# UNITED STATES TAX COURT

> Briefs, arguments, decisions, and more.

URL: https://www.frixlaw.com/law-library/documents/agency%3Atax-court%3A3399bff9c7547541

## Record

- **Collection:** Agency decision
- **Document type:** Agency decision

## Text

505

CT

148 T.C. No. 8

UNITED STATES TAX COURT

AMAZON.COM, INC. & SUBSIDIARIES, Petitioner v.
COMMISSIONER OF INTERNAL REVENUE, Respondent

Docket No. 31197-12.

Filed March 23, 2017.

In 2005 P entered into a cost sharing arrangement (CSA) with
S, its Luxembourg subsidiary. Pursuant to the CSA, P granted S the
right to use certain pre-existing intangible assets in Europe, including
the intangibles required to operate P's European website business.
This arrangement required S to make an upfront "buy-in payment" to
compensate P for the value of the intangible assets that were to be
transferred to S. See sec. 1.482-7(a)(2), (g)(2), Income Tax Regs.
Thereafter S was required to make annual cost sharing payments to
compensate P for ongoing intangible development costs (IDCs), to
the extent those IDCs benefited S. See 4 paras. (a)(1), (d)(1). As
consideration for the transfer of pre-existing intangibles, S made a
$254.5 million buy-in payment to P.
Applying a discounted cash-flow (DCF) methodology to the
expected cash flows from the European business, R determined a buyin payment of $3.6 billion, later reduced to $3.468 billion. P contends that R's DCF methodology is substantially similar to that rejected by this Court in Veritas Software Corp. v. Commissioner, 133

SERVED Mar 23 2017

-2T.C. 297 (2009). P contends that R's determinations are arbitrary,
capricious, and unreasonable and that the comparable uncontrolled
transaction (CUT) method is the best method to calculate the requisite
buy-in payment.
P used a multistep allocation system to allocate costs from its
various cost centers to IDCs. See sec. 1.482-7(d)(1), Income Tax
Regs. (providing that costs "must be allocated between the intangible
development area and the other areas or business activities on a reasonable basis"). While accepting P's allocation method in many respects, R determined that 100% of the costs captured in one important
cost center ("Technology and Content") must be allocated to IDCs. P
contends that R's determination to allocate to IDCs 100% of the
Technology and Content costs is inconsistent with the regulations.
1. Held: R's determination with respect to the buy-in payment
is arbitrary, capricious, and unreasonable. Veritas Software Corp. v.
Commissioner, 133 T.C. 297, followed.
2. Held, further, P's CUT method, with appropriate upward adjustments in numerous respects, is the best method to determine the
requisite buy-in payment.
3. Held, further, R abused his discretion in determining that
100% of Technology and Content costs constitute IDCs.
4. Held, further, P's cost-allocation method, with certain adjustments, supplies a reasonable basis for allocating costs to IDCs.

John B. Magee, Sanford W. Stark, Bryon A. Christensen, Tracey X. Zheng,
Robert S. Kirschenbaum, Beth L. Williams, John G. Ryan, Saul Mezei, Julia
Kazaks, Rajiv Madan, John A. Polito, Michael D. Kummer, Christopher P.

-3Murphy, Royce L. Tidwell, Carl T. Ussing, Hans D. Gerling-Ritters, Nicholas A.

Zemil, and Paul M. McLaughlin, for petitioner.
Jill A. Frisch, Anne O'Brien Hintermeister, Melissa D. Lang, Lloyd T.
Silberzweig, Shannon L. Cohen, Melissa L. Hilty, Scott W. Mentink, David A.

I_e_e, My V. Vo, Jimeel R. Hamud, Aaron T. Vaughan, and Mary E. Wynne, for
respondent.

CONTENTS

FINDINGSOFFACT...............................................10
I.

Overview of Amazon's European Business . . . . . . . . . . . . . . . . . . . . . . . . 12
A.

Initial Expansion Into Europe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

B.

Original Structure of European Business . . . . . . . . . . . . . . . . . . . . . 16

C.

"ProjectGolderest".......................................20

D.
II.

1.

Cost Sharing Arrangement . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.

LicenseAgreement................................. 24

3.

AssignmentAgreement..............................24

4.

European Subsidiary Contribution . . . . . . . . . . . . . . . . . . . . . 25

5.

European Business Contribution . . . . . . . . . . . . . . . . . . . . . . 25

6.

Four-PartyAgreement.............................. 26

Life After Project Golderest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

Retail and Technological Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
A.

Internet Retail Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

B.

Evolution of Amazon's Website Architecture . . . . . . . . . . . . . . . . . 32

C.

Evolution of Amazon's Software Applications . . . . . . . . . . . . . . . . 38

1.

Customer Master Service ............................ 38

2.

Order Master Service............................... 39

3.

Dynamo..........................................40

4.

Payments.........................................41

5.

Item Master....................................... 42

-4-

6.
D.

III.

Personalization.................................... 43

7.
Messaging........................................ 45
New Products and Services After January 1, 2005 . . . . . . . . . . . . . . 46

Amazon's Third-Party Businesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
A.

Merchants.com......................................... 47

B.

Associates and Syndicated Stores . . . . ... . . . . . .... . . . . .... . . . 49

IV.

The Buy-In Payment.......................................... 51

V.

Cost Sharing Payments........................................ 57

VI.

Stock-Based Compensation.....................................65

OPINION........................................................ 67
I.

Cost Sharing Background...................................... 69

II.

Respondent's Determination of the Buy-In Payment . . . . . . . . . . . . . . . . . 73

III.

Petitioner's Determination of the Buy-In Payment . . . . . . . . . . . . . . . . . . . 89
A.

B.

Website Technology..................................... 91

1.

RoyaltyRate...................................... 95

2.

Useful Life and Decay Curve........................ 102

a.

UsefulLife................................. 103

b.

Decay Curve................................ 107

c.

"Tail"Period................................117

3.

RevenueBase.................................... 121

4.

DiscountRate.................................... 124

Marketing Intangibles ................................... 127

1.

RoyaltyRate..................................... 129

2.

UsefulLife...................................... 135

3.

RevenueBase.................................... 143

4.

DiscountRate.................................... 146

5.

EuropeanPortfolio................................ 147
a.
b.

Ownership of the European Portfolio . . . . . . . . . . . . 148
Allocating Value to the European Portfolio . . . . . . . 154

-5-

C.
IV.

Customer Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

Cost Sharing Payments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
A.

Respondent'sPosition...................................174

B.

Petitioner'sPosition.................................... 178

C.

Stock-Based Compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

LAUBER, Judge: The Internal Revenue Service (IRS or respondent) determined, for 2005 and 2006 respectively, deficiencies in petitioner's Federal income
tax of $8,380,790 and $225,653,149.¹ These deficiencies arose from a series of
transactions by which Amazon.com, Inc., and its domestic subsidiaries (collectively, Amazon US) transferred to Amazon Europe Holding Technologies SCS
(AEHT), a Luxembourg subsidiary, the intangible assets required to operate petitioner's European website business. Invoking section 482, the IRS made
substantial transfer-pricing adjustments reallocating income to Amazon US from

AEHT.
From its inception through 2005 Amazon US owned the intellectual property in question. In 2004 Amazon US and AEHT entered into a "cost sharing

¹Unless otherwise indicated, all statutory references are to the Internal
Revenue Code in effect for the tax years at issue, and all Rule references are to the
Tax Court Rules of Practice and Procedure. We round all dollar amounts to the
nearest dollar.

-6arrangement." See sec. 1.482-7(a)(1), (b)(1), Income Tax Regs.2 This arrangement required AEHT to make an up-front "buy-in payment" to compensate Ama-

zon US for the value of the intangible assets that were to be transferred to AEHT.
S_e_e id. paras. (a)(2), (g)(2). Thereafter AEHT was required to make annual cost
sharing payments to compensate Amazon US for ongoing intangible development
costs (IDCs), to the extent those IDCs benefited AEHT.
In a series of transactions in 2005 and 2006 Amazon US transferred to
AEHT three groups of intangible assets: (1) the software and other technology required to operate petitioner's European websites, fulfillment centers, and related
business activities; (2) marketing intangibles, including trademarks, tradenames,
and domain names relevant to the European business; and (3) customer lists and
other information relating to petitioner's European clientele. After concessions,3
this case requires the Court to decide two main issues: the proper amount of
AEHT's buy-in obligation with respect to the assets thus transferred; and the vol-

2Section 1.482-7, Income Tax Regs., was redesignated section 1.482-7A,
Income Tax Regs., with the promulgation of new regulations effective January 5,

2009. See T.D. 9441, 2009-7 I.R.B. 460.
3On May 30, 2014, the parties filed a stipulation of settled issues reflecting
their settlement of an issue captioned "Cost Sharing Acquired Intangibles Buy-In."
Petitioner conceded this issue in full and agreed that its taxable income would

accordingly be increased by $4,881,993 for 2005 and $2,548,165 for 2006.

-7ume of petitioner's costs properly treated as IDCs (the larger the volume, the
larger the cost sharing payments that AEHT must make).
On the first issue, petitioner originally reported a buy-in payment from
AEHT of $254.5 million, to be paid over seven years. In determining the value of
the transferred assets, petitioner assumed that each group of assets (website technology, marketing intangibles, and customer information) had a seven-year useful
life.
On examination of petitioner's returns, the IRS concluded that the buy-in

payment had not been determined at arm's length. See sec. 1.482-7(g)(2), Income
Tax Regs. In respondent's view, the transferred property had an indeterminate
useful life, and it had to be valued, not as three distinct groups of assets, but as integrated components of an operating business. Applying a discounted cash-flow
(DCF) methodology to the expected cash flows from the European business, the
IRS determined a buy-in payment of $3.6 billion, later reduced to $3.468 billion.
Petitioner contends that respondent's DCF methodology is substantially
similar to that rejected by this Court in Veritas Software Corp. v. Commissioner,

133 T.C. 297 (2009), nonacq., 2010-49 I.R.B. (2010). Petitioner disputes the conceptual soundness of that methodology as applied here, contending that it treats
short-lived intangibles as if they had perpetual useful lives. The result, according

-8to petitioner, is to inflate the buy-in payment by improperly including in it the

value of subsequently developed intangible property. See sec. 1.482-7(g)(2), Income Tax Regs. (requiring that the buy-in payment reflect only pre-existing intangibles).
To value the pre-existing intangibles properly, petitioner contends that each
group of transferred assets must be valued separately under the "comparable uncontrolled transaction" (CUT) method. On the basis of expert testimony at trial,4
petitioner contends that the website and related technology had a value when
transferred between $117 million and $182 million; that the marketing intangibles
had a value when transferred between $115 million and $165 million; and that the
customer information had a value when transferred between $52 million and $66
million. This would yield a buy-in payment ranging from $284 million to $413
million. Citing testimony by his trial experts, respondent urges substantially higher values for each group of assets in the event we reject his DCF method.
With respect to ongoing cost sharing payments, we must address two distinct issues. The first is essentially a tax accounting question, requiring that we
determine "all of the costs incurred * * * related to the intangible development

4An alphabetical list of petitioner's and respondent's expert witnesses,
together with a short resume of each, appears in the appendix to this Opinion.

-9area." Id. paras. (b)(2), (d)(1), (f)(1). Petitioner used a multistep allocation system
to allocate costs from its various cost centers to IDCs. See id. para. (d)(1) (providing that costs "must be allocated between the intangible development area and
the other areas or business activities on a reasonable basis"). While accepting
petitioner's allocation system in many respects, the IRS determined that 100% of
the costs captured in one important high-level cost center ("Technology and Content") must be allocated to IDCs. The result of this determination was to increase
by $23 million and $109.9 million the cost sharing payments that AEHT was required to make in 2005 and 2006, respectively. Asserting that this determination
was arbitrary and capricious, petitioner contends that only about half of Technology and Content costs should be allocated to IDCs.
The second question involves section 1.482-7(d)(2), Income Tax Regs.,
which requires that stock-based compensation be included in the IDC pool upon
which cost sharing payments are based. Petitioner complied with this regulation
in preparing its 2005-2006 tax returns; because the vast bulk of petitioner's stockbased compensation was paid to its domestic employees, the result was to increase
accordingly the cost sharing payments required of AEHT. Our treatment of this
issue is governed by the terms of the parties' cost sharing agreement, and in light

- 10 of our decision in Altera v. Commissioner, 145 T.C. 91 (2015), which invalidated
the regulation in question.

FINDINGS OF FACT
Amazon.com, Inc. (ACI), is the common parent of a group of affiliated corporations that join in the filing of a consolidated return (Amazon US) and of numerous foreign subsidiaries (collectively, Amazon or petitioner). ACI was incorporated in 1994 in Washington and was reincorporated in 1996 in Delaware. It
began operations in 1995 and completed an initial public offering of common
stock in 1997. Its principal place of business when it filed the petition was in
Seattle, Washington.5
Amazon is an online retailer. It does not have brick-and-mortar stores, but
rather sells products exclusively through Amazon.com and related websites. Initially Amazon sold only books, but by 2000 it had expanded its offerings into
many other product categories, including music, video, electronics, toys, software,
video games, cameras, kitchen items, tools/hardware, and home/garden. Amazon

5The Court issued protective orders adopting procedures to protect petitioner's trade secrets and proprietary technology (collectively confidential information) during the pre-trial, trial, and post-trial phases of this case. The facts set
forth in this Opinion have been adapted accordingly. All information included
herein has been determined by the Court not to constitute "trade secrets or other
confidential information" within the meaning of section 7461(b).

- 11 would identify unrelated product vendors; buy products from them; manage
inventory and price the products; list the products for sale on Amazon.com; and
ship the items to customers from its warehouses.
Amazon is committed to continuous growth in the number and variety of
products it offers for sale. With a view to increasing selection further, Amazon in
2000 began allowing third parties to sell items on its websites. Amazon made
available a set of eCommerce platforms, services, and tools that enabled third parties to list their own products and services for sale on Amazon.com and related
websites. This branch of Amazon's business was called Marketplace. Through
Marketplace, third-party merchants set their own prices for their products and services, which appeared alongside the products that Amazon itself sold. Amazon
received commissions on these third-party sales and recorded the commission
amounts (not the sale prices) as revenue.
Some third parties desired to use Amazon's technology but did not wish to
sell their products on an Amazon-branded website. To meet this demand, Amazon
built and operated eCommerce websites, custom made for a particular merchant,
through which that merchant could make online sales of its products under its own
brand name. This branch of Amazon's business was called Merchants.com or
M.com. Whereas Marketplace retailers sold items on Amazon's websites, M.com

- 12 retailers sold items on their own branded websites, which were built by Amazon
using the entire suite of eCommerce technology that ran Amazon's own websites.
Although Amazon built websites for numerous customers, the most successful
implementation of its M.com business was the website it built for Target Corp.,
which enabled Target to sell its retail products through Target.com.
Since its founding Amazon has committed itself to the "three pillars" of
selection, price, and convenience. Under the selection pillar Amazon aims to offer
its customers the widest possible selection by ensuring continuous growth in the
number and variety of products offered. Under the price pillar Amazon endeavors
to keep its prices as low as possible at all times. The convenience pillar encompasses a range of goals associated with improving the customer experience, including: (1) helping customers find what they seek as quickly as possible; (2) delivering to them as quickly and accurately as possible all items that they purchase;
and (3) ensuring that potential customers have all information useful in making
purchasing decisions, even if their initial decision is not to make a purchase.
I.

Overview of Amazon's European Business
A.

Initial Expansion Into Europe

In April 1998 Amazon acquired Bookpages, Ltd., an online bookstore in the
UK, and Telebook, Inc., an online bookstore in Germany. Later that year Amazon

- 13 re-launched these websites under the respective domain names Amazon.co.uk and
Amazon.de. In August 2000 Amazon launched its French business organically
rather than by acquisition, adopting the domain name Amazon.fr. At the outset,
Amazon.co.uk and Amazon.de offered only books; Amazon.fr offered books, music, and video products. During the tax years at issue, Germany, the UK, and
France were the only European countries in which Amazon operated.
By 1999 Amazon.co.uk, Amazon.de, and Amazon.com had become the
three most popular online retail domains in Europe. By 2005 Amazon's share of
total online retail spending approached or exceeded double digits in Germany and
the UK, while lagging in France. In each country, however, Internet retailing was
a small fraction of the overall retail market segment. By one account, Internet
sales in 2005 represented only 2.4%, 5.6%, and 1.1% respectively of total retail
spending in Germany, the UK, and France.
Although all three nations were EU members, the manner in which Amazon
operated its business in each country was in many respects local. A local country
manager supervised a staff with responsibility for vendor and customer relationships, fulfillment, pricing, and financial management. Due to differing cultural

preferences, retail traditions, and national regulations, the details of these opera-

- 14 tions--and the technology required to make them happen--often varied from country to country.

As of January 1, 2005, vendors and the terms on which they sold goods to
Amazon differed in each country. Local Amazon employees in each country identified and recruited vendors; this recruitment process was intensely personal and
could take years. For example, although Canon is a multinational corporation selling cameras throughout the world, each of Amazon's European subsidiaries had to
negotiate separately with the Canon team in its country, and those Canon teams
had distinct organizations, pricing policies, and sources of supply. Similar variations existed in the Marketplace program: Local teams speaking the local language identified and recruited prospective Marketplace sellers in each country. As
of January 1, 2005, even the largest of Amazon's vendors and merchants transacted with its European subsidiaries at the local level, rather than transacting with
Amazon on a global or pan-European basis.
Pricing in Europe was also local. Amazon endeavored to offer the best value for customers, which meant matching or beating local prices, both online and
offline. But prevailing prices for a given item could vary considerably from country to country on account of such factors as: (1) unique local competitors and
competitive environments; (2) different vendors with different pricing policies;

- 15 and (3) local laws and regulations that restricted pricing (e.g., by preventing discounting on books and other items). In order to match competitors' prices, each
European subsidiary first had to determine which companies it would treat as competitors; this was a local decision because the markets were all different. Each
subsidiary then sent teams of local employees to check prices by visiting competitors' stores.
Local factors also affected the fulfillment process, that is, the mechanics of
preparing an order for shipment and effecting timely delivery to the customer. As
of January 1, 2005, each European subsidiary filled orders from an in-country fulfillment center; if a French warehouse was short of an item, delivery could be delayed even if UK warehouses had ample supply. Amazon's UK and German fulfillment centers used a "buffer rebin sortation process" that was built independently in Europe; it had a fundamentally different design from that used in Amazon's U.S. and French fulfillment centers. Amazon's transportation costs and the
speed, reliability, and accuracy of its shipping also varied because of local regulations and other economic factors in Germany, the UK, and France.
Surprisingly perhaps, customers' payment preferences also differed geographically. German customers, for example, used credit cards less frequently
than customers in other countries. The German subsidiary thus had to build two

- 16 unique payment systems for use by its customers: "direct debit," which enabled it
to deduct purchase amounts from customers' bank accounts; and "pay by invoice,"
which provided customers a bill along with their shipment. German customers
were also prone to regard items as having been purchased "on approval" and to
return items with which they were dissatisfied. This required the German subsidiary to create unique processes to deal with the high volume of returns, which
could range up to 50% on certain items.
The idiosyncrasies of local markets tripped up many U.S. retailers seeking
to expand into Europe. Best Buy failed in the UK, and Walmart failed in Germany. Amazon itself came close to failing in France; in the early 2000s the
French subsidiary was Amazon's worst performing business. This led to a substantial downsizing in 2004, which required the company to file with French authorities a "social plan" by which it communicated its downsizing plan and the
justification for it.
B.

Original Structure of European Business

Beginning in 1999 and continuing into 2006, the German retail business

was conducted by Amazon.de GmbH and its subsidiary, a disregarded entity for
U.S. income tax purposes (Amazon Germany). See sec. 301.7701-3(b)(2), Proced.
& Admin. Regs. Beginning in 2000 and continuing into 2006, the French retail

- 17 business was conducted by Amazon.fr Holdings SAS and its subsidiaries, which
were likewise disregarded entities for U.S. income tax purposes (Amazon France).
The UK retail business was conducted primarily by Amazon.co.uk Ltd. (Amazon
UK). Collectively, we will refer to these entities as the European Subsidiaries.
The European Subsidiaries were wholly owned by Amazon US.6
Until April 30, 2006, Amazon US was the inventory owner and seller of record for the European businesses. The European Subsidiaries provided services to
Amazon US in operating those businesses; in Germany, these services included
setting up a commissionaire arrangement, whereby Amazon Germany held itself
out to customers as the retail seller for the benefit of Amazon US. The European
Subsidiaries provided Amazon US with retail support, storage facilities, fulfillment, back-office support, and local financial management services.
Until April 30, 2006, Amazon US also operated petitioner's "international
third party" or "international 3PS" businesses. The international 3PS businesses
included: (1) the European activities of Marketplace, which allowed third parties
to offer products for sale on Amazon's websites; (2) the European activities of

6The consolidated group headed by ACI includes a bewildering array of
subsidiaries. In an effort to reduce the alphabet-soup character of this Opinion, we
will generally refrain from noting which particular domestic subsidiary does particular things unless it is material to the legal analysis.

- 18 M.com, whereby Amazon used its technology to build websites enabling European
retailers to sell their products through their own domain names and URLs; and (3)
"Syndicated Stores," essentially the converse of Marketplace, whereby Amazon
used its technology to sell its own products through a European retailer's website.
The European Subsidiaries provided various services to Amazon US in connection
with the international 3PS businesses, including website development and design,
marketing services, and certain commissionaire services.
Before April 30, 2006, Amazon US owned most of the intellectual property
required to operate its European businesses, and it licensed this intellectual property to the European Subsidiaries. This intellectual property included the underlying website technology, all of which had been developed in the United States; all
of the European customer information; and most of the marketing intangibles, including trademarks and domain names. In 1996, for example, Amazon US registered the "Amazon" trademark under Nice classes 9, 37, and 42 in an attempt to
protect that mark throughout the EU. As explained more fully below, some of the
trademarks and domain names used by the European Subsidiaries were titled in
their respective names.
Unlike in the United States, where Amazon operated uniform retail websites
serving a large geography with a large population, Amazon's retail business in

- 19 Europe had a siloed structure. Each European Subsidiary had a distinct website
employing its own national language. Each European Subsidiary had its own fulfillment centers, often processing country-specific inventory, and a distinct universe of customers residing chiefly within its own borders.
This siloed structure resulted in inefficient operations, lack of coordination
among the European businesses, and barriers to pan-European expansion. The
European fulfillment centers had inventory management problems and were reaching their physical capacity. Unlike their U.S. counterparts, European fulfillment
centers could not freely transfer inventory and could not offload volume from one
center to another. The growth of the European business was also stressing Amazon's website technology: All website servers were in the United States, and
increased website traffic exacerbated latency problems. ("Latency" refers to the
speed with which a webpage loads; customers prefer faster speeds.)
Recognizing that serving European customers through poorly coordinated
national silos was inefficient, Amazon during the early 2000s began to investigate
creation of a centralized European headquarters. One goal of this process was to
enhance customer experience by locating servers closer to customers, thus reducing website latency. Other goals were to place top managers in the same time
zones as their customers; to standardize best practices for customer service, traffic

- 20 generation, pricing, and vendor acquisition; to increase efficiency by minimizing
duplicative individual country costs; and to create a pan-European fulfillment
infrastructure that would facilitate expansion into additional EU countries.
Tax issues, including applicable tax rates, also loomed large in Amazon's
thinking. To avoid creating a U.S. permanent establishment, European personnel
could not sign contracts or make ultimate business decisions on behalf of Amazon
US within Europe. By establishing a European headquarters, Amazon could collect and remit value added tax (VAT) at a single rate in a single jurisdiction (the
location of the seller), rather than at multiple rates in multiple jurisdictions (the
locations of the buyers). And Amazon was quite aware that the effective marginal
rate of corporate income tax was (or could be negotiated to be) significantly lower
in certain EU member states--specifically, Luxembourg and Ireland--than it was in
the United States.
C.

"Project Golderest"

Amazon's management decided to establish a European headquarters and
assigned its U.S. tax department the task of developing a tax-efficient strategy for
doing this. It considered several locations for the European headquarters, including Ireland and Luxembourg, and eventually opted for the latter. Amazon representatives met with Luxembourg authorities, including Jean-Claude Juncker,

- 21 Luxembourg's then prime minister, to discuss the potential for Amazon to locate
its European headquarters there.
In broad outline, Amazon's plan was to transfer from Amazon US to the
Luxembourg headquarters affiliate all of the intangible assets required to operate
the European website businesses; to continue using the European Subsidiaries as
service companies earning a nominal rate of return; and to have the vast bulk of
the income from Amazon's European businesses taxed in Luxembourg at a very
low rate. In pursuit of the latter goal, Amazon successfully negotiated an advance
tax agreement with the Government of Luxembourg. After the restructuring, the

Luxembourg entity would function as the operational and administrative headquarters for the European businesses and own virtually all of the intangible assets
required to operate those businesses.
Beginning in 2004 Amazon undertook a series of transactions, dubbed "Project Golderest," to implement this plan. ("Golderest" refers to Luxembourg's national bird.) These transactions were complex; they involved many steps and
many entities. But the basic outline can be stated fairly succinctly. Amazon US
formed AEHT, the Luxembourg headquarters entity that would serve as the holding company for all of the European businesses. AEHT elected to be treated as a
corporation for U.S. income tax purposes from the date of its formation. See sec.

- 22 301.7701-3(c), Proced. & Admin. Regs. Underneath AEHT numerous subsidiaries
were created (collectively, Amazon Luxembourg) to perform various functions
essential to operation of the European businesses. These functions included holding title to the inventory sold in Europe, licensing Amazon's intellectual property,
housing the servers, and maintaining call centers. Amazon Germany, Amazon
UK, and Amazon France thereafter supplied to Amazon Luxembourg the same
types of customer-related, fulfillment, and support services that they had previously furnished to Amazon US.
After forming Amazon Luxembourg, Amazon effected the restructuring by

completing six related transactions: (1) the Cost Sharing Arrangement;7 (2) the
License Agreement For Preexisting Intellectual Property (License Agreement); (3)
the Assignment Agreement For Preexisting Intellectual Property (Assignment
Agreement); (4) the European Subsidiary Contribution; (5) the European Business
Contribution, and (6) the Four-Party Agreement. These agreements may be summarized as follows:
7The Cost Sharing Arrangement actually involved two successive agreements. In December 2004 Amazon US and AEHT entered into an "Agreement to
Share Costs and Risks of Intangible Development" with a stated effective date of
June 7, 2004. On January 11, 2005, Amazon US and AEHT entered into an
"Amended and Restated Agreement to Share Costs and Risks of Intangible Development" with a stated effective date of January 1, 2005. We will generally use the
term "Cost Sharing Arrangement" or "CSA" to refer to the latter.

- 23 1.

Cost Sharing Arrangement

The Cost Sharing Arrangement (CSA) was intended to be a "qualified cost
sharing arrangement" within the meaning of section 1.482-7(a)(1), Income Tax
Regs. Through the License and Assignment Agreements (described more fully
below), AEHT obtained access to pre-existing intangible property of Amazon US,
referred to as "the Amazon Intellectual Property." Through the CSA, the parties
agreed to share the costs of further "research, development, marketing, and other
activities relating to * * * maintaining, improving, enhancing, or extending the
Amazon Intellectual Property."
Through its participation in the CSA, AEHT would assist, by way of financial contribution only, in the ongoing development of technology required to operate the European websites and related activities. The CSA required the parties to
determine "Aggregate Allocable Development Costs," which would then be allocated to the parties according to the ratio of benefits each was projected to derive
from ongoing development. Basically, this meant that Amazon Luxembourg
would pay Amazon US for its ratable share of subsequently incurred intangible
development costs (IDCs).

- 24 2.

License Agreement

On January 1, 2005, Amazon US and AEHT entered into the License Agreement with a stated effective date of January 1, 2005. Amazon US thereby granted
AEHT the rights to use "Amazon Intellectual Property," defined to exclude the
marketing intangibles covered by the Assignment Agreement. The property covered by the License Agreement related to Amazon's website technology. Under
the License Agreement, AEHT agreed to make a buy-in payment for the website
technology in the aggregate amount of $226,520,000, to be paid in installments
during the seven-year period beginning in 2005 and ending in 2011.
3.

Assignment Agreement

In July 2005 Amazon US and AEHT executed the Assignment Agreement.
Amazon US thereby granted AEHT the rights to use Amazon Intellectual Property
not covered by the License Agreement, namely, customer data and previously
developed marketing intangibles including trademarks, trade names, website content, and domain names relating to the European business. Though stated to be
effective as of January 1, 2005, the Assignment Agreement remained executory
until the "Business Transfer Date," which was May 1, 2006. Under the Assignment Agreement, AEHT agreed to make a buy-in payment for the marketing in-

- 25 tangibles and customer data in the aggregate amount of $27,991,000, to be paid in
installments during the six-year period beginning in 2006 and ending in 2011.
4.

European Subsidiary Contribution

In February 2006 Amazon Luxembourg acquired all of the stock of the
European Subsidiaries via tax-free reorganizations under section 368(a)(1)(D).
The reorganizations were accomplished by ACI's transfer of the stock of the European Subsidiaries (valued at approximately $196 million in toto) to AEHT in exchange for AEHT stock and cash, immediately followed by each of the European
Subsidiaries' electing to be disregarded as separate from their owner for U.S. tax
purposes. Amazon Luxembourg thus became the ultimate owner of all of the
European Subsidiaries' property and the ultimate employer of their employees.

S_e_e sec. 301.7701-3(b)(2), Proced. & Admin. Regs.
5.

European Business Contribution

In a series of transactions between April 7 and May 1, 2006, Amazon US
transferred to Amazon Luxembourg, in a section 351 exchange for stock, various
assets (other than intellectual property) required to operate the European businesses. These assets included inventory, accounts receivable/payable, vendor contracts, transportation/delivery contracts, "associates agreements," licenses from
third parties, and service contracts. Amazon US concurrently terminated the prior

- 26 agreements whereby the European Subsidiaries had provided services to and licensed intellectual property from it.
6.

Four-Party Agreement

Under the Four-Party Agreement, effective April 30, 2006, the European
Subsidiaries assigned or licensed exclusively to AEHT certain intellectual property titled in their names. Before 2005 the European Subsidiaries procured protection for, and registered in their own names, certain trademarks and domain
names deployed in Europe. In exchange for these assets AEHT paid in the aggregate about $5 million.
D.

Life After Project Golderest

After April 30, 2006, Amazon Luxembourg reported, for U.S. tax purposes,

all of the income and expenses connected with the European businesses. It entered into agreements with the European Subsidiaries whereby they provided it
with certain fulfillment services, customer and merchant services, and support
services within their respective territories; the European Subsidiaries were paid a
cost-plus fee for these services. In May 2006 AEHT's Irish subsidiary constructed
a facility with numerous servers that provided data storage and hosting services
for the European businesses; it received a cost-plus fee for doing this.

- 27 Amazon Luxembourg was by no means a shell company. Beginning in May
2006 it played a meaningful role in expanding Amazon's existing business in Germany, the UK, and France and extending Amazon's reach elsewhere in Europe.
From 2006 through 2013 AEHT launched 11 new product categories through its
UK and German websites, including apparel/accessories, sports/outdoors, jewelry/
watches, health/personal care, shoes/accessories, auto parts, groceries, and baby
products. It launched 15 new product categories through its French website, including most of the preceding categories and some others. And it launched new
website operations in Italy (Amazon.it) and Spain (Amazon.es). Amazon's European revenues grew very rapidly during this period.
The Luxembourg headquarters also played a nontrivial part in rolling out
new technology--the European Fulfillment Network (EFN)--that implemented
standardized and improved fulfillment operations across Europe. The software
underlying EFN was developed in the United States by Amazon US after 2005. It
addressed the problem of multiple websites with country-specific fulfillment centers located in multiple countries.
The EFN technology successfully converted a three-country silo structure
into a network, leveraging AEHT's status as Amazon's single seller of record
throughout Europe, which simplified the sharing and pooling of inventory. The

- 28 EFN technology enabled customers shopping on one national website to view inventory and acquire products housed in fulfillment centers located in other countries. Implementation of this technology significantly reduced the time that customers waited to receive products, reduced shipment costs, lowered product
prices, and dramatically increased selection.
II.

Retail and Technological Environment
A.

Internet Retail Environment

Internet technology makes retailing a more competitive business than it used
to be. The World Wide Web enables consumers to compare prices in real time and
buy at the lowest price offered on multiple websites. The Internet makes consumers' costs of searching for a product virtually disappear and allows them to switch
from one retailer to another by clicking a mouse. A theory called "stickiness"
posits that a consumer usually will not switch to a competitor after a single bad
experience on a particular site. But Amazon adopted as a guiding principle that
"competition is literally one click away."
The price transparency associated with online retailing leads to lower sales
margins, one factor that makes online retailing so competitive. During the last 20
years, innumerable online retailers have gone out of business or lost significant
value; even today, online retail remains a small fraction of the total retail market

- 29 segment. Because of this competitive environment, constant innovation is
essential to ensure survival. Technological failure damages profitability; a late
delivery or damaged product may also alienate a customer permanently.
Innovative technology underlies every aspect of Amazon's retail business.
It is integral to creating and managing the catalog, displaying items in the catalog
to potential customers, conveying the look and feel of the websites, convincing a
potential customer to buy an item, completing transactions, processing payments,
packaging an item, shipping it to the customer, and preventing fraud. To continue
to deliver on its promises, Amazon in the mid-2000s made massive investments to
ensure a rapid pace of technological innovation. Respondent's principal valuation
expert, Daniel J. Frisch, agreed that Amazon operated in a "highly competitive,
rapidly changing industry" that "requires substantial innovation all the time."
Respondent's technology expert, Edward Felten, noted that "Amazon's entire
existence has been characterized by the challenges of innovating due to running
into unexpected walls and growing so fast that the entire structure seems in danger
of failure at any moment."

Amazon's software engineers, computer scientists, and management team
focused on continuous innovation to provide easy-to-use functionality, rich website content, fast and reliable fulfillment, timely customer service, and a trusted

- 30 transactional environment. Amazon regularly launched software on a test basis,
fully conscious of the need for further improvements; its software and website
content often changed multiple times the same day. Amazon's software development process "leveraged the future": By building a piece of software quickly,
Amazon incurred the risk that it would not be adaptable to future needs. By repeatedly choosing "the expedient path" over "the right path," Amazon built up
"technical debt" that inheres in software with a relatively short useful life.
Amazon's need to innovate was driven by multiple factors. Perhaps the
most important factor was the need to increase "scale." One of Amazon's early
business goals was to "get big fast"; getting big fast meant adding customers, webpage views, vendors, and new products at an extremely rapid pace. Website technology by its nature is subject to scale limitations; if the website cannot "scale up"
to meet the demands placed upon it, it will crash. Rapid technological innovation
was required to overcome the scale challenges posed by rapid growth.
These challenges were especially pronounced during Amazon's peak holiday seasons, when webpage views, transactions consummated, and products packaged increased by 300% to 400% over nonpeak periods. Holiday periods stressed
every aspect of Amazon's system, and each successive peak season entailed higher

- 31 volumes and thus greater scale challenges. Amazon risked system collapse if its
technology could not scale up to these demands.
Amazon's engineers testified that they spent the 1998, 2003, 2004, and
2005 holiday seasons "in crisis mode." Although its website was available to the
public 98% of the time during the fourth quarter of2004, Amazon experienced

outages during critical holiday periods that impaired customers' ability to browse
and shop. These and other outages during 2004, which resulted in lost revenue
approaching $130 million, jeopardized Amazon's relationships with retail customers and Marketplace merchants. Determined to address these scale challenges,
Amazon made 24-7 website availability its chief company goal for 2005.
A related factor driving innovation was the need to reduce "latency." During the early 2000s Amazon's software architecture was built in a siloed fashion
that required particular functions to "call" data from databases. As website traffic
geometrically increased, more nanoseconds were required to call these data, causing webpages to load more slowly and respond less quickly to customer requests.
Amazon believed that website latency frustrated customers, leading them to abandon the page they were viewing or leave Amazon's site altogether.
Another factor driving innovation was the need to protect customer data and
prevent fraud. Computer hackers and other bad actors posed increasing security

- 32 risks through cyber-attacks of various kinds. After 2005 security became a major
focus of investment for Amazon. It dedicated large teams of software engineers to
create innovative security protocols designed to repel cyber-attacks, keep its website up, protect customer information, and maintain customer trust.

The advent of new technologies, such as smartphones, likewise propelled
innovation. Smartphones enabled customers to view products and make purchases
through mobile networks; younger customers were especially prone to doing this.
An online retailer risked losing these customers if it did not meet their demands.
Because the online retail environment was changing so rapidly, Amazon believed
it impossible to anticipate technological advances more than three years away.
B.

Evolution of Amazon's Website Architecture

In January 2005 Amazon knew that many components of its website architecture would not meet its long-term needs. Its engineers, who never lacked selfconfidence, believed that they would succeed in meeting future challenges as they
appeared. But they did not know how they would do this.
When Amazon.com was launched in 1995, its website ran on Obidos, a
single monolithic application atop an Oracle database.8 Its user interface code,
8In software engineering, a monolithic application describes a single-tiered
software application in which the user interface and the data access code are
(continued...)

- 33 database connections, and business logic were heavily interdependent. As traffic
to Amazon's website increased, it added more and more servers, called "onlines,"
each running a web server and an instance of the monolithic Obidos code.
The Obidos application was written in "catsubst," an arcane programming
language that Amazon created. Significant engineering effort was required to perform even basic tasks. The manner in which Obidos was written made it difficult
to identify "dependencies," i.e., instances in which specific parts of the code depended on specific other parts of the code. Since any change to the Obidos code
might affect many applications, a team desiring to modify one application had to
get input from multiple engineering groups. This became a cumbersome process.
Obidos was also subject to "death spirals." Until 2000, every Obidos process, no matter how short lived, created its own database connections by calling
databases directly, without coordinating requests across the different onlines.
When too many onlines were seeking the same database resources, error conditions called "timeouts" would occur. The Obidos process would then throw a fatal
error and restart, creating new database connections without properly releasing the
old ones. As database connections piled up, calls from other Obidos processes to

8(...continued)
combined into a single program from a single computer system or network.

- 34 the databases would time out, throwing more fatal errors. When these "death spirals" occurred, the affected website could become unavailable for an hour or more.
As the number of Amazon's customers grew, such crashes became more common.
By 2002 Amazon had moved six services out of Obidos: By reducing the
load on Obidos it was able to handle a higher volume of web traffic. Although
Amazon's chief technology officer declared in 2005 that the Obidos architecture
had reached its "end of life" in 2001, 98% of Amazon's retail webpages still ran
on Obidos as of January 2005. By May of that year there was real concern that
Amazon's website would not be able to scale through the holiday season if it continued to run on Obidos. Great effort was expended to move a significant amount
of traffic off Obidos and onto a service-oriented architecture known as Gurupa.
Amazon transitioned rapidly away from Obidos during 2005 and ceased using it
altogether on August 31, 2006.
Amazon began its shift toward a service-oriented architecture around 2002.
"Service-oriented architecture" describes software designed as a set of small, independent programs or "services"; the services work together to perform complex
tasks. In this model, each service needs to understand how to interact only with
those other functions with which it has actual contact. This differentiates it from a
monolithic model, where each function must understand how to interact with

- 35 every other function in a larger program. Amazon did not invent service-oriented
architecture, which was a well-known concept in the computer industry.
Amazon built this new architecture on a Gurupa engine, which received requests from a web server, sent the requests to services or applications, and returned responses to the web server. The engine did not create webpage content;
instead, it sent requests to applications that, in coordination with services, created
and returned webpage content to the server. The role of the Gurupa engine, in
effect, was to stitch together dynamic webpages.
The transition from Obidos to Gurupa was an architectural shift from a
monolithic to a distributed system and involved rewriting software in different
programming languages. Amazon decided to abandon catsubst for building webpages and shifted to a programming language known as Perl and a templating language known as Mason (the duo was called "Perl/Mason"). Amazon could reuse
very little of the Obidos software when it moved to Gurupa. And many concepts
that underlay the monolithic architecture were not applicable to a distributed
system. Amazon displayed its first Perl/Mason page in August 2005, but very few
European webpages were generated using Perl/Mason and Gurupa until 2006.
Problems with the new architecture soon emerged. Webpages created in
Perl/Mason and rendered using the Gurupa engine took longer to appear in a

- 36 user's browser than comparable Obidos pages. Gurupa's throughput--the number
of webpages that could be displayed per second--was less than a tenth of Obidos'
in August 2005. One reason was that Gurupa at the time did not support parallel
rendering or "streaming." Rather than loading webpages sequentially, streaming
enables different parts of a webpage to be loaded simultaneously, permitting the
customer to view part of the page while the rest is being delivered to his or her
web browser. Streaming capabilities were increasingly prevalent on Amazon's
competitors' websites, and their absence at Amazon negatively affected its
customers' shopping experience.
Another problem was that the services called by the Gurupa engine were
developed in "silos" rather than as an integrated system. In certain situations,
these services suffered from "circular dependencies," that is, links that eventually
connect code back to itself, which can cause the software to crash after a code
change. Because of lack of coordination between service "silos," code was often
written that duplicated existing functionality. These structural problems increased
as Amazon created new services and expanded its website capabilities.
Because the Gurupa architecture was a set of disconnected silos rather than
a technology platform, many of Amazon's engineers regarded it as "dramatically
unreliable from the first day." Amazon found it increasingly difficult to hire pro-

- 37 grammers eager to work in this environment: Typically trained in Java and C++,
they came to regard Perl/Mason as a "dead language." In October 2009 Gurupa
instances running the U.S. website were expected to crash with unacceptable frequency if not killed and restarted.
In September 2006 Amazon hired Brian Valentine to envision and lead the
development of a new eCommerce platform. He regarded Gurupa as an unreliable
application and put it on "life support" shortly after he arrived. He testified that
his unit had to devote most of its efforts to "keeping the lights on"--that is, keeping Amazon's websites from crashing on Gurupa--rather than pursuing innovative
projects. By 2012 Amazon had migrated most of its website platform from Gurupa to Santana, but it was still using Gurupa for some applications as late as No-

vember 2014.
Mr. Valentine's goal was to create a true technology platform, with a central
and standardized set of services and a uniform interface for interacting with those
services. A key feature of this platform was Santana, a Java-based web hosting
service and rendering engine that addressed Gurupa's performance and stability
problems. Santana was first deployed on an Amazon website in 2007; by 2008
Amazon had chosen Santana as its preferred architecture for future software development projects. This shift to Santana required Amazon to rewrite virtually all of

- 38 its eCommerce services, and the new platform thus took much longer to build and
implement than Amazon had expected.
The new platform was vastly superior to its predecessor. Java was more
efficient and performance oriented than Perl/Mason, and it was a modern programming language that made hiring easier. Java reduced the parallel rendering
problem that had plagued Gurupa, and it significantly reduced latency. Java was
superior for streaming digital content and, in conjunction with Santana, it provided
much better throughput.

C.

Evolution of Amazon's Software Applications

Corresponding to these major shifts in Amazon's software architecture, the
applications or "services" that performed specific business functions underwent
constant change. The shift to Santana required almost everything from Gurupa to
be rewritten or thrown out. Representative examples are discussed below.
1.

Customer Master Service

Launched around 2000, Customer Master Service (CMS) was the first service launched in Amazon's move toward a service-oriented architecture. CMS
initially had limited ambitions, simply allowing customers to change their user
names or passwords. As a result of constant revision its functionality was vastly
expanded. CMS eventually functioned as a "broker" between the database storing

- 39 Amazon's customer information and the various retail segments (e.g., service centers, fulfillment centers, and the website) that needed this customer information.
By fall 2004 CMS had grown very large, complex, and fragile as Amazon's
evolving business requirements necessitated constant changes to the software.
CMS became a bottleneck and experienced more than 50 hours of outages during
the December 2004 holiday season. To address these problems, Amazon embarked on a complete re-architecture of CMS during 2005 and 2006. At the time
of trial, CMS no longer existed, having been replaced by new software known as
Identity Service.
2.

Order Master Service

"Ordering" includes the steps whereby a customer places an order, payment
is authorized, a shipment request is prepared, and the order is handed off for fulfillment. This requires interactions with many other parts of Amazon's software,
including the systems managing fraud prevention, payment, sales tax collection,
and fulfillment. Amazon developed Order Master Service (OMS) around 2000 to
handle key aspects of shopping carts, orders, shipments, and related business logic
and messaging. Although it was built outside of Obidos, it resembled Obidos in
being a large monolithic program that was difficult to support.

- 40 By spring 2004 OMS was causing bottlenecks that generated fatal errors in
1% of all customer sessions, frustrating customers and generating other problems.
Amazon's engineers referred to OMS as "the new Obidos" because it "created
long lead times for bug fixes and new features and needlessly tied together unrelated product launches." Amazon concluded that OMS needed to be substantially
rewritten to simplify it and decouple it from the software handling payments, promotion, shopping cart, and fulfillment.

Beginning in 2006 Amazon's engineers "refactored" OMS into several
different services.9 The software was rewritten in a way that allowed multiple
teams of programmers--e.g., those writing software for payments and sales tax
collection--to work in parallel, thereby improving efficiency and avoiding timeconsuming collaboration. By January 1, 2010, there was virtually nothing left of
the OMS software that existed in 2005.
3.

Dynamo

An important component of Amazon's ordering technology is its muchimitated "shopping cart," which allows customers to save items they intend to
purchase in a single location while they continue to shop. Amazon initially stored
9"Refactoring" means changing the internal structure of source code to
improve its efficiency--e.g., by removing dead code and unwanted dependencies-without changing its essential function or external behavior.

- 41 shopping cart information in an Oracle database. By the end of 2004 Amazon's
overloaded shopping carts were driving the Oracle databases to operate beyond the
limits of their capability, causing multi-day outages in 2004. In early 2005 Amazon's shopping cart had "low availability," which meant that customers could put
items in their carts, experience a session crash, and return to find their carts empty.
During 2005-2006 Amazon's engineers developed a high-availability system
called "Dynamo" to replace the Oracle databases. Dynamo was fully implemented
during 2007 and lasted about five years before it needed to be replaced.
4.

Payments

Amazon's Payments software manages and secures payment transactions
with both retail customers and merchants. By late 2004 this software was experiencing stress. To keep Payments running during the 2004 holiday season, Amazon's engineers had to work around the clock, sleeping in hotels close to its datacenters so they could be available in the event of an emergency. By 2005, simply
adding a new payment method required changing multiple services and consumed
several months of an engineer's time. Amazon's engineers described the state of
the Payments system as "pretty dire" in 2006, when slow transaction processing
caused it to lose a significant amount of revenues. Concluding that small-scale
revisions would not do the trick, Amazon decided (in the words of Distinguished

- 42 Engineer Vosshal) to "declare bankruptcy on Payments and just start over." The
engineering team completely rewrote this software between 2006 and 2011, replacing all but "a few lines of code that may persist in the dark corners."
5.

Item Master

Amazon stores information about products offered on its websites in databases referred to as "the catalog." In 2001 Amazon began developing software
called Item Master as a single repository for product information. This software
allowed Merchants and suppliers to add items to Amazon's catalog and maintain
current, accurate descriptions of these products and their attributes.
Item Master went through three major revisions between 2001 and 2010.
By 2004 the software (then in its second version) was encountering scale limitations and lacked functionality that Amazon's expanding business required. For
example, Item Master was unable to identify situations where items listed by different sellers were in fact the same item; this damaged the customer experience by
returning multiple search results in response to a query for a single product. To
address these and other challenges, Amazon in 2006 began developing a third
version of the software, Item Master v3. Item Master v3 rebuilt the process for
listing and describing items in the catalog and entailed a substantial rewrite of

- 43 Item Master v2. Item Master v3 was deployed in 2007 and ran concurrently with
Item Master v2 until 2010, when the latter was shut down.
6.

Personalization

Amazon's personalization technology includes "Recommendations" and
"Similarities." The "Recommendations" software tracks information about a customer and suggests products he or she might wish to buy on the basis of prior purchases. The "Similarities" software suggests items that resemble or are compatible
with an item for which the customer is shopping (e.g., "customers who bought this
item also bought item X," or "this item and Item Y are frequently bought together").

Personalization poses two sets of problems. One set involves the algorithmic (or mathematical) challenges posed by a constant effort to refine the accuracy
of items suggested as "similar" or "recommended." The second set of problems
involves scaling: As the number of customers, prior purchases, and products for
sale grows geometrically, the number of links between these variables, and the
data that must be searched to make "recommendations" and propose "similarities,"
grows almost exponentially.
As of January 2005, Amazon's Similarities software was scaling very poorly
and faced significant challenges related to latency. The technology could not keep

- 44 up with the rapid growth of available data; as it took longer to update data, similarities could not be detected using the most recent data. This inability to scale
reduced the technology's value.
During 2004-2005 Amazon tried to build a new Similarities engine with in-

creased scaling capacity, but that effort failed. During 2005-2006, Amazon built
and launched a third Similarities engine that engineers described as "a complete
rethinking of the problem" and "a total game changer." By employing "vastly
more efficient algorithms," the new Similarities engine achieved speeds 40 times
faster than its predecessor's. It significantly reduced the time needed to assemble
and update Similarities data.
Amazon's Recommendations software likewise faced algorithmic, scale,
and latency challenges. Because Amazon did not know which customers would
visit when, the software had to process large volumes of data very rapidly in real
time. Its algorithms had to select from and account for a wide variety of inputs in
determining exactly which items would be recommended; for example, a purchase
of patio furniture years ago would not yield useful recommendations for a customer now shopping for a computer. In an effort to ameliorate latency problems,
Amazon rewrote Recommendations between May 2005 and May 2006, reducing
its 60,000 lines of code to fewer than 15,000 lines. This redesign significantly

- 45 improved performance. Because of "substantial changes across every major component" of the Personalization technology, Amazon's engineers estimated that the
contribution of that software, as it existed in January 2005, was "no longer material" within six years.
7.

Messaging

Once Amazon moved toward a service-oriented architecture, it required
"messaging" software to enable the technological components underlying its website to communicate with each other in a unified way. In 2003 Amazon was using
commercially available publish-subscribe (or "pub-sub") messaging software supplied by TIBCO, a third party. TIBCO's messaging service at the time was stateof-the-art, but it was being pushed to its limits by the uniquely heavy performance
demands that Amazon was placing upon it.
During the 2003 holiday season, failures in the TIBCO messaging software
caused outages that could last hours at a time. In early 2004 Amazon's engineers
expressed concern that introduction of new services could bring down Amazon's
website because of the problems associated with TIBCO messaging software. Cisco Systems, which supplied Amazon with routers and other networking equipment, informed the company that its routers, in conjunction with the TIBCO soft-

ware, would not support the "scaling up" required for the 2004 holiday season.

- 46 In October 2004 Amazon stopped using TIBCO for its "biggest use cases."
It then explored developing its own pub-sub software to overcome TIBCO's performance and scaling problems. Amazon ultimately created and deployed the
Amazon Messaging Platform (AMP) in late 2007.
D.

New Products and Services After January 1, 2005

Many Amazon products and services familiar to consumers today did not
exist in January 2005. These included Kindle, Amazon Prime, the Fire smartphone, Fire TV, Amazon's digital music and video offerings, cloud computing,
and cloud storage services. Though some of these products and services were in
development during 2005, none generated revenues until later.
Kindle, Amazon's eBook reader, enables customers to view digital books,
newspapers, magazines, and blogs directly on the device. Kindle was prototyped
in May 2005 and launched in the United States in November 2007. In October
2009, AEHT launched Kindle 2, a subsequent version of the device, in France,
Germany, and the UK.
Amazon Prime is a membership program that enables customers to receive
free one-or two-day shipping and discounted overnight shipping along with other
benefits. Amazon introduced Prime in the United States in 2005 and in Europe

between 2007 and 2008. The Prime program required major changes to the tech-

- 47 nology used to operate Amazon's fulfillment centers, especially during peak shopping periods. Items destined for Prime customers had to be picked, sorted, packaged, and shipped in a different manner from other items in order to ensure the
expedited delivery that Amazon promised.
Amazon introduced Amazon Unbox, a digital video download service, in
September 2006, and it introduced Amazon MP3 in the United States in Septem-

ber 2007. AEHT launched MP3 in Europe the following year. Digital media has
become an increasingly important part of Amazon's business.
Amazon Web Services (AWS), the company's cloud-computing business,
was largely developed after January 2005. It was launched in the United States in
2006. It has grown significantly in terms of employee headcount, number of
servers required to operate the business, and revenues generated.
III.

Amazon's Third-Party Businesses
A.

Merchants.com

In its M.com business, Amazon used the technology that powered its own
retail websites to build and operate eCommerce websites for other merchants.
Amazon's principal M.com clients were large retailers doing business in the
United States and abroad; there were no material differences between the technology "packages" that domestic and foreign clients received. Amazon's major

- 48 clients included Target in the United States, Marks & Spencer in the UK, Mothercare in the UK, and Sears Canada. Amazon's contract with Target went into effect
in 2001 and was initially set to expire in 2006; it was later extended several times
via amendments. Amazon eventually decided to terminate the M.com program
and shut it down altogether after 2010.
M.com retailers sold their own products on their own branded websites; the
third-party retailer was the seller of record and owned all the inventory. Although
Amazon built and operated these sites, it remained "behind the scenes" as far as
the retailer's customers were concerned. The M.com sites employed the retailer's
web address, URL, trade name, and trademarks; M.com clients had no right to use
Amazon's marketing intangibles and received no information about Amazon's
own retail customers.
The clients obtained, as part of the M.com "package," the complete suite of
Amazon's website technology, including the customer service, fulfillment, and related software that Amazon itself used. M.com clients received all of the software
updates and upgrades that benefited Amazon's own websites. There was no extra
charge for these upgrades; they were included within the basic deal structure regardless of whether Amazon's contract with the client included an explicit "feature parity" clause.

- 49 For additional fees, M.com clients could also arrange to have Amazon perform certain services for them, using Amazon's own facilities. These services
could include fulfillment, transportation, customer service, payment processing,
fraud prevention, and/or marketing. Not all M.com clients chose to avail themselves of these additional services. In some cases Amazon agreed to develop for
M.com clients (again for additional fees) specific technologies that Amazon did
not use in its own business.
B.

Associates and Syndicated Stores

Through the M.com program, Amazon functioned essentially as a supplier
of technology to other retailers. Through the Associates and Syndicated Stores
programs, Amazon entered into relationships with third parties with the goal of enhancing its own customer base. In both of the latter programs, Amazon paid referral fees to the third party when its customers or website visitors made purchases
from Amazon. Both programs were in operation in 2005-2006; only the Associates program continues to operate today.
In the Associates program, the third party or "associate" includes on its own
website a link to Amazon's site. When a visitor to the associate's website "clicks
through" to Amazon and makes a purchase, the associate earns a referral fee. Associates could be online retailers themselves, but more typically they maintain

- 50 content-specific sites or blogs that offer product reviews or similar information.
Some have as their principal goal capturing web traffic and linking their viewers
to Amazon and other sellers in order to earn commissions. Generally speaking,
the associate earns a commission on any purchase the customer makes from Amazon within 24 hours of "clicking through" to Amazon's site.
Through Syndicated Stores, Amazon used its eCommerce technology to sell
Amazon products through a third-party retailer's website (called the "mirror site").
In this scenario Amazon was the merchant making the sale; it kept the associated
retail markup and paid a referral fee to the Syndicated Stores partner. Although
Amazon derived benefits from Syndicated Stores apart from customer referral-because the partner was usually a web retailer, the program eliminated some competition--the fees Amazon paid were essentially payments for customer referrals.
Notwithstanding their differences, petitioner viewed the Associates and
Syndicated Stores programs similarly: The primary purpose of both was to drive
customers to Amazon. It signed agreements with about 20 Syndicated Stores partners (including European partners) before discontinuing the program. Whereas
these agreements were individually negotiated, agreements in the Associates program were not. The process by which an associate joined the program was sub-

- 51 stantially automated and the compensation terms were fixed and largely nonnegotiable.
The stated commission rates under both programs depended on product mix
and sales volumes. Commission rates generally ranged from 4% to 8% in the
Associates program and from 4% to 6% in the Syndicated Stores program. Referral fees for Syndicated Stores partners generally had a per-unit cap; this meant that
the effective commission rate could be lower than the nominal rate reflected in the
contract. Amazon expected that most people referred to it under these programs
would be converted into Amazon customers, on whose subsequent purchases no
referral fees would be due. It was accordingly willing to pay relatively high upfront commissions for customer referrals. The average referral fee Amazon actually paid under the Associates and Syndicated Stores programs was approximately
5.9% of referred sales.¹°

IV.

The Buy-In Payment
Petitioner knew that it had a duty to report the Project Golderest transac-

tions on its Federal income tax returns. To assist it in discharging this responsi¹°In July 2001 Amazon entered into a Syndicated Stores agreement with
Waterstone's, a UK retailer. It provided for a 5% referral fee, which could increase to 6% if certain aggregate sales thresholds were met. This agreement had
an unusual feature whereby Amazon would also pay Waterstone's a one-time
bounty of £7 for new purchases by certain customers.

- 52 bility, it hired Deloitte LLP to calculate the required buy-in payment. In order to
do this, Deloitte needed multi-year financial projections for the European business. Given the unpredictability of eCommerce revenue growth, Amazon for internal budgeting purposes did not make financial projections more than 12 or 18
months out. It assigned its tax department the task of creating longer term
projections for this occasion.
The tax department started with Amazon's historical financial data and the
most recent income-statement forecast, which included projections for the next 18
months. In consultation with business and finance personnel, the team forecast
growth rates for revenue, gross margins, operating expenses, and operating margins for the U.S. and the European businesses. These projections covered calendar

years 2005 through 2010.
In 2006 Deloitte supplied Amazon with a "Transfer Pricing Documentation
Report" to calculate the required buy-in payment. Deloitte determined that the
best method for calculating the buy-in price was "an unspecified income-based
method." See sec. 1.482-4(a)(4), Income Tax Regs. Although the method Deloitte employed was not "specified" in the regulations, Deloitte regarded it as similar in many respects to the specified residual profit split method. See sec. 1.482-6,
Income Tax Regs.

- 53 Deloitte determined that the intangible assets Amazon US transferred to
AEHT had a seven-year useful life. Relying on Amazon's 2005-2010 projections
(which Deloitte extrapolated to 2011), Deloitte determined the future income
streams of AEHT reasonably attributable to these assets, then allocated those income streams between pre-existing and subsequently-developed intangible property. Deloitte determined that the appropriate buy-in price was $254.5 million, to be
paid over a seven-year period commencing in 2005.
At trial petitioner supported its position with respect to the buy-in payment
principally on the basis of the comparable uncontrolled transaction (CUT) method.
S_e_e sec. 1.482-4(c), Income Tax Regs. It contended that each species of intangible property--the website technology, the marketing intangibles, and the European customer information--had to be valued separately. For each species of property, Amazon submitted expert reports that estimated the property's useful life and
valued it on the basis of available CUTs.
Respondent contended that the best method for determining an arm's-length
buy-in payment was the discounted cash flow (DCF) methodology employed by
Dr. Frisch. In the event the Court rejects that methodology, respondent submitted
expert reports that employed a CUT methodology. For each species of property,

- 54 respondent's experts supported values substantially higher than those determined
by Amazon's experts.
For the website technology, petitioner's experts derived a CUT by reference
to the prices Amazon charged its M.com clients for the technology needed to run
those clients' eCommerce websites. Petitioner's expert John Wills testified as to
his belief that Amazon offered these M.com customers its full suite of website
technologies together with all necessary services. He opined that the M.com
transactions thus furnished appropriate internal CUTs for the technology that
Amazon US made available to AEHT.
Four of petitioner's other technology experts--Ken Birman and Lorenzo
Alvisi, David Parkes, and Alan MacCormack--employed various approaches to
ascertain the useful life of Amazon's website technology and the rate at which it
would "ramp down" or decay in its utility. Applying these useful life conclusions
to the pricing data derived from the M.com transactions, Dr. Wills concluded that
the value of the website technology transferred by Amazon US to AEHT ranged
between $117 million and $182 million.
Respondent's expert Harlow Higinbotham agreed that the CUT method
could be used to value the pre-existing technology and that the M.com transactions
supplied a reliable source of CUTs. However, he concluded that Amazon's web-

- 55 site technology had an indefinite useful life. On the basis of his useful life determinations, Dr. Higinbotham valued the website technology transferred by Amazon

US to AEHT at $3.34 billion.
For the marketing intangibles, Robert Reilly (petitioner's expert) and David
Haigh (respondent's expert) both used an external CUT methodology to determine
an arm's-length buy-in price. In selecting their CUTs, both experts also relied on
the same sources of public information. But the two experts came to disparate
value determinations, chiefly because of very different conclusions as to the useful
life of the transferred property and the proper royalty rate to apply over the property's useful life. Mr. Reilly concluded that the arm's-length value of the marketing
intangibles ranged from $251 million to $312 million; Mr. Haigh determined a
value of $3.13 billion for the same intangible property.
The customer information that Amazon US transferred to AEHT consisted
of data about European retail customers who had transacted with the European
Subsidiaries before May 1, 2006. These data included names, email addresses,
phone numbers, purchasing history, and credit card information. Amazon viewed
this customer information as having a short useful life: People change their phone
numbers and email addresses often, and their buying habits change significantly

- 56 over time. For its Similarities software, Amazon uses only relatively recent data
because it regards older data as having little or no value.
Given the relatively short useful life of the customer information, Amazon's
experts regarded Amazon US as having in essence "referred" its European customers to AEHT, which then benefited from having a base of inherited customers
when it began operations on May 1, 2006. Dr. Wills accordingly used as CUTs
the referral fees that Amazon paid its business partners in the Associates and Syndicated Stores programs.
Two of petitioner's other experts--Wendy Moe and Robert Wentland-performed analyses that estimated future purchases by the referred European
customers and the rate at which these individuals would convert to direct customers of AEHT. Using these estimates, Dr. Wills applied a referral fee of 5.9%-the average referral fee Amazon itself paid to its Associates and Syndicated Stores
partners--for all purchases by customers deemed to arrive at AEHT's websites by
referral. Dr. Wills assumed that AEHT would pay referral fees for only six years,
and he discounted the resulting revenue stream at 18%. This generated a buy-in
payment of $52 million for the customer information that Amazon US made
available to AEHT. Dr. Wills determined that this value would rise to $66 million
if AEHT paid referral fees for 10 years.

- 57 Dr. Higinbotham agreed that the commissions Amazon paid third parties for
customer referrals supplied appropriate CUTs. However, he gave particular
weight to Amazon's agreement with Waterstone's, a Syndicated Stores partner in
the UK. It provided not only for referral fees ranging from 5% to 6%, but also for
a one-time bounty of £7 for new purchases by certain customers. Using these
parameters, Dr. Higinbotham valued the customer information at $215 million.
V.

Cost Sharing Payments
The CSA required AEHT to make annual cost sharing payments to compen-

sate Amazon US for ongoing intangible development costs (IDCs), to the extent
those IDCs (as determined by a revenue ratio) benefited the Luxembourg headquarters. See sec. 1.482-7(a)(1), (d)(1), Income Tax Regs. Virtually all technological innovation occurred within Amazon US. Thus, the larger the volume of
IDCs that Amazon US is treated as having incurred, the larger the cost sharing
payments that AEHT was required to make.
The regulations define IDCs and provide that costs that contribute both to
intangible development activity and to other business activities must be allocated
"on a reasonable basis." See 4 para. (d)(1). Petitioner's cost accounting system
during 2005-2006 did not specifically segregate IDCs or R&D expenses from
other operating costs. Petitioner therefore developed a formula and applied it to

- 58 allocate to IDCs a portion of the costs accumulated in various "cost centers" under
its method of accounting.
"Cost centers" are accounting classifications that enable a business to manage and measure operating expenses. Petitioner tracked expenses in six high-level
cost centers: (1) Cost of Sales, (2) Fulfillment, (3) Marketing, (4) Technology and
Content (T&C), (5) General and Administrative (G&A), and (6) Other. Each of
these high-level cost centers is a "rollup" of numerous subsidiary cost centers. For
some calendar quarters, more than 200 individual cost centers, each recording a
specific type of expense, "rolled up" into intermediate cost centers and ultimately
into one of the six top-level cost centers. For example, cost center 7710, "Systems
and Network Engineering," rolls up into C210 ("Product Development") and C250
("Technology/External"). All costs accumulated in "Product Development" and
"Technology/External" roll up into the T&C category.
The parties agree that none of the costs accumulated in the "Cost of Sales"
and "Other" categories are allocable to IDCs. Respondent accepts petitioner's
formula-based allocation to IDCs of costs accumulated in the "Fulfillment" and
"Marketing" categories, and he accepts petitioner's decision to allocate G&A costs
to IDCs on the basis of the IDC outcomes for the other five categories.

- 59 The parties' dispute focuses on the T&C category. Respondent contends
that 100% of the costs accumulated in the T&C category constitute IDCs; as a
corollary, this would produce a commensurate increase in the percentage of G&A
costs allocable to IDCs. Petitioner urges that T&C costs must be allocated between IDCs and other activities "on a reasonable basis" and that its allocation
formula accomplishes this result.
Broadly speaking, the costs accumulated in the T&C category include expenses related to technological development and website content. According to
petitioner's SEC filings, its T&C category expenses "consist principally of payroll
and related expenses for employees involved in research and development, including application development, editorial content, merchandising selection, systems
and telecommunications support, and costs associated with the systems and telecommunications infrastructure." T&C costs included costs associated with acquired website content; payroll and related expenses for employees involved in
research, website development, and telecommunications support; and payroll and
related expenses for employees involved in category expansion (i.e., expanding
Amazon's product offerings) and buying.
During 2005-2006 many employees whose time was captured in T&C cost
centers engaged solely in intangible development activity; certain employees en-

- 60 gaged solely in other types of activity; and certain employees engaged in both.
This diversity of tasks was reflected in their job classifications. All Amazon employees have job codes beginning with a capital letter, one of which is "T," which
stands for "Technical." Most cost centers that rolled up into T&C have a mix of
T-coded and non-T-coded workers. During the first quarter of 2006, for example,
35 T&C cost centers had no T-coded employees; 27 had all T-coded employees;
and 69 had a mix. On average during 2005-2006, there were almost as many T&C
cost centers with no T-coded employees (31.4 cost centers) as there were with all
T-coded employees (31.9 cost centers).
The subsidiary cost centers that "rolled up" into T&C captured a significant
volume of non-IDC personnel costs. The diversity of the employees' tasks, which
is reflected in their annual performance evaluations," is illustrated by these examples:
042
Employees in cost center 5155 manage third-party digital content that is
viewed on or downloaded from Amazon.com. Certain employees manage the
"customer purchasing experience." Other employees spend time negotiating and
"The performance evaluations show that certain employees whose time was
captured in T&C cost centers engaged in substantial non-IDC activity. Relevant
cost centers include 7334, New Product Development; 7823, Data Warehouse
Development; 5355, Large Accounts Services Account Management; and 7402,
Payments Platform.

- 61 managing the logistics of acquiring website content from third parties and determining how this website content will be displayed to Amazon's customers.
042
Employees in cost center 5357 build and improve technology that helps
sellers integrate their products into Amazon.com. Some employees spend significant time helping sellers list their products. This includes assisting sellers in
filling out spreadsheets and directing their submissions to other Amazon staff
members.

042
Employees in cost center 7723 design, expand, and maintain Amazon's
product catalog. Some write code to compile and display the catalog or to allow
customers to place orders. Others engage in routine maintenance and make minor
code adjustments to alter the manner in which website content is displayed.
Because petitioner's T&C cost centers, like its Fulfillment, Marketing, and
G&A cost centers, captured both IDCs and other costs, it devised a complex, multi-step formula to allocate costs "between the intangible development area and the
other areas or business activities." See sec. 1.482-7(d)(1), Income Tax Regs. The
details of this formula have varied over time. The formula that petitioner urged at
trial employed data derived from a PricewaterhouseCoopers (PwC) study that
identified qualifying research activities for purposes of claiming section 41 research and experimentation (R&E) credits for 2005 and 2006.

- 62 Because Amazon's employees did not record time specifically to R&E
activities, the PwC study relied heavily on employee surveys. Most employees in
T&C cost centers, regardless ofjob code, were surveyed. Some employees outside
of T&C cost centers were surveyed if petitioner believed they engaged in qualifying research.
In these surveys PwC asked employees to complete questionnaires on which
they divided their time among 14 specified activities. Eight of these activities
involved software development, from the initial "requirements" phase through
deployment and testing, together with direct supervision and direct support of
software development. The time devoted to these eight activities was deemed to
yield qualifying R&E expenses for section 41 purposes. Time devoted to five
types of activities--specifically, routine engineering, routine data collection, reverse engineering, human resources/training, and activities outside the United
States--was deemed not to yield qualifying R&E expenses for section 41 purposes.
The 14th category captured days when no work was done, such as vacation days,
sick days, and holidays.
From these surveys, PwC derived estimates to allocate employee time to
section 41 qualifying activities. For each surveyed employee, PwC computed a
"qualified research expenditure" (QRE) percentage, reflecting the portion of that

- 63 person's time that was spent on qualified research. In the fourth quarter of 2006,
for example, the average QRE percentage for employees in the 7000 series of cost
centers (capturing costs related to technology development, maintenance, and
management) was 61.15%. In that same quarter the average QRE percentage for
employees in the 5000 series of cost centers (capturing costs related to business
line management) was 53.41%. PwC used these percentages in determining the
section 41 credit to which it believed Amazon was entitled.
Noting the similarity between section 41 qualified research expenditures
and IDCs, Amazon employed the PwC survey data as a central component of its
formula for allocating costs under section 1.482-7(d)(1), Income Tax Regs. Simplifying somewhat, petitioner's formula for allocating T&C category costs between IDCs and other activities proceeds in several steps. As the first step, petitioner eliminated from the T&C cost centers all costs captured in 26 general ledger
accounts that petitioner determined to be unrelated to intangible development.
The resulting sum may be called "modified T&C category costs."
As the second step, petitioner identified the employees within the T&C category who were likely to have engaged in intangible development. Assuming that
only T-coded employees were likely to have done this, petitioner divided the number of such employees by the total number of employees in the T&C category.

- 64 This yielded what petitioner called the "T-ratio." Petitioner calculated a distinct
T-ratio for the T&C category for each calendar quarter during 2005-2006.
The next step was to examine the QRE survey results. Petitioner adjusted
the PwC data to reflect the fact that certain costs ineligible for the section 41 credit
(e.g., costs attributable to reverse engineering and non-U.S. activities) may properly be includible in IDCs. Petitioner accordingly determined an "adjusted QRE
percentage" for each person in the T&C category, representing the portion of that
person's time spent on intangible development. Petitioner computed the arithmetic average of these "adjusted QRE percentages," which it called the "adjusted
QRE ratio" or "A-ratio."¹² Petitioner then multiplied the T-ratio by the A-ratio to

yield a "development ratio" for the T&C category. Finally, petitioner multiplied
"modified T&C category costs" (as determined at step 1) by the "development
ratio" to determine the dollar volume of T&C category costs properly allocable to
IDCs. Petitioner made separate computations for each calendar quarter and
summed these results to produce annual IDC figures for the T&C category.¹³
¹²If there was no PwC survey data for a particular T&C cost center--e.g., because that cost center was recently created or because it was determined not to
contribute to section 41 qualifying research--petitioner applied the average A-ratio
for cost centers with available survey data.
¹³Petitioner employed a similar methodology to determine the percentage of
(continued...)

- 65 On its 2005 and 2006 Federal income tax returns, petitioner reported cost
sharing payments from AEHT of $116,092,584 and $77,297,000, respectively.
(These amounts were reported as reimbursed R&E expenses, thus reducing otherwise-allowable deductions.) The cost sharing payment for 2006 was determined
using the PwC survey data as described previously. Because the PwC data were
not available when petitioner filed its 2005 return, it initially used a different system to determine the 2005 cost sharing payment. It subsequently filed an affirmative claim for 2005, reporting a lower cost sharing payment computed under the
methodology used for 2006. By this affirmative claim, petitioner sought to reduce
the 2005 cost sharing payment by approximately $59 million, or almost 50%.
VI.

Stock-Based Compensation
The CSA executed by Amazon US and AEHT defined IDCs to "include all

direct and indirect costs (including Stock-Based Compensation Costs)" relating to
intangible development. Specifically, IDCs were defined to include "compensation provided by a Party to its employees or independent contractors in the form of
equity instruments, options to acquire stock, or rights with respect to * * * equity

¹³(...continued)
G&A category costs allocable to IDCs. Respondent does not object to the G&A
methodology, except to the extent that it employs what respondent views as an
unduly low allocation of IDCs to the T&C category.

- 66 instruments or stock options as defined in Treasury Regulation § 1.482-7(d)(2)(i)
(as amended by T.D. 9088)." The parties further elected, pursuant to section
1.482-7(d)(2)(iii)(B), Income Tax Regs., to take into account "all stock-based
compensation in the form of stock options in the same amount, and as of the same
time, as the fair value of the stock options reflected as a charge against income in
the audited financial statements of a Party." This election was made "without prejudice to the Party's right to challenge the validity of Treasury Regulation § 1.482-

7(d)(2)."
In filing its 2005 and 2006 returns, petitioner thus complied with the reg-

ulation requiring that stock-based compensation be included in the IDC "cost
pool" upon which cost sharing payments are determined. Like many technology
companies, petitioner questioned the validity of this regulation. The CSA accordingly included a "clawback" provision that will apply in the event section

1.482-7(d)(2), Income Tax Regs., is
held to be an invalid regulation by a final decision in a court of law
with respect to pending litigation involving another taxpayer
including a U.S. Supreme Court decision, U.S. Court of Appeals
decision upon denial of a writ of certiorari or lapse of time for filing
such writ, or a decision by a federal trial court upon lapse of time for
filing a notice of appeal, or * * * [is] revised or withdrawn by the
Treasury Department such that the costs of stock-based compensation
are not required to be included as costs for qualified cost sharing
arrangements.

- 67 In the event this regulation is ultimately invalidated or withdrawn, the CSA
provides that "stock-based compensation shall not be included in the determination of * * * [IDCs] in any Year to which this Agreement applies." For any year in
which stock-based compensation turns out to have been "improperly" included in
IDCs, "the Cost Share shall be recomputed without the inclusion of stock-based
compensation in * * * [IDCs]," and "the Cost Share less the Recomputed Cost
Share shall be refunded * * * [to the proper party]." The CSA provides that any
such refund shall be "treated as an adjustment to the Cost Share for the Year in
which the Triggering Event occurs * * *, and to the extent that such adjustment
exceeds the Cost Share, the adjustment shall be applied to subsequent Years until
fully exhausted."
In Altera Corp. v. Commissioner, 145 T.C. 91 (2015), this Court invalidated
section 1.482-7(d)(2), Income Tax Regs., the provision that requires stock-based
compensation costs to be included in the IDC pool. Our decision in that case was
appealed to the U.S. Court of Appeals for the Ninth Circuit on February 19, 2016.
The case remains pending on appeal.

OPINION
Section 482 gives the Commissioner broad authority to allocate gross income and deductions among commonly controlled entities if he determines that it

- 68 is necessary "to prevent evasion of taxes or clearly to reflect the income." The
purpose of section 482 is prevent artificial shifting of income by placing a controlled taxpayer on a tax parity with an uncontrolled taxpayer. Sec. 1.482-1(a)(1),
Income Tax Regs. The statute empowers the Commissioner to determine the "true
taxable income" of a controlled taxpayer by ascertaining the income it would have
earned if it had dealt with unrelated parties at arm's length. See Commissioner v.

First Sec. Bank, 405 U.S. 394, 400 (1972); Seagate Tech., Inc. & Consol. Subs. v.
Commissioner, 102 T.C. 149, 163-164 (1994). "In determining the true taxable
income, 'the standard to be applied in every case is that of a taxpayer dealing at
arm's length with an uncontrolled taxpayer.'" Veritas, 133 T.C. at 317 (quoting
section 1.482-1(b)(1), Income Tax Regs.).
The Commissioner has broad discretion in applying section 482, and we
will uphold his determination unless the taxpayer shows it to be arbitrary, capricious, or unreasonable. Seagate Tech., 102 T.C. at 164; Sundstrand Corp. v. Commissioner, 96 T.C. 226, 353 (1991). Whether respondent has abused his discretion

is a question of fact. Sundstrand Corp., 96 T.C. at 353-354; Am. Terrazzo Strip
Co. v. Commissioner, 56 T.C. 961, 971 (1971).
In a series of transactions in 2005 and 2006, Amazon US transferred intangible property to AEHT. These transfers required AEHT to make an upfront buy-

- 69 in payment to compensate Amazon US for the value of the assets thus transferred.
S_e_e sec. 1.482-7(a)(2), (g)(2), Income Tax Regs. Respondent urges that a DCF
methodology, as implemented by Dr. Frisch, supplies the best method for
determining an arm's-length buy-in payment, and that the required payment is

$3.468 billion. The first question we must answer is whether respondent abused
his discretion in making this determination. We conclude that he did.
I.

Cost Sharing Background

We begin our analysis with the regulations in effect in 2005 when Amazon
US and AEHT entered into the CSA. Where parties have entered into a qualified
cost sharing arrangement (QCSA), they share the cost of developing intangible
property. See id. para. (a)(1). When one participant (here, Amazon US) makes
pre-existing intangible property available for purposes of research under a QCSA,
that party is deemed to have transferred an interest in such property to the other
participant. This requires the other participant (here, AEHT) to make a "buy-in
payment" to the transferor. R para. (g)(1) and (2).
The required buy-in payment "is the arm's length charge for the use of the
intangible" multiplied by the controlled participant's share of reasonably anticipated benefits. E subpara. (2). The best-method rule, set forth elsewhere in the
regulations, "seeks the most reliable measure of an arm's-length result." Veritas,

- 70 133 T.C. at 327; see sec. 1.482-1(c)(1), Income Tax Regs. The regulations provide that an arm's-length charge must be determined under one of four methods:
(1) the comparable uncontrolled transaction (CUT) method; (2) the "comparable
profits" method; (3) the "profit split" method; or (4) an "unspecified" method. See
sec. 1.482-1(c), Income Tax Regs. "[T]here is no strict priority of methods, and no
method will invariably be considered to be more reliable than others." Veritas, 133
T.C. at 327; sec. 1.482-1(c)(1), Income Tax Regs.
The regulations make clear that the buy-in payment represents compensation solely for the use of pre-existing intangibles. Section 1.482-7(g)(2), Income
Tax Regs., captioned "Pre-existing intangibles," states:
If a controlled participant makes pre-existing intangible property in
which it owns an interest available to other controlled participants for
purposes of research in the intangible development area under a
qualified cost sharing arrangement, then each such other controlled
participant must make a buy-in payment to the owner. * * *
By definition, compensation for subsequently developed intangible property is not
covered by the buy-in payment. Rather, it is covered by future cost sharing payments, whereby each QCSA participant pays its ratable share of ongoing IDCs.
As in effect during 2005-2006, the regulations provided that the Commissioner "shall not make allocations with respect to a qualified cost sharing arrangement" except in two respects. Id. para. (a)(2). Specifically, adjustments are per-

- 71 mitted only to ensure: (1) that an arm's-length buy-in payment is made for preexisting intangible property and (2) that each participant pays its appropriate share
of ongoing IDCs.
This Court interpreted and applied these cost sharing regulations in Veritas,
supra. The taxpayer there, a domestic corporation (Veritas US), developed, manufactured, and sold throughout the world advanced storage-management software
products. Pursuant to a QCSA executed in November 1999, Veritas US granted a
European subsidiary (Veritas Ireland) the right to use pre-existing intangible property overseas. The assets thus transferred consisted of short-lived technology,
including source code for software products to be sold outside the United States,
as well as the right to use outside the United States trademarks, trade names, and
service marks owned by Veritas US. Veritas Ireland made a buy-in payment of
$118 million for use of these pre-existing intangibles.
The IRS challenged the buy-in payment and ultimately determined that the
required buy-in payment was $1.675 billion. Veritas, 133 T.C. at 312. The Commissioner's expert, Dr. Hatch, "assumed that the preexisting intangibles ha[d] a
perpetual useful life" and "characterized the CSA as 'akin' to a sale or geographic
spinoff" of the U.S. parent's international business operations. R at 313. Dr.
Hatch "rejected the comparable uncontrolled transaction method" and "employ[ed]

- 72 a discounted cashflow analysis"--specifically, "the income method"--to determine
the requisite buy-in payment. R at 312, 313.
Dr. Hatch defined the buy-in payment as "the present value of royalty obligations" expected to be paid in perpetuity under arm's-length terms. R at 313.
He did not value individually any of the specific intangible assets that Veritas US
transferred to its Irish subsidiary. Instead, "he employed an 'aggregate' valuation
approach" that proceeded in three steps. First, he estimated the arm's-length
royalty that would be due after November 1999 "on a go-forward basis." M
Second, he chose a discount rate to convert these estimated future royalty payments into November 1999 dollars. Third, he "calculated the buy-in payment as
equal to the present value of the royalty payments estimated in step 1, discounted
at the rate determined in step 2." IbR This produced a buy-in payment of $1.675
billion, which Dr. Hatch determined to be economically equivalent to "a 22.2 percent perpetual annual royalty." Ibid.
We held that the buy-in payment thus determined represented an abuse of
discretion. R at 327. Although we found several deficiencies in Dr. Hatch's
methodology, his core error was to value "short-lived intangibles * * * as if they
have a perpetual life." R at 321. We found that Veritas US "was in a perpetual
mode of innovation" and that the useful life of the pre-existing technology-related

- 73 intangibles was four years. E at 324, 336. We found that the useful life of the
trademarks, brand names, and other marketing intangibles was seven years. R at
338. It was unreasonable, we concluded, for respondent to determine the buy-in
payment by assuming that a third party, acting at arm's length, would pay royalties
in perpetuity for use of these short-lived assets.
As we emphasized in Veritas, the cost sharing regulations "unequivocally
require[] a buy-in payment to be made with respect to transfers of 'pre-existing
intangible property.' No buy-in payment is required for subsequently developed
intangibles." R at 323. By valuing "short-lived intangibles * * * as if they have a
perpetual life," the Commissioner's buy-in computation improperly took into account the value of "intangibles that were subsequently developed rather than preexisting." R at 321. We concluded in Veritas that reliable CUTs existed for each
form of intangible property transferred pursuant to the QCSA and that, with certain adjustments, "the CUT method is the best method for determining the requisite buy-in payment." R at 339.
II.

Respondent's Determination of the Buy-In Payment
In this case respondent's primary position as to the appropriate buy-in pay-

ment was set forth in the expert report of Dr. Frisch. Like Dr. Hatch in Veritas,
Dr. Frisch rejected use of the CUT method and applied a DCF methodology to

- 74 determine AEHT's buy-in obligation. His valuation, like that of Dr. Hatch, proceeded in three main steps. First, he estimated the future cash flows of AEHT's
European business. Second, he selected a discount rate to convert these estimated
cash flows into 2005 dollars. Third, he calculated the buy-in payment as equal to
the present value of the cash flows determined at step 1, discounted at the rate determined in step 2.

For step 1, Dr. Frisch started with the projections by Amazon's management
of revenues, expenses, and operating income for the European business for
calendar years 2005 through 2011.¹4 For years after 2011 Dr. Frisch assumed that
the revenues, expenses, and operating income of the European business would
grow at 3.8% per year, the rate at which the EU economy as a whole was projected
to grow.¹5 With these assumptions, Dr. Frisch projected AEHT's operating income into the indefinite future.

Believing that future cash flows, rather than operating income, would yield
the best estimate of an appropriate buy-in payment, Dr. Frisch made several ad¹4We will refer to these seven-year figures as "projections by Amazon's
management" even though the figures for 2011 were extrapolated by Deloitte.
¹sWe find no fault with this assumption. As will be discussed later, various
experts expressed divergent views as to how AEHT's growth rate should be calculated after 2011. Dr. Frisch's decision to drop down to the long-term projected
growth rate of the European economy was conservative and reasonable.

- 75 justments to AEHT's projected operating income. First, he subtracted the cost
sharing payments that AEHT was projected to make to Amazon US under the
QCSA. Second, he added back expected depreciation because it involves no outlay of cash. Third, he subtracted AEHT's projected future capital expenditures.
Finally, he made adjustments based on AEHT's projected working capital.
After estimating AEHT's future free cash flows, Dr. Frisch moved to step 2
of his analysis: selecting an appropriate discount rate. Generally speaking, the
higher the discount rate, the lower the present value of future cash flows. For internal budgeting purposes, Amazon's treasury department employed a weighted
average cost of capital (WACC) of 13%. Using market data, Dr. Frisch concluded
that a discount rate of 18% was appropriate.
At step 3, Dr. Frisch discounted AEHT's future cash flows through 2024-i.e., for 20 years out--at 18%. He determined that the present value of these cash
flows was $3.067 billion. He then computed a discounted "terminal value" of
$399 million for post-2024 cash flows; this calculation was necessary because he
assumed that the intangible property subject to the buy-in payment had a perpetual
(he called it an "indeterminate") useful life. Finally, to isolate the future cash
flows attributable to AEHT's intangible assets, Dr. Frisch subtracted the value of
the tangible assets it owned as of year-end 2004, which he determined to be neg-

- 76 ative $1.8 million (representing a negative cash position). This yielded a buy-in

payment of $3.468 billion ($3.067 billion + $399 million + $1.8 million = $3.468
billion).
One does not need a Ph.D. in economics to appreciate the essential similarity between the DCF methodology that Dr. Hatch employed in Veritas and the
DCF methodology that Dr. Frisch employed here. Both assumed that the preexisting intangibles transferred under the QCSA had a perpetual useful life; both
determined the buy-in payment by valuing into perpetuity the cash flows supposedly attributable to these pre-existing intangibles; and both in effect treated the
transfer of pre-existing intangibles as economically equivalent to the sale of an entire business. Respondent admitted as much on brief, urging that "Project Goldcrest transferred all cash flows relating to the European operation to AEHT in
what was, as an economic matter, a transfer of the European Websites Business for
the indefinite future."

By assuming a perpetual useful life, Dr. Frisch failed to restrict his valuation
to the "pre-existing intangible property," sec. 1.482-7(g)(2), Income Tax Regs.,
that Amazon US actually transferred to AEHT in 2005. As explained in connection with petitioner's CUT methodology, see infra pp. 103-107, the website
technology that Amazon US initially transferred to AEHT had a useful life of

- 77 about seven years. Thus, after decaying or "ramping down" in value over a sevenyear period, Amazon's website technology as it existed in January 2005 would
have had relatively little value left by year-end 2011. But approximately 58% of
Dr. Frisch's proposed buy-in payment, or roughly $2 billion, is attributable to cash
flows beginning in 2012 and continuing in perpetuity.¹6

Conversely, it is clear that Dr. Frisch, like Dr. Hatch, improperly included in
the buy-in payment the value of "subsequently developed intangibles." Veritas,
133 T.C. at 323. Management's projections for the European business were based
on the (extremely high) growth rates that Amazon had achieved in the past. Such
growth rates could be sustained only through constant innovation, including innovations concerning products and services that did not yet exist in marketable
form.¹7 Projects that would contribute to high future growth rates, some of which
were in early development as of January 2005, included the Kindle, Amazon
¹6As explained more fully infta pp. 117-120, we do not conclude that the
website technology transferred in January 2005 had become wholly worthless at
year-end 2011. It continued to have some residual value; it could be used for
research and as a springboard for developing replacement technology. We
conclude that a "tail" of 3-1/2 years is appropriate to capture this residual value.
But this residual value, when added to the initial seven-year value, does not come
close to Dr. Frisch's $3.468 billion figure.
¹7Dr. Frisch admitted that he did not investigate the manner in which Amazon constructed its revenue projections, testifying that he "couldn't do a detailed
analysis of what was in the projections in terms of individual product."

- 78 Prime, the Fire smartphone, Fire TV, digital music/video offerings, cloud computing and storage, and the EFN, which implemented standardized and improved
fulfillment operations across Europe.
These new products and services, as well as the next generation of Amazon's website platform, would be created thanks to massive projected IDC investments by Amazon in years after 2004. But AEHT would have paid, via cost sharing, its ratable share of these future IDCs, and it would thus co-own these subsequently-developed intangibles. Because AEHT would pay for these assets via cost
sharing, it was not required to pay for them through the upfront buy-in payment.

As we held in Veritas, 133 T.C. at 323: "No buy-in payment is required for subsequently developed intangibles."
Dr. Frisch's DCF methodology resembles Dr. Hatch's in another respect: It
is based in essence on an "akin to a sale" theory. See id. at 313, 316, 320-321. Dr.
Frisch computes the buy-in payment not by valuing the specific intangible assets
transferred under the CSA but by determining an enterprise value for Amazon's

entire European business. He deems the pre-existing intangibles to have a value
equal to AEHT's enterprise value, less its initial tangible assets. By employing an
enterprise valuation, Dr. Frisch necessarily sweeps into his calculation assets that

- 79 were not transferred under the CSA and assets that were not compensable "intangibles" to begin with.
"[F]or the definition of an intangible," the cost sharing regulations in effect
for 2005-2006 refer to the definition set forth in section 1.482-4(b), Income Tax
Regs. See id. sec. 1.482-7(a)(2) (last sentence). This definition is essentially the
same as that set forth in section 936(h)(3)(B). In both provisions, "intangible" is
defined to include five enumerated categories of assets, each of which has "substantial value independent of the services of any individual." These include patents, inventions, copyrights, know-how, trademarks, trade names, and 20 other
specified intangibles. Each definition also includes a sixth category, consisting of
"other similar items" in the regulatory definition and "any similar item" in the statutory definition. The former elaborates slightly on the latter by stating that, "[f]or
purposes of section 482, an item is considered similar to those * * * [enumerated]
if it derives its value not from its physical attributes but from its intellectual content or other intangible properties." Sec. 1.482-4(b)(6), Income Tax Regs.
An enterprise valuation of a business includes many items of value that are
not "intangibles" as defined above. These include workforce in place, going concern value, goodwill, and what trial witnesses described as "growth options" and
corporate "resources" or "opportunities." Unlike the "intangibles" listed in the

- 80 statutory and regulatory definitions, these items cannot be bought and sold independently; they are an inseparable component of an enterprise's residual business
value. These items often do not have "substantial value independent of the services of any individual." Sec. 936(h)(3)(B) (last clause); sec. 1.482-4(b), Income
Tax Regs. And these contributors to value are not "similar" to the enumerated
intangibles because they do not derive their value from their "intellectual content
or other intangible properties." Sec. 1.482-4(b)(6), Income Tax Regs.¹8 Thus, as
we concluded in Veritas, there was for the tax years at issue "no explicit authorization" in the cost sharing regulations for "respondent's 'akin' to a sale theory or
* * * [his] inclusion of workforce in place, goodwill, or going-concern value" in
determining the buy-in payment for pre-existing intangibles. 133 T.C. at 316.¹9
¹ªThis conclusion is supported by the regulations' history. See DHL Corp.

v. Commissioner, 285 F.3d 1210 (9th Cir. 2002), aff'g in part, rev'g in part T.C.
Memo. 1998-461; Austin v. Commissioner, 141 T.C. 551, 561 (2013) ("The
history of a regulation may be helpful in resolving ambiguities in it."). In 1993 the
IRS considered updating the 1968 transfer pricing regulations. Via temporary and
proposed regulations, the Secretary requested comments specifically as to whether
"the definition of intangible property * * * should be expanded to include items
not normally considered to be items of intellectual property, such as work force in
place, goodwill or going concern value." 58 Fed. Reg. 5312 (Jan. 21, 1993).
After receiving numerous comments opposing such an expansion, the IRS decided
not to change the definition of intangible property. See T.D. 8552, 1994-2 C.B. 93
(explaining that the revised regulations merely clarified the 1968 definition).
¹°Even if certain items discussed in the text were thought to constitute "in-

(continued...)

- 81 Respondent urges that Dr. Frisch's adoption of a business enterprise valuation is supported by the "aggregation" principle. Under the caption "Aggregation
of transactions," the regulations in effect during 2006 provided that the combined
effect of multiple transactions may be considered "if such transactions, taken as a
whole, are so interrelated that consideration of multiple transactions is the most reliable means of determining the arm's length consideration for the controlled
transactions." Sec. 1.482-1(f)(2)(i)(A), Income Tax Regs.
We rejected respondent's "aggregation" argument in Veritas, 133 T.C. at
321, and we likewise reject it here. The buy-in payment represents compensation
for "pre-existing intangible property" owned by the transferor. Sec. 1.482-7(g)(2),
Income Tax Regs. For at least two reasons, the type of "aggregation" proposed by
respondent does not yield a reasonable means, much less the most reliable means,
of determining an arm's-length buy-in payment. See id. sec. 1.482-1(f)(2)(i)(A).

"(...continued)
tangibles" for cost sharing purposes, they were not necessarily owned directly by
Amazon US or covered by the CSA. The European Subsidiaries had been in business--in Germany and the UK, very successfully--for more than six years as of
January 1, 2005. They had substantial assets apart from their tangible property, including workforce in place and going-concern value. AEHT inherited these
values when it acquired the stock and assets of the European Subsidiaries pursuant
to the European Subsidiary Contribution, the European Business Contribution, and
the Four-Party Agreement. See supra pp. 25-26. AEHT was not required to
compensate Amazon US for these assets via the buy-in payment.

- 82 First, Dr. Frisch's business-enterprise approach improperly aggregates pre-existing
intangibles (which are subject to the buy-in payment) and subsequently developed
intangibles (which are not). Second, his business-enterprise approach improperly
aggregates compensable "intangibles" (such as software programs and trademarks)
and residual business assets (such as workforce in place and growth options) that
do not constitute "pre-existing intangible property" under the cost sharing regulations in effect during 2005-2006. See Veritas, 133 T.C. at 321-323; see also

Guidant LLC v. Commissioner, 146 T.C. 60, 82-83 (2016).2°
In a related vein, respondent defends Dr. Frisch's approach under the "realistic alternatives" principle. The regulations authorize the Commissioner to "consider the alternatives available to the taxpayer in determining whether the terms of

2°In 2011 the Secretary finalized new cost sharing regulations that replaced
the 1995 regulations involved in this case. T.D. 9568, 2012-12 I.R.B. 499. Issued
in temporary form in 2009 and effective (as relevant here) in January 2009, these
new regulations replaced the buy-in payment with the concept of a "platform
contribution transaction" (PCT). 74 Fed. Reg. 341-342 (Jan. 5, 2009). The
preamble to the temporary regulations noted objections from commenters that the
PCT "included elements such as workforce, goodwill or going concern value, or
business opportunity, which in the commentators' view either do not constitute
intangibles, or are not being transferred, and so, in the commentators' view, are
not compensable." 74 Fed. Reg. 342. The Treasury Department replied by stating
that the new regulations "do not limit platform contributions that must be
compensated * * * to the transfer of intangibles defined in section 936(h)(3)(B)."
IA As we noted in Veritas, 133 T.C. at 315-316, 329-330, the 2009 regulations
did not apply in that case, and they have no application to this case either.

- 83 the controlled transaction would be acceptable to an uncontrolled taxpayer faced
with the same alternatives and operating under comparable circumstances." Sec.
1.482-1(f)(2)(ii)(A), Income Tax Regs. Respondent contends that Amazon US
had a "realistic alternative" available to it, namely, continued ownership of all the
intangibles in the United States. If dealing with an unrelated party, respondent
urges, Amazon US would clearly have preferred this alternative to a cost sharing
arrangement that would give a competitor access to its "crown jewels." Respondent asserts that Dr. Frisch's analysis captures the value of this "realistic alternative" by estimating the future cash flows of Amazon's entire European business.
We find this argument unpersuasive for many reasons, but it suffices to
mention two. First, respondent's argument proves too much. Whenever related
parties enter into a QCSA, they presumably have the "realistic alternative" of not
entering into a QCSA. From this truism respondent concludes that the buy-in payment must be determined as if the parties had not elected cost sharing but had instead continued to operate the business as they had done previously. This would
make the cost sharing election, which the regulations explicitly make available to
taxpayers, altogether meaningless. See 4 sec. 1.482-7(a)(3) (providing that a
QCSA "produces results that are consistent with an arm's length result" provided
that all requirements of the cost sharing regulations are satisfied).

- 84 Second, as w

[Text truncated at 120,000 characters. The full text is on the page linked above.]

---

Source: Frix Law Library, https://www.frixlaw.com/law-library/documents/agency%3Atax-court%3A3399bff9c7547541. Public record. Not legal advice.
