Forgive me for beginning with a cliché, a chunk of finance jargon that has not too long ago slipped into the tech lexicon, however I’m afraid I have to speak about “moats.” Popularized many years in the past by Warren Buffett to discuss with an organization’s aggressive benefit, the phrase discovered its approach into Silicon Valley pitch decks when a memo purportedly leaked from Google, titled “We Have No Moat, and Neither Does OpenAI,” fretted that open-source AI would pillage Massive Tech’s fort.
Just a few years on, the fort partitions stay protected. Other than a quick bout of panic when DeepSeek first appeared, open-source AI fashions haven’t vastly outperformed proprietary fashions. Nonetheless, not one of the frontier labs—OpenAI, Anthropic, Google—has a moat to talk of.
The corporate that does have a moat is Nvidia. CEO Jensen Huang has referred to as it his most treasured “treasure.” It’s not, as you would possibly assume for a chip firm, a chunk of {hardware}. It’s one thing referred to as CUDA. What appears like a chemical compound banned by the FDA will be the one true moat in AI.
CUDA technically stands for Compute Unified Gadget Structure, however very like laser or scuba, nobody bothers to increase the acronym; we simply say “KOO-duh.” So what is that this all-important treasure good for? If pressured to provide a one-word reply: parallelization.
Right here’s a easy instance. Let’s say we activity a machine with filling out a 9×9 multiplication desk. Utilizing a pc with a single core, all 81 operations are executed dutifully one after the other. However a GPU with 9 cores can assign duties so that every core takes a distinct column—one from 1×1 to 1×9, one other from 2×1 to 2×9, and so forth—for a ninefold velocity acquire. Fashionable GPUs will be even cleverer. For instance, if programmed to acknowledge commutativity—7×9 = 9×7—they will keep away from duplicate work, lowering 81 operations to 45, practically halving the workload. When a single coaching run prices 100 million {dollars}, each optimization counts.
Nvidia’s GPUs had been initially constructed to render graphics for video video games. Within the early 2000s, a Stanford PhD scholar named Ian Buck, who first acquired into GPUs as a gamer, realized their structure could possibly be repurposed for normal high-performance computing. He created a programming language referred to as Brook, was employed by Nvidia, and, with John Nickolls, led the event of CUDA. If AI ushers within the age of a everlasting white-collar underclass and autonomous weapons, simply know that it will all be as a result of somebody someplace enjoying Doom thought a demon’s scrotum ought to jiggle at 60 frames per second.
CUDA is just not a programming language in itself however a “platform.” I exploit that weasel phrase as a result of, not in contrast to how The New York Occasions is a newspaper that’s additionally a gaming firm, CUDA has, over time, develop into a nested bundle of software program libraries for AI. Every operate shaves nanoseconds off single mathematical operations—added up, they make GPUs, in business parlance, go brrr.
A contemporary graphics card isn’t just a circuit board full of chips and reminiscence and followers. It’s an elaborate confection of cache hierarchies and specialised models referred to as “tensor cores” and “streaming multiprocessors.” In that sense, what chip firms promote is sort of a skilled kitchen, and extra cores are akin to extra grilling stations. However even a kitchen with 30 grilling stations received’t run any sooner with no succesful head chef deftly assigning duties—as CUDA does for GPU cores.
To increase the metaphor, hand-tuned CUDA libraries optimized for one matrix operation are the equal of kitchen instruments designed for a single job and nothing extra—a cherry pitter, a shrimp deveiner—that are indulgences for house cooks however not if in case you have 10,000 shrimp guts to yank out. Which brings us again to DeepSeek. Its engineers went under this already deep layer of abstraction to work straight in PTX, a type of meeting language for Nvidia GPUs. Let’s say the duty is peeling garlic. An unoptimized GPU would go: “Peel the pores and skin along with your fingernails.” CUDA can instruct: “Smash the clove with the flat of a knife.” PTX helps you to dictate each sub-instruction: “Elevate the blade 2.35 inches above the slicing board, make it parallel to the clove’s equator, and strike downward along with your palm at a power of 36.2 newtons.”
