How OpenAI ends and takes Oracle with it | Ed Zitron

The Unsustainable AI Bubble and the Shift to Token-Based Billing [00:00]

• The AI industry is facing an inevitable transition to token-based billing as a way to avoid "annihilating billions of dollars" in operational costs.
• Valuations for companies like Anthropic, reaching up to $1 trillion in secondary markets, are described as "extremely detached from reality" due to low liquidity and volume in private secondary sales.
• The excitement surrounding OpenAI seems to be waning, with its Spud 5.5 model described as "mediocre," while Anthropic still garners some excitement, leading to a shift in market interest.
• Anthropic's valuation is characterized as a "sham built on venture capitalist subsidies," with a lack of transparency regarding the actual running costs of their AI services.
• Users are spending significantly more on token usage ($2,000-$4,000 per month on a $200 plan) than anticipated, suggesting that current subscription models are not reflective of actual costs or are heavily subsidized.
• There is no proof of the claimed 50% gross margins for companies like Anthropic; it is speculated that their margins are likely "in the toilet."
• The move by GitHub Copilot to token-based billing is a significant indicator that even Microsoft, a highly profitable and well-positioned AI player, is feeling the financial pressure to lower costs.
• The transition from fixed monthly subscriptions to token-based billing for GitHub Copilot ($19/month for $19 worth of tokens) signifies a tightening of financial belts.
• Companies have trained users to expect subsidized products, making a transition to paying for actual token costs a potentially unpopular move, leading to user backlash.
• The true value and popularity of AI products will be revealed when the underlying compute costs are no longer subsidized.

"The problem with this whole bubble, one of many I should say, is that they never should have done monthly subscriptions. It was a mistake from the beginning."

The Financial Peril of Oracle and OpenAI's Data Center Ambitions [22:10]

• Oracle's Stargate data center project in Abilene, with a planned 1.2 gigawatt capacity, is significantly behind schedule, with only two of eight buildings operational and minimal hardware installed, impacting Oracle's ability to collect project management fees from OpenAI.
• Oracle is projected to spend approximately $2.14 billion annually on the fully built Stargate facility, with theoretical revenue of $10 billion, but completion is now estimated for Q1 2027, far beyond initial targets.
• The overall Stargate data center project, totaling 7.1 gigawatts, is expected to cost $340 billion and generate $75 billion in annual revenue, but completion is now pushed to late 2030 or later, with predictions of significant delays.
• OpenAI requires $852 billion in funding or debt over the next four years, a figure that exceeds their projected revenue of $673 billion, indicating a massive funding shortfall.
• If OpenAI fails to pay Oracle, Oracle faces potential bankruptcy, as it has no other tenant capable of affording the projected $75 billion annual cost and has experienced negative cash flow of nearly $25 billion in the previous quarter.
• Larry Ellison has pledged $60 billion worth of his Oracle shares for personal loans, which could lead to severe margin calls if Oracle's deal with OpenAI collapses and its stock price plummets.
• Oracle's cloud business, while growing, has negative to low margins, and its other businesses are plateauing, making the OpenAI deal critical for its survival.
• OpenAI owes substantial amounts to other cloud providers, including $138 billion to Amazon Web Services, $250 billion to Microsoft Azure, and significant sums to Cerebrus and Corewave, further straining its financial resources.
• The planned installation of 10 GW of Broadcom GPUs by the end of 2030 for OpenAI seems infeasible given the current construction and installation timelines.

"And by the way, if OpenAI cannot pay Oracle, Oracle dies. This is not hyperbole. This is the financial situation Oracle has got itself into."

GPU Obsolescence and the Looming Data Center Crisis [28:53]

• The GPUs intended for the Stargate Abilene data center, such as GB200 and future Vera Rubin GPUs, will be significantly outdated (two to three years old) by the time the facilities are operational, due to lengthy construction timelines.
• Nvidia's transition to newer, more powerful "Kyber racks" requires entirely new infrastructure, making older Blackwell-based data centers and their GPUs potentially obsolete shortly after deployment.
• Millions of Blackwell B200 GPUs are in inventory, with companies like Super Micro holding over $1 billion worth that they cannot sell, signaling a potential glut and a first-ever write-down of Nvidia AI GPU inventory.
• The rental rates for Blackwell GPUs are reportedly as low as $3.70 per hour, potentially not covering the infrastructure costs and debt associated with housing them.
• The prospect of GPUs being dumped onto the secondary market is a significant concern, potentially leading to massive impairments of value and billions of dollars in losses for hyperscalers and their lenders.
• A wave of refinancing and renegotiation of loans within the data center industry is anticipated, with companies like Corewave and Nebius facing potential financial distress and construction halts.
• The true profitability of these massive data center investments remains largely obfuscated, a worrying sign that the underlying economics may not be favorable.
• Nvidia is unlikely to buy back GPUs, suggesting that companies will attempt to use them even at a loss, increasing the risk of financial failure for those unable to cover their operational costs and debt.
• The financial survival of Oracle is precarious, heavily reliant on the AI GPU market and OpenAI's ability to pay for its services; without this, Oracle faces significant restructuring or potential dissolution.

"So yeah, everything that you're hearing about data centers is already obsolete and it just means that and the worst thing is is that with Blackwell especially, all of the data centers you've heard announced recently are all Blackwell."