Aug 27, 2026 · 55 min
AI debt boom tests the limits of real demand
‘Things Are Going to Get Crazier’: AI Debt Deluge Recalls 2007 Risks
The episode examines whether debt, vendor financing, and circular funding are turning genuine AI infrastructure demand into a fragile credit cycle.
- 1AI infrastructure spending may be productive, but circular financing and concentrated demand make the boom difficult to assess.
- 2NVIDIA’s moat rests on software, networking, systems integration, and ecosystem strength—not chips alone.
- 3Highly leveraged neoclouds could face the sharpest stress if hyperscalers flatten capital spending or financing conditions tighten.
Don't miss
Jay Goldberg compares the emerging AI credit cycle with the period before the 2008 crisis, while predicting casualties without necessarily forecasting a systemic collapse.
The brief
Jay Goldberg of Seaport Research Partners joins James Crombie and Rob Schiffman to examine whether AI infrastructure is becoming a debt-fueled bubble reminiscent of the years before 2008.
The central tension is real enterprise demand versus financing-driven demand: AI models may be useful, while the infrastructure built to generate tokens risks becoming overcapacity.
NVIDIA’s advantage extends beyond GPU specifications to software, networking, systems integration, and ecosystem depth, even as custom chips from major cloud companies advance.
Goldberg argues that GPUs are awkward collateral because physical longevity does not guarantee economic usefulness, making repossession and redeployment difficult when demand shifts.
The most exposed businesses may be smaller neocloud providers dependent on debt, long-term contracts, and continued hyperscaler spending rather than durable bargaining power.
Goldberg expects more issuance, spending, and casualties in vulnerable corners of the market, followed by cheaper compute and potentially transformative innovation.
What was said on this episode
51 statements · 10 positive · 26 negative · 9 mixed · 6 neutral
Parts of the data-center complex show bubble characteristics.
“I think there are definitely some signs of bubble. percolating through the data center complex.”
Listen at 2:26
The AI data-center complex will become more bubble-like in coming months.
“the whole thing is poised to get even bubblier in the coming months”
Listen at 2:39
NVIDIA financing is approaching the line between enabling and creating demand.
“I don't think it's entirely clear which side we're on yet, but we're getting pretty close to that line.”
Listen at 3:25
AI infrastructure investment will ultimately face timing issues rather than total waste.
“I think we end up with timing issues.”
Listen at 4:01
AI investment timing problems could be severe for some companies.
“some of those timing issues could be pretty drastic for some of the companies involved.”
Listen at 4:06
Hyperscalers and the AI complex will spend about $800 billion in CapEx this year.
“hyperscalers and the AI complex this year is going to spend, what, $800 billion in CapEx.”
Listen at 4:13
Hyperscalers and the AI complex will spend at least $1 trillion next year.
“Looks like they're going to spend at least a trillion dollars next year.”
Listen at 4:20
AI progress toward transformative outcomes will be nonlinear.
“I do think it's going to be a non-linear path to get there.”
Listen at 5:52
AI will achieve major advances, but progress will be uneven.
“we're going to get those advances from AI, those accomplishments from AI, but it's going to be a rocky path”
Listen at 6:20
NVIDIA and Broadcom financing the AI ecosystem carries considerable risks.
“there are considerable risks, what we're talking about here.”
Listen at 8:42
Neoclouds will seek NVIDIA support when AI demand tapers and capacity becomes excessive.
“that's exactly when the NeoClouds are going to be putting those options back to NVIDIA saying, hey, we need you to support us.”
Listen at 10:43
NVIDIA cash flow will decline once AI demand becomes saturated.
“NVIDIA's own cash flows are going to decline because demand is saturated.”
Listen at 10:55
Seven companies account for the AI complex’s capital expenditure.
“the entire AI complex, all of this CapEx, comes from seven companies.”
Listen at 12:56
AI lacks a compelling consumer compute use case.
“We don't have a compelling compute consumer use case for AI”
Listen at 14:45
Most enterprise AI pilots currently produce roughly 10% cost optimization.
“most of those pilots tend to do things like optimize costs by 10%.”
Listen at 15:32
Current enterprise AI pilot gains do not justify present investment levels.
“they're nowhere near the scale needed to justify the size of the investment now.”
Listen at 15:50
Debt limits AI adoption when benefits arrive on a different timeline.
“debt becomes a limiting factor”
Listen at 17:43
A server’s physical life is at least five years.
“a life of a server is at least five years.”
Listen at 19:23
GPU demand is strong while supply remains tight.
“Demand is very strong and supply is very tight.”
Listen at 20:37
Compute trading will develop but be more difficult than expected.
“the trading of compute is going to be not impossible. People will do it, but it's going to be trickier than you'd think.”
Listen at 22:28
AMD currently offers chips comparable to NVIDIA’s.
“AMD, who's sort of a pure play merchant chip vendor, has chips that are on par with NVIDIA.”
Listen at 23:20
NVIDIA’s CUDA moat is weaker than it was three or four years ago.
“It's much less potent than it was three or four years ago.”
Listen at 24:09
NVIDIA continues expanding its competitive advantages beyond chips.
“NVIDIA is not standing still, though.”
Listen at 24:13
Broadcom helps NVIDIA’s largest customers develop custom chips.
“NVIDIA's biggest customers designing their own chips. And Broadcom is helping them do that.”
Listen at 26:23
China poses a genuine competitive threat in AI chips.
“So yes, China's a real threat.”
Listen at 27:27
AI chip demand is real but somewhat inflated by financing.
“I think there's real demand, but it is to some degree a little bit, I don't want to say exaggerated, but there is some inflation taking place.”
Listen at 27:46
Major hyperscalers have reached or are approaching negative cash flow.
“They've all gone or are headed towards negative cashflow.”
Listen at 30:45
Cloud revenue from AI labs can resemble circular financing.
“it does start to look a little bit circular”
Listen at 32:14
Technology and chip companies depend critically on product cycles.
“tech companies, chip companies, live and die by the product cycle”
Listen at 36:01
Missing a product cycle while carrying debt can prevent funding the next cycle and end a company.
“if you miss a product cycle and you have a debt bill coming due, then you can't afford the R&D to get the next product cycle, and then you're done.”
Listen at 36:08
NVIDIA and Broadcom currently do not trade at extremely high multiples.
“NVIDIA and Broadcom do not have sky high multiples.”
Listen at 39:02
The semiconductor industry is substantially different from the dot-com era.
“The semis industry has changed radically since then”
Listen at 39:33
Semiconductor companies’ current balance-sheet use is sensible.
“they can absolutely making use of their balance sheet is is sensible right now”
Listen at 40:05
A 20% hyperscaler CapEx cut would sharply reduce Goldberg’s semiconductor stocks.
“If a hyperscale is cut CapEx by 20%, all my stocks fall pretty sharply.”
Listen at 44:00
Flat or reduced future hyperscaler CapEx guidance would unsettle semiconductor investors.
“they don't even have to cut it. They just have to say, oh, you know, 2028, we're going to be cautious about it. We're going to take it down. We're going to, we're just keep it flat. I think even that would get people nervous.”
Listen at 44:10
Major AI companies will return to positive operating free cash flow by end-2028.
“I actually believe that the inflection point is... the end of 2028, where all these companies once again return to operating free cash flow positive.”
Listen at 44:52
Hyperscaler CapEx is Goldberg’s key indicator for the AI boom.
“I obsessively track hyperscaler capex.”
Listen at 45:38
CoreWeave and Nebius are important indicators of AI demand weakness.
“the other place to keep an eye on are these neoclouds, CoreWeave, Nebius”
Listen at 46:25
Financing disruptions will appear first among smaller neoclouds.
“If there's any sort of hiccups, you're going to see it at the NEOs first”
Listen at 47:09
Neocloud contracts with OpenAI, Microsoft, and Meta will last five years.
“those are all going to be fixed left. Like those will last five years.”
Listen at 47:51
Neoclouds’ major contracts will eventually expire, exposing their remaining businesses.
“eventually that goes away. And then trying to figure out what's left after that.”
Listen at 47:56
Neoclouds will be absorbed into larger hyperscalers within five to seven years.
“I think five, seven years from now, there's no neoclouds. They're just part of these bigger hyperscalers.”
Listen at 49:40
The current number of neoclouds appears excessive.
“There are 200 neoclouds today. That seems like too many.”
Listen at 49:49
Many neoclouds will disappear, but some will survive.
“A lot of them are going to go away, but I don't know that all of them will.”
Listen at 50:42
The AI debt and infrastructure cycle will become more extreme.
“I think things are going to get crazier.”
Listen at 51:11
New AI debt issuance will intensify the infrastructure boom.
“all this debt is coming on stream. It's going to amp up things even further.”
Listen at 51:14
Further AI debt growth will produce bubble-like activity at the industry’s fringes.
“things will get even bigger, and then we'll start to see real bubbly things on the fringes of that.”
Listen at 51:30
Semiconductor supply constraints will persist until 2028 or 2029.
“There are real constraints in the industry right now that don't get worked out until 2028, 2029.”
Listen at 51:45
AI financing failures will hurt some companies but not become systemic.
“I don't think it'll be systemic, but I do think there will be casualties along the way.”
Listen at 52:12
The AI boom will become extreme, collapse, and make compute much cheaper.
“things are going to get very crazy and then it will fall and then compute will get very cheap.”
Listen at 52:52
Cheaper post-crash compute will enable a small team to discover a cancer cure.
“that crazy team in somebody's garage is going to find that cure to cancer”
Listen at 53:00
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
The Circuit
The related podcast is mentioned at the close and offers another route into the technology and semiconductor discussion.