The Bezos Letters

This shelf holds books on how to read a market. These are not a book. They are the 24 letters Jeff Bezos wrote to Amazon’s owners, one a year, covering 1997 to 2020, and they are the only primary source for how Amazon judged its own capital spending while it was spending it. Each was filed with the SEC when it was sent. A memoir written later can tidy the record; these cannot.

We pulled them because of a widely shared post in August 2026 that set CoreWeave against a “teleportation machine” from Bezos’s 2004 letter. The machine is real, but the post got the word wrong (Bezos wrote “transport”) and credited him with a short position he never took. The original says something narrower and more useful. Chapter 4 sets it out in full.

The corpus: 24 letters and 49,294 words, every one taken from an SEC filing — the annual report for 1997–1999, an 8-K exhibit for 2000–2020. All but one of the later filings reprint the 1997 letter; we cut those 22 reprints, of 10,245 to 10,991 characters each, so the founding letter appears once. The 2006 letter says the reprint is attached, but the filing omits it. The first letter went out in about 13,000 printed copies; the second in more than 200,000.

The first letter reports 1997 sales of $147.8 million, up 838%, and 1.5 million customers, then calls the year “Day 1 for the Internet.” Most of it is a list of nine rules for how the company will decide. Bezos asked shareholders to read them so they “may confirm that it is consistent with your investment philosophy.” Every later letter pointed back to that list.

One rule matters most for this shelf: “When forced to choose between optimizing the appearance of our GAAP accounting and maximizing the present value of future cash flows, we’ll take the cash flows.” It is the seed of the 2004 machine. In 1997 it was a promise from “a business incurring net losses,” which chose growth “because we believe that scale is central to achieving the potential of our business model.”

The rules also set the order of measurement: customers, revenue growth, repeat purchases and brand first, profit later. And they tied pay to ownership. Staff would be paid in stock options “rather than cash,” because each employee “must think like, and therefore must actually be, an owner.”

The balance sheet behind it was small. Cash and investments at year-end were $125 million, “thanks to our initial public offering in May 1997 and our $75 million loan.” Staff grew from 158 to 614 and distribution space from 50,000 to 285,000 square feet. Bezos quoted the 158 and the 614 again in his last letter, 23 years later.

The 1998 letter calls Amazon “a business model that is cash-favored and capital efficient.” It reached a billion-dollar sales rate “with just $30 million in inventory and $30 million in net plant and equipment,” and its $31 million of operating cash flow “more than offset net fixed asset additions of $28 million.” The same letter raised cash to over $1.5 billion with high-yield and convertible debt.

1999 was the build-out. Distribution space went from about 300,000 square feet to over 5 million in under twelve months, and revenue grew 90% in the fourth quarter alone. Even so, the letter counts $318 million of fixed assets against a $2 billion sales rate, and says five years of operations had “cumulatively used just $62 million, net, in operating cash.”

The 2000 letter opens with one word: “Ouch.” The shares were down more than 80% in a year. Bezos answered with Benjamin Graham: in the short term the market is “a voting machine,” in the long term “a weighing machine.” He named his own losses on Pets.com and living.com, and said the “land rush” metaphor that had justified them had stopped working.

The lesson of 2000 is about financing. Online selling is “a scale business characterized by high fixed costs and relatively low variable costs,” which “makes it difficult to be a medium-sized e-commerce company.” Pets.com and living.com needed “a long enough financing runway” to reach scale, and “when the capital markets closed the door on financing Internet companies, these companies simply had no choice but to close their doors.”

In 2001 Amazon stopped choosing between growth and cost. Lower costs pay for lower prices, lower prices drive growth, and growth spreads fixed costs over more sales: “Please expect us to repeat this loop.” Inventory turns rose from 12 to 16, and the fourth quarter brought the first pro forma operating profit, $59 million.

The same letter explains the cash rule. Why focus on cash flows? “Because a share of stock is a share of a company’s future cash flows.” Know a company’s future cash flows and its future share count, Bezos wrote, and “you would have an excellent idea of the fair value of a share.” So he set a target for dilution from employee options: an average of 3% a year.

2002 is the first letter to call free cash flow “our most important financial measure.” It was $135 million, $305 million better than the year before. The footnote shows the sum: $174 million of operating cash less $39 million spent on fixed assets. A year earlier, operations had used $120 million and fixed assets another $50 million.

The 2002 letter also names the trade that made it work: “We trade real estate for technology (which gets cheaper and more capable every year).” Much of the customer experience became “largely a fixed expense,” costed “more like a publishing model than a retailing model,” so costs shrink as a share of sales as sales grow. Inventory turned 19 times that year.

2003 is about time. “Owners are different from tenants,” Bezos wrote, citing a family who rented a house and “nailed their Christmas tree to the hardwood floors.” Many investors, he said, are “really just renting the stocks.” His examples are choices that cost sales today: allowing negative reviews, and a warning that a customer already owns an item, which “slightly reduced sales.”

The 2004 letter opens: “Our ultimate financial measure, and the one we most want to drive over the long-term, is free cash flow per share.” The reason: “shares are worth only the present value of their future cash flows, not the present value of their future earnings.” A company “can actually impair shareholder value” by growing earnings, “when the capital investments required for growth exceed the present value of the cash flow derived from those investments.”

The example: a machine that can “quickly transport people from one location to another.” It costs $160 million, carries 100,000 trips a year and lasts four years. Each trip sells for $1,000 and costs $450 in energy and materials and $50 in labor. Full in its first year, it earns $10 million after $40 million of depreciation: “a 10% net margin.”

The owner doubles the fleet each year. Earnings grow 100% a year to $80 million, $150 million in all. “Investors considering only the above income statement would be delighted.” Over the same four years the business produces negative free cash flow of $530 million, because the machines cost $1.28 billion. EBITDA doubles too, and Bezos dismisses it in four words: “EBITDA isn’t cash flow.”

Slower growth helps and does not save it. With no growth at all, one machine returns $40 million of cumulative free cash flow, but “the gross cumulative cash flow doesn’t surpass the initial machine cost until Year 4,” and at a 12% cost of capital the present value “is still negative.” The verdict: “There is no growth rate at which it makes sense to invest initial or subsequent capital to operate the business.”

The same letter then describes Amazon, which is the machine’s opposite. It collected “payments from our customers before payments are due to suppliers,” held $480 million of inventory on nearly $7 billion of sales, and owned $246 million of fixed assets, “4% of 2004 sales.” Free cash flow grew 38% to $477 million, and Amazon repaid more than $600 million of convertible debt. The machine was a lesson, and Amazon was the counter-example.

2005 splits decisions in two. Some “can be made with data,” such as where to put a fulfillment center. Others need judgment, and the largest is price. The short-term math “always says that the smart move is to raise prices”: a cut never sells enough extra units at first to pay for itself. Amazon cut anyway, judging it would produce “a much larger dollar amount of free cash flow” over five or ten years.

2006 sets the bar for a new business. It must earn “the returns on capital our investors expected when they invested in Amazon” and grow large enough to matter. Physical stores failed: “we don’t know how to do it with low capital and high returns.” Amazon Web Services passed, with ten services and 240,000 registered developers. A new business, the letter warns, takes “three to seven years” to matter.

2008, written in the financial crisis: “Our primary financial goal remains maximizing long-term free cash flow and doing so with high rates of return on invested capital.” The letter names where the money was going — Web Services, seller tools, digital media, China — and the condition attached: each had to “clear our high bar for returns.”

2009 shows how the company steered. Sales were $24.51 billion and free cash flow $2.92 billion, up 114%. Yet of 452 goals set for 2010, 360 concerned customers; “free cash flow” appears four times, and “net income,” “gross profit,” “margin” and “operating profit” not once. Bezos’s claim: manage the inputs and the financial results follow.

The 2010 letter says bandwidth, disk space and processing power “continue to get cheap fast.” In 2000 Bezos had done the sum: with bandwidth doubling in price performance every nine months, Amazon could use 60 times the bandwidth per customer in five years at the same cost. For a company that buys technology to sell something else, a falling price is a gift.

2011: AWS services “radically transform capital expense into a variable cost.” Customers stop owning servers and rent them by the hour, which moves the capital, and the risk of owning it, onto Amazon. By 2012 AWS had cut its prices 27 times in seven years; by 2015, 51 times.

2012: Amazon sold its devices “at roughly breakeven prices” because “we want to make money when people use our devices – not when people buy our devices.” The same letter returns to Graham’s weighing machine: “We want to be weighed, and we’re always working to build a heavier company.”

2014 is where Bezos defends the capital in AWS. The team checks returns “because AWS is capital intensive,” he writes, and he expected “strong returns on capital” for a stated reason: companies’ own data centers “have low utilization rates, almost always below 20%,” while “pooling of workloads across customers gives AWS much higher utilization rates.” Scale, he added, “can provide us a relative advantage on capital efficiency.”

The same letter defines “a dreamy business offering”: “Customers love it, it can grow to very large size, it has strong returns on capital, and it’s durable in time.” He named three: Marketplace, Prime and AWS. Return on capital is one of the four tests, and lasting for decades is another.

2015: “Given a ten percent chance of a 100 times payoff, you should take that bet every time.” Business, unlike baseball, has a long tail: “every once in a while, when you step up to the plate, you can score 1,000 runs.” That year Amazon became the fastest company to reach $100 billion in annual sales, and AWS reached $10 billion.

Also 2015: decisions come in two kinds. Type 1 are “one-way doors,” “consequential and irreversible or nearly irreversible,” and need slow care. Type 2 are two-way doors and should be fast. The failure Bezos warned of was using the slow process for everything. A footnote names the opposite failure: firms that make one-way decisions lightly “go extinct before they get large.”

2016: “Day 2 is stasis. Followed by irrelevance. Followed by excruciating, painful decline. Followed by death.” His defenses include resisting proxies, since “the process can become the thing,” and “the eager adoption of external trends.” The trend he named was “machine learning and artificial intelligence,” and AWS was already renting instances built for it, which it called P2.

2017: high standards can be taught, belong to one field at a time, and need a realistic sense of “scope,” how long the work really takes. The list of milestones shows the size of the company by then: more than $150 billion invested since 2011 in fulfillment, transport and “technology infrastructure, including AWS data centers,” and AWS at a $20 billion revenue run rate.

2018: third-party sellers’ share of Amazon’s physical sales rose from 3% in 1999 to 58%. “Third-party sellers are kicking our first party butt. Badly.” Amazon had paid for the tools that let them, Fulfillment by Amazon and Prime, “at significant financial risk and after much internal debate.”

Also 2018: “Failure needs to scale too.” A company of Amazon’s size should expect “occasionally” to have “multibillion-dollar failures,” because “a single big winning bet can more than cover the cost of many losers.” The Fire phone failed, and its engineers went on to build Echo and Alexa. The argument assumes a company with many bets and the cash to carry the losers.

2019, written during COVID, repeats the data-center sum: “Typical single-company data centers operate at roughly 18% server utilization.” AWS runs “at far higher server utilization rates” because many customers share it. The letter makes the point about carbon, but it is the capital argument of 2014 again.

2020, the last letter as CEO: “If you want to be successful in business (in life, actually), you have to create more than you consume.” Bezos estimated $301 billion of value created in 2020, of which $21 billion was net income for shareholders. And he restated the rule of 2001: stock prices “are a prediction of future cash flows discounted back to the present.”

The 2004 test is the one to apply: does the asset return more cash than it costs, at the price of the money that bought it? Bezos’s machine earned a 10% margin after depreciation and still failed at 12%. In the quarter to June 2026, CoreWeave reported an operating loss of $49 million on $2.575 billion of revenue, before $640 million of net interest. On the letter’s terms, growth does not settle the question; the margin over the cost of money does.

Bezos treated falling technology prices as a gift because Amazon bought technology to sell something else: “We trade real estate for technology (which gets cheaper and more capable every year).” A company whose product is the technology stands on the other side of that curve. Each cheaper, faster chip lowers what the older one can earn, and the depreciation schedule is where that shows.

His case that AWS could earn on its capital rested on pooling: company data centers below 20% utilization, AWS far higher because many customers share it. A cluster built for one AI lab under one contract gets its utilization from that contract and its risk from that one customer. The 2014 argument is about a shared cloud; it does not carry over to a dedicated build by itself.

The 2000 letter is about who survives a funding gap. Pets.com needed “a long enough financing runway” to reach scale, and closed when the markets did. The 2018 letter allows “multibillion-dollar failures” for a company whose winners pay for its losers. Read together, they split the AI capex spenders in two: those paying from operating cash, and those who need the debt markets to stay open.

The letters offer no verdict on any stock, and Bezos’s own record argues against judging an asset by its first years: the 2000 letter was written with Amazon’s shares down more than 80%. The test he left is narrower and harder to fake: free cash flow per share, over the life of the asset, at the real cost of money.