ICT Asset Prices: Marshaling Evidence into New Measures

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. ICT Asset Prices: Marshaling Evidence into New Measures David Byrne and Carol Corrado 2017-016 Please cite this paper as: Byrne, David, and Carol Corrado (2017). “ICT Asset Prices: Marshaling Evidence into New Measures,” Finance and Economics Discussion Series 2017-016. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2017.016r1. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

ICT Asset Prices: Marshaling Evidence into New Measures David Byrne∗ and Carol Corrado†‡ July 10, 2016 (Revised, October 2, 2017) Abstract This paper is a companion to our recent paper, “ICT Services and their Prices: What do they tell us about Productivity and Technology?” It provides the sources and methods used to construct national accounts-style price deflators for the major components of ICT investment—communications equipment, computer equipment, and software—that were analyzed and used in that paper. The ICT equipment measures described herein were also used in Byrne, Fernald, and Reinsdorf (2016). Keywords: Information and Communication Technology (ICT); ICT asset prices; Price measurement. ∗Board of Governors of the Federal Reserve System. This paper reflects the sole opinion of the authors and does not reflectopinionsoftheBoardofGovernorsoftheFederalReserveSystemorothermembersofitsstaff. Theauthorsthank John Fernald and Marshall Reinsdorf for helpful comments. †The Conference Board and Center for Business and Public Policy, McDonough School, Georgetown University. ‡Corresponding author: carol.corrado@tcb.org

Contents 1 Introduction 1 2 Communication Equipment 2 2.1 Equipment components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2.2 Capitalized services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.3 Shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3 Computers and peripheral equipment 9 3.1 Multiuser computers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 3.2 Personal computers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 3.3 Storage, other peripherals, and capitalized services . . . . . . . . . . . . . . . . . . . . . 16 3.4 Shares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 4 Computer Software 18 4.1 Software products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4.2 Custom and own-account software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 5 Conclusion 23 List of Figures 1 Commodity composition of communication equipment . . . . . . . . . . . . . . . . . . . 3 2 Communications equipment investment shares by asset type . . . . . . . . . . . . . . . . 5 3 Commodity composition of computers and periphal equipment . . . . . . . . . . . . . . 9 4 Server price changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 5 Computers and peripheral equipment investment shares by asset type . . . . . . . . . . 14 6 Personal computer price changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 7 BLS software price indexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 8 Software and software components price change . . . . . . . . . . . . . . . . . . . . . . . 23 1

List of Tables 1 ICT Investment Price Change (annual rate) . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 Communication Equipment Investment Price Index and Components . . . . . . . . . . . 4 3 Sources and Methods Used to Estimate Price Change for AV Equipment . . . . . . . . . 6 4 Sources and Methods Used to Estimate Price Change for Wired Telecom Services . . . . 8 5 Computers and Peripherals Investment Price Index and Components . . . . . . . . . . . 10 6 IBM Enterprise Mainframe Prices and Performance Scores . . . . . . . . . . . . . . . . . 13 7 Sources and Methods Used to Estimate Price Change for Computers . . . . . . . . . . . 17 8 Sources and Methods Used to Estimate Price Change for Software Products . . . . . . . 22 i

1 Introduction National accounts-style research price indexes for information and communication technology (ICT) assets have been used and analyzed in recent works (Byrne and Corrado, 2017; Byrne et al., 2016). The sources and methods used to construct these prices are documented and described in this paper. The table below, reproduced from Byrne and Corrado (2017), reports these new price indexes. The primary approach used to construct the ICT price indexes shown below was to take stock of existing literature on high-tech price change. Although much of our own new work also is introduced, the emphasis is on integrating available information into existing measures. The evidence marshaled suggests official ICT prices suffer from substantial mismeasurement. As the discussion in Byrne and Corrado (2017) highlighted, ICT has entered a new era in which mobile and cloud platforms are becoming the predominate means organizations use to interact with both customers and employees, and skepticism abounds as to whether statistical agencies are capturing all that is going on. The new ICT asset prices in fact suggest a rapid pace of change. That said, many types of ICT products have not been studied or examined in-depth for many years—or ever in the case of enterprise systems softwareandnewertypesofenterpriseapplications(e.g., “businessintelligenceapps”)—whichstrongly suggests new work and new evidence are needed to better assess all that may be going on. Therearethreesectionsinthispaper,onecorrespondingtoeachmajorcomponentofICTinvestment— communications equipment, computers and peripherals, and software. Table 1: ICT Investment Price Change (annual rate) 1963 to 1987 to 2004 to 1994 to 2004 to 2008 to 1987 2004 2015 2004 2008 2015 (1) (2) (3) (4) (5) (6) 1. ICT investment -4.9 -10.6 -8.0 -12.4 -8.9 -7.5 2. Communication equipment .4 -7.3 -8.7 -9.1 -7.4 -9.5 3. Telecom -.3 -11.7 -12.4 -14.3 -10.1 -13.7 4. Other equipment .4 -8.3 -9.3 -10.3 -8.1 -10.0 5. Capitalized services – 1.1 -3.7 -.1 -2.5 -4.3 6. Computers and peripherals -17.1 -21.2 -17.0 -24.0 -21.8 -14.1 7. Servers and storage -18.1 -25.2 -25.7 -31.0 -30.6 -22.7 8. PCs – -27.9 -23.4 -30.3 -30.2 -19.2 9. Other equipment -9.0 -9.3 -3.3 -8.8 -5.4 -2.0 10. Capitalized services – -2.0 -2.2 -3.1 -1.5 -2.6 11. Software -1.0 -4.4 -3.9 -5.5 -3.5 -4.1 12. Prepackaged -9.8 -9.0 -7.0 -9.6 -6.8 -7.2 13. Custom and own-account .0 -2.0 -2.2 -3.1 -1.5 -2.6 Memos: 14. ICT excluding PCs -4.5 -8.4 -6.5 -9.9 -6.6 -6.4 15. Computers excluding PCs -16.6 -17.1 -11.6 -19.8 -14.5 -9.9 16. BEA ICT -2.7 -6.4 -2.1 -7.5 -3.3 -1.4 17. BEA ICT excluding PCs -2.6 -4.5 -1.4 -5.2 -1.9 -1.2 18. Computers excluding PCs -16.6 -11.0 -3.6 -12.7 -6.6 -1.8 Note: Figuresreportedas“BEA”areauthors’calculationsbasedonBEAdata.

2 Communication Equipment The primary source for this index is Byrne and Corrado (2015a,b), who developed prices for many of the detailedproductsof the communicationsequipmentmanufacturing industry(NAICS 3342). Byrne andCorrado(2015a)alsoreviewedtheextensiveliteraturethathasstudiedtechnologicaldevelopments in telecommunications equipment. The U.S. BEA’s detailed commodity composition of communications equipment investment is reproducedinthefigureshownonpage3. ThetableisstructuredaccordingtoI-Oitemcodes, whichlargely align with NAICS industry codes. As may be seen, communication equipment investment has both equipment and capitalized services commodity types, and the equipment components include other NAICS 334 industries, notably, audio and video equipment manufacturing (NAICS 3343) and search, detection, and navigation equipment manufacturing (NAICS 3345). Price indexes and shares for each relevant equipment component, as well as for the capitalized services components, are needed to obtain an overall price index for communications equipment. The estimated rates of change for these price indexes are shown in the table on page 4, and shares of the key components in total communication equipment investment are shown in the figure on page 5. A description of the derivation of these estimates follows. 2.1 Equipment components To build a new price index for communication equipment investment, the equipment components are grouped into four types (some very small items are ignored): • telecommunications and data networking equipment (line 4 plus lines 14–18) • audio/video equipment (line 23) • broadcasting and other communications systems (lines 12–13) • search & navigation equipment, nondefense (line 26) The telecommunications and data networking equipment grouping captures systems enabling twoway communications, a classification scheme the authors developed for measuring the products of the communicationsequipmentmanufacturingindustry(NAICS3342)inearlierwork(ByrneandCorrado, 2015a,table1,page7). Thebroadcastingandothercommunicationssystemsgroupingistheindustry’s one-way systems. Telecommunications/data networking equipment. Telecommunicationsanddatanetworkingequipment (hereafter telecom equipment) is the major component of communication equipment investment (see figure 2), and its price index is an important driver of estimated price change for other equipment components (except A/V). Table 2 in Byrne and Corrado (2017) showed some of the more dynamic products of the communications equipment industry, but all told, Byrne and Corrado (2015a) estimated constant-quality spending price indexes for 14 major product lines of the industry based on item-level price indexes for about 150 individual products. Comprehensive coverage began in 1963. Two modifications are made to the Byrne-Corrado NAICS 3342 spending price index before including it in the communications equipment investment price index constructed for this paper. First, the contribution of consumer purchases of cell phones and terminal equipment (telephone handsets, telephone 2

Figure 1: Commodity Composition of Communication Equipment Investment 2007 Value I-O item code Equipment by type of product (millions of dollars) 1 Communication equipment 106,169 2 33411A Computer terminals and other computer peripheral equipment manufacturing 339 3 Computer terminals (excl. parts/attachments/accessories/etc.) 4 334210 Telephone apparatus manufacturing 20,652 5 Telephone switching equipment 6 Carrier line equipment & nonconsumer modems 7 Telephone sets, including wireless phone sets, exclude cell phones 8 Wireline voice equipment Data communications equipment (including routers, gateways, bridges, 9 terminal servers, and concentrators) 10 Telephone apparatus, nsk Radio and television broadcasting and wireless communications equipment 11 334220 manufacturing 41,112 12 Other communication systems and equipment 13 Broadcast, studio, and related electronic equipment 14 Cellular handsets (cell phones) 15 Wireless networking equipment Radio station equipment including satelite, airborne and earth-based (fixed and 16 mobile) 17 Antenna systems, sold separately 18 Radio and TV broadcasting and wireless communications equipment, nsk 19 334290 Other communications equipment manufacturing 1,035 Vehicular and pedestrian traffic control equipment, including electric railway 20 signals and attachments Intercommunications systems, including inductive paging systems (selective 21 paging), except telephone and telegraph 22 Other Communications Equipment, nsk 23 334300 Audio and video equipment manufacturing 12,294 24 Television receivers, including combination models Speakers, including loudspeakers systems and loudspeakers sold separately, 25 and commercial sound equipment 26 334511 Search, detection, and navigation instruments manufacturing 16,419 27 Search, detection, navigation, and guidance systems 28 Search, detection, navigation, and guidance systems, nsk 29 335920 Communication and energy wire and cable manufacturing 380 Telephone and telegraph wire and cable, made of nonferrous metals 30 (purchased wire) 31 336414 Guided missile and space vehicle manufacturing 119 32 All other services on complete space vehicles for other customers 33 517110 Wired Telecommunications Carriers 7,712 34 Force account, telephone equipment installation 35 541300 Architectural, engineering, and related services 6,084 36 Engineering services 37 S00402 Used and secondhand goods 23 38 Used communication equipment Source: BEA Private Fixed Equipment Investment Bridge Table messaging units, and fax machines) is removed.1 Second, to obtain a telecom equipment price index that spans as many years as possible, the Byrne-Corrado results are spliced with a telecom equipment price index based on Flamm (1989) as complied by (Gordon, 1990, table 9.7, column 8) that begins in 1947. The Flamm-Gordon index does not include terminal equipment. The resulting price index is shown on line 3 of table 2, a calculation that drives our main results. Note first, thatbecausepricesforcellphonesdropmorerapidlythanpricesformostothertelecomproducts, the rate of decline in the telecom equipment investment price index shown on line 3 is less than that of the Byrne-Corrado spending index shown on line 10. Second, line 3 suggests that the pace of change in communications technology was swift in the immediate post-WWII period (column 1) but slowed 1CorrespondencelastyearbetweentheauthorsandBEAconfirmedthatBEAusesa60(household)–40(businessand government) ratio to split total spending on terminal equipment and cell phones. This split is apparently based on the Census Bureau’s results for sources of wireline telecom carrier revenue by type of customer. 3

Table 2: Communications Equipment Investment Price Index and Components (average annual percent change) 1947 to 1963 to 1985 to 2010 to 1963 1985 2015 2015 (1) (2) (3) (4) 1. Communications equipment, private fixed investment -6.1 .9 -7.7 -9.8 2. Equipment -6.1 .9 -8.4 -10.3 3. Telecom/data networking1 -6.4 .3 -11.6 -14.4 4. Audio/video -7.2 -3.2 -10.4 -14.3 5. Broadcast and other -1.1 2.9 -2.5 -2.6 6. Search and navigation -1.8 1.9 -3.8 -4.0 7. Capitalized services2 – – -.8 -5.1 8. Engineering – – 3.0 1.5 9. Telecommunications3 – – -3.1 -6.6 Memos: 10. Byrne-Corrado NAICS 3342 spending5 – .8 -12.3 -16.1 11. Official BEA price index .2 4.0 -2.7 -1.8 12. Difference (row 1 less row 11)4 -6.3 -3.1 -5.0 -8.0 Notes: Figuresfor2015arepreliminaryestimates. 1.Excludesconsumerequipment. 2. Column3 is from1987. 3. Nonresidential component of NAICS 51711. 4. Percentage points. 5. Comprehensive spending price index including consumerequipmentfromByrneandCorrado(2015a,b). Endingdateis2014. in the 1960s/early 1970s (column 2), as discussed by Flamm and Gordon and confirmed by Byrne and Corrado’s more comprehensive estimates. As previously mentioned and reviewed in more depth in Byrne and Corrado (2015a,b), telecom equipment prices began to decline at a historically rapid clip after 1985 (column 3) and have been declining even more rapidly in recent years (column 4). Audio and Video Equipment. Audio and Video (AV) equipment is a specific type of equipment included in communication equipment investment. The construction of a quality-adjusted price index for private investment in AV equipment from 1947 on is set out in table 3. As may be seen, a key ingredient to its history is a series developed by Gordon (1990). The bias implied by this series (4 percentage points per year) is carried forward in time even though certain findings suggest it could be increased: Gordon prepared a rough update of his earlier work (Gordon, 2015) and suggested price declines from 1985 to 2015 averaged 6 percentage points faster than the declines posted by the CPI. A matched-model calculation using detailed data from the Consumer Electronics Association for the years 2004 to 2008 yields price drops 11 percentage points faster than shown by the CPI for those years. Owing to the fragmentary and coarse nature of these findings, the bias is kept at the value implied by Gordon’s long history. The results for this AV equipment price index are shown on line 4 of table 2 and suggest a pace of change similar to that of telecom equipment. Broadcast and other NAICS 3342 communications equipment. Byrne and Corrado did not study broadcast equipment in detail, but Hilbert and L´opez (2011) compiled information on the world’s technologicaleffectivecapacitytobroadcastandtotelecommuteinformation. From1986to2007,based 4

Figure 2: Communications Equipment Investment Shares by Asset Type, 1958 to 2014 Note: Authors’elaborationofdatafromtheU.S.BEAandCensusBureau. on their estimates adjusted to exclude non-electronic broadcast means (e.g., newspapers), effective broadcast capacity grew at 1/4 the rate of growth in telecommunicate capacity (7.3 percent per year versus 29.6 percent). Although this result mixes demand (via use rates) and capacity and pertains to the entire world, it reveals a historically wide gap in the pace of change in these two means of communication. Video compression, digital modulation, and forward error correction are key drivers of capacity increases in radio-wave based broadcasting (e.g., the more computationally complex the compression algorithm, the more video can be compressed). A recent analysis of U.S. broadcast capacity (satellite and cable) suggests digital video compression increased 7 percent per year through 2013 but was expected to accelerate 9-1/2 percent thereafter (Crowley, 2013). Convergence between video broadcasting and telecommunication services has been ongoing, and broadcasting quality is also improving along with better receiving equipment (i.e., TVs) and deployment of fiber.2 To construct quality-adjusted price indexes for this equipment grouping, a production cost index is constructed from 1947 on and a simple bias adjustment is applied to the pre- and post-1985 segments. Prior to 1985, the production cost index is the communication equipment price index BEA derived from regulatory information on costs (i.e., prior to adjustments for quality of equipment produced); PPIs are used from 1985 on, the year their initial availability. The bias adjustment is 25 percent of the differencebetweentherateofchangeoftheproductioncostindexandtherateofchangeofthequality- 2Convergence is also exemplified by the growth of digital TV services and video-on-demand. Statistics on the use of two-way video channels are not available, but according to a report from The Nielsen Company, over 40 percent of U.S. TV homes had subscription video-on-demand access service in 2014Q4. See http://s1.q4cdn.com/199638165/files/ doc_presentations/2015/total-audience-report-q4-2014.pdf, accessed Febuary 8, 2016. 5

Table 3: Sources and Methods Used to Estimate Price Change for Audio and Video Equipment 1947 to 1960 to 1982 to 1985 to Components of the index: 1959 1981 1984 2015 Television receivers A A A B Speakers and other audio D C C C Other equipment D D E E Source data: A Gordon TV price indexa B PCE price index for TVs (≈ CPI for TVs) C PCE price index for speakers D PPI for audio and video manufacturing E PPI for audio and video manufacturing, excl.TVs and speakers Methods for TV component: 1947 to 1984 Series A 1985 to 2015 Bias-adjusted Series B; bias is -4.0 percentage points per yearb Methods for speaker and other audio component: 1947 to 1959 Series D 1960 to 2015 Series C Methods for other equipment component: 1947 to 1981 Series D 1982 to 2015 Series E Weighting: Production weights for 1987 were developed from product shipments reported in the Census Bureau’s Current Industrial Reports and Annual Survey of Manufactures and used to represent investment in all years. These shares are 68, 18, and 14 percent, respectively. Notes: a. Gordon (1990, page 306, table 7.19, last column). b. Bias is the average difference between changes in Series AandtheCPIforTVsfromto1951to1984. adjusted telecom equipment investment price index (e.g., the post-1985 adjustment is -3.3 percentage points per year). Search&navigationequipment. ByrneandCorrado(2015a)didnotstudythesearchandnavigation component of communication equipment investment in a comprehensive manner, but they did look at prices for GPS devices. A matched-model price index for GPS equipment built using detailed modellevel data from The NPD Group and federal government (GSA) price schedules fell at a rate similar to overall telecomm equipment (13 percent) from 2003 to 2010. GPS technology is fundamental to a substantial segment of the search & navigation equipment manufacturing industry, which sells 70 percent of its output to the U.S. Defense Department and long has been one of the U.S. economy’s most technologically sophisticated sectors. Most of the industry’s product shipments are for equipment and systems that arguably exploit the same advances in solid-state circuitry and radio wave transmission that lie behind the rapid change found for telecom 6

equipment (which recall includes satellite communication systems).3 Moreover, the search & navigation and computer and communications equipment industries are similar with respect to the share of intermediate inputs accounted for by electronic components (semiconductors), for which prices fall very rapidly (Byrne, 2015a). Information on defense/nondefense splits by detailed product line is not readily available, and with little further information to go on, we assume that 50 percent of nondefense purchases are for high-end equipment. To construct quality-adjusted price indexes for this equipment grouping, the same procedure as that used for broadcast and other is used, only for search & navigation equipment, the bias adjustment is 50 percent of the difference between the rate of change in production costs and the rate of change of the telecom equipment investment price index. The bias adjustment for the post-1985 segment is -5.7 percentage points per year. 2.2 Capitalized services Capitalized services are the systems design, engineering, and installation services associated with configuring and installing a new system of ICT equipment for a customer. Today’s specialized markets for providing these services evolved from companies such as AT&T and IBM that included design and configuration services in the sales price of their manufactured systems. Capitalized services are included in ICT investment because they are part of the purchasers’ price of a system whether such services are provided by specialists or original equipment makers. According to national accounting practice, however, only separately priced services are included as components with their own deflators (e.g., wholesale and retail services typically are not considered separately priced services). For communications equipment, capitalized services consist of engineering services and telecommunications services used in installation. BEA introduced a separately priced component for engineering services in communication equipment investment in 1987 for which BEA documentation indicates that a PPI for engineering services has been used. More precisely, the commodity composition table (line 35 of table 1) suggests the BEA’s price index for NAICS 5413 (which includes architectural as well as engineering services) is appropriate for this purpose. With regard to telecommunications services (line 33 of table 1), table 1 suggests a price index for wired telecommunications carriers (NAICS 517110) may be used. A price index for the output of wired telecom carrier services from 1987 on was newly constructed for the Byrne and Corrado (2017) paper. Its nonresidential segment was used to deflate capitalized telecom services in the communications investment index, and results are shown on line 9 of table 2. The new price index incorporates information on prices of enterprise services from Telegeography; details of its construction are summarized in table 4 (page 8). 2.3 Shares Shares in purchaser’s prices are calculated from information found in BEA’s “bridge” tables that map commodities to final demand. BEA provides this information for private equipment investment annually from 1997 on and in periodic benchmark I-O tables before then. The shares used to construct 3The high-end segment of search & navigation equipment includes reconnaissance and surveillance systems; radar systems and equipment; sonar search, detection, tracking, and communications equipment; specialized command and control data processing and display equipment; electronic warfare systems and equipment; and navigation systems and equipment. 7

Table 4: Sources and Methods Used to Estimate Price Change for Wired Telecommunications Services (NAICS 51711) 1987 to 2002 to 2007 to Components of the index: 2001 2006 2014 Residential A A A Nonresidential B B C, B Source data: A Price index for residential wireline telecom services, obtained by aggregating BEA PCE price indexes for internet access and wireline telephony (local and long-distance) B BEA implied nonresidential price index obtained by chain-stripping Series A from BEA’s gross output price index for wireline telecommunications carriers (NAICS 51711) C Enterprise wireline services price index, this papera Methods for nonresidential component: 1987 to 2001 Series Bb 2002 to 2006 Bias-adjusted Series B; bias is -5.2 percentage points wedged to 0 in 2001c 2007 to 2014 Weighted average of Series C (75 percent) and Series B (25 percent)d Weighting: Aggregation (and stripping) of components uses the net of own-use weights implied by the estimates shown in figure 6 of Byrne and Corrado(2017). Notes: a. Matched-model price index of Telegeography prices for four groups of enterprise business services (virtual private network; dedicated internet access; IP private line, domestic; IP private line, international). b.SeriesBforthissegmentisessentiallyidenticaltoSeriesA.c. Biasistheaveragedifferencebetweenchanges inSeriesBandchangesinthenonresidentialpriceindexfrom2007to2014. d. SeriesBisusedastheimplicit pricefornon-enterprisenonresidentialservicesforthissegment;weightingisapproximate. the overall price index for communications investment uses detailed information from the 1963, 1967, 1972, 1977, 1987, 1992, and 1997, 2002, and 2007 benchmark tables, as well as the annual information from 1997 on. From 1987 on, this I-O information suggests that the telecom equipment share has been 60 percent of total investment spending, broadcast, AV and other equipment 11 percent, and search & navigation 19 percent. Capitalized services average 11 percent. Prior to 1987 the situation is somewhat different: The telecom services share of communication equipment investment is much larger—it averaged 25 percent prior to the break-up of AT&T—but appears to have been partially offset by a larger wholesale margin after 1987. This is consistent with treating telecom services in the period prior to the early 1980s as an unpriced services margin. Capitalized services are therefore assumed to have edged from zero in the year after the break up of AT&T (1983) to the value (3.8 percent) reported for engineering services in its introduction year (1987). Subsequent years are interpolated values until 1997, when annual share information is available. Equipment shares for telecom equipment and other equipment prior to 1997 are also obtained from the periodic bridge tables. Annual information on shipments of search and navigation equipment (nondefense) as well as detailed information in the Census Bureau’s CIRs compiled in Byrne and Corrado (2015a) are used to interpolate the telecom/non-telecom equipment split where needed. The estimated shares were previously displayed in figure 2. 8

3 Computers and peripheral equipment This index is constructed using price indexes from the existing literature supplemented by biasadjustments suggested by a reading of additional indicators of price-performance trends. The U.S. BEA’s detailed commodity composition of computers and peripheral equipment investment is reproduced in the figure shown on page 9. As in the case of communications equipment, the table is structured according to I-O item codes, which largely align with NAICS industry codes. Computers and peripheral equipment investment has both equipment and capitalized services commodity types. We develop price indexes for the largest product lines in the electronic computer manufacturing category as shown in table 5—multiuser computers (line 2) and personal computers (line 3)–using the products for which research is available. In the case of multiuser computers, we employ an index for servers as a proxy for other types of multiuser computers, such as mainframes and supercomputers. In the case of personal computers, we employ and index for desktop PCs and notebook PCs as a proxy for other types of single-user computers, such as workstations and tablets. The computer storage device index is taken from Byrne (2015b) and is relatively comprehensive in scope. (Used equipment is “ignored”, meaning it is effectively deflated by a weighted average of the price indexes for the other categories.) The estimated rates of change for the price indexes corresponding to key components are shown in the table on page 10 and shares of the key components in total computers and peripheral equipment investment are shown in the figure on page 14. A description of the derivation of these estimates follows. Figure 3: Commodity composition of computers and peripheral equipment investment 2007 Value I-O item code Equipment by type of product (millions of Dollars) 1 Computers and peripheral equipment 8 7,725 2 334111 Electronic computer manufacturing 5 0,365 Host computers, multiusers (mainframes, super computers, medium scale 3 systems, UNIX servers, PC servers) Single user computers, microprocessor-based, capabale of supporting 4 attached peripherals (personal computers, workstations, portable computers) 5 Other computers, including array and other analog, hybrid and special purpose 6 Electronic computers nsk, total 7 334112 Computer storage device manufacturing 4 ,688 8 Computer storage devices (excl. parts/attachments/accessories/etc.) 9 Parts, attachments, and accessories for computer storage devices 10 Computer storage devices, nsk, total 11 33411A Computer terminals and other computer peripheral equipment manufacturing 1 4,266 12 Computer terminals (excl. parts/attachments/accessories/etc.) 13 Computer terminals, nsk, total 14 Input devices, all types 15 Impact printers 16 Nonimpact printers 17 Digital cameras 18 Optical scanning devices 19 Monitors, accessories, and other peripheral equipment 20 Other computer peripheral equipment, nsk, total 21 541512 Computer systems design services 1 5,152 22 S00402 Used and secondhand goods 3 ,254 23 Used computing equipment Source: BEA Private Fixed Equipment Investment Bridge Table 9

Table 5: Computers and Peripheral Equipment Investment Price Index and Components (average annual percent change) 1959 to 1985 to 1985 to 2010 to 1985 2015 2010 2015 (1) (2) (3) (4) 1. Computers and peripherals, private fixed investment -18.0 -19.4 -20.7 -12.7 2. Multiuser computers -21.4 -22.5 -23.3 -18.5 3. Personal computers – -26.4 -28.0 -18.0 4. Storage equipment -16.3 -26.7 -27.3 -23.4 5. Other peripherals1 -13.0 -8.0 -9.3 -1.1 6. Capitalized services2 – -2.1 -1.8 -2.3 Memos: 7. Official BEA price index -16.9 -11.6 -13.5 -1.6 8. Difference (row 1 less row 7)3 -1.1 -7.8 -7.2 -11.1 Notes: Figuresfor2015arepreliminaryestimates. 1.Includesterminals. 2. Column3isfrom1987. 3. Percentagepoints. 3.1 Multiuser computers Our analysis proceeds as follows. Multiuser computers (MCs) are divided into three classes—servers, mainframes and supercomputers—according to the role they play in the IT ecosystem. This allows separate discussion of computation, storage, transmission, and power consumption trends, which differ across classes. In contrast to PCs, for which highly granular price data are abundant, the information available for estimating quality-adjusted price trends for MCs is limited; consequently, we rely on judgmental adjustment of existing price indexes informed by data on engineering advances. We adopt the price index employed in the national income and product accounts for the period from the late 1950s to the mid-1990s, a period for which research is robust. In the 1980s, following on early work by Chow (1967), researchers at IBM conducted path-breaking research on multiuser computer prices in collaboration with BEA (Cole, Chen, Barquin-Stolleman, Dulberger, Halvacian, and Hodge, 1986). Price indexes were constructed based on list prices for IBM and IBM-compatible computer processors (CPUs) using hedonic regressions that accounted for key technical features as well as the disequilibrium in the market resulting from the dominant role played by IBM. Performance from a user value perspective was included in the form of million instructions per second (MIPS). Although MIPS does not control for the translation of those instructions into the performance of actual tasks, this nevertheless was likely a reasonable proxy for performance given the homogeneity of the computer market at the time. A composite hedonic/matched-model price index for multiuser computers (MCs) based on this research was introduced into the national accounts for 1969-1984 (Cartwright, 1986).4 The indexes were later extended back to 1959 and updated through the mid-1990s. The results for the BEA price index for MCs are largely corroborated by Gordon (1990) and Triplett (1989), and the computer price index shown in table 5 adopts the BEA MC index through 1996. 4Composite hedonic indexes aggregate item prices using a matched-model approach, but for periods where an item has just appeared in the sample, prices imputed by hedonic regression are used to construct relative prices. Triplett (1989)notesthatthisapproachistheoreticallyequivalenttothetime-seriesregressionapproachmorecommonlyusedin academic research, where time dummies are chained together to generate a price index. 10

MCs in the national accounts are deflated from 1996 forward by the Bureau of Labor Statistics (BLS) producer price index (Grimm, Moulton, and Wasshausen, 2005). For this period, we use the server price index as discussed below.5 Servers. Servers traditionally focussed on providing a single function for client computers, such as managing print resources, processing email, or delivering web pages. Monolithic multi-user computers, most notably mainframes, accounted for the bulk of the commercial MC market from its inception until the mid-1980s. A surge in PC investment and networks to connect them–including both local networks and the internet–followed in the 1990s, raising the importance of servers providing data, communications, and applications to PC clients. Data from a high-tech consultancy (IDC) suggest that servers now account for about 80 percent of total MC investment. Not surprisingly given the importance of servers and the difficulty noted here of measuring prices and quality for other types of MCs, servers appear to be the sole focus of the PPI.6 Hedonic regressions are estimated by BLS for servers annually with an extensive set of control variables for technical features, such as the amount of memory, and for engineering performance, such as clockspeed. However, there is reason to believe that the specifications employed may not adequately account for performance trends. In particular, no controls are used for directly-measured efficiency of user task completion, which may be problematic in light of the divergence between clockspeed and performance trends that opened up in the early 2000s (Byrne, Oliner, and Sichel, 2017a).7 To address this concern, we examine system performance trends. From 2002 to the present, we use scores from by the widely-referenced benchmark scores published by Systems Performance Evaluation Corporation (SPEC).8 Specifically we employ the median SPEC performance for systems employing Intel Xeon processors, which dominated the server market in that period.9 For performance in earlier years, we use price per million floating point operations (megaflops) reported in Hilbert and L´opez (2011) which extend through 2007. The Hilbert and L´opez-based series has a similar contour to the SPEC-based series but falls about 4 percentage point faster in the overlap period. Because the SPEC score is based on a suite of user applications, we prefer that index and extrapolate it back to 1995 using the Hilbert and L´opez series adjusted for this bias. We use the performance index to quality-adjust a server unit price series constructed using data from high-tech consultancies (Gartner and IDC) on revenue and units for the U.S. market available beginning in 1996. In particular, we divide U.S. server revenue by the product of U.S. unit sales and our performance index. Beginning in 1997, our indicator falls markedly faster than the PPI, with an average difference of 20 percentage points (figure 4). For our final index, we blend our indicator with the PPI, assigning 1/3 weight to the PPI. We combine the indexes because each has distinct merits. We have concerns about the control for computing performance in the PPI, but we also note that the PPI controls for other important features of servers such as the operating system, memory, and 5The BLS producer price index (PPI) also uses a hybrid hedonic/matched-model approach. See http://www.bls. gov/ppi/ppicomqa.htm(datedJune2011,accessedJuly1,2016)forillustrativeexamplesofBLShedonicregressionsand quality adjustment of computer prices. 6We deduce this from the fact that BLS doucumentation provides the specification for a regression used in 2011 that only references Xeon MPUs, which suggests that mainframes, which do not use Xeons, are not in scope. 7Theabsenceofdirectmeasuresdoesnotnecessarilyimplythatquality-adjustmentisinadequateifalinearcombination of technical features predicts performance. 8Scores for SPEC® CPU2006 and SPEC® CPU2000 retrieved from http://www.spec.org on June 6, 2016. Comparison between the two benchmark suites based score ratios for models reporting results for both. 9The BLS hedonic regressions for servers also concentrate exclusively on Xeon processor systems. 11

interconnect technology.10 Our final index, a blend of the indicator and the PPI, falls 11 percentage points faster than the BEA investment index. (Note that the BEA investment index is a weighted average of the BLS PPI and the BLS import price index.) Figure 4: Alternative server price indexes (percent change), 1993 to 2014 Mainframes. Mainframes are designed to rapidly handle a diverse and shifting workload of computationaltasks—especiallybusinesstransactions,looselydefined—inahighlysecure,fail-safeenvironment. For mainframes, the analysis is confined to IBM, long the most relevant and dominant player in the market.11 IBM publishes performance scores for its systems whose usefulness is enhanced by two features: they are based on time to complete commonly used applications (rather than an artificial basket of instructions) and IBM is deeply committed to backward compatibility, making comparisons over time informative. Table 6 below shows benchmark scores, known as Large System Performance Reference ratios (LSPRs), for IBM’s flagship mainframe models.12 Our goal with this table is a simple examination of the pace of advance of the frontier for mainframes in the “enterprise” (i.e., large business) market. In the mid-1990s, six generations of mainframes using the System/390 architecture were introduced in quick succession and performance shot up 90 percent per year. In the 2005–2015 period, a series of mainframes known as the zSeries was released for which performance rose 18 percent per year on average. This performance slowdown may well reflect the challenges that emerged in single-core performance as excessive heat generation restrained the pace of increase in processor clockspeed; as is 10In some cases, the server PPI controls for power consumption as well, a crucial consideration. (Conversation with Fred Merkel, BLS April, 2016.). 11The IBM System 360 line of mainframes had a 65 percent market share in 1977. Competitors, such as Amdahl and Hitachi, primarily produced systems intended to mimic IBM. 12See https://www-304.ibm.com/servers/resourcelink/lib03060.nsf/pages/lsprindex/$file/SC28118719.pdf (accessedJuly1,2016)formoreinformationonLSPRs. Modelsfullyloadedwithprocessorsarechosenforthisanalysis. 12

Table 6: IBM Enterprise Mainframe Prices and Performance Scores Model Year Processors Performance Price Price/Perf. System/390: G1 1994 6 1.1 1,260,000 1,200,000 G2 1995 10 2.7 2,550,000 940,959 G3 1996 10 5.9 3,600,000 607,083 G4 1997 10 7.1 3,240,750 457,733 G5 1998 10 17.6 4,000,000 227,015 G6 1999 10 26.1 4,842,000 185,588 Annual growth rate 90% 31% -31% zSeries: Z9 2005 54 39.6 22,251,000 562,462 Z10 2008 64 50.9 25,949,995 509,523 Z196 2010 80 93.4 28,546,000 305,632 EC12 2012 101 140.1 33,096,000 236,231 Z13 2015 141 199.3 n/a n/a Annual growth rate 18% 6% -12% Notes: System/390: Large-scale Performance Ratio for OS/390 operating system, indexed toperformanceofSystem/390SeriesG49672R15=1. zSeries: Large-scalePerformanceRatio,Multi-Image(MI)forz/OSoperatingsystem,indexed to System z9 2094-701 = 1. z13 score is extrapolated using ratio z13 Performance Capacity Index(MIPS)toEC12. apparent in the rise of the number of processors used, IBM turned increasingly to parallel processing to deliver greater performance. Prices for selected IBM mainframe models from 1990 to 2012 were obtained from an analyst’s compilation available online.13 Only prices on introduction are available and we do not have relative importance weights or an exhaustive set of model characteristics.14 Table 6 shows that prices rose substantially less than performance in both periods. The price-performance ratio fell 31 percent per year for System/390 frontier models in the late 1990s, but by a more moderate 12 percent per year for zSeries models from 2005 to 2012. These results are not comprehensive and should be interpreted with caution. Also, it is important to note that IBM’s business model is focused on the joint sale of hardware, software and services using long term contracts, and quality-adjusting prices is difficult without information on the terms of sale. That being said, the index for frontier mainframes in the late 1990s falls about 30 percent per year, then slows to a bit more than 10 percent per year in the latter part of the first decade of the 2000s. Interestingly, theserates falloneitherside oftheroughly20percent averagerate ofdeclineofthe BEA mainframe investment index from 1959 to 1995, suggesting that this long-run trend has persisted on average, though the recent slowing is substantial and the outlook is unclear. BEA has not conducted research on mainframe prices for its official computer index since the early 1990s. The BLS PPI was adoptedbyBEAfordeflatingMCinvestmentatthatpointandasnotedpreviously, thatindexappears to be based exclusively on servers. 13See Technology News, http://www.tech-news.com. We thank Hesh Werner for generously providing these data to the public and for guidance on interpreting the IBM mainframe market. 14In ongoing work, we match these prices to characteristics and update the results of Cole et al. (1986) through 2012. 13

Figure 5: Computers and peripheral equipment investment shares by asset type, 1958 to 2014 Note: Authors’elaborationofdatafromtheU.S.BEAandCensusBureau. Supercomputers. Supercomputers are used to solve immense computational problems, such as simulating complex phenomena (e.g., weather) or processing seismic data for geological exploration, using massively parallel architecture. Although we do not have the information to construct a deflator for investment in supercomputers, recent trends in performance are examined to shed light on the pace of technical change in advanced computing and adoption by industry. Supercomputers account for 5-10 percent of MC spending.15 Performance—as measured by the speed of inversion for a very large matrix with the LINPACK benchmark—of the top U.S. supercomputer rose 75 percent per year on average from the inception of the Top 500 list in 1993 to 2015, but since 2000 there have been two distinct periods (Byrne and Corrado, 2016, table 2, line 24). From 2000 to 2008, performance of the median Top500 computer rose 95-1/2 percent per year, but has slowed to 55.2 percent per year since then. Performance of the median U.S. industrial supercomputer appearing on the TOP500 list, rises at the same pace as the overall median, on average, but typically attains the level of performance of the concurrent leading edge machine five years later. The slowdown in computational performance gains at the leading edge and for industrial machines is striking. This apparent regime change may be attributable to increasing challenges with power consumption that emerged in the mid-2000s. This led to a shift of focus away from raising the speed of individual processors and toward greater reliance on use of processors in parallel and experimentation with more energy-efficient processors, such as graphics processing units (GPUs) and “embedded” MPUs. These architecture changes have paid off in greater power efficiency. The median megaflopsper-Watt score for supercomputers as recorded the The GREEN500 list of the world’s most energyefficient supercomputers (also available at http://www.top500.org, rose 47 percent per year from 2012 to 2015, a notable step-up from the previous pace of 34 percent since the inception of the list in 15GartnerestimatedU.S.supercomputingspendingofroughly$2billionperyearintheearly2000s,whichincludesthe federalgovernmentanduniversities. AccordingtotheTOP500listofhigh-performancecomputers(http://www.top500. org), however, about 2/3 of US supercomputers on the global top 500 list were industrial. 14

2007.16 PayoffstotheGPU-basedarchitectureintermsofcalculationperformancerequirenewsystems and new programming tools and thus may take longer to arrive than gains previously associated with better and faster CPUs. 3.2 Personal computers Desktop personal computers (PCs) emerged as a significant share of U.S. computer investment beginning in the mid-1980s, followed by portable personal computers (laptops or notebooks) in the late 1990s.17 The emergence of PCs as an important business tool has led to a situation in which the PC price index has notable impact on the overall computer investment deflator. BEA introduced a PC price index covering 1982–1987 in 1988 (Cartwright and Smith, 1988). The agency maintained this index until adopting the BLS composite hedonic/matched-model price index (Sinclair and Catron, 1990) that begins in January 1992. More precisely, BEA’s annual PC investment price index is a chain-weighted average of the BLS PPI and the BLS import price index (IPI) for computers from 1993 on. The ready availability of data and interest in the role of the PC in productivity has led to a host of studies of PC prices. As acknowledged in Landefeld and Grimm (2000) and the review in Berndt and Rappaport (2001), price indexes from some outside studies fall significantly faster than the BEA PC price index. One study cited was an early version of results reported in Berndt and Rappaport (2003) and is the basis of the PC price index shown on line 8 of table 1. Compared with earlier research, Berndt and Rappaport (2003) use a very flexible form for the hedonic function that relates prices and characteristics of PCs. They include dummies for processor type and interactions of those dummies with clock speed to allow the regression to better account for the rapid increase in MPU performance in the 1990s. Berndt and Rappaport (2003, hereafter BR) adjacent-year laptop and desktop price indexes are used to develop the history for the price index shown on line 8 of table 1. Specifically, the adjacent year indexes reported in BR table 3 with flexible processor controls that cover 1989 to 2003 are extended back to 1983 using the adjacent-year indexes reported in BR table 2; the indexes are smoothed using a three-year moving average. The smoothed laptop and desktop indexes are aggregated using platform investment shares constructed with consultancy data (IDC and Gartner) beginning in 1992 and using CIR (production) detail in prior years. Recent research shows that the performance of MPUs for PCs is not well proxied by clockspeed and other technical features in hedonic regressions after 2002, and that direct (benchmark) measures of performanceareneededtofullyadjustforperformance(Byrneetal.,2017a). TheBLShedonicanalysis of PCs relies primarily on engineering features and does not employ benchmarks, and the gap between the BEA investment index and average prices has dwindled appreciably since the late 1990s (figure 6). Indeed, the average price of a PC sold in the U.S. business market fell 5 percent per year on average from 2010 to 2015, the same rate as the decline in BEA PC investment price index, indicating the BEA is picking up no quality improvement in PCs over that time period. 16Lookingatthebroadercomputerindustry, Koomey, Berard, Sanchez, andWong(2011)reportthatpowerefficiency has historically grown 26 percent per year, implying that the recent efficiency gains for supercomputers have been very rapid in comparison. 17Tablet computers have risen in importance since 2010 and are not addressed in this analysis. Byrne, Oliner, and Sichel (2017b) and Byrne, Dunn, and Pinto (2016) estimate price indexes for tablets, but this work is preliminary. 15

ToadjustthePPIforimprovementsinqualityduetoincreasesinperformance,weaddthedifference(20 percentage points) found by Byrne et al. (2015) between MPU hedonic indexes accounting for quality and indexes that do not times the approximate cost of MPUs as a share of PC value (15 percent) to arrive at a correction to the PPI (3 percentage points per year from 2002 on). The corrected PPI is used as PC investment price deflator from 2003 on.18 Figure 6: Personal computer price change, 1996 to 2014 Source. IDC, Inc.; Bureau of Economic Analysis 3.3 Storage, other peripherals, and capitalized services For storage equipment, the official BEA investment price is a weighted average of the BLS PPI and import price index. This index falls roughly 20 percentage points slower than a measure of price per megabyte for hard drives developed and updated by McCallum (2002) from 1985 to 2014.19 Work by Byrne(2015b)usesmodel-levelpricesforstorageequipmenttocreateamatched-modelpriceindexthat falls at nearly the rate of the raw price-per-megabyte series from 2002 to 2014. From 2002 forward, we use the Byrne (2015b) index. We use the BEA investment price index for storage equipment from 1958 to 1984 and from 1985 to 2001 use the geometric mean of this index and the McCallum price-per-gigabyte series. Very little research has been conducted on price trends for other computer peripherals. Although Cole et al. (1986) and subsequent extensions were based on hedonic analysis, the BLS price indexes employed in BEA investment deflators from the early 1990s forward use standard matched-model techniqueswithouthedonicadjustment. Althoughinprinciplethestandardmethodologycouldcapture quality change, Aizcorbe and Pho (2005) estimate price indexes for a range of peripherals using NPD model-level scanner data and report prices that fall substantially faster than the BLS PPI. This work is suggestive but only covers a three year period, which we view as insufficient to justify bias-adjusting the official investment index. 18The shortcomings of BEA’s procedure of averaging the PPI and the IPI are set out in Byrne and Pinto (2015). 19Updates can be found at http://www.jcmit.com/diskprice.htm. 16

Table 7: Sources and Methods Used to Estimate Price Change for Computers and Peripheral Equipment Methods for multiuser computer component: 1959 to 1996 BEA investment price deflator for “mainframes”a 1997 to 2015 Weighted average of quality-adjusted server unit value (2/3) and BLS PPI for “host computers, multiuser (mainframes, medium scale systems, Unix and PC servers)” (1/3). Methods for personal computer component: 1983 to 1988 Chained adjacent-year hedonic indexes for desktop and mobile PCsb, Berndt and Rappaport (2003) table 2. 1989 to 2003 Hedonic estimates w/ flexible MPU speed/family interactions, Berndt and Rappaport (2003) table 3. 2004 to 2015 U.S. revenue-weighted average of BLS PPIs for “portable computers, laptops, PDAs and other “single-user computers” and for “personal computers and workstations (excluding portable computers),” bias-adjusted using MPU index bias from Byrne et al. (2017a) by authors’ estimate of MPU input cost share for PCs. Methods for storage equipment component: 1959 to 1984 BEA investment price deflator for “storage devices.” 1985 to 2001 Equal-weighted average of BEA investment deflator and price-per-megabyte series based on McCallum (2002). 2002 to 2015 Byrne (2015b) as extended by Federal Reserve Board.c Weighting: Investment weights for computers (as a whole), storage equipment, other peripherals, and systems are from BEA’s “Detailed Data for Fixed Assets and Consumer Durable Goods”, available at https://bea.gov/national/FA2004/Details/Index.htm. BEA formerly published further detail on investment by type of computer beginning in 1995, including PCs (as a whole) and MCs, in a spreadsheet entitled ”Final Sales of Domestic Computers”. These data were extended to 2015 using reports on U.S. sales from IDC, and backward to 1977 (the emergence of PCs) using domestic production by product from Census Current Industrial Reports. Investment weights for portable and other personal computers is based on U.S. sales shares from IDC for 1995–2015 extrapolated from an arbritrary portable revenue share of 2% in 1981 to 1995 using a fixed growth rate. Notes: a. “Mainframes”inBEAnomenclaturereferstoallmulti-usercomputers. b. Investmentinworkstationsis deflatedbypriceindexfordesktopPCs;investmentinluggables,laptops,notebooks,andtabletsisdeflatedbyprice indexfor“mobile”PCs;smartphonesareclassifiedascommunicationsequipment. c. Availableat https://www.federalreserve.gov/releases/g17/ComputerStoragePriceIndex.htm 17

The price index for custom software investment, whose corresponding NAICS industry is 541511 (custom computer programming services), is used as the price index for the capitalized services component of computer investment, NAICS 541512 (computer systems design services). The construction of the price index for custom software is described in the following section. 3.4 Shares The shares used for aggregating the component price indexes into an investment index for computers and peripheral equipment are found on the BEA website under “Detailed Data for Fixed Assets and Consumer Durable Goods” (https://bea.gov/national/FA2004/Details/Index.htm). We use the nominal weights included in the “Investment in Private Nonresidential Fixed Assets” table. This table is not included in the standard fixed asset tables because “the detailed estimates are more likely to be either based on judgmental trends, on trends in the higher level aggregate, or on less reliable source data.” That being said, the broad contours of these series are quite similar, so the contour of the overall price index for computers and peripheral equipment is not likely to be driven by the weighting scheme. 4 Computer Software ThisindexisconstructedusingthebasicapproachusedbyBEAbutadifferentstructureisintroduced. The novelty in the structure is that it includes an explicit enterprise software component consisting of systems and applications software components. As of the writing of this paper, neither desktop nor enterprise systems software are included in BEA’s investment price index for software. What do we know about systems software? Greenstein and Nagle (2014) looked at the economic benefitsofApache,awidely-usedopensourcesoftwaresystemusedtomanagee-commercetransactions inthe1990s,andfoundsubstantialbenefitsevenafteraccountingforsubstitutionwithpricedproducts. This suggests there could be large price declines in the systems software in wide use today, but we cannot really know without additional research. The PPI does not disclose its index for systems software, but we calculate an implied index as detailed below. This implied index in fact drops relatively rapidly, and while its incorporation in an ICT asset price index is a major step forward (it has a large weight), its impact on the new price index for enterprise software is partially offset by the risingPPIforenterpriseandnetworkapplicationsoftware. Alltold,muchadditionalresearchisneeded to further improve the new price measures for enterprise application and systems software developed in this paper. The U.S. BEA’s software price index has three components: • Prepackaged software products (hereafter, software products), • Custom programmed software, and • Software produced on own-account. where own-account software is specialized software developed or improved in-house rather than purchased as custom-made software from a software development company. The relative proportions of these components were shown in figure 7 (b) of Byrne and Corrado (2017). BEA’s approach to measuring current prices for these components may be summarized as follows: 18

• The price index for software products from 1998 on is a bias-adjusted BLS PPI for application software publishing; prior to 1998, a research price index is used. • The price indexes for custom and own-account software prices are productivity-adjusted input cost measures, where productivity in custom and own-account software production is essentially 1/2 that of productivity in software products production. The BEA approach was developed by Parker and Grimm (2000), who opined “...the information available on price indexes for prepackaged software is limited, and no price indexes are available for the prices of custom or own-account software.” Information on software products was limited in that mostoftheresearchavailabletoParkerandGrimmpertainedtoPCdesktopapplicationsoftware—this situation has not changed.20 Nor has the fact that the industry whose subcomponent largely reflects purchases of custom programmed software (NAICS 5415) remains out-of-scope in the BLS PPI. For the software price index developed for this paper, the BEA’s basic approach is retained but four steps are taken to improve it: • A more granular structure is developed for the software products index. • Information on system software prices is incorporated in the software product price measures. • TheBEAbiasadjustmenttothePPIisupdatedbasedonresearchavailablesubsequenttoParker and Grimm’s work. • The input cost measure used by BEA for the custom and own-account components is refined to reflect capital costs as well as wages of programmers. Further details follow. 4.1 Software products The sources and methods used to construct this index are summarized in table 8. There are two major categories to this index: (1) desktop and portable device software, and (2) enterprise and related software. Structure and coverage. The Census Bureau uses two classification schemes to present its data on software. In each scheme, the high level components are application software and systems software. In the periodic economic census, application software consists of: • general business productivity and home use software • game software • cross-industry application software • vertical market application software • utilities software • other application software whereas system software consists of: 20One study available to Parker and Grimm (Harhoff and Moch, 1997) covered a PC database platform for business applications, but prices for PC desktop operating systems were not studied until much later (Abel, Berndt, and White, 2007; Copeland, 2013). 19

• operating systems software • network software • database management software • development tools and programming languages • other system software. In the annual surveys, application and system software are each disaggregated according to platform: • personal computer software • enterprise or network (i.e., server) software • mainframe computer software • other application/system software. The table below figure 7 shows the structure the BLS has used for its software products PPI (NAICS 5112) Figure 7: BLS software price indexes, since 2006. As may be seen, the BLS is using a blend published and implied, 1998=100 oftheCensusschemesinpresentingitssoftwareproducts price index. The blended structure is useful for building an investment price index because it strips out games, which are products purchased almost exclusivelyforhomeuse. Annualdataongamesoftware salesarenotreportedintheCensusBureau’sServices Annual Survey. The software PPI structure table highlights that a price index for software products excluding games is not issued by the BLS. A price index for application softwareproducts(excludinggames)ispublished,but the BLS price index for system software does not meet the agency’s standards for disclosure and therefore is not published. Price developments in a very Structure of BLS Software PPI important area of software product development are 1. Software publishing, except games therefore (unintentionally) obscured: cloud comput- 2. Software, except games and other receipts ing platforms are enabled by systems software, high 3. Application software publishing performance computing is enabled by specialized sys- 4. Desktop and portable device apps tems development tools; big data analytics are en- 5. Enterprise and other apps 6. Systems software publishing abled by database management software, etc. 7. Desktop and portable device systems Becauseapriceindexforsoftwareproductsexcluding 8. Enterprise and other systems 9. Other receipts games is the most relevant driver of a software prod- 10. Game software publishing ucts investment price index, an implied BLS price in- Note: Price indexes on lines 2, 6, 7, and 8 are not disclosed. dex for this and its systems software component are calculated from 1998 on.21 The bottom line is that BEA has been using the dark blue line in figure 7 21Note that the calculations used to create the implied price indexes for software products excluding games and systems software are sensitive to the weights used. Furthermore, the calculations entail a series of steps because the PPI for software publishing excluding games only begins in 2006 and the components of the application software PPI also changed in that year. Weights for games, other receipts, and system software are needed to calculate the relevant price indexes from 1998 on and were obtained from the periodic economic censuses from 1997 on. Because weights from 20

todriveitsinvestmentsoftwarepriceindex, whereastheindexsetoutintable8isdrivenessentiallyby the dashed red line. In fact, the implied systems PPI is used directly and aggregated with appropriate annual weights, not the lagged census weights used to construct the PPI (see previous footnote). Systems software averaged 49 percent of total software products excluding games from 2010 to 2014 according to the 2014 Services Annual Survey.22 The largest component within both systems and applicationsoftwareisenterpriseandmainframesoftware,whichrosefrom62.6percentoftotalsoftware products excluding games in 2010 to 66.5 percent in 2014. This increase is driven by movement within the system component—sales of systems software for personal computers were lower in 2014 than they were in 2007. Information on prices for desktop vs enterprise/other components of system software is unavailable. For this reason, and also because desktop system software does not loom very large in total desktop software (as well as in total system software), software investment price change is built from the three available component measures from 2007 on (desktop application, other application, and system software) where desktop application software prices are used to represent all desktop software prices, and all system software prices are used to represent enterprise system software prices. Prior to 2007, thethreecomponentsareestimatedusingmethodsconsistentwiththemethodsusedtodevelopBEA’s existing history. Bias adjustment. When Parker and Grimm developed BEA’s current approach, experience with BLS’s software PPI was very limited (the series began in January 1998). The bias adjustment they recommended (-3.15 percentage points per year) was thus based on limited information. On the basis of subsequently available work, including PC desktop operating systems software (Abel et al., 2007; Copeland, 2013), the bias adjustment has been changed, to -4.15 percentage points per year. The adjustment is based on the difference between a spliced research series for PC productivity and O/S software and BEA’s existing series (the application software PPI adjusted by -3.15 percent per year) from 1998 to 2003, i.e., the research series fell 1 percentage point per year faster from 1998 to 2003. Strictly speaking, the newly calibrated bias adjustment pertains only to the desktop component of software products investment, but it is used for the other components due to lack of other pertinent information. Not only are there no recent studies of software prices, there are no studies of enterprise, mainframe, or server software prices, period—making it next to impossible to determine whether the software products that power today’s mobile and cloud business platforms are being adequately captured in our price statistics. periodiccensusesbecomeavailablewithalag,tocorrespondwiththePPI,thecalculationsusedtoconstructtheimplied price indexes including system software employ 1997 weights beginning 2004, 2002 weights beginning 2008, and 2007 weights from 2012 on. Owing to the lack of detail in the 1992 census, 1997 weights are used in lieu of actual weights for this period. 22Game software is not separately reported in annual surveys. Game software sales are assumed to be included in the “other application software” category, and annual estimates for game software sales were based on the 2012 economic census and press reports issued by NPD. 21

Table 8: Sources and Methods Used to Estimate Price Change for software products investment 1959 to 1986 to 1999 to 2007 to Components of the index: 1985 1998 2006 2015 Desktop and portable computer device software – A B C Enterprise and server/mainframe software D D – – Applications – – B E Systems – – F F Source data: A BEA prepackaged software price index derived from research estimates of prices for selected types of PC desktop application and database software (Parker and Grimm, 2000) B BLS application software price index (excluding games) C BLS desktop and portable device application software price index D BEA prepackaged software price index based on .6 times hardware price changea E BLS other applications software price index F Implied BLS systems software price index, this paper Methods for desktop component: 1986 to 1998 Series A 1999 to 2006 Bias-adjusted Series B; bias is -4.15 percentage points per yearb 2007 to 2014 Bias-adjusted Series C; bias is -4.15 percentage points per yearb Methods for enterprise components: 1959 to 1998 Series D 1999 to 2015 Bias-adjusted Series B, E and F, depending on segment and component, where the bias is -4.15 percentage points per year Weighting: Desktop and enterprise components are weighted based on information on the composition of domestic sales from the Census Bureau from 1997 on and from trade reports for prior years. The sub-components of the enterprise price index are equally weighted. Notes: a. The hardware price index is BEA’s computer price index to 1994 and this paper’s server/mainframe and data networking price indexes from 1995 on (equally weighted). b. Bias is the average difference between changes in Series B and changesinaresearchpriceindexavailablefrom1998to2003. Theresearchpriceindexlinksmatched-modelpriceindexesfor PCdesktopproductivityandO/SsoftwaredevelopedbyAbeletal.(2007)andCopeland(2013). 4.2 Custom and own-account software The BEA currently applies a productivity adjustment to programmer wage costs whereas using the combined costs of programmer labor, capital, and intermediate inputs is a more appropriate approach. There are at several drawbacks to pursuing the more theoretically appropriate approach. First, precise informationonthecapitalandintermediateinputcostsofcustomprogrammingfirmsandown-account software production within firms is unavailable. Second, even if industry-level measures of combined input costs were to be used as a relevant approximation of the unobserved combined costs, such measures usually are extremely volatile. Indeed, BLS estimates of the combined input costs for the software products and computer design services industries (NAICS 5112 and NAICS 5415, where NAICS 5415 includes custom programming services) are found to be rather volatile. Third, while the 22

BLS makes use of essentially all available data, their estimates begin in 1987, and it is necessary to have price deflator for (own-account) software that begins in 1959. Two steps are taken to circumvent these drawbacks: (1) Changes in the GDP deflator are used as a combined cost indicator. The trend in the average (log) change in BLS’s combined cost indexes for NAICS 5112 and N5415 industries is essentially the same as the trend in the GDP price deflator from 1987 to 2010. (2) To tie the indicator more closely to computer program development costs, changes in the GDP deflator are averaged with changes in programmer wages. TheBEAassumptionthatcustomandown-accountsoftwareoutputisproducedwith1/2theefficiency of mass-marketed software products is retained, that is, the final custom and own-account price index is an average of changes in the combined cost indicator and the software products price index. Results are shown in figure 8. Figure 8: Software and software components price change, 1972 to 2015 5 Conclusion This paper supplements our recent paper “ICT Services and their Prices: What do they tell us about ProductivityandTechnology?”byprovidingdetaileddescriptionsoftheconstructionofthealternative national accounts-style ICT investment price deflators that are reported and analyzed in that paper. The ICT equipment prices described herein were also used in Byrne et al. (2016). As seen in columns 2 and 3 of table 1, according to our estimates, the understatement of the pace of change in ICT asset prices is substantial. Declines in ICT investment prices are estimated to have been 10.6 percent per year from 1987 to 2003 and 8.0 percent per year from 2004 to 2015 (4.2 and 5.9 percentage points per year faster, respectively, than official estimates). Further implications are explored in the above-cited papers. 23

References Abel, J. R., E. R. Berndt, and A. G. White (2007). Price indexes for Microsoft’s personal computer software products. In E. R. Berndt and C. R. Hulten (Eds.), Hard-to-Measure Goods and Services, Volume 67 of NBER Studies in Income and Wealth, pp. 269–289. Chicago: University of Chicago Press. Aizcorbe, A. and Y. Pho (2005). Differences in Hedonic and Matched-Model price indexes: Do the weights matter? Technical report. Berndt,E.R.andN.J.Rappaport(2001). Priceandqualityofdesktopandmobilepersonalcomputers: A quarter-century historical overview. American Economic Review 91(2), 268–273. Berndt, E. R. and N. J. Rappaport (2003). Hedonics for personal computers: A reexamination of selected econometric issues. Presented at R&D, Education and Productivity, an international conference in memory of Zvi Griliches (1930–1999), August 25-27, Paris, France. Byrne, D. (2015a). Domestic electronics manufacturing: Medical, military, and aerospace equipment and what we don’t know about high-tech productivity. FED Notes June 2, Federal Reserve Board. Byrne,D.,W.Dunn,andE.Pinto(2016). PricesandDepreciationintheMarketforTabletComputers. FEDS Notes 1885 (December), Federal Reserve Board, Washington, D.C. Byrne, D. M. (2015b). Prices for data storage equipment and the state of IT innovation. FEDS Notes (July 15), Federal Reserve Board, Washington, D.C. Byrne, D. M. and C. A. Corrado (2015a). Prices for communications equipment: Rewriting the record. FEDS Working Paper 2015-069 (September), Federal Reserve Board, Washington, D.C. Byrne, D. M. and C. A. Corrado (2015b). Recent trends in communications equipment prices. FEDS Notes (September 29), Federal Reserve Board, Washington, D.C. Byrne, D. M. and C. A. Corrado (2017). ICT Services and their Prices: What do they tell us about Productivity and Technology? International Productivity Monitor (29, forthcoming). Byrne, D. M., J. G. Fernald, and M. B. Reinsdorf (2016). Does the United States have a productivity slowdown or a measurement problem? Brookings Papers on Economic Activity (Spring), 109–182. Byrne, D. M., S. D. Oliner, and D. E. Sichel (2017a). How fast are semiconductor prices falling? Review of Income and Wealth (forthcoming). Byrne, D. M., S. D. Oliner, and D. E. Sichel (2017b). A New look at prices of personal computers: Desktops, laptops, and tablets. https://www.aeaweb.org/conference/2017/preliminary. Byrne, D.M.andE.Pinto(2015). Therecentslowdowninhigh-techequipmentpricedeclineandsome implications for business investment and productivity. FEDS Notes (March 26), Federal Reserve Board, Washington, D.C. Cartwright, D. W. (1986, March). Improved Deflation of Purchases of Computers. Survey of Current Business 66(3), 7–10. Cartwright, D. W. and S. D. Smith (1988). Deflators for purchases of computers in GNP: revised and extended estimates, 1983-88. Survey of Current Business 68(11), 23–24. 24

Chow, G. C. (1967). Technological change and the demand for computers. The American Economic Review 57(5), 1117–1130. Cole, R., Y. Chen, J. A. Barquin-Stolleman, E. Dulberger, N. Halvacian, and J. H. Hodge (1986, January). Quality-adjusted Price Indexes for Computer Processors and Selected Peripheral Equipment. Survey of Current Business 66(1), 41–50. Copeland, A. (2013). Seasonality, consumer heterogeneity and price indexes: the case of prepackaged software. Journal of Productivity Analysis 39, 47–59. Crowley, S. J. (2013). Capacity trends in direct broadcast satellite and cable television services. Technical report prepared for the National Association of Broadcasters, National Religious Broadcasters, and National Black Religious Broadcasters (October). Flamm,K.(1989). Technologicaladvanceandcosts: Computersversuscommunications. InR.Crandall and K. Flamm (Eds.), Changing the Rules: Technological Change, International Competition, and Regulation in Communications, pp. 13–61, 371–410. The Brookings Institution. Gordon, R. J. (1990). The Measurement of Durable Goods Prices. Chicago: University of Chicago Press. Gordon, R. J. (2015). Productivity, prices, and measurement. For presentation at Boskinfest, Stanford University (October). Greenstein, S. and F. Nagle (2014). Digital dark matter and the economic contribution of Apache. Research Policy 43, 623–631. Grimm, B. T., B. P. Moulton, and D. B. Wasshausen (2005). Information processing equipment and software in the national accounts. In C. Corrado, J. Haltiwanger, and D. Sichel (Eds.), Measuring Capital in the New Economy, Volume 66 of NBER Studies in Income and Wealth, pp. 363–398. Chicago: University of Chicago Press. Harhoff, D. and D. Moch (1997). Price indexes for PC database software and the value of code compatibility. Research Policy 26(4-5), 509–520. Hilbert, M. and P. L´opez (2011). The world’s technological capacity to store, communicate, and compute information. Science 332(6025), 60–65. Koomey, J., S. Berard, M. Sanchez, and H. Wong (2011). Implications of historical trends in the electrical efficiency of computing. IEEE Annals of the History of Computing 33(3), 46–54. Landefeld, J. S. and B. T. Grimm (2000). A Note on the Impact of Hedonics and Computers on Real GDP. Survey of Current Business 80(12), 17–22. McCallum, J. C. (2002). Price-performance of computer technology. In V. Oklobdzija (Ed.), The Computer Engineering Handbook, Chapter 4, pp. 4–1 to 4–18. CRC Press. Parker, R.andB.T.Grimm(2000). Recognitionofbusinessandgovernmentexpendituresforsoftware as investment: Methodology and quantitatvie iimpact, 1959–98. Technical report, U.S. Bureau of Economic Analysis. Available at http://www.bea.gov/papers/pdf/software.pdf. Sinclair, J. and B. Catron (1990). An experimental price index for the computer industry. Monthly Labor Review 113(10), 16–24. 25

Triplett, J. E. (1989). Price and technological change in a capital good: A survey of research on computers. In D. W. Jorgenson (Ed.), Technology and Capital Formation, pp. 127–213. Cambridge: MIT Press. 26