In partnership with
OpenAI plans to spend more on compute than it expects to earn. Lenders have started pricing it in.
OpenAI expects to earn $840 billion between now and 2030.
It plans to spend $856 billion on computing power over the same stretch.
Both numbers come from a company presentation seen by the Financial Times this month. The gap works out to a projected cash burn of $278 billion. Revenue is supposed to climb from $36 billion this year to $350 billion in 2030. The spending comes first. The revenue is supposed to follow.
Where the money comes from
Last week SoftBank sold $11.1 billion in bonds, the largest high-yield corporate bond sale on record. The money covers SoftBank's final payment to OpenAI, bringing its total commitment to $64.6 billion.
The price was up to 9.75% a year. In 2021, SoftBank borrowed at rates as low as 2.125%. The cost of insuring its debt against default has climbed past 400 basis points, up from around 280 in June.
Here is the chain. Bond buyers lend to SoftBank. SoftBank funds OpenAI. OpenAI promises to buy compute. Companies like Oracle build the data centers with their own borrowed money and wait for OpenAI's revenue to arrive.
Each step makes sense on its own. The whole thing works only if the last step pays.
The agentic era needs a different CRM. That’s Attio.
Parallel, Turbopuffer, and Wordsmith run their entire GTM motion on Attio, with agents that chase every buying signal, build pipeline, and move deals forward, 24/7.
The 90-cent loan
In New Mexico, Oracle is the anchor tenant of Project Jupiter, a campus built to draw 2.45 gigawatts and deliver compute to OpenAI. About 20 banks put up an $18 billion loan to build it.
The banks planned to sell that loan on to other investors. The buyers did not show up at full price. The debt is now quoted at 89 to 91 cents on the dollar.
Last week Oracle sent a force majeure notice to the developer. A permit for a gas pipeline to power the site had been denied, and the schedule slipped. The notice gives Oracle room to delay payments if the campus misses its 2028 target. Oracle says the site is on track and no payment has been missed.
A data center without power is a building.
The valuation gap
This month, investors offered to buy into OpenAI at a $1.2 trillion valuation. OpenAI says it is worth at least $1.5 trillion. Sam Altman also said the company will not go public in 2026, calling this an ill-advised moment.
Same company. Same numbers. $300 billion apart.
What it signals
Nobody serious is saying AI is fake. Oracle's cloud revenue is growing fast $ORCL ( ▼ 3.12% ) . OpenAI's revenue is real. What changed is the price of money. A year ago capital arrived cheap. Now it arrives at 9.75%, or at 90 cents on the dollar, or with a force majeure clause attached.
The binding constraint in New Mexico was not chips. It was a pipeline permit.
What it means
If you build AI products for European clients, your customers already ask where their workloads run and who controls the stack. Now there is a second question: how stable is the financing behind the provider they depend on?
Most business work does not need a frontier model running in a 2.45-gigawatt campus. Document processing, classification and routine code often run fine on smaller models, hosted on infrastructure the client can audit and budget for.
The frontier labs are borrowing at record rates to build capacity. The opening is on the other side: products that route each task to the smallest model that can do it, on European infrastructure, priced so the client never carries someone else's debt structure. The buyers for that have started reading the bond market.
404 Found covers AI developments from a European Insider, three times a week.
Analytics on Live Data Without Leaving Postgres
When analytics on Postgres slows down, most teams add a second database. Then come the pipelines, the sync jobs, and a copy of your data that's always a little behind.
TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage up to 95%. Continuous aggregates keep dashboards live without re-querying everything.
CERN runs Postgres this way for sensor data from the Large Hadron Collider.
No split architecture, no pipeline lag, no new query language to learn. Same SQL, same drivers, same tools.
Start on Tiger Cloud and get $1000 in credits.