What it means for mortgage lenders if DeepSeek helps innovation costs collapse
This Bloomberg TV segment with Ark Investment Management CEO Cathie Wood is a good primer on what really matters in the DeepSeek saga. The big deal isn’t about a new AI model competing with Llama, ChatGPT, Gemini, etc. It’s that DeepSeek represents a key inflection point on innovation costs collapsing. This is great for mortgage for lenders in a market where it costs $10,716 to close each loan.
To recap, DeepSeek’s R1 model can mimic the way humans reason, and for now seems to outperform Llama, ChatGPT, etc. on key benchmarks like mathematical tasks and general knowledge of a topic for a fraction of the cost.
In the mortgage space — where The Basis Point advises banks, lenders, and the fintechs that power them — this starts to get exciting.
Our industry is so highly regulated, lenders have been slow to embrace AI for two reasons.
First, human reasoning plays an outsized role in credit risk management (aka loan underwriting an approval) and compliance (machines can help, but licensed activity has to be done by licensed humans).
Second, putting armies of humans on these high volume tasks is super expensive. The mortgage industry will fund about 5.74 million loans this year, per MBA, at a cost of $10,716 per loan. And to actually fund this many units, lenders must perform these processes on at least double that amount of units.
So … just 2 years into the generative AI era, new players can match today’s functionalities at 10% of the cost.
How? Because this headline-grabbing moment is representative of going from the training phase of AI to the inference phase.
That’s a fancy way of saying the app layer is going mainstream
Training is teaching models about your data.
Inference is models making decisions with your data.
Training is way more resource-intensive (aka expensive) than inference.
But DeepSeek represents an inflection point for mortgage lenders on the inference (aka app) phase getting way cheaper.
For mortgage lenders, this hopefully means more streamlined, automated loan processing and underwriting — something we call loan manufacturing.
Cleaner manufacturing is cheaper.
But lenders will proceed slowly to comply with the thicket of Federal and state regs.
Still, it’s exciting.
And in the meantime, tech costs and managing tech vendors will actually get more — not less — complicated as new AI players flood the market even faster.
But The Basis Point’s advice to lenders is to keep experimenting with emerging generative AI tools.
And not just the marketing stuff. Better to test the income analysis, and other tools that help your loan manufacturing early in the process.
This will start to trim sales and fulfillment costs, which right now represent 42% of the $10,716 per loan cost.
Please comment below or reach out with questions.
___
Reference:
– Ark’s Cathie Wood on DeepSeek collapsing tech costs (Bloomberg TV)