Formula 1: The GOAT

”The idea is pretty solid.”

”What is the second?”

”A GPU compute infrastructure software company,” he said, and immediately understood from the look on his mother’s face that she was completely clueless as to what he was talking about. So, he had to start from the beginning to explain to her what he wanted the company to be, what it should do, and why.

As someone who had lived through the AI bubble, he knew very clearly that the largest winner of it all was NVIDIA, who, instead of rushing to dig for gold, just manufactured and sold the shovels that everyone who wanted to dig for gold needed to buy in order to be competitive.

In the past, he didn’t understand why NVIDIA had become the biggest winner among the GPU makers when there was also AMD and Intel, who, although they benefited from the boom, did not reach the trillions of dollars in valuation that NVIDIA did. After taking the time to look into it in-depth in this lifetime, he had finally discovered what exactly made NVIDIA the winner: CUDA.

Compute Unified Device Architecture. It is a proprietary parallel computing platform and programming model created by NVIDIA that allows software developers to use NVIDIA GPUs for general-purpose processing. More importantly, it was the platform on which all of the AI giants had built their systems, locking them into NVIDIA and giving them pricing power beyond what others like AMD and Intel could achieve.

The pricing difference was double what the other two could offer, and NVIDIA still sold millions of them, while AMD and Intel, despite pricing their products at nearly half, failed to find customers due to the software being proprietary to NVIDIA.

So, his idea was to create a company that would be solely focused on creating a competitor for CUDA, with the difference being that it would be GPU-agnostic, meaning it could be used by any GPU and would provide flexibility to AI companies on the type of GPU they needed to buy during the large GPU shortage that would arise from the release of ChatGPT, Anthropic, and other major AIs.

If it succeeded, it would be on the route to becoming the Microsoft of AI. But he knew it was going to be expensive to start, at least a hundred million dollars for the initial year and even more for the following years. He also knew that they only needed to survive the first year and would then have as much cash as they would need to rapidly develop everything in time for the 2022 release of ChatGPT, which would then unleash the rush into the AI market and the beginning of the bubble.

”That sounds expensive,” Rümeysa, who was yet to know how much it would cost, said as she finished noting everything he had explained.

”If we succeed, this company will be very big and a good one, so we need to have the best foundation possible from the start. After all, the better the foundation, the higher we can build, and the return should be at least fifty times what we put in if it succeeds,” he said, trying to make sure she understood that he was serious, as a system like that needed at least three or four years before it became competitive.

”But isn’t this then dependent on the other GPU companies accepting it as the standard and actively implementing it into their software? Why would AMD and Intel agree to do that?” Although she wasn’t fully knowledgeable about the field, the details Fatih had given her were enough for her to see the possible weaknesses or blind spots that he might have missed.

”They will agree if it is a good enough system. Those two are mostly hardware powerhouses and build GPUs of the same quality; the only thing is that their own software isn’t competitive enough to compete with CUDA because it has a fifteen-year head start, and many of the people in this field are already locked into the CUDA infrastructure.

If they want to migrate to another GPU, they would need to rewrite their codes from the beginning and would, in return, once again be locked into a new infrastructure with even less support than CUDA. If we demonstrate the capabilities, I believe they are going to embed their own engineers into this system to optimize the software for their GPUs.

And if the company I have in mind enters the market, we would also build a conversion software to convert CUDA code into our own code, requiring little touch-up. This would turn what would have been an 18-month rewriting job into about a week of work, allowing a vendor who still has NVIDIA hardware and wants to increase their number of GPUs to go and buy AMD’s GPUs, which are cheaper, migrate to the new software, and still retain the usefulness of their already-owned GPUs.

They could even go further and have a multi-vendor GPU infrastructure. We will also be building a system that will assign the best GPU for a task. The target is to tax the entire ten-billion-dollar AI market,” he added, knowing full well that this was going to become a multi-trillion-dollar market, and if they managed to embed themselves, they would find themselves taking about two to three percent of that market, which in itself would be billions.

”Isn’t CUDA already embedded? How do you plan to get it adapted by people in this field who are already used to it?” she asked, once again hitting a very important point.

”In my mind, I have two routes we will need to take. The first is to open-source some parts of the system, just like Linux, and keep some parts proprietary, like the part that assigns tasks to the best GPU in a cluster, the management system, and more.

This will also have companies like Google, who don’t want to rely on NVIDIA’s monopolistic pricing, supporting it, as they would improve the open-source code to better fit their systems, and other companies will do the same.

The second one is to mirror what NVIDIA did for CUDA, which was supporting universities by providing them with free computing and funding to do research using our system. With the open-source nature of ours, it would be adapted even faster and also teach the next generation of engineers how to use our system, just like NVIDIA did with CUDA.” He was doing his best to make it clear that he had thought deeply about it before suggesting any of the plans, to reduce the chances of her denying them to near zero.

If she denied it, he would be in a bind and would have to wait until he was eighteen to then be able to make those decisions, but by then, it would be too late to get into this market, and it would require more money for the same results.

This CUDA competitor company was just step one of the multi-step plan he had in mind to fully exploit the upcoming AI bubble by selling golden shovels and watching as people kept fighting for theoretical gold.

”I will forward it to them, and we should have a completed investigation and research document by the end of the week for at least the first one. This second one should take at least a month,” she said as she took a picture of both the notes she had taken and the information in the peaice of paper Fatih had given her and forwarded it to the head of the team, having them start looking into it before she had a meeting with them to ask Fatih some questions if they had any, and for him to answer if he had the answers, or for them to look for the answers if he didn’t.

”Thank you, Mom,” Fatih said with a bright smile, as he knew the first step had been taken, and although the hardest step still remained, it was something he would handle when it came.