What Training Actually Is
Training is the step where a model learns. You feed it enormous amounts of data, it makes predictions, you measure how wrong those predictions are, and you nudge the model’s internal weights to be a little less wrong next time. Repeat that loop billions of times and a pile of random numbers becomes something that can hold a conversation.
That loop is staggeringly expensive in compute. Every pass over the data requires a massive number of arithmetic operations, which is why training runs happen on thousands of specialized chips working in parallel for weeks or months at a time.
This is the key distinction between training and the other half of AI. Training is a one-time, compute-hungry build phase, while inference is the cheap, repeated act of using the finished model. The AI inference vs training piece unpacks that split in detail.
The Chips That Do the Work
Training has been dominated by graphics processing units, and Nvidia built its AI franchise on the fact that its chips happen to be very good at exactly the kind of parallel arithmetic training requires. The Nvidia GPUs article explains how that lead was built.
There are challengers. Cerebras, the wafer-scale company behind this promo’s “50X faster” claim, attacks the same problem from a different angle by building one enormous chip instead of many wired-together ones. The architectures differ, but they are all chasing the same thing: more arithmetic per dollar per watt.
What every one of these designs has in common is that none of them springs fully formed from a napkin. They are designed in software first.
Where EDA Software Comes In
This is the link the promo leans on. Every training chip, whether from Nvidia, Cerebras, or anyone else, is designed using electronic design automation software. You cannot tape out a modern processor by hand, and the design tools are not optional.
Synopsys (SNPS) is one of the two dominant EDA companies, and its tools are used across nearly every advanced chip design flow. The business model is elegant for investors: it earns revenue on the design activity itself, which rises with the total number of chips being built, without betting on which architecture wins.
That is why an EDA name shows up in a pitch about training. Training drives demand for chips, chips drive demand for design tools, and the design tools are a durable, high-margin layer sitting upstream of all of it.
Why the Training Angle Matters
For an investor, the training story is the reason the demand feels so durable. Companies are not just building one model; they are continuously retraining and building specialized agents, which means the compute and the design work do not stop after a single launch.
The honest caveat is that training is also where the hype concentrates. The headline speed claims attached to any given chip are usually narrow benchmark figures, and a fast inference result does not automatically translate into training dominance. The AI workloads piece explains why the two are different businesses.
If you want the full accounting of the promo’s claims, the AI Super Chip teardown is the place to start.
The Cost of a Training Run
It helps to put a rough number on why training drives so much of the AI hardware market. A single large-model training run can consume thousands of chips for weeks or months, and the bill runs into the tens of millions of dollars before the model ever answers its first prompt.
That cost is why training is a strategic decision rather than an experiment. Companies budget for it, plan around it, and buy compute well in advance, which is exactly the kind of demand that flows through to chip designers and, one step up, to the EDA tools those designers use to build the chips.
It is also why training and inference behave differently as businesses. Training is lumpy and concentrated among a relatively small number of large buyers. Inference spreads across every company that deploys a model, which is where the long-term volume growth actually lives.
Knowing which of the two a given chip claim is aimed at is half the job of reading the marketing. A number that sounds spectacular on an inference benchmark says little about training economics, and the pitches rarely volunteer which one they are quoting.
For the investor, the training step is where the spending is most concentrated and most durable, and it is the reason the design-tool layer underneath it keeps compounding. Every new model means new chips, and every new chip means more design work.
Ready to see the research? Click here to access the Brownstone Research report.
NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.