AI's cheap tokens just hit a hard stop

The AI buildout ran out of cheap money, not chips.

Three things landed on Saturday and they all answer the same question: who pays for the next data centre. OpenAI took money off its own fundraising target. Microsoft, with $92 billion of revenue in the quarter, went to the bond market anyway.

Advertisement

1. OpenAI took $10 billion off its own round

OpenAI is now seeking about $20 billion in its current round, down $10 billion, the Financial Times reported on Saturday. The targeted valuation stands at $1.25 trillion against roughly $1.4 trillion before, a cut we covered on 24 October.

The detail under the headline is worse than the headline. The first close is expected to raise less than even the reduced target after some investors pulled back their allocations. A consortium of UAE investors including the state-backed fund MGX is expected to be the largest participant, and it gets an option to lease new computing capacity in the region.

The Nasdaq Composite fell 2.9% that Friday and sits 7.8% below its closing high of 10 October.

The same day, CNBC wrote that Anthropic's path to the Nasdaq just got a lot bumpier, and CTech in Tel Aviv had Google, Amazon and Intel stumbling as investors began to question the spending.

I do not think this is the bubble going. I think it is the end of free optionality. For two years the price of a token kept falling because somebody else was carrying the capacity, and a round that shrinks by a third is the first visible edge of that subsidy. My expectation is that list prices for enterprise inference stop falling before the end of Q1 2027, which makes the cost controls everyone deferred because inference kept getting cheaper a spring problem rather than a 2028 one.

Read the source →

2. Microsoft is borrowing $25 billion to pay for capacity

Microsoft announced a $25 billion debt offering to fund capital expenditure, acquisitions and refinancing. In the same set of numbers: quarterly revenue of $92.28 billion. It booked $92 billion in three months and still went out for a loan.

The Wall Street Journal, the same day, called the shortage of cheap money the new bottleneck in the American buildout and Wall Street the party now calling the shots. On the revenue side of the same ledger, Perplexity has signed a $750 million agreement with Microsoft for Azure.

Borrowing against a cash machine is arithmetic, not panic: bonds cost less than the equity Microsoft would otherwise issue, and the halls have to go up now. What interests me is the shape of the obligation. A coupon gets paid whether or not anybody opens Copilot, so the pressure to grow paid seats does not soften next year, it hardens. I would read the next Copilot renewal as a conversation about how long you commit for, not about what discount you get.

Read the source →

Advertisement

3. Cipher Digital is paying for its data centres with borrowed money

Three moves from one operator, written up on Saturday:

• A planned $850 million zero-coupon convertible note offering.
• Its Barber Lake lease extended to 20 years, lifting contracted revenue at the site.
• A 15-year agreement with Amazon Web Services.

On 13 October the same outlet traced Cipher's route out of bitcoin mining into AI infrastructure, and on 2 October Forbes counted the trillions already committed to the data centre buildout.

A zero-coupon convertible is a polite way of saying the lender takes shares if this works and a place in the queue if it does not, and the twenty-year lease next to the fifteen-year AWS contract is what made anybody willing to sign it. That matters when you buy inference from an operator this size, because you are buying the building's financing as much as its racks: who owns the site decides what happens to your workload when the refinancing goes badly. I would want to know who holds the paper on a site before committing to anything longer than a year on it.

Read the source →

Names to watch

Seen this week by Ins7ghts.

• Machenta · teaches a robot a new task from one demonstration
  Data quality. It raised $40 million on 28 October on the argument that it needs far less training data than the usual approach, and two days later the Center for Data Innovation put the single-demonstration system on its data news list.
• Anglera · tidies up messy product data
  Data quality. Coveo announced the acquisition on 29 October, describing it as AI-native product data enrichment, and a funding roundup the next day has it as Y Combinator backed.

Companies that later raised money had been in the news about twice as often in the quarter before. Worth a second look in three months.

Advertisement

Also today

• Microsoft made automation flow evaluations generally available in Copilot Studio. (8 min read)
• Amazon's chief executive expects six to eight AI models to be “good enough” for most enterprise work. (4 min read)
• Oracle said it will integrate OpenAI's models, ChatGPT and Codex deeper across its applications, database and cloud. (5 min read)