When a $45B AI Fund Meets the Margin Clock

CNBC reported that Situational Awareness had grown as large as roughly $45 billion by early July. Bloomberg reported on July 30 that its assets had fallen to about $10 billion after it liquidated positions to meet margin calls. The publications did not disclose a common denominator, so the figures are not a clean AUM comparison. On June 17, the Financial Times reported that Oxford Economics counted nearly 850 data centers worth about $7 trillion under construction worldwide. The portfolio and the infrastructure ran on different clocks.

Cloud platforms buy chips and power; developers borrow from lenders against buildings and contracts. A manager can buy one layer and short another, but the trade still carries the valuation, liquidity, financing and basis risks shared by both legs.

Capex totals obscure who carries the load

Microsoft, Alphabet, Amazon and Meta reported a combined $106 billion in capital expenditure for the first half of 2024, 50% more than a year earlier. Their filings put cloud operators in the buyer’s seat, but the total did not identify which suppliers or financing counterparties would earn the highest return.

Bloomberg’s December 2025 analysis tracked thousands of newcomers across global data-center ownership and $178.5 billion in U.S. data-center credit deals. Lenders, developers, utilities, equipment vendors and operators each held claims on the buildout through debt, leases, power agreements and equipment sales.

Reuters reported in November 2025 that Oxford Economics estimated roughly 60% of U.S. data-center capital expenditure went to imported IT equipment, principally from Taiwan, South Korea and Vietnam. An American operator’s dollar can become revenue for an Asian manufacturer, collateral for a lender, rent for a building owner or depreciation on a cloud operator’s balance sheet. Investors still have to trace those payments before deciding who captures the return.

Developers and cloud operators commit to buildings and equipment before customer usage determines the payoff. Bloomberg’s $178.5 billion credit tally documents the scale of that financing bridge, whose payment schedules can arrive years before AI utilization matures.

Customers inherit more of the AI bottleneck

Dell and Qualcomm’s 2023 product announcements offered customers on-premises AI hardware, putting accelerators, networking, memory and physical location back on the buyer’s procurement list.

In 2025, AWS launched AI Factories, placing Trainium chips and Nvidia GPUs inside customers’ data centers. Customers supplied the sites; AWS supplied the infrastructure and operating relationship, extending its cloud system into facilities it did not own.

The Financial Times reported in March 2025 that Barclays expected inference capital expenditure to surpass training spending and reach $208.2 billion in 2026. Frontier-model developers train in large clusters, while application providers run inference across customers and locations. Barclays’s forecast would shift more spending decisions toward the operators choosing processors, measuring utilization and controlling deployment.

A developer can buy accelerators before a utility delivers power. Until the interconnection arrives, those chips produce no billable compute, so the developer can read demand correctly and still miss the revenue date.

Two hedge legs can inherit the same load

A manager who buys a chip supplier and shorts a software company may intend to isolate the difference between scarce hardware and vulnerable applications. That spread works only while the relationship between the two instruments remains stable enough for the short to offset risks carried by the long.

When rates rise, investors can mark down both companies. Lenders can tighten financing for their customers, while forced sellers can drain liquidity from both legs. The manager may place one security on each side of the book even as capex expectations, interest rates and investor positioning move both prices.

An unlevered investor can wait for a bottleneck thesis to mature. A prime broker can demand cash while the spread moves the wrong way because the broker marks today’s collateral under contract, not Barclays’s 2026 spending forecast.

The liquidation reveals a deadline, not a strategy

The Wall Street Journal reported on June 8 that Situational Awareness managed more than $20 billion; CNBC said the fund had grown as large as $45 billion by early July. Bloomberg reported on July 30 that assets had fallen to about $10 billion after positions were liquidated to meet margin calls. The Journal also reported that Citadel bought the bulk of the stock portfolio after the losses.

Those reports do not disclose the size, direction, timing or financing of the fund’s complete positions, nor do they establish that it held an AI-infrastructure basis trade. They document margin-driven liquidation but cannot show whether a long position failed, a short position failed, correlations changed, options expired, financing tightened or several mechanisms arrived together.

SEC records show that Situational Awareness filed six Form 13Fs covering Q4 2024 through Q1 2026. Those filings omit its complete short book, leverage, borrowing terms, option premiums and realized profit and loss, so they cannot reconstruct portfolio-level protection or identify what produced the reported losses.

A manager who shorts a semiconductor security offsets only the exposure that tracks it. The short can leave a portfolio exposed to cloud spending, power constraints, duration, credit availability or a common unwind in AI-linked valuations, especially when the long owns a bottleneck supplier and the short owns a broad basket.

A put buyer receives protection through the contract’s strike, expiry and size. The buyer pays a premium and takes exposure to volatility’s path; if the portfolio and the reference asset stop moving together, tracking error becomes a separate position and protection may expire before the thesis pays.

A lender can demand collateral even when a hedge reduces estimated market exposure. In a stressed market, protection may gain quoted value while losing tradable liquidity, so the hedge and its financing must work on the date the lender tests collateral.

More compute does not settle who gets paid

TrendForce forecast in March 2026 that Asia’s leading chipmakers would spend more than $136 billion that year, an increase above 25%. Suppliers were committing capital to continued AI demand even as investors disputed where the returns would land.

Citadel argued that broader AI deployment requires far more compute as computation becomes cheaper relative to human labor. Investors still must divide the resulting revenue among chipmakers, cloud platforms, data-center owners, power suppliers and application companies.

An Nvidia accelerator needs memory, networking and power before a cloud operator can bill for compute. A data-center developer needs financing, skilled labor and an interconnection before opening a site. A model provider needs distribution and customers whose savings exceed inference costs. Demand can rise across all three businesses while their financing costs, margins and cash receipts diverge.

The reported $45 billion and $10 billion figures cannot diagnose an AI-buildout thesis because the reports disclose neither a common denominator nor the fund’s full positions. Racks, generators and inference demand may validate a buildout view over years. The margin agreement demanded cash first.