Jensen Huang

The pattern they actually represent

In 2006 NVIDIA unveiled CUDA and opened the parallel processing inside a graphics chip to science and general computation. Huang then put it on the consumer GeForce line — the chips gamers were buying — knowing what that would cost. His own account is that he recognised CUDA "was going to add so much cost, but it was something we believed in". The return took about a decade. Somewhere in the middle of it, by his telling, the company's market value fell to around a billion and a half dollars. Mark Stevens, the Sequoia partner who sat on NVIDIA's board, describes the demand side without any softening: the market for AI chips in 2012, 2014, 2015 was a zero billion dollar market.

Read the shape of that rather than the outcome. For roughly ten years, one architecture was paid for out of the revenue of a completely different business, shipped inside a product customers were buying for another reason, and paid for by a listed company whose results were public every quarter. There was no private phase. The bet was visible while it was still being made, and for most of that decade anyone could argue it was a mistake.

The half most retellings drop is the test. Asked at Stanford about a share price falling by eighty per cent, Huang does not answer with resolve. He answers with a checklist of external facts: did physics change, did any of the assumptions behind the original decision change. If none of them changed, nothing changes.

Those two halves are one pattern, and they only work bolted together. Fund a position long enough that you can be visibly wrong about it, and hold a written condition that could end it. On its own, the first half is stubbornness and the second half is an opinion.

The blockage it speaks to

Read this one if you have a graveyard of six-week projects.

Each was dropped at roughly the moment it stopped being new, and before it produced evidence either way. From the inside you cannot tell a bet that needs more time from a bet that is dead, so you settle the question by switching — because switching is the only move that feels like information.

Two things are doing that, and neither is a discipline failure. The first is a timing mismatch: a platform bet pays out later than your reward loop will tolerate going unpaid, and a new framework supplies the signal the old project stopped supplying. The second is quieter. Every restart moves the evaluation date forward, so you never have to be the person who backed the wrong thing. You cannot lose a position you have not held long enough to lose.

Both of those protect something real. Both stop working the moment the outcome you actually want takes longer than one switch cycle.

Huang is the right study for that exact stuckness, and not for the reason usually given. He is not evidence that holding works. He is evidence that holding needs a criterion. His was written out of facts about the world and checked on a schedule. Yours does not exist, so the abandon-or-continue decision gets made by whichever week you happen to look at it in.

With a criterion, holding is a position. Without one, holding and switching are the same coin flip, performed at different frequencies.

Three moves you can steal

Write down the fact that would make you stop. Take the bet you are currently doubting and name one observable thing that would end it — a customer behaviour, a cost curve, a technical limit, and not a feeling. Put a date on when you will check it, and leave the question shut until then. It also gives you permission to quit, which is what keeps it honest.

Name the payer and the bet on two separate lines. This pays the bills. This is the bet. NVIDIA sold graphics chips to gamers for years while CUDA cost money and returned none. If you cannot fill in the first line, the second is not a bet, it is exposure — and most of your restarts were cashflow decisions wearing a strategy costume.

Ship the bet inside the product you already sell. Not as a side project. As a feature, this week, smaller than you want it to be. CUDA did not sit in a lab; it went onto the chips gamers were already buying. An unshipped bet generates no usage, so it generates no evidence, so on the week you get discouraged there is nothing to argue back with.

Where the pattern breaks

The pattern has a documented limit, and it is not impatience. In September 2020 NVIDIA signed an agreement to buy Arm and expected to close it in about eighteen months. On 7 February 2022 the two companies terminated it, citing "significant regulatory challenges preventing the consummation of the transaction"; SoftBank kept the $1.25 billion NVIDIA had prepaid at signing. Holding is a tool for outlasting a clock. It does nothing against a gatekeeper who is not on your clock. Before you commit to a long bet, ask what stands between you and the payoff. If it is time, waiting is a strategy. If it is somebody else's decision, waiting is a bill.

What to do this week

Pick the bet you keep almost abandoning. Write one sentence: the observable fact that would make you stop. It has to be about the world — a customer stops renewing, a cost does not come down, a technical limit holds — and not about how much energy you had that week. Put a date beside it, far enough out to mean something, then close the question until that date arrives.

Then take the smallest shippable piece of that same bet and put it inside something you already sell. A feature, a section, an option at checkout. Not a separate launch and not a new repository.

That is an afternoon of work. What it buys you is a wait that collects evidence, instead of a wait that only collects doubt.