Elon Musk Predicts Cash Shortage For AI Buildout Next Year

The debate over whether the AI buildout is a bubble has mostly been about demand: will customers keep paying for all the compute being built? Elon Musk thinks the more immediate problem may be cash.

“Getting this amount of compute turned on, cooled and working reliably is incredibly hard,” Musk posted on X. “The cash shortage is also going to be nuts next year. Even cumulative industry delays of a few months between paying for compute and getting it reliable enough to generate revenue will mean hundreds of billions of incremental dollars just for working capital. Where will that come from?”

The Context: A Bull Case For Compute

Musk was responding to a post from Keras creator and ARC Prize co-founder François Chollet, who had pushed back against the recent gloom around AI stocks. Chollet was quoting a chart showing that memory makers SanDisk, SK Hynix and Micron were trading at forward price-to-earnings ratios of just 2.2x, 2.6x and 2.8x respectively, well below NVIDIA’s 7.6x and far below the multiples of companies like AMD, Intel and Arm.

“Either you believe the AI buildout is over, that AI was a flash in the pan, and that the need for compute has already peaked… or this a rare buying opportunity,” Chollet wrote. He pointed out that GPU spot prices were still rising and capacity was hard to secure, that HBM and DRAM production was effectively sold out through 2029, and that chip and memory supply was expected to stay tight well into the 2030s. He added that frontier training compute was doubling roughly every five months, that inference was profitable at 70–85% margins, and that demand was growing geometrically, with each new model generation unlocking new use cases. “I’m not even an AI booster, just being realistic here,” he said.

The memory supply point is hard to argue with. Producers have diverted so much capacity to high-bandwidth memory for AI accelerators that SK Hynix and Micron both joined the $1 trillion club earlier this year, while smartphone sales have dropped globally as device makers fight over what’s left. The squeeze has got bad enough that Sam Altman changed his X username to “Ram Altman” after his remarks about consumer devices hogging scarce DRAM went viral.

Musk’s Point: Demand Isn’t The Only Constraint

Musk didn’t dispute Chollet’s case on demand. His argument was about what happens between buying compute and getting paid for it.

Building a frontier AI cluster means paying upfront for GPUs, memory, networking, power and cooling, and then waiting until the whole thing runs reliably enough to serve customers or train models. If that gap stretches by even a few months across the industry, Musk says, companies will need hundreds of billions of extra dollars just to stay afloat while they wait for revenue to arrive. That’s working capital, not new investment, and it has to come from somewhere.

Musk knows this problem from the inside. His AI operations run Colossus 2 in Memphis, currently the largest AI datacenter in the world by compute capacity. His teams also have a reputation for speed: NVIDIA CEO Jensen Huang once said xAI took only 19 days to set up a datacenter that would usually take a year. If Musk says getting compute online is “incredibly hard,” it’s worth paying attention to what that means for labs and neoclouds that move slower.

Financing Is Already Getting Tighter

The warning comes as cracks are showing in how the buildout is being paid for. Earlier this month, Musk claimed that SoftBank is defaulting on its turbine orders. That claim came in reply to a report that an anonymous seller was trying to offload about $500 million of data center loans, including financing tied to a Stargate site in New Mexico. SemiAnalysis has reportedly cut its 2028 AI accelerator spending forecast by 28%, citing slower revenue growth at frontier labs and tighter financing for data centers.

Markets have noticed too. AI stocks have been hammered through November, and Anthropic’s shares fell nearly 8% on only their second day of trading as the selloff deepened. The company’s IPO filings show around $895 billion in compute and infrastructure commitments. Most of Anthropic’s 2026 compute has gone to training rather than serving paying customers, which shows how much spending sits well ahead of revenue.

The scale of the spending doesn’t help. The four largest hyperscalers alone guided to roughly $715 billion in AI capex for 2026, up more than 70% from 2025. At those numbers, even a small slip in deployment schedules ties up a huge amount of money.

Both Can Be True

Chollet and Musk aren’t really disagreeing. Chollet is saying the long-term demand for compute is real and that markets are mispricing the companies that supply it. Musk is saying that even if all of that is right, the industry could still run short of cash in the near term, because money goes out long before it comes back in.

Huang has proposed a simple test for whether AI is a bubble: try renting a GPU. By that measure, demand is still strong. But Musk’s post points to a different kind of risk, one that has hit plenty of capital-heavy industries before: a company can have more demand than it can serve and still fail because it can’t fund the time it takes to deliver. If Musk is right, 2027 could be the year the question changes from “is there enough demand?” to “can anyone afford to wait for it?”

Posted in AI