Here's what's on our plate today:

• 🧪 Wall Street now lends against AI chips, repaid by rent.

• 📰 Privacy fears stall smart glasses; China probes DeepSeek; Greece's PM gets candid.

• 🧰 Three tools worth trying: Silicon Data, Vast.ai, WhaleWisdom.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

Wall Street has started treating AI chips as property, and ordinary savers may end up carrying the bet.

• The shift: investors who once bought shares in AI's suppliers are now lending against the chips themselves, with loans repaid from the rent those chips earn.

• The warning: Aschenbrenner's fund bet early on AI's hardware demand, but falling chip stocks and leverage of up to 400% forced a sale to Citadel in July.

• The push: NVIDIA is working with six Wall Street firms to raise more than $500B for its buyers and is pitching computing power as an asset class.

• The pricing: CME plans to launch contracts on October 5 that set the cost of a month of chip rent, pending CFTC review.

• The stakes: the loans are headed for insurers and retirement funds, so if chips lose value faster than the loans are repaid, savers absorb the gap.

How Wall Street turned AI chips into an asset class

In January 1848, a carpenter building a sawmill in northern California found flakes of gold in the American River. Samuel Brannan, a San Francisco shopkeeper, bought up the region’s picks, pans, and shovels before walking through the city with a bottle of gold dust, shouting news of the discovery. Tens of thousands rushed west over the following years, and while most of them found little gold, Brannan grew rich selling them the tools to look for it.

The lesson became familiar investing advice, which holds that in a rush the steadier money lies in selling what every prospector needs. The advice rests on simple arithmetic, because a miner’s success depends on luck while the shopkeeper earns something from every miner who tries.

The AI boom is a rush of the same kind, and its picks and shovels are the chips, data centers, and electricity that every AI company must pay for, whoever wins. Until now, investors mostly bought shares in the companies selling those supplies. This summer, the largest firms on Wall Street began treating the chips themselves as property, something that can be borrowed against and traded like a building. That shift reaches further than any one company’s fortunes, because it moves the cost of the AI boom into loans that ordinary people’s savings may end up holding.

A right bet still ended in a forced sale

Leopold Aschenbrenner, a former OpenAI researcher, became the best-known public champion of the shovel trade with a 2024 essay predicting that trillions of dollars would go into chips, data centers, and power. He then built a hedge fund, Situational Awareness, around that prediction, so instead of guessing which AI company would win, it bought shares in the businesses selling what they all need. Early this year, its bets ran through power companies and bitcoin miners, which own the power connections data centers want. By June 30, two memory chip makers accounted for more than half of its disclosed holdings.

The fund enlarged its bets with borrowed money, and borrowing makes a bet bigger in both directions: a fund that adds a bank’s money to its own gains more when prices rise and loses more when they fall. Situational Awareness borrowed heavily from banks, and CNBC reported that it used leverage of up to 400%, meaning the fund’s bets were far larger than the money its investors had put in.

Banks that lend against shares ask for more cash when those shares fall, and a fund that cannot find the cash has to sell, even at a bad price. That is what happened when chip stocks dropped in July, and the fund lost heavily: on July 30, it sold most of its stock holdings to Ken Griffin's hedge fund Citadel. A seller in a hurry gives up part of the value, and Citadel paid more than 10% below market value, after which the fund’s assets fell to about $10B.

The demand Aschenbrenner had bet on did not disappear, and the clearest sign is the price of renting an AI chip, which kept climbing through the year. NVIDIA says the typical hourly rent for its chips rose from about $2.00 last October to $2.70 in June. Falling chip stocks caused the fund’s losses, and its borrowing helped force the sale, since being right about AI’s needs did not help once its banks wanted cash. Wall Street’s new lending rests on the same belief that AI demand will last, but it works differently, since the money goes to chip buyers and is repaid from the rent the chips earn.

NVIDIA wants its chips treated like rental property

NVIDIA, which makes most of the chips used for AI, wants big investors to see those chips as something worth owning and lending against, the way they already see buildings or government bonds. On August 10, it announced a plan with six of Wall Street’s largest investment firms to raise more than $500B for the companies that buy its chips, though the agreements are not yet final. Its chief executive, Jensen Huang, told CNBC that chips are “revenue-generating assets now,” and a company blog post pitched computing power as an ‘asset class’. An asset class is a category of investment, such as bonds or property, that pension funds and insurers own as part of their routine.

A chip earns money every hour someone rents it, much like an apartment earns rent, so lenders can treat it as ‘collateral’, something they can seize and sell if a borrower stops paying. Companies need these loans because their AI spending has grown faster than the cash their businesses produce. Alphabet, Google’s parent company and one of the most profitable businesses in the world, spent more cash than its business brought in last quarter for the first time. Lenders have already shown they will accept chips as security for a loan. CoreWeave, a company that rents out AI chips, was the first to borrow against them, and this spring it raised $8.5B in a chip-backed loan that credit raters judged to be low risk.

A lender’s biggest worry is that the chips will be worth little by the time the loan ends, so NVIDIA has offered to carry part of that worry itself. Huang wrote that NVIDIA may guarantee part of the chips' value at that point, up to 25% of any one deal. Lenders also need a way to check what chip time is worth on any given day, much as a homeowner can look up recent sale prices on their street. On October 5, the exchange operator CME Group plans to start trading contracts that fix today the cost of renting an NVIDIA chip for a month, giving lenders that public price. The contracts are still under review by the Commodity Futures Trading Commission (CFTC), the U.S. regulator for such markets, which opened a public comment period on August 19.

The risk moves from one fund to ordinary savers

The money for these loans is expected to come from institutions that manage ordinary people’s savings. Fortune identified insurers and retirement funds as the real destination of Nvidia’s plan, and those institutions hold the pensions and life insurance of ordinary workers. If the loans go bad, the losses would land on savers who never chose to lend against AI chips.

Everything depends on how long a chip keeps earning, because a loan is only safe if the chip is still worth something when the loan is repaid. Michael Burry, whose bet against U.S. housing before 2008 became the subject of the film The Big Short, doubts chips will hold their value that long. Cloud companies count a chip’s cost as if it remains useful for five or six years, while he argues that the hardware loses most of its value within two or three years. If he is right, lenders could hold chips worth less than the money they lent, the kind of gap between value and debt that forced Aschenbrenner to sell, spread across far more people.

The shift is real even if the loans prove risky

Jim Osman, a Forbes contributor, argued in an opinion column that Aschenbrenner’s lesson is about borrowing, which can force a sale even when the idea behind a bet is sound. Ships and aircraft have long been financed against what they earn, so on that reading the chip loans are nothing new either. That view is right about the tools, but it struggles to explain how quickly the largest institutions moved. NVIDIA, CME, the CFTC, and six major investment firms all acted within 10 days of one another in August, which is hard to square with a name change alone.

Brannan’s shovels paid off only as long as miners kept arriving in California. Chip rent holds up the same way, only as long as AI companies keep needing more of it, and the new loans assume that demand will last for years. The first public price for chip rent should appear once CME’s contracts start trading on October 5, and CFTC comments close on October 20. Whose savings will be holding these loans if the rush slows before they are repaid is a question those dates will not settle.

Thursday Poll

3 Things Worth Trying

• Silicon Data: Daily GPU rental benchmark sitting under CME's October 5 contracts, the public price lenders have been missing.

• Vast.ai: Live marketplace of GPU listings, the fastest way to see what an hour of compute actually costs today.

• WhaleWisdom: Free 13F tracker that shows what Situational Awareness and Citadel held, quarter by quarter.

Quick Bits, No Fluff

• Privacy fears stall the smart glasses bet: Meta's Ray-Bans have drawn lawsuits and UK venue bans, pushing device makers toward camera-free pins, pendants and clip-on recorders.

• China probes DeepSeek and Moonshot: Z.AI and MiniMax fell as much as 7.4% and 6.3% in Hong Kong after a report of a data-security investigation.

• Greece's PM says the rules already lag: Kyriakos Mitsotakis called his January under-15 social media ban yesterday's battle, with AI companions the problem nobody has legislated yet.

Meme Of The Day

The Toolkit

• Baseten: Deploy and scale AI models in production without managing GPUs, autoscaling, or serving infrastructure yourself.

• Fireworks AI: Fast inference platform for open models, built for teams that need low latency at low cost.

• Granola: AI notetaker that sits in your meetings and turns your rough notes into clean summaries.

Rate This Edition