The Spreadsheet Revolution

Abstract

A Harvard student no longer wanted to recalculate numbers on paper every time an assumption changed. What he built with a friend over two months in the winter of 1978–79 became the first “killer app” in computing history, a program that sold hardware rather than the other way around. The idea had older roots in accounting research and mainframe budgeting tools, and it went on through VisiCalc, Lotus 1-2-3, Multiplan and Microsoft Excel, through two landmark copyright cases, into the browser and, with the LAMBDA function, into full programming-language territory. Hundreds of millions of people now write spreadsheet formulas, and their mistakes have cost banks billions, lost COVID test results and forced geneticists to rename human genes.

Before VisiCalc

The electronic spreadsheet had a paper ancestor and several mainframe ones. In July 1961 Richard Mattessich, an Austrian accounting professor then at the University of California, Berkeley, published “Budgeting Models and System Simulation” in The Accounting Review. He described a firm’s budget as a set of linked matrices in which each cell was computed from others, so that a change in one assumption flowed through the whole model. In 1964 he followed with a book, Simulation of the Firm Through a Budget Computer Program, whose model was written in FORTRAN IV for a mainframe. Later reviewers credited the work with the basic principles of the spreadsheet: matrices, budget simulation, and a calculation behind every cell.

Commercial tools followed on time-sharing systems. AutoPlan, written in 1968 by three former General Electric employees, ran on GE’s time-sharing service; an IBM mainframe version was sold as AutoTab, and by 1975 Autotab-II advertised models of up to 1,500 rows and columns.

In the summer of 1969 Rene Pardo and Remy Landau, two recent Harvard graduates, wrote LANPAR (“LANguage for Programming Arrays at Random”). Bell Canada, AT&T and 18 US telephone operating companies used it for budgeting, and General Motors adopted it as well. Its key feature was forward referencing, now called natural order recalculation: a formula could refer to a cell that was computed later, and the program worked out the correct order itself. VisiCalc, a decade later, still recalculated row by row or column by column and could need several passes to settle.

Pardo and Landau filed a patent on the natural-order algorithm in 1970. The Patent Office rejected it as a purely mathematical invention; after twelve years of appeals they won at the Court of Customs and Patent Appeals in 1983 and received US patent 4,398,249. In 1995 a federal district court ruled the patent unenforceable for inequitable conduct, and the Federal Circuit upheld that in 1996.

None of these tools reached ordinary office workers. They ran on shared mainframes, were operated through printed reports and batch runs, and were priced for telephone companies. The missing piece was a screen on a desk.

The Paper Problem

Dan Bricklin was a student at Harvard Business School in 1978. Like all MBA students, he spent hours doing “what-if analyses” on large sheets of paper (the greenbar sheets): a company has these costs, that revenue, this growth rate. What happens if you change one assumption? You erase, recalculate, fill in, and discover that three cells further up you forgot a dependency.

Bricklin imagined the paper pad existing on a screen: cells that recalculate automatically when you change a number. He spoke with his friend Bob Frankston, an experienced programmer then living near MIT. Frankston was enthusiastic. They founded the Software Arts Corporation and split the work: Bricklin designed the interface and the logic, Frankston wrote the actual code.

Frankston programmed VisiCalc over two months in the winter of 1978–79 in 6502 assembler on the Multics time-sharing system, working at night, when computer time cost about a dollar an hour. He had hoped to fit the program into 16 KB; the finished code took about 20 kilobytes and needed an Apple II with 32 KB of RAM.

On October 17, 1979, VisiCalc went on sale, produced and distributed by the publisher Personal Software. The price: just under 100 US dollars (it had risen to 250 by 1982).

The First Killer App

What happened next had no precedent in the personal computer business: a software product drove the sale of hardware.

VisiCalc had not been written for hobbyist tinkerers or academics. It was built for accountants, financial analysts and entrepreneurs, people who worked with columns of numbers every day and who bought an Apple II because they wanted VisiCalc.

The Apple II had come to market in 1977; sales moved, but slowly. With VisiCalc they took off: more than a quarter of the Apple IIs sold in 1979 were reportedly bought to run it, a windfall for Steve Jobs and his company.

Info

Software drives hardware. VisiCalc was the first clear historical example of this principle. The Apple II was technically interesting, but VisiCalc made it necessary. The pattern repeated: Lotus 1-2-3 sold IBM PCs, Microsoft Office sold Windows PCs. The lesson (that a computer’s value comes from its software) sounds trivial, but in 1979 it surprised an industry that still thought hardware-first.

Within six years VisiCalc sold more than 700,000 copies, and roughly a million across all versions. Apple dealers said that VisiCalc had caused more business computer sales than all other software combined.

Bricklin and Frankston nevertheless profited less than hoped. They had not patented VisiCalc (software patents were considered legally uncertain at the time) and quickly faced imitators. Above all, the IBM PC arrived in 1981, and VisiCalc for the PC was a quick port, not a new development. A competitor seized the opportunity.

Multiplan and SuperCalc

The first imitators came from the other 8-bit world. SuperCalc, published by Sorcim in 1980, became the standard spreadsheet on CP/M machines (see Gary Kildall) and was bundled with the Osborne 1 (see Dead End: Osborne).

Microsoft answered in August 1982 with Multiplan, developed by Doug Klunder. Microsoft compiled it to its own portable p-code, which let the company ship about a hundred versions for different machines: CP/M, MS-DOS, the Apple II, the Commodore 64, the TI-99/4A, the TRS-80 line and the early Macintosh among them. Multiplan also used a different cell notation. Where VisiCalc wrote B3, Multiplan wrote R3C2, row 3, column 2, which made relative references easier to express. Users preferred the A1 style; Excel still offers R1C1 as an option, and almost nobody switches it on.

Multiplan sold more than a million copies, but on the IBM PC it lost to Lotus. By 1990 only 7 percent of the members of the American Institute of CPAs used it. Bill Gates later blamed the decision to spread the effort over so many 8-bit versions instead of concentrating on the 16-bit PC. Microsoft drew its own conclusion: the next spreadsheet would be written for a graphical interface.

Lotus 1-2-3 and the IBM PC Era

Mitch Kapor was, in the mid-1970s, a transcendental meditation teacher and psychology student who began taking an interest in computers as a hobby. He learned to program, worked briefly for Personal Software (the VisiCalc publisher), and made a precise observation: VisiCalc could tabulate, but it could not draw charts, query a database or produce reports.

Kapor founded the Lotus Development Corporation in 1982 and launched Lotus 1-2-3 on January 26, 1983. The name described the product: three functions in one program (spreadsheet, charting, database). The price was 495 dollars, twice what VisiCalc then cost, and it still sold 175,000 copies in its first year.

Jonathan Sachs, who wrote the program, spent ten months coding it for the IBM PC: directly on the hardware, in x86 assembly, exploiting every available clock cycle. On an IBM PC with an 8088 processor it computed in seconds what took VisiCalc minutes.

Lotus 1-2-3 dominated the market from 1983 into the late 1980s. It was the best-selling PC software product of its time: 3.5 million copies by 1988. With 53 million dollars in sales, Lotus was the third-largest microcomputer software company of 1983, its first year. Many people bought an IBM PC for Lotus, the way an earlier wave had bought the Apple II for VisiCalc.

The Look-and-Feel Lawsuits

Lotus defended its lead in court. Several competitors sold cheaper spreadsheets that used the 1-2-3 menus and keystrokes, so that users (and their macros) could switch without retraining. Lotus argued that this command structure was protected by copyright, even if not a line of its code had been copied.

The first test was VP-Planner, a low-cost clone from Paperback Software, the company of Osborne computer founder Adam Osborne. On June 28, 1990, Judge Robert Keeton of the federal district court in Massachusetts ruled for Lotus: the user interface of 1-2-3, he wrote, was “its most unique element,” and Paperback’s copying of the menus, commands and screen displays was “overwhelming and pervasive.” Before the damages trial, Paperback settled in October 1990: its insurer paid Lotus 500,000 dollars and VP-Planner was taken off the market.

The second case was harder. Borland’s Quattro Pro had its own interface but offered a compatibility mode with the 1-2-3 menu tree and a “Key Reader” that could run Lotus macros. Again no code had been copied. Keeton ruled for Lotus in 1993. In 1995 the Court of Appeals for the First Circuit reversed him: the menu command hierarchy was a “method of operation,” the means by which users control the program, and section 102(b) of the Copyright Act excludes methods of operation from protection. The court compared it to the buttons on a video recorder. On January 16, 1996, the Supreme Court split 4–4, with Justice Stevens recused, which affirmed the reversal without setting a national precedent.

By then the question had lost its commercial weight: both Lotus and Quattro Pro were losing to a program that had never needed to copy anyone’s menus. The case still matters for software law: it is a standard citation for the principle that an interface a user must learn in order to operate a program is not in itself protected expression (see Copyright and IP in the Digital Age). Borland’s developers from that period include Anders Hejlsberg.

Excel on the Macintosh

Microsoft Excel did not first appear for Windows or the IBM PC. It appeared on September 30, 1985, exclusively for the Apple Macintosh.

The project had the code name Odyssey. Jabe Blumenthal led the design and Doug Klunder, the principal developer of Multiplan, did much of the programming. Steve Jobs and Bill Gates announced it together on May 2, 1985, at a time when Apple needed business software to sell the Mac and Microsoft needed a platform where Lotus was absent.

Gates had recognized that the Macintosh’s graphical interface allowed a different kind of spreadsheet: one operated with the mouse, one that rendered fonts, one that placed charts next to the numbers. Lotus 1-2-3 was built for the character-based display of the IBM PC; on the Mac it had no counterpart. Excel also recalculated intelligently: only the cells that depended on a changed value were updated, instead of the whole sheet.

Excel was a success on the Mac and showed where things were heading: a spreadsheet that looked like a printed document instead of a terminal window.

Excel on Windows and Office

On November 19, 1987, Microsoft shipped Excel 2.0 for Windows, a month before Windows 2.0 itself. Few PCs ran Windows yet, so Excel came with a runtime version of Windows 1.0 that let customers without it run the program. Microsoft could develop Excel and Windows together, and each made the case for the other.

Lotus reacted late. Lotus 1-2-3 for Windows appeared only in 1991, four years after Excel for Windows, and was slow and buggy. Lotus had spent those years optimizing for DOS.

Excel 3.0 (1990) added toolbars, drawing tools, outlining and 3D charts, and Excel 4.0 followed in 1992. By the early 1990s Excel had started to outsell 1-2-3. Excel 5.0 in 1993 brought multi-sheet workbooks and Visual Basic for Applications (VBA), a full programming language inside the spreadsheet. VBA turned Excel into a platform: banks, engineering firms and government offices built applications in it that outlived the people who wrote them. It also made Excel a carrier for macro viruses later in the decade.

The contest was settled by Microsoft Office (1990). By bundling Excel, Word and PowerPoint into one package and selling it to companies at a discount, Microsoft displaced Lotus through distribution more than technology. IBM acquired Lotus Development Corporation in 1995 for 3.5 billion dollars, mainly for its Notes groupware, not for 1-2-3.

Dead End: Lotus Improv

In 1991 Lotus released a product that reviewers praised as conceptually ahead of everything that had come before: Lotus Improv.

The problem Improv set out to solve is real. A classical spreadsheet model is bound to its geometry. Rows are one thing, columns another, but an analysis often has three or four dimensions: quarterly revenue broken down by region, product category and sales channel. In the classical grid that forces either endless copying or hard-to-follow 3D worksheets.

Pito Salas, the lead designer, built Improv around items, categories and views. Instead of cells in a fixed geometry, the user arranged named data points in dimensions and then defined different views on that data, each view a different arrangement of the dimensions. Formulas were written in terms of names (“Profit = Revenue - Costs”) and applied to every matching item, instead of being copied cell by cell. The model separated the structure of the data from its presentation.

A financial analyst could display the same revenue data once as a quarter×region matrix, once as quarter×product and once as region×product, without copying anything.

Improv failed on switching cost. Users who knew Lotus 1-2-3 or Excel had mental models of rows, columns, formulas and cell references, and Improv discarded all of them. Anyone with years of work invested in Excel models had little reason to relearn.

Improv first appeared for the NeXT computer, Steve Jobs’ post-Apple platform, which sold fewer than 100,000 units in total. The Windows version came in 1993, when Excel already controlled the category.

Warning

Better loses to established. Improv was the more coherent design, and it lost to a grid that people already knew. The mechanism recurs across computing history: a technology that asks users to abandon their habits needs an advantage large enough to pay for the retraining, and Improv’s advantage showed only in large models that most users never built.

Lotus discontinued Improv in 1996. Its ideas lived on in OLAP cubes, business-intelligence tools, the pivot tables of later spreadsheets, and modern tools like Airtable or Notion. The spreadsheet grid itself stayed.

Bigger Grids and Formulas as Code

From Excel 97 to Excel 2003 a sheet had at most 65,536 rows (2¹⁶) and 256 columns; earlier versions stopped at 16,384 rows. Excel 2007 replaced the binary .xls format with the XML-based .xlsx (Office Open XML) and raised the limits to 1,048,576 rows (2²⁰) and 16,384 columns (2¹⁴). Old .xls files kept the old limit, which mattered again in 2020 (see below).

Later versions moved the formula language toward programming. Dynamic arrays, introduced from 2018, let a single formula return a whole range of results that “spill” into neighbouring cells. On December 3, 2020, Microsoft announced the LAMBDA function, which lets users define their own named functions, including recursive ones, in the formula language alone. With it, Excel’s formula language became Turing complete without any VBA. In 2023 Microsoft announced Python in Excel, which lets a cell run Python code; since September 2024 it has been generally available to business and enterprise customers on Windows.

Spreadsheets in the Browser

The first serious challenge to Excel’s desktop monopoly came from the web. XL2Web, a browser spreadsheet built by Jonathan Rochelle and Farzad “Fuzzy” Khosrowshahi at 2Web Technologies, was bought by Google in 2006. On June 6, 2006, it opened as a limited beta called Google Labs Spreadsheets. What it offered was not more features than Excel but simultaneous editing: several people working in the same sheet at the same time, with no files emailed back and forth. Google bought DocVerse in 2010 to improve Office compatibility and renamed the product Google Sheets in October 2012.

Microsoft responded with web and co-authoring versions of Excel, and the two products now compete more on collaboration and integration than on the grid itself, which has barely changed since 1979.

The Spreadsheet as a Cognitive Tool

The spreadsheet offers a metaphor that most people grasp at once: a table with rows and columns. The metaphor existed before the computer, as accounting journals on paper, and survived the move onto the screen as a digital grid. It lets non-programmers write executable logic. An Excel formula is code. A user who writes =IF(A2>B2,"Target met","No") is programming without ever using the word.

Spreadsheet users number in the hundreds of millions, far more than any programming language has. A 2022 survey found that two thirds of office workers opened Excel at least once an hour. Since 2021 the program even has a competitive scene: the Microsoft Excel World Championship sets players 30-minute modelling cases and eliminates the lowest scorers round by round. Its finals are held in Las Vegas and broadcast on ESPN’s alternative channel ESPN8: The Ocho; in December 2023 the Australian Andrew “The Annihilator” Ngai won his third title in a row and took home 3,000 dollars and a wrestling-style champion’s belt.

Errors in these spreadsheets have real consequences. In 2012, JP Morgan Chase lost 6.2 billion dollars in the “London Whale” trades. Its internal task force traced part of the failure to a value-at-risk model run in Excel with manual copy-and-paste steps, in which one formula divided by a sum instead of an average and so halved the reported risk. In 2013 a graduate student found that an AVERAGE formula in an influential Reinhart-Rogoff austerity paper had silently omitted five countries, turning a claimed -0.1% growth figure into +2.2% (see Fun Fact: The Excel Error Behind an Austerity Argument). In 2020, the English government lost 15,841 positive COVID-19 test results because a file in the old .xls format ran out of rows.

Biology produced the strangest case. Excel converts text that looks like a date into a date, so the gene symbol MARCH1 became 1 March and SEPT2 became 2 September. A 2016 study in Genome Biology checked more than 3,500 papers from 18 journals that shipped gene lists as Excel files and found such errors in about one in five. Microsoft did not change the default. In August 2020 the HUGO Gene Nomenclature Committee changed the genes instead: 27 human genes were renamed, MARCH1 to MARCHF1 and SEPT1 to SEPTIN1, and new symbols may no longer look like dates. In 2023 Microsoft finally added an option to switch the automatic conversion off.

Richard Mattessich, who had described the calculating budget matrix in 1961, died on September 30, 2019, thirty-four years to the day after Excel went on sale.