He Flew to Austin With No Return Ticket to Win a Photo Customer
Luminal says it now runs almost all of AutoHDR's five million monthly images after working on site and shifting work to cheaper chips.

Standard Capital · Luminal Raises $25M Series A to Break AI's Hardware Bottleneck, Led by Standard CapitalA founder of Luminal, a startup that helps AI models run more cheaply, boarded a flight to Austin with no return ticket.
That detail comes from his own account of the trip. He said his plan was to sit next to the chief technology officer of AutoHDR, an Austin company that cleans up listing photos for real estate photographers, and keep working until Luminal handled the work.
AutoHDR manages the full photo pipeline for real estate media businesses. The photographers shoot houses going up for sale, and AutoHDR touches up the pictures before they are posted to sites like Zillow.
The founder said the stakes were scale. He put AutoHDR's volume at about five million images a month, and estimated that it accounts for about 25 percent of all images posted to Zillow.
Luminal was pitching what is called hosted inference, meaning it would run the AI part of that pipeline itself, renting and managing the specialized computer chips needed to do it.
In his telling, his pitch was blunt. He recalled asking who AutoHDR was using for that AI work, and saying he could do it better, then showing up in person to prove it.
He said he did what he promised, sitting with the chief technology officer and working until Luminal's software could take on the load. One major part of the effort, he said, was Luminal's compiler, software that translates an AI model into efficient instructions for a specific chip.
He paired that with cheaper hardware. He said Luminal ran the work across a mix of chips and was moving much of it to chips from AMD, which he said cost about half as much as comparable chips from Nvidia for the same amount of calculation.
As he described the outcome, Luminal ended up processing almost all of those five million monthly images. Powered by us, the founder said.
Questions people ask
What does a compiler do in AI?
A compiler translates an AI model written by researchers into detailed instructions a specific chip can execute. A good compiler uses more of the chip's capacity, through steps like managing memory and scheduling calculations, so the same hardware produces more answers for the same cost.
Why would a startup move work from Nvidia to AMD chips?
Nvidia dominates AI computing, but its chips are expensive and in high demand. Chips from AMD and other makers can cost less for the same amount of calculation, but they are often harder to program well, so startups that can make them run efficiently can pass savings to customers.
How is AI used to touch up real estate photos?
Listing photos are often brightened, straightened, and cleaned up to make rooms look larger and clearer. Newer tools use AI to do jobs like adjusting light, replacing gray skies, or removing clutter, which is faster than editing each photo by hand when millions arrive each month.


