In the first quarter of 2000, Amazon’s shares traded as high as $91.50. By the third quarter of 2001 they had traded as low as $5.97. Early that year the company had announced a restructuring that cut about 1,300 positions. It had lost money since it began doing business, and its annual report warned that the losses might go on for the foreseeable future.
In July 2001, in the middle of all that, Amazon cut its prices. Books over $20 would sell at 30 percent off list.
A money-losing company cutting its prices can look like a fire sale, the kind of move made once, to get through a bad quarter.
Jeff Bezos described it differently a few months later, in his letter to shareholders. “Focus on cost improvement makes it possible for us to afford to lower prices, which drives growth. Growth spreads fixed costs across more sales, reducing cost per unit, which makes possible more price reductions.” Then he added a sentence no fire sale would carry: “Please expect us to repeat this loop.”
A name for the loop
That autumn, between the price cut and the letter, Amazon invited the business writer Jim Collins to talk with Bezos and a few of his executives. By Collins’s account it was “right in the middle of the dot-com bust, when some wondered how (or if) Amazon could recover.” He taught them an idea from Good to Great, the book he had just published: the flywheel.
Picture, as Collins asks, a metal disk about 30 feet across, 2 feet thick, weighing about 5,000 pounds. The first push barely moves it. After hours you complete one turn. You keep pushing in the same direction, and each turn starts from the speed of the last, until the wheel’s own weight is working for you. Ask afterward which push made it fly and there’s no answer. It was all of them.
No single push turns the wheel. Each turn makes the next one a little easier.
Collins found the opposite pattern in his comparison companies. They launched big programs with fanfare, then stopped, changed course, and threw the wheel in a new direction. He called it the doom loop.
The order of events matters here. Amazon cut prices first; Collins supplied the word later. The team, he wrote, “grabbed onto the flywheel concept.” What mattered, in his telling, was the way Bezos and his team turned good intent into a repeating loop. In the letter it reads as a chain of links: lower costs, lower prices, growth, fixed costs spread over more sales, lower costs again. Each arrow says that one thing causes the next. Each arrow can be checked.
The arrow that doesn’t hold
That is where flywheels fail. A loop drawn on a whiteboard is a hypothesis, and one weak arrow can stop the wheel however hard you push.
Collins’s own evidence deserves the same care. Good to Great picked companies that had already succeeded and looked back for what they shared. The economist Steven Levitt pointed out in 2008 that a portfolio of the book’s eleven companies looked like it would have underperformed the S&P 500 since publication. The flywheel is a good description of momentum where it exists. It is not proof that pushing creates it.
Imagine you lead a small support team that launches a new initiative every quarter: a survey, then a chatbot pilot, then a new ticket form. Each one starts from zero. So you draw a loop instead. Each answered question becomes a help-center article. Articles cut repeat tickets. Fewer repeats free an hour a day. That hour goes into more articles.
Then you check the arrows, one a month. If repeat tickets on topics with an article don’t fall, the second arrow is broken, and you know before you’ve spent a year pushing.
When the wheel is made of data
AI products come with a flywheel story of their own. As the investors Martin Casado and Peter Lauten summarized it in 2019: “More data means better recommendations, which means more customers, and even more data… the famous ‘flywheel’.”
Their essay, written about enterprise startups, argued that the arrow from more data to a better product often weakens. They cited a study by Arun Chaganty of Eloquent Labs on questions sent to a customer-support chatbot: the share of customer intents the data covered approached an asymptote of 40 percent. Beyond some point, they wrote, the cost of adding unique data may go up while the value of each new piece goes down. They also noted that the point varies by domain.
If your plan depends on a data loop, test that arrow the way you’d test any other. Hold back a fixed test set, retrain or re-evaluate as each new batch of data arrives, and plot quality against volume. A curve that keeps climbing means the arrow holds for now. A curve that flattens means more data has stopped turning the wheel, and the next push has to come from somewhere else.
Bezos kept the word. In his letter for 2014, writing about letting outside sellers compete on Amazon’s own product pages, he wrote: “The success of this hybrid model accelerated the Amazon flywheel.”