The Steward

How Jeff Wilke Rebuilt Amazon by Giving It Away

Feb 24, 2026

A reconstructed history, 2021–2030

Prologue: The Other Jeff

The call came on a Tuesday in February 2021. Jeff Bezos, sitting in his Medina living room with a legal pad full of notes he’d been making for three weeks, told Jeff Wilke he’d changed his mind.

“I’m not giving it to Andy.”

Wilke had been out of the building for two months. He’d announced his retirement in August 2020, effective in early 2021, after 22 years at the company. The press release had been warm and corporate. Wilke was going to spend time with family, pursue personal interests, the usual euphemisms for a senior executive who’d lost whatever internal argument led to his departure.

The truth was simpler and uglier. Wilke had spent two decades building the physical machine that made Amazon work (the fulfillment centers, the logistics network, the operational architecture that turned a bookstore website into the backbone of American commerce) and he’d watched the company begin to eat itself. The mechanisms he’d helped build alongside Bezos were being hollowed out. The 6-pager process had become a performance. The Leadership Principles had become interview trivia. The customer obsession flywheel, which was the actual engine of the company’s value, was being strip-mined for quarterly margin improvements by people who could recite the principles but had never understood why they worked.

He’d said as much to Bezos, privately, more than once. Bezos had listened. Bezos always listened. And then Bezos had done what Bezos always did when confronted with an organizational problem that didn’t have a clean technical solution: he’d moved on to something more interesting. Blue Origin. The Washington Post. The $500 million yacht.

So Wilke left. And then Bezos called him back.

The specifics of that conversation have never been fully reported. Wilke has declined every interview request on the subject. What’s known comes from three people who spoke to one or both of them in the days that followed: Bezos had watched the first eight weeks of the post-announcement transition and concluded that Andy Jassy, his chosen successor, was going to optimize the company into a margin machine. That Jassy would do exactly what Wall Street wanted, which was to apply AWS-level margins to the retail business, and that the result would be a company that looked great on a spreadsheet for five years and then collapsed, because you can’t run a customer-obsession flywheel on margin optimization fuel.

This was not a criticism of Jassy’s intelligence. Jassy was (and is) exceptionally smart. It was a recognition that Jassy was an AWS guy, that AWS had always been a margin business, and that the instincts that made someone great at running a margin business were exactly the wrong instincts for running a company whose core engine was “sacrifice margin today to build something customers can’t live without tomorrow.”

Bezos wanted Wilke because Wilke had built the physical flywheel with his own hands. He understood it at a mechanical level. He knew where the bolts were.

Wilke said no. Twice.

The third time, Bezos offered something he’d never offered anyone: a public commitment, in the shareholder letter, that Amazon would not pursue a retail margin convergence strategy for at least five years. It was an extraordinary concession. Bezos was essentially pre-committing his successor to a strategy that Wall Street would hate, and doing it in the one document that Amazon’s institutional investors actually read.

Wilke said yes.

The announcement came on March 15, 2021. Jassy would run AWS as an independent division. Wilke would be CEO of Amazon. Bezos would remain Executive Chairman but (and this was the part nobody believed at the time) would genuinely step back from operations.

Amazon’s stock dropped 4% on the announcement. The analyst notes were skeptical. “An operations guy, not a visionary.” “Wilke lacks the strategic imagination that drove Amazon’s biggest bets.” “We see this as a caretaker appointment.”

They were right about one thing. Wilke was a caretaker. They were wrong about what that meant.

Part One: The Shareholder Letter

Wilke’s first annual shareholder letter landed in April 2022, and it broke several conventions simultaneously.

It was six pages long. Not because Amazon’s shareholder letters had a mandated length, but because Wilke had spent 22 years reading and writing 6-pagers, and writing anything else would have felt like wearing someone else’s suit. The format was the format. You stated the problem, you stated the hypothesis, you presented the evidence, you described what you were going to do, and you explained how you’d know if it worked.

The letter opened: “Amazon’s job is to be the most customer-obsessed company on Earth. If your investment thesis requires us to become a margin-optimization company, you have the wrong thesis. I won’t tell you to sell your shares. But I will tell you what we’re going to do, and you can decide for yourself whether you want to own this company while we do it.”

The reactions split cleanly along a line that would define the next three years. Institutional investors (BlackRock, Vanguard, State Street, the index funds that collectively owned roughly 20% of the company) read it as a challenge. Retail investors (the ones who’d bought Amazon stock because they used Amazon every day and believed the company was good at what it did) read it as a promise.

BlackRock’s consumer sector team requested a call. Wilke took it himself, which was unusual. CEOs of trillion-dollar companies generally have their CFO or head of IR handle institutional investor calls. Wilke’s view, which he stated to his chief of staff before picking up the phone, was: “If I’m going to tell them no, I should tell them myself.”

The call lasted eleven minutes. The substance of it leaked to Bloomberg within a week (from the BlackRock side, not from Amazon). The key exchange, as reported:

“Jeff, the street wants to see retail margins converge toward AWS margins over a five-year horizon.”

“Then the street is wrong. Retail margins fund growth. If you want 35% operating margins, Microsoft is right there. We’re going to keep investing in same-day delivery, and you’re going to like it in 2027.”

The stock dropped 14% over the following two days. CNBC ran the BlackRock call as breaking news. Jim Cramer, in a segment that has since become a meme, called Wilke “the anti-Bezos” and meant it as an insult.

Wilke sent an internal all-hands email that evening: “If you’re here because of your RSU price, I understand if you need to leave. If you’re here because you want to build, stay. The stock price is going to do what it does. Our job is to build things customers love. The stock will follow. It always does.”

Attrition did spike that quarter (12% annualized among L6+ engineering staff, up from a baseline of 8%). But it was a specific kind of attrition. The people who left were disproportionately mid-tenure (3-5 years), hired during the 2019-2021 stock run-up, and concentrated in teams that had more process than product (internal tools, some AWS services, advertising optimization). The people who stayed were disproportionately either very senior (10+ years, deeply embedded in the institutional knowledge) or very junior (1-2 years, still genuinely excited about the work).

The following quarter, attrition dropped below the pre-letter baseline.

Wilke didn’t design this as a talent strategy. He was just being honest. But honesty, it turned out, was a remarkably effective filter.

Part Two: The 6-Pager Purge

The story that became legend inside Amazon (and has since leaked into the broader tech press through approximately forty different “insider account” pieces, most of them substantially accurate) began in June 2022, when Wilke did something no Amazon CEO had done in at least five years: he started reading 6-pagers from individual product managers.

Not the executive summary versions. Not the S-team review documents. The actual 6-pagers, written by L6 and L7 PMs, the ones that were supposed to be the intellectual engine of the company.

He read fourteen of them in the first week. He later told a small group at an internal offsite that reading those documents was the moment he understood how far the rot had spread.

“Every one of them opened with ‘We believe’ or ‘We are excited to announce.’ Every one of them had three pages of background that said nothing, two pages of metrics that measured nothing useful, and a final page that proposed something so vague it couldn’t be evaluated, let alone tested. They weren’t proposals. They were applications for continued funding disguised as strategic documents.”

Wilke wrote a memo. It was, naturally, a 6-pager. The title was “What a 6-Pager Is For.” The core directives were simple: A 6-pager that does not contain a falsifiable hypothesis in the first paragraph will be returned unread. A 6-pager that does not name the customer segment and the specific pain point by page two is not a 6-pager. It’s a press release for an internal audience.

He distributed it to every product manager and engineering manager at L6 and above.

The immediate aftermath was exactly what you’d expect. Half the product organization panicked. They panicked because their entire workflow was built around producing documents that looked like 6-pagers but functioned as political cover. The document was the deliverable. The approval was the product. Actually building something was a secondary concern, and for some teams, not a concern at all.

Three Directors and a VP quit within sixty days. Their departures were quiet and barely noted in the press. Inside the company, their exits were seismic, because every senior IC knew exactly why they’d left. Their value proposition had been “I can navigate the document review process.” With the process demanding substance again, they had nothing to navigate.

The half of the product org that didn’t panic started shipping. Within six months, Amazon’s internal product velocity (measured by the number of externally launched features per quarter) increased by 40%. Not because Wilke had hired anyone. Not because he’d reorganized anything. He’d just changed what was required to get a document through review, and the people who could only produce process-documents left while the people who could produce product-documents started getting their ideas funded.

A senior principal engineer (who has since left Amazon for a startup and speaks about this period freely) described it: “For three years before Wilke, I’d been writing 6-pagers that got stuck in review for months because some Director wanted to add their name to the review chain. After Wilke’s memo, my next 6-pager went from submission to funded in eleven days. The Director who’d been bottlenecking my reviews had quit the week before. Nobody replaced him. Nothing broke.”

Part Three: The Advertising War

In October 2022, Amazon’s advertising team presented their annual roadmap to the CEO. This was a presentation that, under Jassy in the real timeline, resulted in one of the most profitable and most destructive decisions in Amazon’s history: the systematic degradation of search results on Amazon.com to maximize advertising revenue.

The roadmap was aggressive. Sponsored placements in the first four search results. Banner advertising on product detail pages. Pre-roll ads on Prime Video content. The projected revenue was staggering: a path to $40 billion in annual advertising revenue by 2025, with operating margins north of 50%.

Wilke read the 6-pager. His written comment in the margin (which was photographed and circulated internally within hours, because nothing stays secret in a company of 1.5 million people) read: “This document proposes making our search results worse so that brands will pay us to make them slightly less worse. Explain to me how this is different from a protection racket.”

He didn’t kill advertising. He wasn’t naive about the economics, and Amazon had been running a modest advertising business for years. What he did was impose constraints.

Sponsored results were capped at one per search results page. Sponsored products had to pass the same relevance threshold as organic results (meaning you couldn’t buy your way to the top of a search for a product category your product didn’t belong in). Product detail pages remained ad-free. Prime Video would carry no pre-roll advertising.

The advertising team pushed back hard. Their argument was financial: the constrained model would generate perhaps $18 billion annually instead of $40 billion. The unconstrained model was “leaving $22 billion on the table.”

Wilke’s response, in a meeting that three people have described with near-identical details: “That $22 billion comes from somewhere. It comes from customers seeing search results they don’t trust. It comes from customers going to Reddit to check whether the top Amazon result is actually the best product or just the one that paid the most. It comes from the flywheel slowing down. We’re not leaving $22 billion on the table. We’re choosing not to extract $22 billion from our customers’ trust, because their trust is worth more than $22 billion.”

BlackRock downgraded Amazon to neutral. Wilke’s comment at the next earnings call: “We appreciate the feedback. We disagree.”

Part Four: The Flat Years

From mid-2022 through late 2025, Amazon’s stock price was, by the standards of a company that had doubled roughly every three years for two decades, essentially dead.

It wasn’t falling. It just wasn’t going up. The stock traded in a range between $115 and $145 while Google, Microsoft, and Nvidia rode the generative AI hype cycle to stratospheric valuations. “Amazon’s Lost Decade” pieces ran in the Wall Street Journal and Bloomberg (the decade had technically lasted 18 months, but headlines have their own relationship with time). Comparisons to Ballmer-era Microsoft were everywhere.

Inside the company, the flat stock was doing exactly what flat stock always does: sorting the organization.

Amazon’s RSU vesting schedule was backloaded (5% in year one, 15% in year two, 40% in years three and four). This meant the people whose big vest was coming due in 2023 and 2024 (people who’d been hired in 2020 and 2021, at peak stock prices, during the pandemic hiring binge) were watching their expected payouts stagnate or decline. These were disproportionately the people who’d been hired through what Wilke privately called “the degraded loop” (the interview process that had, through years of incremental dilution, evolved from a test of building ability into a pattern-matching exercise for people who could recite Leadership Principles).

They left. Not all at once, but steadily. Attrition among the 2019-2021 hiring cohort ran at 18% annualized for three years.

Meanwhile, everyone hired during the flat period (2022-2025) received RSU grants priced between $120 and $140. These people had not joined for the stock. There was no stock story to tell in recruiting. Amazon’s recruiters, stripped of the “and your RSUs will probably double” pitch that had anchored offers for a decade, were forced to sell the work itself. The people who said yes to those offers were, by definition, people who wanted to build things at Amazon scale and found that proposition compelling even without a guaranteed financial windfall.

This was not a new phenomenon in the technology industry. It was, in fact, one of the most reliable patterns in the history of Silicon Valley.

After Steve Jobs returned to Apple in 1997, the stock languished for years while the company rebuilt. Everyone who’d joined for the stock left. Everyone who joined during the trough (1998-2002) got grants at single-digit prices. By 2012, those grants were worth 50 to 100 times their issue price.

After the dot-com crash, Amazon’s own stock fell from $107 to $7 between 1999 and 2001. The people who stayed (or joined during the crash) formed the operational core that built AWS, Prime, the Kindle, and the fulfillment network. Their equity was worth life-changing money by 2010.

Satya Nadella inherited a Microsoft whose stock had been flat for over a decade. Everyone who’d joined for a Ballmer-era payout had long since left. The people who stayed (or joined during the “Microsoft is dead” narrative) got grants in the $30-50 range. Microsoft’s stock eventually exceeded $400. Those people were millionaires who attributed their fortune to Nadella’s leadership, and they weren’t entirely wrong, but the flat period had done half the work by filtering the organization before Nadella’s strategy had time to pay off.

Wilke understood this mechanism. It’s not clear he designed it deliberately. What is clear is that he didn’t fight it. He didn’t do layoffs-as-theater (the performative headcount reductions that signal “fiscal discipline” to Wall Street while actually gutting institutional capability). He didn’t freeze hiring to “tighten the belt.” He let the stock price do the sorting.

By late 2025, Amazon’s headcount was roughly 15% lower than its 2021 peak, but the ratio of ICs (individual contributors who build things) to managers (people who manage the process by which things get built) had shifted from approximately 4:1 to 6:1. The company was smaller and denser. The people who remained were, on average, more technically capable, more motivated by the work itself, and less distracted by compensation anxiety.

And their RSU grants were priced at $130.

Part Five: Operation Clean Shelf

The catalog purge was the move that defined Wilke’s first era. It was the one that caused the most immediate pain, generated the most external controversy, and did the most to restore the thing that actually powered Amazon’s business: customer trust that when you searched for something on Amazon, the results were real products that worked.

In early 2023, Wilke commissioned a 6-pager from a team that, in another timeline, would have already been dissolved: the retail quality organization. This was a small group that had been responsible for monitoring product quality across the Amazon catalog (testing products for safety compliance, investigating counterfeit complaints, tracking return rates as a signal of quality problems). Under the previous leadership, their budget had been cut repeatedly because their work generated costs rather than revenue. Several senior members had been reassigned to the advertising team, where their analytical skills could be applied to something that showed up on the P&L as revenue rather than expense.

Wilke reconstituted the team, gave them a direct reporting line to his office (bypassing the retail VP who had defunded them), and asked for a single document: “Tell me the truth about what’s in our catalog.”

The 6-pager, when it arrived, was sixty-three pages long. The team lead had included a cover note: “I know this is supposed to be six pages. The problem is not six pages big.”

The findings: 65% of new ASINs added in the prior 24 months originated from sellers based in mainland China, shipping directly from Shenzhen-area warehouses. Of these, roughly 40% had no verifiable product liability insurance (the insurance documents on file were from entities that, upon investigation, did not exist or were shell companies with no underwriting capacity). Approximately 30% of electronics products from these sellers had not undergone any third-party safety testing. Return rates for this cohort were 3.2 times the catalog average. Fake review activity (identified through statistical analysis of review timing, language patterns, and reviewer account behavior) was concentrated overwhelmingly in this seller cohort.

The team’s framing was careful. They did not frame it as a China problem. They framed it as a trust problem. The concluding section read: “Every counterfeit USB-C cable that catches fire is a customer who stops trusting Amazon search results. Every fake review that tricks a customer into buying a product that doesn’t work is a customer who starts checking Reddit or Wirecutter before purchasing. We are subsidizing the destruction of our own competitive moat.”

Wilke’s directive was straightforward: all new seller onboarding would require product liability insurance from a US- or EU-domiciled insurer with verifiable underwriting capacity. Electronics and consumable products would require third-party safety certification from an accredited testing laboratory. All sellers would need to demonstrate supply chain traceability (where the product was manufactured, by whom, under what quality controls).

Existing sellers were given 180 days to comply. Sellers who could not demonstrate compliance would have their listings suspended.

The catalog shrank by roughly 40% over the following 18 months.

Revenue per transaction went up. Return rates dropped by almost half. The average review score across the catalog rose from 3.8 to 4.3 stars, not because Amazon was manipulating reviews, but because the floor of product quality had been raised high enough that genuinely bad products were no longer in the catalog to be reviewed.

The second-order effect was the one that made the analysts recalculate. Brand-name manufacturers that had been pulling back from Amazon (because their products were drowning in a sea of knockoffs and counterfeit listings) began re-engaging. Nike, which had pulled its official Amazon storefront in 2019, returned. Apple expanded its direct presence. Patagonia, which had never sold on Amazon, opened a storefront. The catalog was smaller and the revenue was bigger, because the products that remained were products people actually wanted to buy and didn’t regret buying.

The trade press ran stories about “Amazon’s retreat from global sellers.” The CCP’s commerce ministry issued a statement expressing concern about “discriminatory trade practices.” Wilke didn’t engage publicly.

Internally, the framing was one line, and it was pure Bezos-era thinking: “We are not in the business of providing shelf space. We are in the business of helping customers find things they want to buy and not regret buying.”

Part Six: The Flywheel Breaks Free

Q4 2025 earnings arrived like a thunderclap.

Revenue growth reaccelerated to 18% year-over-year. Not from financial engineering, not from advertising revenue, not from accounting changes. The underlying retail business was bigger. Same-day delivery had reached 40% penetration among Prime orders. The customers who had stuck around (because search results still worked, because the catalog wasn’t a landfill of counterfeits) were spending 30% more per year than they had three years earlier. Customer retention was at historic highs.

The flywheel had never stopped spinning. Wilke had never stripped it for parts. And now, after three years of compounding investment without extraction, it was spinning fast enough that the financial results became undeniable.

Analysts who had downgraded sheepishly upgraded with notes like “we underestimated the durability of the consumer flywheel.” The stock broke out of its trading range and began climbing.

Wilke’s comment on the earnings call was four words: “We told you so.”

And then the RSU math hit.

Every engineer, product manager, and operations leader who had been hired between 2022 and 2025 (people who had joined because the work was interesting, not because the stock was a sure thing) watched their grants double in value within a year. People who had accepted offers at $130/share were suddenly sitting on grants worth $260 and climbing.

The psychological effect was enormous. These weren’t people who felt entitled to the gains (the way peak-cycle hires always did). These were people who had joined during the flat years, when joining Amazon was mildly unfashionable, when the press was writing obituaries and the comparisons to Ballmer were flowing freely. They had joined anyway, because the work was real. Now they were accidentally rich, and the gratitude (to the company, to each other, to the CEO who had held the line) created a loyalty that no compensation structure could have engineered deliberately.

BlackRock quietly upgraded to overweight. Nobody at Amazon acknowledged this.

Part Seven: Forge

While the press was writing “Amazon’s Lost Decade” pieces during the flat years, Wilke had been funding something that wouldn’t show up on the P&L for two years: Amazon Forge.

The idea originated from an unlikely source. An L6 product manager in the AWS serverless team had written a 6-pager (a real one, with a falsifiable hypothesis on the first page, per Wilke’s edict) proposing that Amazon build a system that could compose AWS services into working applications based on natural language descriptions of business problems.

The hypothesis: “Most businesses that could benefit from custom software do not have custom software, because the cost of hiring developers exceeds the expected return. If we can reduce the cost of building a cloud-native application by 95%, the addressable market expands by 10-50x.”

Wilke read it. His margin note was one word: “Yes.”

Amazon Forge launched in early 2026. The pitch was deceptively simple. Describe the business problem. Forge builds the SaaS application, provisions the infrastructure, connects to your data sources, and runs it on AWS. You pay per use.

A small business owner could say: “I need a system that takes customer orders from my website, checks inventory in my warehouse system, and texts me when something is running low.” Forge would build it. A working system, deployed, monitored, billing live, in minutes.

An enterprise customer could describe a compliance monitoring system that watched transactions, flagged anomalies, generated reports, and integrated with existing SAP infrastructure. Forge would build that too. Not a demo, not a proof of concept. A running system with logging, alerting, and access controls already wired up.

The strategic architecture underneath was what mattered. Every Forge-built application was deeply entangled with AWS services (Lambda for compute, DynamoDB for data, S3 for storage, Bedrock for AI capabilities, API Gateway for connectivity). The more applications Forge built for a customer, the more their AWS infrastructure bill grew, and the harder it became to leave. But customers didn’t want to leave, because the applications worked and the total cost was a fraction of hiring a development team.

Within eighteen months, Forge-built applications began talking to each other. A logistics company’s Forge app automatically negotiated capacity with a warehouse company’s Forge app. An insurance company’s system pulled data from a healthcare provider’s system through permissioned API connections that Forge had configured automatically. An ecosystem of AI-built, AI-operated SaaS applications emerged on AWS, and Amazon took a cut of every transaction.

The AI companies that had been building “copilots” for individual developers suddenly realized they’d been solving the wrong problem. The bottleneck was never “developers need help writing code.” The bottleneck was “most businesses can’t afford developers at all, and the ones that can are waiting eighteen months for a project that Forge builds in an afternoon.”

Microsoft, which had invested tens of billions in Copilot (a tool that helped individual programmers write individual functions), was playing chess. Wilke was playing infrastructure. The difference, as it always was, was that the infrastructure play was invisible until it was everywhere, and then it was too late to catch up.

Part Eight: SkyGrid

Jeff Wilke’s background was operations. He’d spent two decades thinking about how physical objects move through space and time, how to optimize routes, how to position inventory, how to shave seconds off delivery times. So when the question of drone air traffic control arose, Wilke didn’t see a regulatory problem or a technology bet. He saw a logistics infrastructure gap that somebody was going to fill.

The FAA was never going to build drone air traffic control. The agency could barely maintain the existing ATC system for manned aviation, a system built on 1960s-era radar technology and sustained by a workforce of controllers operating under chronic understaffing. Adding autonomous low-altitude traffic management to the FAA’s mandate was like asking a hospital running on paper charts to also build an electronic health records system. Technically possible. Never going to happen.

Amazon already had the four things you needed to build it: logistics data (where packages needed to go), mapping data (from millions of delivery routes), compute infrastructure (AWS), and real-time sensor processing capability (from the warehouse robotics division). Nobody else had all four.

SkyGrid launched as an AWS service in 2026. Amazon’s own drone delivery fleet was the anchor tenant. The platform was open. Zipline (medical delivery in Africa, expanding to US hospital networks) plugged in. Wing (Google’s drone program, which quietly became an Amazon infrastructure customer in what employees at both companies described as “hilarious”) plugged in. Agricultural drone operators, power-line inspection companies, real estate photography services, emergency response agencies (they all connected).

The FAA, which had been stalling on drone integration rules for a decade because the problem was genuinely hard and they lacked the technical capacity to solve it, granted Amazon’s SkyGrid system “approved infrastructure provider” status. In regulatory language, this meant Amazon met the safety standards. In practical terms, it meant Amazon was the de facto regulator of low-altitude autonomous airspace, and the FAA was going to pretend it was overseeing them while quietly being grateful that someone had solved the problem.

The EU followed eighteen months later, adopting a framework that was functionally identical.

Revenue from SkyGrid itself was modest in its first year. The strategic value was something else entirely. Every autonomous aerial vehicle in North America (and eventually Europe) was now an AWS customer by default. The pattern was identical to the original AWS: build infrastructure you need, realize everyone else needs it too, rent it out. The difference was that this time, the infrastructure wasn’t virtual. It was the sky.

Part Nine: The Fulfillment Reformation

Amazon Fulfillment Services (AFS) might have been the most important thing Wilke built, and it was the thing that required something no technology or strategy could substitute for: credibility.

The problem was old and obvious. Amazon had been offering fulfillment services to third-party sellers for years through FBA (Fulfillment by Amazon). It was always poisoned by the same structural conflict. The team fulfilling your orders reported to the same organization that was competing with you.

Every third-party seller on Amazon had the same story. You’d find a profitable niche. Your sales data would flow through Amazon’s systems. Eighteen months later, Amazon Basics would launch a suspiciously similar product at a lower price point, warehoused in the same fulfillment center that handled your inventory, informed by the demand data your sales had generated.

Amazon’s official position was “information barriers.” Policy documents. Training modules. Legal language. Everyone in e-commerce knew it was theater.

Shopify had tried to build the alternative. The Shopify Fulfillment Network launched with exactly the right pitch (”we’ll handle your logistics and never compete with you”) and exactly the wrong capabilities. Shopify was a software company trying to run warehouses. They outsourced to third-party logistics providers who were unreliable. The service never reached Amazon-level speed or consistency. Shopify eventually abandoned the effort.

The market was desperate for someone who could deliver Amazon-quality fulfillment without the Amazon-will-steal-your-business risk. Wilke was the one person in the world positioned to offer it, because he had built the fulfillment network and understood both why it was best-in-class and why the trust problem was real.

His solution was architectural, not political. AFS became a separate P&L within Amazon, reporting directly to the CEO’s office. AFS customer data (what was selling, at what velocity, to which zip codes, at what margins) lived in a separate data environment with infrastructure-level access controls. Not a policy document. IAM permissions enforced at the infrastructure level, continuously audited, with access logs flowing to AFS’s own compliance team.

Wilke’s internal memo stated it plainly: “If I find out that anyone on the retail side accessed AFS customer data, I will fire the person who accessed it and the VP who manages them. This is not a suggestion.”

Within the first year, he fired two VPs.

The first was in the private-label organization. A direct report had requested (and received) a data export of top-selling third-party products in the home goods category from someone on the old FBA analytics team. The VP claimed ignorance. Wilke’s response: “Your job is to know what your team is doing. You didn’t know. That’s also a firing offense.”

The second was more politically expensive. A VP in the retail organization, fourteen years at the company, well-liked, strong performance history, close relationships with two board members. He’d built an informal back-channel with a colleague on the marketplace insights team to get “category trend reports” that included third-party seller performance data. Nothing formal. Just two people who’d worked together for years sharing spreadsheets over email.

Two S-team members argued for leniency. He’d been there since the early days. It was the old culture. Give him a warning.

Wilke’s response (reconstructed from three independent accounts): “The old culture is why sellers don’t trust us. If we don’t fire him, we’re telling every AFS customer that our data isolation guarantee is negotiable. It’s not.”

The VP was fired. The story leaked to Bloomberg (almost certainly deliberately, almost certainly through Wilke’s communications team). The headline was: “Amazon Fires Senior Executive Over Data Access Violation.”

The third-party seller community noticed. Not the headline. The firing. They’d spent fifteen years developing a deep and justified paranoia about Amazon’s data practices. They’d heard “we have information barriers” a thousand times and knew it was meaningless. But watching Amazon fire a politically connected fourteen-year veteran over a spreadsheet (that was different). That was cost. That was credibility purchased with internal political capital, which is the only currency that actually buys trust.

Shopify integrated with AFS within six months. Tobi Lütke, Shopify’s CEO, had been trying to solve the fulfillment problem for years. The AFS deal gave Shopify’s merchants access to Amazon’s fulfillment network (same warehouses, same robots, same delivery speed) without being on Amazon’s marketplace. The key contractual term: AFS customers’ products would never be listed on Amazon.com unless the customer explicitly opted in.

Shopify’s stock jumped 15% on the announcement. WooCommerce, BigCommerce, Magento, Wix (they all built AFS integrations within eighteen months). AFS became the default fulfillment layer for e-commerce the way AWS had become the default compute layer for software.

AFS pricing was published, transparent, and simple: per-unit pick, pack, and ship; storage per cubic foot per day; returns processing per unit. Volume discounts at published thresholds. No hidden fees, no surge pricing, no algorithmic pricing. A published rate card.

By 2028, AFS revenue was $22 billion annually.

The two fired VPs had cost Amazon perhaps $5 million in severance and legal fees. They had generated $22 billion in annual revenue by making the fulfillment-as-a-service business credible.

Part Ten: The Platform Independence Play

Fire OS was always a hack. A fork of Android with Google Play Services torn out and Amazon’s services duct-taped on. Every major Android version bump meant Amazon’s Fire team had to reverse-engineer whatever Google had changed, patch around it, and hope nothing broke. The team spent 70% of its engineering cycles keeping up with Google rather than building anything differentiated.

Wilke read the Fire OS team’s 6-pager. His margin note: “We are spending $400M/year to be a second-class citizen on someone else’s platform. Explain why.”

Nobody could.

The alternative was OpenHarmony, the operating system Huawei had built after US sanctions forced them off Android entirely. Huawei had open-sourced it through the Eclipse Foundation, which meant it was genuinely community-governed rather than controlled by a single company (the way AOSP was controlled by Google in practice, whatever the license said in theory).

OpenHarmony had properties that mattered for Amazon’s use case. A microkernel architecture (better security isolation, better resource efficiency on low-power hardware). Native distributed device interaction (the ecosystem play Amazon had been trying to build with Alexa, Fire, and Echo, and failing at because Android was never designed for it). Zero Google dependency at any level of the stack.

Wilke migrated in phases. Echo devices first (2026), because nobody noticed or cared what OS an Echo ran. The Echo team reported 40% lower memory usage and 60% faster boot times. Fire TV next (2026-2027), harder because of the app ecosystem, but an Android compatibility layer handled most APKs in a sandboxed container. The streaming apps (Netflix, Disney+, the rest) didn’t care what the underlying OS was, as long as their apps ran and DRM worked. Fire tablets last (2027), where the compatibility layer handled 90% of the app catalog and Amazon’s native apps ran noticeably better.

The real move was underneath. Every OpenHarmony device needed cloud services (OTA updates, push notifications, app distribution, telemetry, identity management, content delivery). On Android, all of that ran through Google’s infrastructure. Amazon built Amazon Device Services (ADS) on AWS. It handled everything an OpenHarmony device needed from the cloud side. And then Amazon offered it to every other OpenHarmony device manufacturer.

Samsung, which had funded Tizen for years as an Android escape hatch that never worked, was interested. European OEMs facing EU regulatory pressure about Google’s platform dominance were interested. Indian OEMs wary of both Google and Huawei dependency were interested. ADS gave them a turnkey backend: app store infrastructure, push notifications, identity management, analytics, OTA pipeline, all running on AWS, all available as a service.

By 2028, there were roughly 800 million active OpenHarmony devices worldwide. About 60 million were Amazon’s own hardware. Of the remaining 740 million (third-party OEMs), approximately 300 million (the non-Chinese ones) ran on ADS. Three hundred million devices generating recurring AWS revenue and pre-loaded with Amazon’s commerce and content apps.

Google saw it coming eighteen months too late. Their response was to tighten AOSP further and increase Google Play Services licensing fees, which the EU had already been punishing them for. This accelerated the migration rather than slowing it.

Wilke hadn’t beaten Google at the platform game. He’d made the platform game irrelevant by owning the layer underneath.

Part Eleven: The Library of Alexandria

The announcement that Amazon was spinning Kindle into a Swiss-domiciled international nonprofit called books.org was, measured by column inches of confused business journalism, the most bewildering move of Wilke’s tenure.

Amazon contributed the Kindle e-reader hardware design (schematics, firmware, industrial design files) under open hardware licenses. The Kindle rendering engine and format tooling, open-sourced. The Whispersync protocol specification, published as an open standard. A $500 million endowment. Five years of guaranteed AWS backend services at cost.

Amazon retained its ebook store, its publisher relationships, its DRM infrastructure, and its customer data. It became one of several storefronts operating on the books.org platform.

The foundation’s governance was designed to be boring: two seats for publishers, two for hardware manufacturers, two for library systems, two for authors’ guilds, two for rotating public interest representatives. Amazon got one observer seat with no vote.

The press was baffled. “Has Wilke lost his mind?” was the general thrust.

One layer down, the logic was straightforward. Amazon was spending enormous political and legal capital defending a monopoly position (roughly 80% of the US ebook market) in a business that generated relatively modest margin and attracted constant regulatory scrutiny. The EU was building a platform-dominance case that would eventually force interoperability anyway. Publishers treated every contract negotiation as a hostage situation. Wilke’s view: “What if we just stopped?”

Every participant in the ebook ecosystem fell into the open platform. Kobo, which had been dying slowly for years, pivoted from failing at hardware-plus-software vertical integration to succeeding as a storefront. Sony, which had abandoned e-readers but still ran a digital bookstore in Japan, plugged into books.org and got global reach for free. OverDrive and Libby, the library ebook platforms, got a unified lending protocol that tripled library ebook usage within two years. Independent bookstores (through Bookshop.org and others) could sell ebooks for the first time in a way that was technically feasible and economically viable.

The strategic logic was the same as every other Wilke play: give away the top layer, own the infrastructure underneath. books.org needed cloud infrastructure (document storage, content delivery, authentication, sync services, storefront hosting, analytics, payment processing, search and recommendation engines). All of it ran on AWS. The $500 million endowment flowed substantially back to Amazon as AWS fees. And as the open ecosystem attracted more participants than the closed one ever could, the AWS bill grew with it.

Amazon’s ebook retail market share settled around 52% (down from 80%). Amazon’s infrastructure revenue from the books.org ecosystem was $4.2 billion per year (up from zero). Antitrust exposure in digital publishing: zero. Political liability from book censorship controversies: zero (books.org handled it, and a Swiss nonprofit with distributed governance was structurally resistant to the kind of political pressure a US corporation was vulnerable to). Publisher negotiations for Amazon’s remaining retail operations became dramatically less hostile.

The reference hardware design was the sleeper hit. Open-sourced Kindle schematics meant manufacturers in Shenzhen could build books.org-compatible devices at price points Amazon had never targeted. A $30 e-reader for developing markets became feasible. Education ministries in Sub-Saharan Africa and South Asia could procure purpose-built reading devices for digital textbook distribution. books.org became the de facto platform for digital education content in the developing world.

By 2029, books.org had two billion registered reading devices worldwide. The largest digital book distribution platform in human history. A nonprofit. Amazon didn’t control it.

Every byte of every ebook delivered to every device moved through AWS.

Part Twelve: The Silicon Plays

Two moves, executed in parallel, that most of the press covered as separate stories but that were, architecturally, a single play.

The first was Inferentia 3, the next generation of Amazon’s custom AI inference chip, built on RISC-V instead of proprietary cores.

Wilke’s insight (which was really Forge’s insight) was that Nvidia’s monopoly on AI compute rested not on superior silicon but on CUDA, the software ecosystem that made developers dependent on Nvidia hardware. If the customer never touched the hardware layer (because Forge was building and deploying applications automatically), then the customer didn’t care whether the underlying silicon was Nvidia, Inferentia, or anything else. Forge abstracted the hardware. The customer described a problem. Forge built a solution. Forge decided what to run it on based on cost and performance.

Inferentia 3, built on RISC-V with custom vector and matrix instruction extensions, delivered roughly 80% of Nvidia’s inference performance at roughly 20% of the cost. The customer didn’t see “RISC-V” or “Inferentia.” The customer saw their AWS bill drop by 80% on inference workloads, because Forge automatically routed to whatever was cheapest.

Nvidia’s response options were structurally constrained. They could cut prices, but not to Amazon’s cost basis without destroying their own margin structure. They could push CUDA lock-in messaging, but CUDA was irrelevant when no human developer was writing CUDA code for Forge-deployed workloads. They could pursue exclusive partnerships with other cloud providers, but that just accelerated Google’s TPU efforts and Microsoft’s Maia chip program.

Nvidia didn’t die. They retained dominance in training workloads. But inference was where the volume growth was, and their share of inference compute dropped from 90%+ to roughly 40% within three years.

The second move was Graviton 4, the next generation of Amazon’s general-purpose server processor, also rebuilt on RISC-V.

The logic was simpler. Amazon was the world’s largest ARM server customer through Graviton. They were paying ARM Holdings licensing fees on every chip. ARM, newly public and under pressure from SoftBank to grow revenue, was shifting toward more aggressive licensing structures. Wilke’s question to Annapurna Labs: “Why are we paying rent?”

The same team that had designed Graviton 1, 2, and 3 on ARM ported their microarchitecture to RISC-V. The out-of-order execution engine, the branch predictor, the cache coherency protocol (all of that carried over). The instruction decoder changed. The compiler backend changed. The performance per watt improved by 20%. The cost per instance-hour dropped by 30%.

AWS announced Silicon as a Service: customers who needed custom server SOCs (telecom companies, defense contractors, national cloud providers pursuing compute sovereignty) could work with Annapurna Labs to design custom RISC-V SOCs based on the Graviton 4 core. Amazon handled design, validation, and fab relationships. The customer got chips with custom accelerators and I/O configurations built on a proven general-purpose core, at a fraction of what ARM charged for an architectural license.

European governments that had been pursuing “digital sovereignty” suddenly had an option: RISC-V based, open ISA, with a reference design from the world’s most experienced cloud chip team. The fact that it ran best on AWS was something they’d worry about later. (They didn’t.)

By 2029: Amazon’s annual spend on ARM licensing was zero. Amazon’s annual spend on Nvidia GPUs for inference was declining toward zero. Amazon’s annual revenue from silicon-adjacent services was $8 billion and growing.

Amazon published the RISC-V AI instruction set extensions as an open specification through RISC-V International. Other chip designers (SiFive, Tenstorrent, startups in India and Europe) could implement the same extensions. Applications compiled for Inferentia 3 ran on any conforming chip. Amazon didn’t need to be the only manufacturer. They needed the ecosystem to standardize on the instruction set and the software layer, both of which ran best on AWS.

Part Thirteen: The Last Mile

In 2028, the United States Postal Service signed a contract with Amazon for endpoint delivery services in designated suburban and exurban zones.

Congressional hearings were scheduled within 48 hours. “AMAZON PRIVATIZING THE POST OFFICE” ran as a chyron on Fox News and MSNBC simultaneously, one of those rare moments of bipartisan panic.

The actual deal, when examined, was almost aggressively boring. USPS trucks would bring Express Mail and Priority Mail packages to Amazon’s local distribution points (which were already everywhere). Amazon’s existing last-mile fleet would drop them at doorsteps along routes those vans were already driving. Amazon would be paid per package. The rate was lower than USPS’s cost-per-package in those zones but higher than Amazon’s marginal cost. First-class mail, marketing mail, periodicals: all still USPS. The universal service obligation was untouched. The mailbox (legally a federal receptacle accessible only to USPS carriers) was untouched.

The deal was possible only because of what had come before. Three years of AFS had demonstrated that Amazon could operate as a neutral infrastructure provider without exploiting the relationship. The fired VPs. The data isolation architecture. The published rate cards. USPS leadership looked at that track record and concluded that Amazon was, in fact, trustworthy enough to carry federal mail packages.

The congressional hearing lasted most of a day. The first several hours were theater. The hearing turned when the Postmaster General testified: “We are spending $2.40 to deliver a Priority Mail package in suburban Williamson County, Texas. Amazon will deliver it for $1.10. The alternative to this contract is not ‘USPS delivers it better.’ The alternative is ‘USPS continues losing money on every package in these zones and Congress yells at us for running a deficit.’”

The postal workers’ union was furious. Wilke, through intermediaries, made a concession: Amazon DSP drivers on USPS-contract routes would be paid at or above USPS carrier rates. This cost Amazon almost nothing (the routes were marginal additions to existing routes) and defused the wage-depression argument.

The FTC investigated. They found that Amazon’s delivery rates to USPS were lower than the rates Amazon charged its own retail division for the same routes (the USPS contract had been negotiated competitively). This was the opposite of the self-preferencing that antitrust enforcers usually found. The investigation quietly closed.

Royal Mail, Australia Post, Canada Post, and Deutsche Post called within the year.

Part Fourteen: The Pattern

The pattern was the same every time. You could write it on an index card:

Build something you need. Make it best-in-class because you’re your own most demanding customer. Realize everyone else needs it too. Externalize it as a service. Give away the visible layer. Own the infrastructure underneath.

Needed servers: built AWS, sold it to everyone. Needed delivery logistics: built the fulfillment network, opened it as AFS. Needed an ebook platform: built books.org, ran it on AWS. Needed AI capabilities: built Forge, ran it on Inferentia. Needed server CPUs: built Graviton on RISC-V, sold SOC designs to the world. Needed drone airspace management: built SkyGrid, made every drone operator an AWS customer. Needed a device OS: adopted OpenHarmony, built ADS, made every OEM an AWS customer.

Wilke wasn’t a visionary. Visionaries build new things from imagination. Wilke was a pattern recognizer. He had watched Bezos execute this play five times over twenty years, and he understood it at a structural level that Bezos himself may not have fully articulated. He just kept running it.

The one thing Wilke added to the Bezos playbook was the willingness to give away the top layer. Bezos had always wanted to own both the infrastructure and the application. He wanted the AWS margin and the retail margin and the marketplace fees and the advertising revenue and the data advantage and the private-label arbitrage. Every dollar on every layer of the stack.

For twenty years it worked, because Bezos was smart enough and the company was growing fast enough that the contradictions didn’t matter. But the contradictions were always there. You cannot be a trusted neutral platform and also compete with your platform customers. You cannot promise sellers a fair marketplace and also launch Amazon Basics using their sales data. You cannot claim customer obsession while filling search results with paid advertisements. You cannot be the infrastructure provider everyone depends on while trying to eat their lunch on the application layer.

Wilke succeeded by giving up what Bezos could never give up: the visible layer. The retail credit. The direct customer relationship. The brand recognition. The ego gratification of being the face.

books.org meant Amazon didn’t get credit for ebooks. AFS meant merchants owned their customer relationships. Forge meant the customer never saw AWS. SkyGrid meant drone operators thought of themselves as independent. The open RISC-V extensions meant the chip ecosystem didn’t have Amazon’s name on it.

Wilke let everyone else be the face. He owned the pipe.

Part Fifteen: The Kindergarten Lesson

There is a thing that the business press, and the analyst community, and the MBA programs, and the boards of directors of Fortune 500 companies cannot say and cannot process, because it sounds like something a child would say, and the entire professional apparatus of corporate governance is built on the assumption that the world is more complicated than children think it is.

The thing is: the honest version of the business is bigger than the dishonest version.

Not because honesty is virtuous, though it may be. Because dishonesty is a tax on your own addressable market. Every seller who wouldn’t use FBA because they feared Amazon Basics was revenue Amazon left on the table. Every enterprise that built on Azure instead of AWS because they feared Amazon would enter their market was revenue Amazon left on the table. Every publisher who invested in alternatives to Kindle because they saw Amazon as an existential threat was ecosystem growth Amazon left on the table.

Wilke stopped leaving it on the table. Not through cleverness or strategic sophistication. Through the radical unsophistication of stating the deal plainly: “We run the infrastructure. You pay us for infrastructure. We will not compete with you on the layers above. If our people violate that, we fire them.”

The honest version of Amazon’s business turned out to be worth roughly eight trillion dollars.

Bezos built the first trillion. He couldn’t build the next seven, because building the next seven required giving away the thing that the first trillion was built on (control of every layer of the stack), and Jeff Bezos does not give things away.

Jeff Wilke does. And that turned out to be worth seven trillion dollars.

Epilogue: The 6-Pager

In early 2030, at an internal offsite that has since been described by four attendees, Wilke was asked by a young L5 product manager what he considered his most important decision as CEO.

The room expected him to say Forge, or AFS, or Operation Clean Shelf, or the BlackRock call.

Wilke said: “I read the 6-pagers.”

The room was quiet.

“Everything else followed from that. If you read what people actually write (not the summaries, not the executive reviews, the actual documents) you know what the company knows. You know what it doesn’t know. You know who’s building and who’s performing. You know where the rot is and where the life is. The CEO’s only irreplaceable job is to know the difference.”

He paused.

“Jeff built the mechanism. The 6-pager, the Leadership Principles, the flywheel, the customer obsession framework. He built all of it, and it was brilliant. The problem was that after he left, nobody was reading the actual documents. They were reading the reviews of the documents. They were reading the summaries of the reviews. They were three layers of abstraction away from the actual thinking, and at three layers of abstraction, everything looks fine. Everything always looks fine in a summary.”

He picked up his coffee.

“I just removed the abstraction layers. The rest was easy.”

It wasn’t easy. Nothing about the previous nine years had been easy. But Wilke had an operator’s instinct for understatement, and an operator’s faith that if you get the process right, the outcomes follow.

The stock, on the day of that offsite, was $847 per share. It had been $130 when he started.

He was, by that point, reading a 6-pager from an L6 product manager about applying Forge to municipal water infrastructure management. He had his pen out. He was making notes in the margin.

Author’s note: This is a work of alternate history fiction. Jeff Wilke retired from Amazon in early 2021. Andy Jassy became CEO. The events described here did not happen. The structural analysis of Amazon’s institutional dynamics, the mechanisms of stock-price-driven talent sorting, and the pattern of infrastructure externalization are drawn from real organizational behavior and real precedents at Amazon, Microsoft, Apple, and other technology companies. The question the narrative asks (whether a non-founder CEO could resist the structural pressure to optimize a company into fragility) remains open. The base rate says no. This story explores what “yes” might look like, and what it would cost.