08Code
He kept a founder's stake in one company for thirty-three straight years while quietly taxing every chip he shipped to fund a software platform that, sixteen years later, made his hardware the only practical substrate for deep learning — so when the AI boom arrived, he owned the toll booth.
- Started with
- A professional Taiwanese family who bought their sons an American education via a relative. Real hardship in the middle of it — sent at nine to a rural school for troubled kids, cleaning restrooms daily. Then an engineering degree and ten years of paid apprenticeship inside the industry he would later disrupt.
- The decisive move
- Deciding, around 2004–06, to make every chip capable of general-purpose computing — paying silicon area, unit cost and gross margin on every unit shipped, for a customer base of approximately zero. Roughly $475m and a third of R&D over four years. Gross margin fell from 45% to 35%. He then refused to reverse it for a decade while the market punished him.
- How the money was realised
- Never. No sale, no buyout, no liquidity event. His percentage declined for twenty-seven years, from 10–13% to 3.58%, almost entirely through employee equity issuance rather than selling. The fortune is the market repricing an unsold position. Roughly 97% of it arrived in the final four years of a thirty-three-year hold.
- Elapsed time
- Thirty-three years, honestly counted. An 84% drawdown in 2008–09, and a stock that did not reclaim its 2007 peak until May 2016 — eight and a half years underwater.
Transfers
- Do the apprenticeship before you found anything. Ten paid years inside the industry meant the company was built on knowledge someone else funded him to acquire. He was thirty, not twenty. This is free and available anywhere — get hired into the thing you intend to build.
- Kill your own architecture the moment the market picks a different standard. The cost of that decision is ego, not capital. Most founders pay capital to protect ego instead.
- Pay a small recurring cost on every unit you ship for an option nobody is pricing yet. For a small operator: instrumenting every transaction, owning the customer list, building the internal tool nobody asked for.
- Hold. The most copyable act here required no genius. A co-founder with the same shares and the same information sold in 2006 and is estimated at around $30m. The only difference was duration.
Does not transfer
- Venture money with no product and no revenue, because his employer's CEO personally phoned the most powerful investor in Silicon Valley. No amount of grit manufactures that call.
- A stranger's mercy. In 1996, roughly thirty days from missing payroll, a customer executive gave him $5m on personal judgement — a customer whose project he had just failed to deliver. Without it there is no company and no profile.
- Surviving thirty-three years through an 84% drawdown and a decade of Wall Street calling his core programme a waste, with no board revolt and no activist campaign. Almost no public-company CEO survives that sequence.
- The accident that made him right. Three researchers independently decided to train a neural network on two of his consumer cards. He did not cause it and could not have commissioned it. His thesis was scientific computing; deep learning arrived and retroactively vindicated him.
The misreading
That this is a story about betting the company on a vision. It is close to the opposite. He never bet the company — he bet the margin. A profitable, unglamorous gaming-graphics business paid the bills for sixteen years, which is the only reason the option survived long enough to be exercised. Beginners read "conviction through a decade of being wrong" and copy the decade of losses while skipping the boring cash-generating core that funded it.