AI Giants' Earnings Look Like Enron: The Compute Bubble Backed by Nvidia
AI executives no longer talk tech, only money and gigawatts. Nvidia props up customer revenue with investments and guarantees, playing financial engineering like Enron—but not yet illegal.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Executives no longer talk tech, only money
Steve Burke points out that Jensen, Lisa Su, Sam Altman, and other executives now rarely discuss any technical details on podcasts and TV—it's all tens of billions, hundreds of billions, trillions in TAM and gigawatt numbers. Previously, they appeared only at major keynotes and then vanished for six months; now they show up frequently but only talk finance. Ed adds that even Lisa Su, once considered pragmatic and not given to marketing stunts, now sounds like she's huffing paint when she speaks.
— Steve BurkeCompute pricing is far below cost
Steve reveals that Oracle rents H100 GPUs for about $4 per hour, but according to Oracle insiders he's spoken with, just the operating cost (opex) to run an H100 is over $6 per hour, not even counting depreciation. Long-term contracts are cheaper: Oracle's H100/H200 contract price for OpenAI is just over $2 per hour. At that price, even at full utilization, it would take ten years to break even, with a gross margin of only 30%—worse than running an airline.
— Steve BurkeNvidia's financial engineering resembles Enron
Steve breaks down Nvidia's financial maneuvers: it never borrows money itself but instead invests in customers or guarantees their debt. For example, it invested $1.5 billion in SB Energy and guaranteed the data center's value at $10.8 billion once built. If OpenAI can't lease out capacity, Nvidia covers the difference. It also helps CoreWeave secure revenue by committing to purchase compute from them, while Jensen's foundation rents compute from CoreWeave. These moves stay within legal bounds, but they resemble Enron's SPVs to offload bad debt—except Nvidia is smarter, never becoming a direct creditor.
— Steve BurkeJensen dodges the bubble question
When Axios asked Jensen whether this is a bubble, he paused noticeably and chose his words carefully, saying a bubble will come someday, but not now—rather in five to ten years. He used a word starting with B; Ed thought he was going to say 'bubble,' but he said 'build out.' Steve also notes Jensen said compute should become an investable asset class, even traded like a commodity. Ed sarcastically remarks that's like treating GPUs in a warehouse as a separate asset class for tax purposes.
— Ed ZitronThe PR scam of data centers generating their own power
Ed mentions the Trump administration's push for data centers to generate their own power, ostensibly to prevent residential electricity rates from rising, but Trump said at a meeting that 'data centers need good PR right now.' Steve points out that energy is a hazardous material and cannot be deregulated, and that AI practitioners severely underestimate the difficulty of power. He also notes that gas turbines, meant for emergency backup, are now being run around the clock as primary power for data centers, causing them to wear out prematurely.
— Steve BurkeOpenAI's surveillance dream
Ed contrasts Microsoft's Recall feature, shelved due to strong user backlash, with OpenAI's high-profile launch of a similar feature, with Sam Altman saying AI can watch everything you do. Steve adds that smart TV manufacturers are collecting massive amounts of data, even using ACR technology to capture screenshots of users' screens, currently analyzed by humans but eventually intended for LLM automation. This kind of surveillance use is exactly the ideal customer for OpenAI and Anthropic, but Steve doubts that if such an efficient surveillance tool existed, someone would have already bought it.
— Ed ZitronThe absurdity of mandatory AI use in companies
Ed shares a friend's experience at a major game company: internal policy makes AI use part of the job, and not using AI counts as dereliction of duty. Management requires employees to figure out how to make AI help them do their work and then report back to management—if AI truly can replace them, the job might disappear. Steve adds that this is like what Nick Sireesh said before: everyone in the company is loudly proclaiming they use AI, afraid of being outdone, creating a culture of collective irresponsibility.
— Ed ZitronSam Altman's AI podcast advice
Ed mentions Sam Altman tweeted that you could ask your child how their day went and then use AI to generate a podcast for them to listen to on the way to school. Ed sarcastically asks, why not just talk to the child? The tweet got many likes, with one reply saying 'or you could talk to your child.' Ed believes that making kids listen to such a sycophantic machine is terrible parenting; AI is bad for adults, and even more dangerous for children.
— Ed ZitronIn their own words · checked verbatim
the whole thing feels like we've completely lost the plot of any of the technology
Steve Burke6:16
That is not enough to cover the costs. That's not even fucking close enough.
Steve Burke12:27
the next word started with the B and I was like, please be bubble, but he said build out unfortunately
Ed Zitron25:48
President Trump literally in the meeting said data centers need some good pr right now.
Ed Zitron28:51
if they don't find ways to burn tokens on queries, then they're not doing their jobs.
Ed Zitron39:04
Figures
| Oracle H100 hourly rental price | about $4 | 12:27 |
| Oracle H100 hourly operating cost | over $6 | 12:27 |
| Oracle contract price for OpenAI H100/H200 per hour | just over $2 | 13:28 |
| Nvidia investment in SB Energy | $1.5 billion | 20:41 |
| Share of CoreWeave revenue from Nvidia | about 12% | 23:44 |
| Value of Jensen's foundation donated CoreWeave compute | $108 million | 22:44 |
| Broadcom's share of revenue from AI in coming years | over 60% | 32:55 |
Glossary
- ACR (Automated Content Recognition)
- Technology that identifies content by capturing screenshots or fingerprints on smart TVs.
- SPV (Special Purpose Vehicle)
- An entity used to isolate financial risk; Enron used it to hide bad debt.
- Opex
- Operating expenditure: day-to-day costs like electricity and maintenance.
How to listen
Investors, tech journalists, and hardware enthusiasts focused on the AI bubble and semiconductor financial engineering.
The opening minutes of chatter about local data center protests can be skipped.