Part A decision time: August 24, 2006, 12:00 a.m. PDT
Part A knowledge cutoff: August 23, 2006, 11:59:59 p.m. PDT
Outcome horizons: October 2008, December 2015 and December 2020
Outcome classification: Success, with causal attribution unresolved
This note reveals evidence unavailable to the Part A decision maker. It is bound to canonical Part A digest f2f77797c383b2661a00c742a4ef574eaff53aedfc4f0dfc54b3cb57a280986c. Structured ledgers, reported facts and exact excerpts are authoritative over this narrative.
Outcome in one sentence
Amazon launched EC2 in beta, opened access and added instance types, moved the service to general availability with an SLA, and later reported AWS as a large profitable segment. That sequence is consistent with the bounded-learning logic recommended in Part A, but it does not identify how much of the later outcome was caused by the initial beta rather than subsequent execution, investment and external adoption. source · aws.ec2.beta-announcement source · aws.ec2.unlimited-beta source · aws.ec2.general-availability source · amazon.2015.10k source · amazon.2020.10k
What Amazon actually did
On August 24, 2006, Amazon announced EC2 in beta as compute in the cloud alongside S3 storage. It framed the customer economics as paying only for capacity actually used. The product announcement verifies the action and proposition; it does not disclose Amazon's capacity cost, customer retention or contribution margin. source · aws.ec2.beta-announcement
Independent reporting in November said hundreds of programmers took all test slots in less than five hours. The same publication described investor concern about Amazon's multiyear technology spending. Together, those observations show demand for scarce beta access and real capital-allocation skepticism; neither establishes durable demand or attractive returns. source · businessweek.amazon-risky-bet.2006
The 2006 shareholder letter, filed in April 2007, made the platform thesis explicit. Amazon said AWS served a new developer customer group, had more than 240,000 registered developers, and targeted broad storage and compute needs by using expertise gained while scaling Amazon.com. Management also asserted that the business could become significant and financially attractive. Those are attributed management claims, not contemporaneous segment economics. source · amazon.2006.shareholder-letter
Amazon then expanded in stages. In October 2007 it opened the EC2 beta to all developers and introduced larger instance types. In October 2008 it said EC2 entered general availability after just over two years in beta and added a 99.95 percent regional availability service-level commitment. source · aws.ec2.unlimited-beta source · aws.ec2.general-availability
Reported financial outcomes
Every value below is an as-reported segment fact. No growth rate, margin, return on capital or valuation is model-generated in this episode.
| USD millions | FY 2013 | FY 2014 | FY 2015 | FY 2019 | FY 2020 |
|---|---|---|---|---|---|
| AWS net sales | 3,108 | 4,644 | 7,880 | 35,026 | 45,370 |
| AWS operating income | 673 | 660 | 1,863 | 9,201 | 13,531 |
Source lineage: table.amazon.outcome.aws-segment-results. The 2013-2015 values trace to Amazon's 2015 Form 10-K, and the 2019-2020 values trace to its 2020 Form 10-K. source · amazon.2015.10k source · amazon.2020.10k
These values establish that AWS later became financially material and reported positive segment operating income in all five displayed periods. They do not isolate EC2, exclude later AWS products, or measure the capital committed over the intervening years.
Amazon attributed 2015 AWS sales growth primarily to increased customer usage, partly offset by pricing changes. For 2020 it attributed higher AWS operating income primarily to customer usage and cost-structure productivity, while also citing a server useful-life estimate, payroll, infrastructure investment and customer price reductions. This disclosed mix is exactly why the endpoint cannot be assigned to one launch decision. source · amazon.2015.10k source · amazon.2020.10k
Process quality versus outcome quality
The Part A recommendation did not predict the displayed segment results. It recommended buying information through a bounded compute beta because early S3 adoption and developer interest supported an adjacency test while standalone economics and capacity requirements were unknown.
The observed sequence is aligned with that process:
- Amazon launched the compute product as a beta rather than presenting cutoff evidence for an immediate full-scale commitment.
- Scarce test access produced an early external demand signal.
- Access and instance scope broadened before general availability.
- A service-level commitment accompanied the transition out of beta.
- Later filings reported substantial AWS sales and positive operating income.
This is good evidence that staged platform learning was workable. It is not proof that the original capacity, pricing, rollout speed or spending envelope was optimal. Public sources here do not show the internal gates, losses, capacity decisions or rejected alternatives. The later successful outcome must not retroactively convert the Part A unknowns into facts.
Causal assessment
Primary hypothesis: staged learning enabled platform expansion
The best-supported explanation is that S3 supplied an initial developer and operating signal, the EC2 beta exposed real workloads, and subsequent access, product and reliability expansion accumulated information and platform capability before the segment reached material scale. [hypothesis.amazon.staged-learning-platform]
Support comes from the cutoff S3 object count, independent developer-access assessment, rapid use of scarce EC2 test slots, later open beta, general availability and reported AWS segment outcomes. [claim.amazon.cutoff.s3-adoption-signal; claim.amazon.cutoff.external-developer-value; claim.amazon.outcome.early-beta-demand; claim.amazon.outcome.ec2-expanded; claim.amazon.outcome.ec2-general-availability; claim.amazon.outcome.aws-2020-reported]
Confidence is moderate. The records establish chronology and consistency with the mechanism. They do not provide a control group or a deterministic counterfactual.
Rival hypothesis: later execution and secular adoption dominate
A credible rival is that later service development, capacity investment, pricing, geographic reach, enterprise adoption and general growth in cloud demand explain more of the endpoint than the initial beta. Amazon's own 2015 and 2020 explanations name usage, productivity, pricing, payroll, infrastructure spending and an accounting-estimate change. [hypothesis.amazon.later-execution-and-adoption; claim.amazon.outcome.aws-growth-drivers]
The long interval and aggregation across AWS services make this rival impossible to dismiss with the selected evidence. The bounded conclusion is that the beta plausibly enabled learning and platform entry; no percentage of 2020 sales or operating income is attributed to it. [assumption.amazon.outcome.partial-attribution]
Counterfactual
A feasible cutoff action was to keep S3 and other APIs operating while delaying external compute. S3 already had disclosed pricing and an object-count signal, while AWS economics and the cost of fixed infrastructure remained unknown. [counterfactual.amazon.storage-api-led; evidence.amazon.s3.launch.pricing; evidence.amazon.s3.july-2006.objects; evidence.amazon.q1-2006.segments; evidence.amazon.2005.10k.fixed-investment-risk]
The effect is not quantified. Delay might have preserved capital and allowed storage evidence to mature. It might also have sacrificed workload learning, developer relationships and time in market. No cutoff-valid compute cohort, capacity budget, competitor response or deterministic causal model supports a financial estimate for either path. [assumption.amazon.outcome.storage-led-counterfactual]
Ex ante signals an agent should retain
- Adjacent usage, carefully labeled. More than 800 million S3 objects showed service use, not compute demand, customer retention or profitability. [evidence.amazon.s3.july-2006.objects]
- Independent customer-problem evidence. A contemporary observer said sophisticated, scalable storage had been out of reach for small developers. The observation supported problem relevance, but not market size or moat. [evidence.techcrunch.s3.infrastructure-access]
- Low-commitment customer pricing. S3 had no minimum fee and charged by usage. This reduced a customer's trial barrier but said nothing about Amazon's contribution margin. [evidence.amazon.s3.launch.pricing]
- A funding base with constraints. Consolidated cash generation and liquidity made a capped test plausible, while debt, operating-income pressure and fixed-investment risk limited the inference. [evidence.amazon.fy2005.cash-flow; evidence.amazon.q1-2006.liquidity; evidence.amazon.q1-2006.results; evidence.amazon.2005.10k.fixed-investment-risk]
- Missing segment economics. Reporting only North America and International segments made the standalone infrastructure economics explicitly unknown. Missing evidence was itself a decision-relevant signal to limit commitment. [evidence.amazon.q1-2006.segments]
An agent should preserve each signal's false-positive risk. Otherwise hindsight turns ambiguous early evidence into supposedly obvious proof.
Candidate decision rule
rule.stage-platform-adjacency-under-unknown-unit-economics is a candidate, not corpus-validated, rule:
When an incumbent has a transferable internal capability and a plausible adjacent customer problem, but standalone economics and retention are unknown, use a bounded beta to generate cohort, contribution and reliability evidence before making the largest fixed commitments.
The watch list is retained workload usage, normalized contribution, reliability and core isolation, support burden, and funding health. Reverse or kill expansion when usage fails to deepen, losses breach the approved envelope, severe incidents spill into the core, support remains structurally high or the funding cushion deteriorates.
This rule does not transfer automatically to safety-critical systems, regulated approvals, indivisible infrastructure or markets in which delaying a scarce right has independently verified costs. One successful episode cannot validate a universal rule.
Evidence gaps and abstentions
The selected sources establish the launch sequence and reported segment values, but they do not establish:
- EC2-specific revenue, costs, capital expenditure, margins or returns;
- the initial beta's customer retention, workload intensity, utilization or contribution;
- Amazon's internal capacity, spending, reliability and kill gates;
- the causal share of later AWS outcomes attributable to the August 2006 decision;
- the value created or preserved versus a storage-led delay;
- a complete capital bridge from the 2006 beta to the displayed segment periods; or
- an intrinsic value or target price for Amazon.
No material accounting adjustment, valuation or counterfactual model is presented. Human publication approval remains pending, so this episode must not be represented as approved for external use.
Source map
- Amazon 2005 Form 10-K — cutoff business model, investment and fixed-cost risk (
src.amazon.2005.10k). - Amazon full-year 2005 results — cutoff consolidated financial reconstruction (
src.amazon.fy2005.results). - S3 launch and July adoption update — primary launch terms and object-count signal (
src.amazon.s3.launch;src.amazon.s3.adoption-july-2006). - TechCrunch S3 coverage — contemporaneous independent developer-access and pricing context (
src.techcrunch.amazon-s3.2006-03-14). - EC2 beta announcement, unlimited beta and general availability — primary action and staged expansion (
src.aws.ec2.beta-announcement;src.aws.ec2.unlimited-beta;src.aws.ec2.general-availability). - Bloomberg Businessweek, “Jeff Bezos' Risky Bet” — contemporaneous independent early demand and investor skepticism (
src.businessweek.amazon-risky-bet.2006). - Amazon 2006 shareholder letter — management platform and long-run economics claims (
src.amazon.2006.shareholder-letter). - Amazon 2015 Form 10-K and 2020 Form 10-K — reported AWS segment outcomes and cited drivers (
src.amazon.2015.10k;src.amazon.2020.10k).
Every displayed number resolves through a fact to an exact, hash-bound excerpt. Interpretive conclusions remain typed as assumptions, conflicts, judgments, causal hypotheses and a candidate rule rather than facts.