Essay / Nate Busa

AI’s Internet Moment: From Infrastructure to Useful Applications

Building the technology and discovering what people will do with it are different kinds of invention. Their adoption and productivity curves run on different clocks.

AIProductivityTechnology adoption

In 1995, Bill Gates sat opposite David Letterman and tried to explain why the internet mattered. Letterman described hearing that a baseball game could be listened to online. “Does radio ring a bell?” he asked. When Gates explained the convenience of listening later, Letterman reached for another familiar comparison: “Do tape recorders ring a bell?”1

Bill Gates and David Letterman discuss the internet in 1995. Watch on YouTube.

The exchange remains funny because both comparisons made sense. A radio already delivered live audio. A tape recorder already let someone listen later. What those comparisons could not capture was the possibility of a common network through which almost anyone could publish, distribute, discover and eventually transact.

AI invites the same kind of judgment. A generated email looks like an email. A summary looks like something an employee could write. A chatbot looks like another search box. Evaluated only as substitutes for those familiar activities, the larger possibilities are easy to miss.

The internet’s infrastructure and its successful applications were different kinds of invention. AI has the same distinction to navigate. Faster models, cheaper inference and better tools expand what developers can build. The economic question is which applications turn those capabilities into something people repeatedly use, trust and find worth paying for.

Financial expectations can move faster still, pricing in future benefits before useful applications reach scale. The dot-com boom and crash show why that speculative curve deserves its own place in the comparison.

Servers made the internet possible. Applications made it useful

Consider the distance between the engineering that enabled the internet and the applications that brought it into everyday life.

Servers supplied compute and storage. WAN routing moved traffic between networks. Web protocols and APIs let software exchange information and invoke services. Payment gateways connected online merchants to payment processing. These components answered questions about connectivity, programmability, scale and transactions. They did not, by themselves, specify a compelling shopping experience or a way to find the most useful page on the Web.

The historical record makes the distinction tangible. BGP’s first RFC appeared in June 1989. Authorize.net dates its payment gateway service to 1996.1415 Meanwhile, applications emerged through distinct technical and commercial experiments:

Application Illustrative milestone Additional invention between infrastructure and usefulness
Online shopping Amazon began selling books online in July 1995 A searchable catalog, ordering experience and fulfillment operation that made remote purchasing attractive.16
Search Brin and Page’s 1998 paper described Google’s large-scale search architecture Crawling, indexing and ranking turned a reachable collection of pages into a system for finding relevant information.17
Social applications Facebook launched in 2004 Identity, connections, sharing and interaction gave people reasons to create content and return.18
Online maps Google Maps launched in February 2005 Geographic data, an interactive interface and directions connected the network to decisions about physical places.19
Video on demand Netflix introduced instant viewing on PCs in January 2007 Playback, content rights, a usable catalog and a subscription offering made delivered video a service.20

These are selected milestones, not claims that the named companies invented each category. Nor did infrastructure finish before applications began. Amazon’s launch predates the payment gateway example above; application demand and enabling technology evolved together. Successful applications also became platforms themselves: a mapping service could expose APIs for other developers.

The distinction is about the problem being solved. An infrastructure team can demonstrate progress through throughput, availability, latency or successful transactions. An application team has to establish that a particular combination solves a meaningful problem, fits people’s behavior and supports a sustainable service. A dependable payment API establishes that a payment can be processed. A successful online store has to make the whole purchase worthwhile.

Netflix’s original streaming announcement is an unusually clear example. It described bandwidth requirements, a phased rollout, a limited initial catalog, licensing constraints and integration with an existing subscription. The product depended on those choices together. Faster delivery of video bytes was only part of the invention.20

Two parallel stacks compare the internet and AI. Internet foundations include servers, WAN routing, APIs and payments; applications include search, maps, streaming, shopping and social services. AI foundations include models, inference, context, tools and evaluation; proposed applications deliver verified business outcomes. Product discovery connects each foundation to useful services, with feedback in both directions.
The missing step is application invention.Conceptual comparison. Layers coevolve, and products can become infrastructure for other products. The AI application examples are proposed directions, not claims of established market success.

AI needs an application discovery curve too

The corresponding AI infrastructure includes accelerators, model training, inference serving, context retrieval, tool APIs, orchestration and evaluation. Identity, permissions and observability make that infrastructure usable inside an organization. Better performance at this layer expands the feasible product space.

Application invention asks different questions. Which customer problem can now be solved? What should the interaction feel like? What evidence establishes completion? Who handles an exception? Does the outcome cost less, improve quality or make useful work possible that previously was unaffordable?

For example, a model that explains an invoice supplies a capability. A service that investigates a billing dispute, checks the applicable policy, performs an authorized correction, updates the records and confirms the result combines capabilities into a business application. Its value depends on that complete experience. Similar opportunities include procurement, onboarding and service recovery, provided their economics and reliability hold up in practice.

That is why measuring the application layer only through chatbot use misses an important possibility. Some useful AI applications may become embedded in services people already want. Their users may judge them through a resolved problem, a completed transaction or a shorter wait, without caring which model performed the work.

My expectation is that many durable gains will come from this layer of product and process invention. That is an editorial inference from the historical mechanism and the evidence below. It is not a claim that we already know AI’s equivalents of search, maps or online shopping, or that their emergence is guaranteed.

A technology has to be discovered more than once

The first discovery is technical: a new capability becomes possible. The next is practical: people discover what it is useful for. A further discovery is organizational: someone works out how to arrange people, software, incentives and responsibility around it.

A laboratory result, an impressive demonstration and a productive operating model are therefore different achievements. Their development overlaps, and useful applications can send new requirements back to researchers. They do not advance in a fixed sequence or at a common speed.

I would separate six curves when assessing AI:

Curve What changes Evidence that belongs on it
Technical capability Which problems a system can solve, at what cost and reliability Evaluations, failure rates and cost at a specified quality
Application discovery Which combinations become useful, dependable services Successful experiments, repeat use, retention and viable delivery economics
Hype and expectations How much attention and anticipated future value surround the technology Publicity, financing and valuations, with each proxy’s limitations made explicit
User adoption How many people use the technology, how often and for how much work Comparable surveys, active use and assisted work hours
Organizational integration How much real work depends on maintained, connected applications Production coverage, process changes and accountable owners
Realized productivity Useful output relative to the labor or total resources required Quality-adjusted output, total cost per outcome and a credible comparison
Six separate schematic panels distinguish technical capability, application discovery, hype and expectations, user adoption, organizational integration and realized productivity. Expectations rise and crash independently of the other curves. The axes have no dates and the shapes are explanatory, not fitted data.
Expectations can reverse while the other curves keep developing.Conceptual illustration, not measured trajectories or a forecast. The productivity panel shows one possible path with complementary investment; gains and recovery are not assured. The panels have different vertical meanings.

An S-curve can describe diffusion: early experiments give way to wider use, followed by slower growth as the remaining population becomes harder to reach. Technical progress can produce successive waves rather than one smooth curve. Productivity may follow a J-shaped path when complementary investment precedes its payoff.13

Calling all of this “AI adoption” makes the discussion less precise. A company can have widespread employee use and very little organizational integration. A mature application can produce measurable gains while most of the company remains untouched.

Hype adds another source of confusion. Gartner’s Hype Cycle describes inflated expectations followed by disillusionment and a possible progression toward useful deployment.21 It is a useful interpretive framework, rather than a calibrated law that dates a crash or guarantees a later payoff.

The internet did not begin in 1995

The interview took place during the internet’s commercial emergence, after substantial technical work had already happened. The Web was proposed at CERN in 1989; CERN released its software into the public domain in April 1993. The Web itself was built on an existing internet.2

That history matters when comparing speed. Starting the internet clock with early networking research and the AI clock with a polished consumer chatbot produces a dramatic comparison with little analytical value. Research origins, public release, mass-market access and use at work are different starting points.

Pew’s historical series records U.S. adult internet use rising from 14% in June 1995 to 46% in March 2000, then 87% by January 2014.3 The economic trajectory was less straightforward. In a 2021 retrospective, the Bureau of Labor Statistics reports private nonfarm business labor productivity growth averaging 1.7% annually in 1981–1997, 3.3% in 1997–2005 and 1.4% in 2005–2018.4

Separate historical panels show U.S. adult internet use rising from 14 percent in 1995 to 87 percent in 2014, while private nonfarm business productivity growth averaged 1.7, 3.3 and 1.4 percent annually across three periods from 1981 to 2018.
Diffusion can keep rising while productivity growth slows.Top: selected dated observations from Pew’s historical survey series, joined only to guide the eye. Bottom: period averages from Figure 2 of the BLS April 2021 retrospective, using its private nonfarm business definition and historical data vintage. Different populations, measures and time spans; these panels do not estimate the internet’s causal contribution.

The productivity acceleration cannot be assigned entirely to the internet. Computing equipment, software, capital investment and changes in business practice were intertwined. Federal Reserve research finds a central contribution from information technology to the late-1990s revival, with a significant but smaller role after 2000.5

There is nevertheless a clear lesson for the analogy. A technology can become foundational without producing permanently accelerating measured productivity. The internet’s eventual importance does not establish that every early investment was sound, or that AI must reproduce the same economic history.

Hype can crash while applications keep developing

The internet’s speculative cycle peaked around 2000 and collapsed while its application economy was still taking shape. The Nasdaq Composite closed at 5,048.62 on 10 March 2000 and fell to 1,114.11 on 9 October 2002, a decline of 77.9% between those closes.22 The index is a broad equity-market proxy for financial expectations. It does not directly measure public enthusiasm, internet-only valuations or technological usefulness.

The market reversal and the application revolution overlapped:

Company What was developing through and beyond the crash What this adds to the analogy
Google Its 2004 IPO filing records profitability beginning in 2001 and the shift to cost-per-click AdWords in early 2002 A business model around useful search was being refined during the market collapse.23
Amazon Third-quarter 2002 net sales reached $851 million, up 33% from $639 million a year earlier Growth in actual commerce could coexist with collapsing technology-sector valuations.24
PayPal eBay completed its acquisition on 3 October 2002 Payment convenience and trust were becoming embedded in an existing transaction network.25
Netflix Its website launched in 1998; streaming followed in 2007 The familiar application emerged through years of service development, spanning the boom, crash and later broadband era.2620
Aligned historical panels show the Nasdaq Composite peaking in March 2000 and falling sharply into October 2002, while U.S. adult internet use continued rising. Dated milestones show Google, Amazon, PayPal and Netflix developing real applications through and beyond the crash.
The speculative reversal did not stop application development.Nasdaq data: monthly last available closes, plus the cited daily peak and trough, from Nasdaq via FRED. Adoption: selected Pew survey observations. The shared calendar helps compare timing; the panels have different units. Company milestones are illustrative, not a measure of aggregate productivity or a representative sample of startups.

“Quietly brewing” captures the distance between the market narrative and the work of building useful services. These companies were not all obscure: Amazon was already public in 1997.16 The important point is that customer adoption, product refinement and viable business models continued developing when the surrounding financial story had turned negative.

The distinction also sharpens the AI analogy. A technology can be transformative while particular expectations about its timing, profits or investment returns are badly wrong. A valuation crash would not, by itself, establish that AI applications had stopped improving. Continued technical progress would not, by itself, vindicate every valuation either.

These are survivor cases. They establish that speculation can reverse while useful applications develop; they do not tell us which current AI firms will survive, how quickly productivity will improve, or whether the present market follows the same path. The practical task remains to track evidence at each layer instead of letting the loudest curve stand in for all the others.

Faster AI adoption is already an empirical claim

We have better evidence than viral user-count comparisons. Alexander Bick, Adam Blandin and David Deming compare nationally representative U.S. surveys with historical technology adoption. Their published 2026 paper finds that overall generative AI adoption outpaced both PCs and the internet when aligned to each technology’s first mass-market product launch.6

Its late-2024 observations also reveal the distance between reach and intensity:

Measure in the published study Reported result
U.S. population aged 18–64 using generative AI 45%
Employed respondents using it for work in the previous week 27%
Employed respondents using it every workday 10%
Share of all work hours assisted by generative AI Estimated range of 1–7%

These are dated observations with different denominators, not a funnel or an October 2026 snapshot. They support rapid diffusion; they do not establish an equivalent rate of enterprise transformation. The study’s matched comparison also carries more weight than directly comparing its working-age sample with Pew’s all-adult series.

There is a plausible mechanism for faster diffusion. AI reaches people through browsers, phones, cloud services and software they already use. The internet helped build those distribution channels. An employee can try an AI service without waiting for their company to construct a new physical network.

Organizational dependence is a larger commitment. A company still has to connect records, establish permissions, evaluate failure, change workflows and decide who owns the result. Existing digital infrastructure shortens the route to experimentation. How much it shortens the route to reliable production depends on the work.

The productivity evidence is real, and heterogeneous

The claim that AI never improves productivity is already inconsistent with empirical research. The evidence also gives little support to a single universal uplift. Studies examine different workers, tasks, tools and outcomes.

Study and setting Finding What it establishes, and its boundary
Noy and Zhang, Science, 2023; professional writing experiment Task time fell 40%; assessed output quality rose 18% Gains on specified writing tasks, rather than an estimate for a complete job or firm.7
Brynjolfsson, Li and Raymond, QJE, 2025; 5,172 customer support agents Issues resolved per hour increased 15% on average Evidence from an actual workplace, with substantial differences across workers; AI assisted human agents.8
Dell’Acqua and colleagues, Organization Science, 2026; 791 P&G professionals in a product innovation workshop Individuals with AI matched the assessed performance of teams without AI Some benefits of collaboration could be supplied differently in the studied challenges; this does not imply that teams are generally redundant.9
METR, July 2025; 16 experienced developers completing 246 tasks in familiar repositories Allowing early-2025 AI tools increased completion time by 19% A randomized result for those developers, tasks and tools, with limited grounds for generalizing to later systems or all development.10
Humlum and Vestergaard, March 2026 working-paper revision; Danish surveys linked to administrative records No detectable effects on earnings or recorded hours; effects larger than 2% ruled out over the study horizon Rapid adoption and task restructuring can coexist with small average labor-market effects. Earnings and recorded hours are not direct measures of output per hour.12

These results should not be averaged into an “AI productivity score.” A reduction in writing time, an increase in resolved cases and unchanged recorded hours measure different things.

The P&G result is particularly relevant to the internet analogy. AI may change access to expertise and the composition of work, beyond helping someone produce a familiar artifact faster. Whether that becomes a better organizational design requires further evidence.

The developer result adds another discipline: evidence needs a model and observation date. In February 2026, METR reported that its follow-up experiment had become difficult to interpret. Developers increasingly declined participation or withheld tasks because they wanted to use AI; parallel agent use also complicated time measurement. The researchers believed newer tools were more helpful, but judged the follow-up data an unreliable estimate of the current effect.11

That update does not erase the original randomized finding. It shows why a moving technical frontier requires repeated measurement, and why strong impressions of improvement need careful verification too.

Why productivity can lag behind use

The productivity J-curve offers an explanation for part of the gap. Brynjolfsson, Rock and Syverson show how general purpose technologies require complementary intangible investment that is imperfectly captured in conventional accounts. Early measured productivity can understate progress while those assets are being built; later measurement can overstate growth when their benefits arrive.13

There are also ordinary operating costs. Employees spend time learning. Engineers build connectors. Specialists define acceptable outcomes and resolve ambiguous cases. A workflow may temporarily run through both old and new systems. Those resources have an opportunity cost even when the resulting knowledge becomes valuable.

For AI, the complementary assets include maintained business context, usable interfaces to systems of record, task-specific evaluations, permissions and recovery procedures. These are proposed mechanisms for understanding the adoption gap. Their value has to be demonstrated in the setting where they are deployed.

A J-curve is a conditional explanation, not a promise of eventual recovery. A failed project can remain a failed project. Management needs evidence that the investment is creating a reusable capability and that subsequent outcomes improve after its full costs are counted.

This is also where the Letterman analogy becomes practically useful. Replacing the production of a document may leave all the surrounding work intact. A more consequential application may connect the evidence, decision, permitted action and verification so that an entire case moves differently. I examine that organizational opportunity in How L3 AI autonomy changes the economics of a company.

A task becoming twice as fast can produce a small business gain

Consider an explicitly illustrative process requiring 100 human minutes per completed case. One task accounts for 20 minutes. AI doubles the speed of that task, reducing it to 10 minutes. The other 80 minutes remain.

Illustrative case Human minutes per completed case Labor productivity change at unchanged output quality
Existing process 100 Baseline
Target task becomes twice as fast 90 +11.1%
Add five minutes of extra review and correction 95 +5.3%

The arithmetic is simple: labor productivity is output per labor hour. At the same quality and case mix, reducing human time from 100 to 90 minutes raises it by 100 / 90 − 1 = 11.1%. Including the five additional minutes gives 100 / 95 − 1 = 5.3%.

More generally, if the improved task represents a fraction f of baseline effort and its speed increases by a factor s, the time ratio before new overhead is (1 − f) + f / s. This is the familiar bottleneck logic: the untouched work limits the overall gain.

Inference, integration and supervision costs further affect the financial result. Released employee time becomes economic value when it supports useful additional output, better service or avoided expenditure. A payroll saving requires an actual spending change. A capacity improvement can be valuable without one.

Elapsed service time needs its own measurement. A five-minute task improvement may barely change a case that spends three days waiting for approval. Removing that wait can improve customer experience substantially even if it saves few labor minutes. Labor efficiency, service speed and financial return deserve separate measures.

Where the analogy needs restraint

The two technologies have different failure modes. Successful data transmission does not establish the truth of the transmitted claim; an AI application that interprets evidence or proposes an action needs additional semantic and business validation. Software distribution can be rapid while compute, energy, latency and supervision costs still constrain production.

The clip itself also needs proportion. Letterman was entertaining an audience. It cannot establish his private convictions, and his jokes do not provide a test that classifies every modern skeptic. Hindsight gives us a successful technology to point to; many other promising technologies never achieved comparable impact.

The useful question for a skeptic is which capability or deployment fails to create value, under what conditions. The useful question for an enthusiast is what evidence would change their assessment. Both questions improve investment decisions.

Measure the distance between the curves

For a board or executive team, I would ask how much of the organization’s work lies between trying AI and depending on it. Five measures make that distance concrete:

  1. Reach and intensity. Who uses the tools, how regularly, and for what share of actual work? Purchased licenses alone answer little.
  2. Process coverage. What fraction of incoming business volume passes through a maintained application? Disclose the eligible subset and the total workload.
  3. Verified output. Does the intended result survive review, customer response and subsequent corrections? Count reopened and failed cases.
  4. Complete resource use. Include human preparation, review, exceptions, rework, platform operation and implementation costs.
  5. Realized value. Identify useful additional output, service improvements and actual spending avoided, with an owner accountable for the claim.

Start with a bounded process and a credible baseline. Randomized or phased comparisons can help distinguish the effect of the application from easier cases, more experienced users or changes in demand. Reassess as tools improve. Stop or redesign applications whose gains disappear after quality and overhead are included.

The strategic opportunity is to discover arrangements that become economical as AI capability improves. Some involve automating existing tasks. Others involve reducing coordination, making expertise more accessible, or attempting useful work that previously cost too much. Their importance will be established through outcomes.

The 1995 interview captures how difficult it is to explain a new infrastructure through the examples available at its beginning. The dot-com crash adds a second lesson: inflated expectations can collapse while the application revolution continues. Widespread use, viable businesses and lasting productivity growth remain separate achievements.

AI inherits much of the internet’s distribution. Its lasting value will depend on the applications we invent and the productive ways of working we build around them.

Sources and method

This is a research-based editorial synthesis, not an original causal study or a forecast. Sources were reviewed on 27 October 2026. Historical market and survey observations, published research, a working-paper revision and the illustrative arithmetic are identified separately. The layer and six-curve illustrations are schematic. The historical productivity chart preserves the cited data vintage rather than presenting those values as newly revised statistics. The Nasdaq series is a financial proxy, not a direct measurement of hype.

Download the historical chart data and illustrative calculation inputs and the source register with evidence boundaries. Related reading: The six levels of an AI-native organization and Enterprise automation: from theory to practice.


  1. Late Show with David Letterman, 1995 interview with Bill Gates. Clip supplied with this article’s brief; official Letterman channel version, “Bill Gates Explains the Internet to Dave (1995)”. The exchange concerns baseball audio, radio and tape recorders. ↩

  2. CERN. The birth of the Web. Distinguishes the Web’s 1989 invention and April 1993 public-domain release from the underlying internet. ↩

  3. Pew Research Center. Internet Use Over Time, 10 June 2015 historical series. U.S. adults; selected dated survey observations, not annual averages. ↩

  4. Shawn Sprague. The U.S. productivity slowdown: an economy-wide and industry-level analysis. BLS Monthly Labor Review, 6 April 2021, Figure 2. The private nonfarm business period series is used consistently here. ↩

  5. Stephen D. Oliner, Daniel E. Sichel and Kevin J. Stiroh. Explaining a Productive Decade. Federal Reserve FEDS, 2007. Research attribution concerns broader IT, rather than an isolated internet treatment. ↩

  6. Alexander Bick, Adam Blandin and David J. Deming. The Rapid Adoption of Generative AI. Management Science 72(10), 2026, pp. 8440–8456; published online 20 January 2026. This article uses the published paper’s late-2024 figures; earlier working-paper estimates differ. ↩

  7. Shakked Noy and Whitney Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science 381, 2023, pp. 187–192. A 40% time reduction is a time measure, not a 40% output-per-hour estimate. ↩

  8. Erik Brynjolfsson, Danielle Li and Lindsey Raymond. Generative AI at Work. Quarterly Journal of Economics 140(2), 2025, pp. 889–942. The published study reports 5,172 agents and a 15% average increase in issues resolved per hour. ↩

  9. Fabrizio Dell’Acqua and colleagues. The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork. Organization Science, published online 12 June 2026. The published paper analyzes 791 participants; 776 supplied complete post-task surveys, the sample count commonly quoted from the earlier working paper. ↩

  10. Joel Becker, Nate Rush, Beth Barnes and David Rein. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, 10 July 2025. A task completion-time result for the early-2025 study setting. ↩

  11. Joel Becker and colleagues. We are Changing our Developer Productivity Experiment Design. METR, 24 February 2026. The researchers identify selection and time-measurement problems that limit interpretation of their follow-up estimates. ↩

  12. Anders Humlum and Emilie Vestergaard. Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, March 2026 revision. Previously circulated as Large Language Models, Small Labor Market Effects. This article uses the revision’s 2% bound, not the earlier version’s 1% bound. ↩

  13. Erik Brynjolfsson, Daniel Rock and Chad Syverson. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics 13(1), 2021, pp. 333–372. A model of complementary intangible investment and productivity measurement; it does not guarantee returns to individual AI projects. ↩↩

  14. K. Lougheed and Y. Rekhter. RFC 1105: Border Gateway Protocol, June 1989. Historical protocol specification, subsequently superseded; cited for the development of inter-network routing. ↩

  15. Authorize.net. About us. Company history dates its payment gateway service to 1996; this is an illustrative provider milestone, not the origin of online payments. ↩

  16. Amazon. From Zero to 10 Million in Less Than Four Years, June 1999; and World’s Largest Bookseller Opens on the Web, 4 October 1995. Primary company records for the July 1995 launch and early service design. Its five-year timeline records the May 1997 IPO. ↩↩

  17. Sergey Brin and Lawrence Page. The Anatomy of a Large-Scale Hypertextual Web Search Engine, 1998. Original research on Google’s search architecture; not a claim that Google invented web search. ↩

  18. Meta. Company information and history. Primary company history records Facebook’s 2004 launch. ↩

  19. Google. Google Maps’ biggest moments over the past 15 years, 2020. Records the 8 February 2005 desktop launch and subsequent product development. ↩

  20. Netflix. Netflix Offers Subscribers the Option of Instantly Watching Movies on Their PCs, 16 January 2007. Original announcement describes the initial catalog, licensing, bandwidth requirements, subscription integration and phased release. ↩↩↩

  21. Gartner. Hype Cycle research methodology. An interpretive framework for expectations and technology maturity, rather than a dated empirical trajectory for the whole internet or AI economy. ↩

  22. Nasdaq, Inc. NASDAQ Composite (NASDAQCOM), retrieved through FRED, Federal Reserve Bank of St. Louis, 27 October 2026. Historical daily closes; decline calculated as 1 − 1,114.11 / 5,048.62. The chart uses monthly last available closes plus those two dated daily observations. ↩

  23. Google Inc. Form S-1, filed 29 April 2004, Management’s Discussion and Analysis, “How We Generate Revenue.” The discussion describes profitability from 2001 and cost-per-click AdWords from the first quarter of 2002. ↩

  24. Amazon.com. Amazon.com Announces 33% Sales Growth Fueled by Lower Prices, 24 October 2002. Actual third-quarter sales figures; growth does not by itself establish profitability or investor returns. ↩

  25. eBay Inc. eBay Completes PayPal Acquisition, 3 October 2002, filed as Exhibit 99.1. A completed integration milestone, not evidence that every online payment venture succeeded. ↩

  26. Tara Bitran. Skip Intro: Netflix Turns 25 Today, Netflix Tudum, 29 August 2022. Company history records the 1998 website launch and 2007 introduction of streaming. ↩