In 2016, everyone was iterating on the GPU. One team looked at the same chip and named a problem the rest of the field worked around.
Only about 4% of a GPU’s silicon did the actual AI math. The rest was graphics machinery, carried along out of habit. Andrew Feldman’s read: the hard part of deep learning is moving data around, rather than doing the multiplication. So the right machine gets built around data movement, from a blank sheet.
That idea lived in a Series A deck marked working name, from a company so early its identity was still a working title and its vocabulary predated the word transformer. Most people who saw it thought it was crazy.
In May 2026, that company, Cerebras, IPO’d at a valuation near $56 billion, the biggest US tech IPO since Uber.
Here is the part worth keeping even if you stop here. Two questions decide a bet like this, and the deck answers both:
Is the demand curve about to bend? They saw AI compute going vertical while the broader industry still treated AI as a science project.
Is the incumbent tool accidentally good, or deliberately good? The GPU was accidentally good at AI, a graphics chip pressed into service. That gap is the opening.
When demand is bending and the incumbent is only accidentally good, you have the setup for a generational company. That lens works on any hardware bet you evaluate.
Now the concrete part. We got the confidential Series A deck, and the numbers behind the climb. Below the line:
▫️ The confidential Series A deck, all 20 slides, with what they pitched versus what they shipped
▫️ The 135x bridge, the engineering slide that splits the headline claim into four honest multipliers
▫️ The revenue climb, $25M to $510M, with the growth rates and the swing to profit
▫️ Every round, from the seed off this deck to the $23B private mark to the $56B IPO
▫️ The IPO play-by-play, the pricing, the 68% pop, and the near-halving six weeks later
▫️ The risk file, the customer concentration and OpenAI dependency the headlines skip
▫️ The operator lessons, what this deck teaches anyone building in AI right now
1 subscription unlocks every teardown
This is one breakdown in a growing library. Premium opens all of them:
