Back to News

ByteDance would rent OpenAI models from Microsoft as its chip route narrows

ByteDance is in talks to spend about $500 million more on Azure for OpenAI models, as banks paused a $480 million Nvidia chip loan tied to it. Our model puts renting at about 5.5 times the per-token cost of owning the chips.


ByteDance would rent OpenAI models from Microsoft as its chip route narrows

We built this rent-vs-own model in Compound in minutes. Plug in your own GPU and token prices. Try Compound free, no credit card needed.

ByteDance is in advanced talks to spend about $500 million more on Microsoft's Azure, mostly to use OpenAI's models, The Information reported on Wednesday, citing two people with knowledge of the talks (unconfirmed). The same week, banks paused a $480 million loan for Nvidia chips ByteDance would use, Bloomberg reported.

Put side by side, ByteDance looks set to rent AI output in the US at roughly the price of the chips it is finding harder to own in Southeast Asia. On Monday the US began requiring data-center operators in countries including Singapore and Malaysia to get a licence for advanced AI chips, TNW reported. The $480 million loan was for GPUs at a Malaysian data center run by Singapore-based Aolani Cloud.

Rent vs own

We modeled what each route buys. With our assumptions, $45,000 per GPU all-in, a four-year life, 60% utilization, $0.60 an hour to run and 3 million tokens served per GPU-hour, the $480 million of chips works out to about $0.91 per million tokens. At an assumed blended $5 per million tokens on Azure, renting costs about 5.5 times as much, and the $500 million buys about 100 trillion tokens against about 673 trillion over the life of the chips.

The multiple swings with the inputs: from 0.7x if API prices fall to $2 and chips serve only 1 million tokens an hour, to above 50x at $15 and 10 million. What holds across the table is that losing the chip route makes each unit of AI more expensive for ByteDance, and moves that spending onto Microsoft's books.

Model output: ByteDance's cost per million tokens owning Nvidia GPUs vs renting OpenAI models on Azure, tokens each route buys, and an Excel data table of the renting cost multiple by API price and GPU throughput.

What it means for Microsoft and OpenAI

For Microsoft, the money is small. ByteDance was already on track to spend more than $1 billion a year on Microsoft's AI and cloud services, Bloomberg reported in June, per GuruFocus. With the new deal that is about $1.5 billion, or 0.43% of Microsoft's $346.4 billion in revenue over the latest 12 months.

The friction is with OpenAI. OpenAI blocks most Chinese companies from using its models, but Microsoft, which can sell OpenAI's technology through Azure, does not apply the same restrictions, The Information reported. OpenAI has complained that Chinese firms distill its models; the report does not say ByteDance has. Microsoft's licence to OpenAI's model technology runs through 2032.

Download the model

XLSX fileByteDance rent-vs-own AI model (Excel)Cost per million tokens owning GPUs vs renting models on Azure, tokens each route buys, and a native Excel data table on API price and GPU throughput.Download ยท Free

Download the file

ByteDance rent-vs-own AI model (Excel)

Tell us who you are and the download starts right away.

By downloading, you agree to our Privacy Policy.

What to watch

  • Whether Microsoft and ByteDance sign, and over what period the $500 million is spent.
  • Whether OpenAI seeks limits on Chinese customers in Azure's resale of its models.
  • Whether Aolani Cloud's $480 million loan survives the new US licence rule.