Nvidia had a contract with Sega to build a game console, and Sega’s money was funding the company. A year into the work, Jensen Huang, its chief executive, saw that the design was wrong. Microsoft was about to announce Direct3D, built on a different approach, and many companies were already making chips for it.
Finishing the console meant inferior technology, cut off from Windows and too far behind to catch up. Walking away meant running out of money. “Either way, we would be out of business,” Huang told graduates of National Taiwan University in 2023.
So he contacted Sega’s chief executive, told him Nvidia’s invention was the wrong approach, and said Sega should find another partner. Then he asked to be paid in full anyway. “I was embarrassed to ask.”
Sega’s chief executive agreed. “His understanding and generosity gave us six months to live,” Huang said.
Six months is not long enough to fix a company. It is plenty of time to waste.
Three parts and an argument
In 2011, Richard Rumelt, a management professor at UCLA, used what Nvidia did next to show the kernel of a strategy. It has three parts. A diagnosis: “an explanation of the nature of the challenge.” A guiding policy: “an overall approach chosen to cope with or overcome the obstacles identified in the diagnosis.” And coherent actions: “steps that are coordinated with one another to support the accomplishment of the guiding policy.”
A strategy names what’s wrong, picks a way through, and lines up actions that fit.
In Rumelt’s account, Nvidia’s diagnosis was blunt: “We are losing the performance race.” A rival start-up, 3Dfx, had taken the lead in fast 3D chips for gamers, and Intel was rumored to be planning one of its own. The guiding policy followed from the diagnosis: release a faster chip every six months, three times the industry’s usual pace.
Then came the actions, each one serving that pace. Three development teams worked on overlapping schedules, so no single team had to design a whole chip in six months. Heavy spending on simulation and emulation kept chip fabrication and driver software from causing delays. Over time, Nvidia took control of the driver software back from the companies that built the graphics boards.
A doctor who tells you to get well has set a goal. One who names the illness has made a diagnosis, and the treatment can follow from it. Rumelt’s own summary is that strategy is “coherent action backed by an argument.”
Goals dressed as strategy
Rumelt coined the term “bad strategy” in 2007, at a Washington seminar on national-security strategy. His 2011 book Good Strategy/Bad Strategy grew from it. One chief executive, whom he calls Chad Logan, had a “20/20 plan”: revenue up 20 percent a year, profit margin 20 percent or higher. Asked what would have to happen to get there, Logan talked about the will to win. When Rumelt talked strategy with the mayor of a small city, the planning committee’s plan held 47 strategies and 178 action items. Action item 122 was “create a strategic plan.”
Neither plan said what was in the way, so there was nothing for a policy to answer, and the actions piled up as a list.
The lunch nobody had time for
Imagine you’ve taken over a café beside an office park where lunch sales have fallen for two years. The last owner’s plan, left by the till, reads: grow sales 20 percent, improve service, refresh the menu, post more online.
You spend three lunchtimes at the counter with a stopwatch. Most customers have a 45-minute break, the kitchen cooks each of 22 dishes to order, and food takes 25 minutes to arrive. The people who stopped coming didn’t dislike the food. They ran out of break.
That’s a diagnosis, and a policy can answer it: a lunch you can finish within half an hour of walking in. Then actions that fit: six dishes prepared ahead, pay when you order, trays set at the counter. Any one alone saves a few minutes. Together they make the half hour real.
They also cost something. The regular who came for the slow-braised special finds it gone at noon.
The agent with eleven fixes
Imagine you ask a coding agent to make your app’s search page faster. It comes back with eleven plausible changes: caching, compressed images, a new database index and eight more. Together they are a list of things to do, not a strategy, and an agent may carry out all eleven before anyone asks which one mattered.
Ask for the kernel instead. First the diagnosis, with evidence: where does the time go, according to a profiler trace or timing logs? Then the guiding policy: which part it will change, and which it will leave alone this round. Then only the changes that serve it.
Nvidia tested its chip on an emulator before committing to silicon. You can test the diagnosis before committing the code. Open the trace and check that the slow part is where the agent says. Make the change on a branch and time the page several times before and after. If the agent can’t point to a measurement, its diagnosis is a guess, however fluent it sounds.
An hour per frame
Good strategy, Rumelt wrote, “almost always looks this simple and obvious in retrospect.” The kernel tells you whether you have a strategy. It can’t tell you whether your diagnosis is right.
Huang had one chance to find out. With six months to live, he told the podcast Acquired in 2023, Nvidia could send its next chip to the factory only once. The team ran it on the emulator and sat in the lab “waiting for Windows to paint,” by his estimate about an hour per frame. Asked how he knew the chip would be perfect, he said: “because if it’s not, we’ll be out of business.”
The RIVA 128 shipped in August 1997. Nvidia’s product revenue went from $3.7 million in 1996 to $27.3 million in 1997.