chips.
Meta raised its equity programme to $97.5 billion on 27 October, the largest in its history, Reuters reports. Meta's 8-K shows how the first part works. It sold 40 million Class A shares through forward sale agreements, about $25 billion at the $625 offer price, but receives the cash only when it settles, on dates of its choosing before 19 November 2027. It also sold 250 million depositary shares in 6.50% mandatory convertible preferred stock at $50 each, a sale that closed on 29 October. An at-the-market programme worth as much as $60 billion starts once new non-voting Class C shares trade, expected early in 2027. NVIDIA said it will invest $3 billion and AMD $1 billion. Free cash flow was negative $2.12 billion in the third quarter, against positive $10.63 billion a year earlier. Meta holds $82.45 billion of cash and marketable securities and $80.92 billion of long-term debt. Its capex guidance is $135 billion to $145 billion, and committed future spending totals about $825 billion.
Why this matters:
Meta's suppliers are funding their own orders. NVIDIA and AMD are putting $4 billion into the company that buys their chips, the same loop as Google, Amazon and Broadcom lending to Anthropic.
Meta locked in the price and kept the timing. It can draw about $25 billion of equity at roughly the $625 offer price whenever capex needs the cash, any time before November 2027.
Class C shares let Meta raise money without touching control. The non-voting stock funds the $60 billion programme and leaves Mark Zuckerberg's voting position where it is.
Equinix Raised Guidance for a Third Straight Quarter on Record Bookings
Equinix booked a record $533 million and still spends more than it brings in.
Equinix reported third-quarter revenue of $2.571 billion on 28 October, up 11%. It raised its 2026 revenue guidance for a third consecutive quarter, to between $10.325 billion and $10.365 billion. Annualised gross bookings hit a record $533 million, up 35%, which the company attributed to strong demand and firmer pricing. It now plans to deliver nearly 40% more cabinets in 2026 than it first expected, and guides to $5 billion to $6 billion of capex this year. Adjusted EBITDA was $1.321 billion, a 51% margin. Free cash flow over nine months was negative $801 million, and Equinix carries $25.2 billion of debt principal. The release gives no figure for AI-related bookings.
Why this matters:
The landlord sees demand first. Record bookings and firmer pricing at Equinix point to tight retail capacity, separate from the hyperscale campuses that take the headlines.
Growth still costs more than it earns. Nine months of negative free cash flow at a company with a 51% EBITDA margin shows how much capital the build absorbs.
AI is in the products, not the numbers. Equinix launched an inference exchange and liquid-cooled centres, and put no figure on what AI adds to bookings.
Everything Else
Microsoft booked $10 billion of revenue from OpenAI in its September quarter, 11% of its $92.3 billion total, and is owed $7.2 billion in OpenAI receivables, TechCrunch reports.
Microsoft disclosed Azure revenue in dollars for the first time, $33.2 billion for the quarter, TechCrunch reports.
Google will offer its TPUs on Oracle's cloud from 2027, making Oracle the first hyperscaler outside Google to sell them.
NVIDIA and Oracle plan to expand Vera Rubin to multi-gigawatt scale on Oracle's cloud, with no gigawatt figure or financial terms disclosed.
Alibaba Cloud opened its first public cloud region in the Netherlands, with two Amsterdam data centres, at its first Apsara Conference outside China.
General Compute raised a $100 million Series A for its inference cloud on Cerebras, SambaNova and Positron chips, with first deployments due in the first quarter of 2027.
SK hynix posted an 80% operating margin on record third-quarter revenue of KRW 104.6 trillion, with DRAM and HBM output effectively sold out through 2029, TrendForce reports.
SK hynix shelved a US listing of Solidigm, citing market conditions and its share price, Blocks & Files reports.
Moore Threads reported a third-quarter profit while burning 3.31 billion yuan of operating cash over nine months, more than its 3.29 billion yuan of sales, China AI Dispatch reports.
Apollo's chief economist Torsten Slok said AI has delivered "zero financial payoff to corporate America outside of tech", TechCrunch reports.
AirTrunk secured an A$4.25 billion green loan from 32 lenders for its SYD3 campus in Western Sydney.
The DOE closed a $5.1 billion loan to Alliant Energy and offered Constellation up to $5 billion for nuclear uprates, together covering 3.7GW, Utility Dive reports.
GE Vernova agreed to buy Trench Group from Triton Partners for $9 billion, T&D World reports.
US data centre moratoriums passed 400 this year, covering more than 100GW of planned capacity, as the pipeline added 54GW in the second quarter, according to Wood Mackenzie data reported by T&D World.
