Part IThe Story
Minutes Before the Close
On the afternoon of July 8, 2019, a senior executive at a large technology company, which the Securities and Exchange Commission would later decline to name and call only "Tech Company A," got in touch with Andy Bechtolsheim. He wanted to ask the kind of question people had been bringing to Bechtolsheim for forty years. Should we buy this? The target was Acacia Communications, a manufacturer of high-speed optical interconnect products, the components that turn electrical signals into light and push them between the machines of a data center. Tech Company A and Arista Networks, the company Bechtolsheim chaired, were bound by a nondisclosure agreement, and so, the SEC would say, were the two men. The question also carried an answer. Tech Company A was weighing a bid because somebody else's acquisition of Acacia was imminent.
Minutes before the market closed, Acacia options were traded in two brokerage accounts. One belonged to a close relative of Bechtolsheim's and the other to an associate. The next morning, before the opening bell, Acacia and Cisco Systems announced that Cisco would pay $70 a share. Acacia's stock rose 35.1 percent that day, and the two accounts cleared a combined $415,726.
The buyer was Cisco. Twenty-three years earlier, in September 1996, Cisco had paid $220 million for Granite Systems, the Gigabit Ethernet start-up Bechtolsheim founded after leaving Sun Microsystems. He owned 60 percent of it. He then spent seven years inside Cisco building what his own biography would call "the highest volume modular Ethernet switching platform in history." The company whose deal he allegedly traded ahead of was, in a sense he could hardly have missed, a company he had helped build.
The settlement came in March 2024, and a federal court approved it that June. Bechtolsheim neither admitted nor denied the charges. He agreed to a permanent injunction, a five-year bar from serving as an officer or director of a public company, and a civil penalty of $923,740, roughly twice the alleged gain. The Register did the arithmetic with some relish. Against a net worth then estimated at $18 billion, the fine came to about 0.005 percent of his assets. For an ordinary 68-year-old American, the proportional penalty would have been $84.
The episode resists explanation, and Bechtolsheim has offered none in the public record. That is why it makes a useful place to start. It sits at the end of a fifty-year career built almost entirely on one faculty: seeing, slightly before anyone else, what is about to matter. He saw it in a Motorola 68000 and a few circuit boards in 1979. He saw it in a search engine run by two graduate students in 1998. He saw it in commodity switching chips in 2009, and in the electrical appetite of artificial intelligence a decade after that. He has spent much of that half-century telling audiences that the faculty is not what they think it is. Being first, he likes to say, is not the point. Being first to solve a relevant problem is.
On July 8, 2019, nothing was solved. Something was simply known.
By the Numbers
The Bechtolsheim Ledger
~10SUN workstations built at Stanford in 1981–82 before the university declined to build more
$100,000Check written to Google in September 1998, before the company was incorporated
$220MCisco's price for Granite Systems in 1996; Bechtolsheim owned 60%
$25B+Commercial sales of systems he had designed, by his own 2006 accounting
~14%Stake in Arista Networks held directly or through his family trust, per a 2023 filing
$923,740SEC civil penalty in the 2024 insider trading settlement
$28.9BEstimated net worth in 2025, against $7B in 2018
A House With No Other Houses in View
As a small boy, he did believe in Santa Claus. He also wanted to make sure Santa got the order right. So he wrote to him with a wish list that gave the exact SKU numbers of the experimental kits he wanted. "Magically," Bechtolsheim told the Computer History Museum's Doug Fairbairn in 2015, "they showed up under the Christmas tree." Fairbairn laughed and said something about leaving nothing to chance. Bechtolsheim answered with the line that may be the most complete self-portrait he has ever given: "Yes, one has to plan ahead."
The house where he made that plan stood at Hängeberg am Ammersee, a hamlet now absorbed into the municipality of Finning in Bavaria's Landsberg district. The family behind it was old nobility. His full name is Andreas Maria Maximilian Freiherr von Mauchenheim genannt Bechtolsheim, a baron's name that takes nearly a line of type. The estate itself was more modest than the name. His father, Heinrich Freiherr von Bechtolsheim, who lived from 1924 to 2018, taught elementary school. His mother kept the house. The family did not work the farm and leased it to a farmer. Andreas, born September 30, 1955, was the second of four children.
"The farm was a mile away from the next village," he told Fairbairn, "and we couldn't see another house from our house. We looked out on beautiful meadows, forests and the Alps in the distance." The house had no television. It had no close neighbors either, and in a different child that might have produced loneliness. "There wasn't much interaction with other people," he said, "but that did not bother me since I was mostly interested in understanding how things worked."
That last sentence works as an origin story because it so carefully declines to be one. There is no wound in it, no hunger and no chip on the shoulder. A boy who could not see another house turned his attention to the inside of things. By four or five he was taking objects apart. One of them was his father's reel-to-reel tape recorder, an expensive machine in a schoolteacher's household. "He was getting kind of worried," Bechtolsheim recalled, as the boy spread its parts across the floor. At six he was winding small electric motors from wire and magnets. He wanted to know why a compass needle always swings north, and how phosphor paper glows in the dark without a battery. The questions came before the transistors, which reached him in the late 1960s. When they arrived, the basement of his parents' house became what he called his "hobby department."
The geography kept changing, and the solitude went with him. In 1963 the family moved to Rome, where his father taught at a German school abroad and, for a time, taught his own son as well. Before that Andreas had attended a one-room village school. In 1968 they settled at Nonnenhorn, on the Bavarian shore of Lake Constance near Lindau. He finished his Abitur at the Bodensee-Gymnasium in Lindau in 1973. He had been born beside one lake and raised beside another. Much later, according to Fortune, he would settle in Nevada near a third, Lake Tahoe. Whether that means anything is up to the reader. The record shows only that the water keeps appearing.
At sixteen he designed an industrial controller for a nearby company, built around the Intel 8008, one of the first 8-bit microprocessors. He had no assembler, the program that turns human-readable mnemonics into machine code, so he wrote the software directly in binary, in ones and zeros entered by hand. The company paid him royalties, and the royalties paid for much of his education. This is the first appearance of a pattern that would define his life. He designed a thing, someone else manufactured it, and a stream of money flowed back that let him design the next thing.
He entered Jugend forscht, Germany's national science competition for young researchers, and lost. He entered again and lost again. On the third attempt, in 1974, he won the national physics prize. That same year, by his own later account, he wrote his first serious program: a neural network simulator running on a PDP-11 with 64 kilobytes of memory. "We could only model 256 nodes," he said, forty-four years later, at a Goldman Sachs conference where he had just told the room that artificial intelligence now consumed most of the compute cycles at Google.
He enrolled at the Technical University of Munich on a scholarship from the German Academic Scholarship Foundation, the country's most selective merit award. He studied electrical engineering with a focus on data processing, earned his degree, and was miserable throughout. The coursework bored him, and the access to computers was meager. An older version of his Wikipedia entry puts it more bluntly: he was "bored and angry with the level of his studies." In 1975 a Fulbright scholarship carried him to Pittsburgh, and he finished a master's in computer engineering at Carnegie Mellon in a single year. There he designed and built the network interface for Cm*, a multiprocessor research machine. A later biography called it "perhaps the first microprocessor based compute cluster," and the perhaps is a nice touch, the hedge of an engineer who knows someone will check.
The Megapenny Machine
He came to Silicon Valley in 1977 for a job and stayed in a classroom. Justin Rattner, then at Intel, had offered him an internship. When Rattner relocated to Oregon, Bechtolsheim declined to follow and enrolled at Stanford as a doctoral student in electrical engineering instead. Plenty of other Stanford stories turn on a choice between Intel and the university. In this one the choice was nearly accidental, and the consequences were very large.
Stanford in the late 1970s was where three technologies were meeting, and Bechtolsheim placed himself at the junction. The first was very-large-scale integration, which had begun to fit whole subsystems onto a single chip. The second was Multibus, an industry-standard backplane that let circuit boards from different vendors share an enclosure. The third was electronic computer-aided design, the software that let one person draw, verify and lay out a board that would once have needed a team.
The fourth ingredient came from up the road. In 1979 Xerox donated a number of Alto computers from its Palo Alto Research Center to Stanford's computer science department. The Alto was the machine that had more or less invented personal computing as we now know it: a bitmapped screen, a mouse, a local Ethernet connection to its neighbors. It was also very expensive, and Xerox never sold it commercially. Bechtolsheim negotiated an odd arrangement at PARC as a "no fee consultant." He was paid nothing and in exchange could see everything. Lynn Conway was there at the time, using workstations to design VLSI circuits. He watched her do it.
His adviser, Forest Baskett, suggested the configuration that would make the project famous. Baskett was a Stanford professor whose name now appears mostly in footnotes, though the footnotes are for things like this. The idea was a low-cost personal workstation for computer-aided logic design. The specification became known as the "3M" machine: one million instructions per second, one megabyte of memory, and a one-megapixel bitmapped display. Someone added a fourth M. The estimated price was $10,000, and wags called it a "Megapenny."
Ralph Gorin, Stanford's director of computer facilities, put up the initial money. Vaughan Pratt, a theoretical computer scientist who would later design one of the most elegant corporate logos ever made, became the project's unofficial faculty leader in 1980. DARPA supported it. The name came from the campus network the machines were meant to sit on: the Stanford University Network. SUN.
The hardware was Bechtolsheim's. He designed the boards on the Stanford Artificial Intelligence Laboratory's aging computer using SUDS, the Stanford University Drawing System. SUDS had originally been written for a machine called the Foonly, which is the sort of detail that tells you how improvised all of it was. He verified the designs with a package called SCALD, which also automated the layout and produced first wire-wrapped prototypes, then printed circuit boards. The graphics display controller came first, and he and Baskett published it in 1980. Next came the main CPU board, built around Motorola's 68000, with memory and parallel and serial controllers. The third board was an interface to Stanford's experimental Ethernet, which ran at 2.94 megabits per second because the standard had not yet settled on ten. He built the first machines from spare parts scavenged from the computer science department and bought from Silicon Valley supply houses.
The SUN workstation turned out to be a seed as much as a computer. Because it was modular and built on Multibus, other people could combine its boards into things Bechtolsheim had not designed.
Pair the CPU board with a multiport serial controller and you had a terminal server, which Stanford called a TIP. It connected rows of terminals to the university's Digital Equipment time-sharing machines or to anywhere on the young internet. Pair the CPU board with several Ethernet controllers and you had a router. William Yeager wrote the routing software, and Cisco Systems later adopted and evolved it. Cisco was founded by two Stanford support engineers, Leonard Bosack and Sandy Lerner, a married couple who commercialized the campus's gateway server. Les Earnest licensed the CPU board for one of the first low-cost laser printer controllers, at a company called Imagen. Students of a Stanford professor named James Clark joined the processor board to a prototype high-performance graphics display, and that group became Silicon Graphics.
So the board sits in the family tree of at least three companies that defined the next two decades: Sun, Cisco and Silicon Graphics, along with assorted cousins. Its designer was a German graduate student without a doctorate, and he would never finish one.
About ten SUN workstations were built at Stanford during 1981 and 1982. Then the university declined to build any more. Stanford was not a manufacturer and had no wish to become one.
They Had All the Pieces
What Bechtolsheim did next, before Sun existed, is the part of the story most often skipped, and it may be the most revealing. He went into business, though not the business people remember. He licensed the board designs to outside companies in exchange for royalties, as he had done at sixteen with the Intel 8008 controller, and set up arrangements to have the boards manufactured. It was a reasonable plan for a hardware engineer. The engineer designs the system, manufacturers build it, the market buys it, and royalties come back.
The plan failed in a specific and maddening way. He described it in 2007 to InfoWorld's Paul Krill, and twenty-five years had not dulled the irritation.
— Andy Bechtolsheim, interview with InfoWorld, 2007It dazzles me that none of these companies had any clue about the workstation business. They had all the pieces, but they didn't see the market opportunity, and I was kind of getting frustrated because I was really more the engineer, the hardware guy, than the businessman.
They had all the pieces. The licensees could build a SUN board, but they could not see the SUN workstation. They treated the design as a component when it was really a category. The man who could see the category did not consider himself a businessman.
Vinod Khosla could see the business. He was a Stanford MBA working at Daisy Systems, one of the companies building computers for VLSI design, which made him precisely the kind of customer the SUN board had been born to serve. He called Bechtolsheim. As Bechtolsheim remembered it, Khosla's pitch had the brevity of a man who had already done the math: "Let's just start a company." Khosla brought in a classmate from business school, Scott McNealy, who was managing manufacturing at a Unix start-up called Onyx Systems. McNealy was a hockey-playing operator, later described by Fortune as "notoriously cheap," and the description was meant as praise. The three wrote a short business plan. Venture capitalists, Kleiner Perkins among them, funded it quickly. On February 24, 1982, Sun Microsystems was incorporated, and Bechtolsheim left Stanford without his doctorate.
The fourth founder came from across the Bay. Bill Joy had spent years at Berkeley on the Berkeley Software Distribution of Unix, writing the vi editor and the C shell along the way and doing significant work on the TCP/IP networking stack. He was the most important Unix hacker of his generation, and he arrived the following week. Bechtolsheim, a stickler for the record, later explained the badge numbers. He was number one. Two employees hired on day one took the next slots, and Joy, arriving a few days later, got number six. "But he was considered a founder," Bechtolsheim told Krill, and then argued against his own seniority with the fairness of a man correcting a spec. He had badge one only because he had worked on the workstation for three years before the company existed. "Although I think Bill worked on Berkeley Unix longer than that. So maybe he should have been number 1."
For a while Bechtolsheim and Joy shared an apartment in Palo Alto. It is hard to imagine what that kitchen sounded like.
The first product, the Sun-1, used the Stanford CPU board with improved memory expansion inside a sheet-metal case. It shipped with UniSoft's version of Unix until Sun's own SunOS, a customized Berkeley Unix, replaced it later that year. One small detail from the Sun-1's first year deserves notice. By year's end, the experimental Ethernet interface Bechtolsheim had designed was replaced by a commercial board from 3Com. He did not insist on his own part, because a better part existed that someone else would build in volume. The 2009 essay that would define his late career was already present here as a reflex, nearly three decades before he wrote it down.
Sun was profitable from its first quarter, which ended in July 1982. It was only the second company, after Apollo Computer, to specialize in workstations. Its machines were the only computers besides DEC's VAX that ran Berkeley's 4.2BSD Unix, and they found early markets among chip designers, engineers and, unexpectedly, Wall Street. The company went public in 1986 under the ticker SUNW. It reached $1 billion in sales in 1988, the same year AT&T announced it would buy up to a 20 percent stake. By early 1989 InfoWorld called Sun "the unqualified leader in the workstation arena."
Vaughan Pratt designed the logo: four interleaved copies of the word sun arranged as a rotationally symmetric ambigram, so that it reads the same from any side. It was orange at first, with its edges horizontal and vertical. Later it was turned to stand on one corner and recolored purple, then blue. The marketer John Gage coined the company motto, "The Network is the Computer." It was a slogan that described what Bechtolsheim's third board had been for, years before most of the world had a network to plug into.
Inside Sun, Bechtolsheim held the title of Vice President of Technology from 1982 to 1995 and led workstation development. Around 1988 he started a project code-named UniSun to build a small, inexpensive desktop machine for universities. Internally it was called "campus." It shipped as the SPARCstation 1, which began a product line that filled engineering departments and trading floors for a decade. That made two Megapenny machines for the same constituency, ten years apart.
All the Wood Behind One Arrowhead
McNealy had a phrase for strategy, and the phrase was "all the wood behind one arrowhead." He used it in the early 1990s, when Sun retired its last Motorola-based systems and its short-lived Intel 386 hybrid, the Sun386i, to concentrate on SPARC. SPARC was the company's own RISC processor architecture, in its systems since 1987 and extended to 64 bits in 1995. It was a disciplined decision. It was also a bet that Sun's proprietary silicon would keep outrunning everybody else's.
Bechtolsheim did not believe the bet. "One reason I left in 1995," he told Krill, "is that one of my concerns at the time was that we should be really building an x86-based server, and I was getting worried that the cost performance of those systems was edging up with the SPARC architecture."
No public record from that period shows a fight. Nobody slammed a door. A founder who had designed the company's foundational product quietly noticed a cost curve bending toward Intel and declined to stay and argue about it. He left to build something else, and he took the curve with him.
The something else was Granite Systems, a Gigabit Ethernet switching start-up he co-founded in 1995 with David Cheriton. Cheriton was a Stanford professor who had done research on SUN workstations in the early 1980s, a systems thinker as rigorous as Bechtolsheim, and from this point on he appears again and again in the story, often enough to count as a structural element. Bechtolsheim was CEO and president. Cisco announced the acquisition on September 3, 1996, and closed it on September 27, paying $220 million. The press release described a team headed by Bechtolsheim developing multilayer Gigabit Ethernet switching to relieve congestion in network backbones as users, applications and servers multiplied. That was the logic of the SUN workstation turned inside out. Making each computer faster had created a traffic jam between them, and the fix was to make the roads faster.
He spent seven years at Cisco, first as vice president of engineering and then as general manager of its Gigabit Systems Business Unit. Under him the unit developed the Catalyst 4000 and 4500 family of modular switches. Those were the boxes enterprises bought by the thousands and forgot about, which in infrastructure is the highest compliment available.
Meanwhile Sun had its moment and then lost it. During the dot-com bubble its stock rose as high as $250 a share, as start-ups bought racks of SPARC servers for business that would never arrive. In 2000 it paid $2 billion for Cobalt Networks, a maker of Internet appliances. When the bubble burst, customers shut down and auctioned their high-end servers, and Sun's hardware sales went into freefall. By December 2001 the stock was back at its 1998 level. A year later it was below $10. Layoffs followed in waves. Linux on cheap Intel and AMD boxes took the low end of the server market. The arrowhead had found its target, and the target had moved.
In early 2001, while still at Cisco, Bechtolsheim co-founded Kealia with Cheriton to build advanced servers around AMD's 64-bit Opteron processor. Those were x86 servers, the machine he had wanted Sun to build six years earlier. He left Cisco in December 2003 to run it. In February 2004 Sun announced it would acquire Kealia in a stock swap, and the deal closed on April 12. ZDNet's headline put it plainly: "Sun to buy Opteron server maker, reclaim co-founder." He came back as senior vice president and chief architect.
"Here I am 10 years later or 12 years later," he told InfoWorld, "actually doing what I suggested the company should be doing earlier." People kept asking him whether the reunion had been a prearranged marriage. He let the question hang.
At Sun the second time, he led a team that designed a new range of Opteron servers from scratch. The Sun Fire X2100, X4100 and X4200 were unveiled in September 2005 and built to address the heat and power consumption problems already strangling data centers. The X4500 followed in July 2006. It was a storage server built on Kealia's technology, dense enough that high-energy physics installations used it for their data. He was early once more, and once more the timing did not save the company. In April 2009 Oracle agreed to buy Sun for $7.4 billion, or $5.6 billion net of cash and debt, and the deal closed on January 27, 2010. At the merger price of $9.50 a share, McNealy's 14,566,433 shares were worth $138 million. The operator who had held the arrowhead was paid out in full, at a fraction of what the arrowhead had once been worth.
The coda belongs to someone else. Facebook bought Sun's former Menlo Park campus, a million square feet built around an enclosed courtyard, which its employees promptly nicknamed "Sun Quentin." Mark Zuckerberg did not replace the sign out front. He had it turned around and put Facebook's logo on the other side. The old Sun logo stayed on the back, by one account, as a reminder of "what happens when you take your eye off the ball."
By then Bechtolsheim was long gone. He had left in October 2008 for a company whose name, for a while, had been Arastra.
A Check Made Out to No One
In September 1998 Bechtolsheim wrote a check for $100,000. David Cheriton wrote another for the same amount. The recipients were two Stanford graduate students, Larry Page and Sergey Brin, who had built a search engine that ranked pages by the links pointing at them. Bechtolsheim made his check out to Google Inc. There was no Google Inc. The company had not been legally incorporated, which meant the check could not be cashed until it was. Wired ran the story under a headline that is still the best summary of it: "If the Check Says 'Google Inc.,' We're 'Google Inc.'"
The scene has since been retold so often that it has drifted toward myth. A version on the My First Million podcast, typical of the genre, has Bechtolsheim answering a morning email from the students, meeting them at 8 a.m. the next day, running a single search, walking to his Porsche for a checkbook and announcing that it was "the best idea I've ever seen." The podcast also quotes a $10 million valuation. Neither the Porsche nor the valuation appears in the documentary record cited here. Something that does appear is a correction: the persistent claim that Bechtolsheim coined the name "Google" is untrue. He simply pushed the founders to organize the company under the name they already had. The check worked like a deadline. A company existed in practice, and he forced it to exist on paper.
Deutsche Welle's interview with him ran under a headline in his own words: he invested in Google "to solve my own problem." The phrase is characteristic. He needed better search, and here was better search. The investment was less a prophecy than a purchase order.
The return was enormous. Chris Yeh, a Silicon Valley entrepreneur, later wrote that Bechtolsheim made "$1.7 billion as the first investor in Google." Yeh also noted with some wonder that this put founding and selling Granite Systems only third on Bechtolsheim's highlight reel. The sum mattered less than the reputation it gave him, which has followed him around ever since. Stanford's School of Engineering, in its 2012 citation naming him an Engineering Hero, wrote that "Bechtolsheim's technology foresight is legendary." The event copy for an October 2025 keynote put it more breathlessly: "with Andy's track record, his guesses are better than most people's facts."
The record of those guesses is long, technical and mostly unglamorous. He was an early-stage investor in VMware. Diane Greene, its co-founder, has described him in public talks as one of the technically sophisticated angels who gave the company seed money, $100,000 after a meeting of about an hour. He backed Mellanox and Brocade, both makers of the fabrics that connect computers to storage and to each other. He invested in Magma Design Automation and in Co-Design Automation, the start-up that developed SystemVerilog, the language now used to design almost all digital hardware. Those last two are electronic design automation companies, which means software for chip designers. A man who once drew his circuit boards in SUDS on a Foonly-derived system kept financing better pencils.
Some of his angel investing looks less like category bets and more like serial loyalty. George Haber had been a colleague at Sun. Bechtolsheim invested in Haber's start-ups CompCore, bought by Zoran; GigaPixel, bought by 3Dfx; Mobilygen, bought by Maxim Integrated Products in 2008; and CrestaTech, a wireless chip company, in 2006 and 2008. He also backed Moovweb, a mobile web company, in 2009. He put money into Tapulous, a maker of music games for the iPhone that Disney bought in 2010, and into PerimeterX, which defends websites against automated attacks, between 2015 and 2017.
Not all of it worked. He was reportedly an early investor in Claria Corporation, an adware company the New York Times described in a 2004 story under the headline "Barbarians at the Digital Gate." Claria ceased operating in 2008. In 1999 he and Bill Joy founded a venture firm, HighBAR Ventures, which the general histories mention briefly and then let drop.
In February 2018 he spoke at the Goldman Sachs technology conference about artificial intelligence. "In normal IT, you might see 10-20% improvement per year," he said. "Here you see an order of magnitude improvement each year." He added that analog AI chips would draw the least power, because "the human brain is remarkably power efficient, and it is analog." The venture capitalist Steve Jurvetson was in the audience and had invested in Mythic, a start-up building analog AI chips. He introduced the company to Bechtolsheim. They met the same week, and Bechtolsheim invested.
Small enough to fit in a pocket, there is now a book on the subject, Andy Bechtolsheim and the Seeds of Silicon Valley, whose subtitle asks the question everyone asks: What Did He See in Google Before the Rest of the World Even Knew Its Name? The answer suggested by the record is less mystical than the question. He saw a machine that worked better than the alternatives. He saw it because he was the kind of user who would notice, and he paid before the paperwork existed because to him the paperwork was the least interesting part.
Economics Always Wins
The paperwork for Arista begins in 2004 or 2005, depending on whose records you consult. His oral history dates the founding to 2004, and most other accounts give 2005. It began as Arastra, co-founded with Cheriton, which made it the third company the two men had started together. It was renamed Arista Networks, and in October 2008 Bechtolsheim left Sun to become its chairman and chief development officer. He continued, he said, to advise Sun.
His partner at Arista was Jayshree Ullal, who became its president and chief executive. In 2015 Ullal and Bechtolsheim were jointly named EY's U.S. Entrepreneur of the Year. The division of labor matched the one at Sun. Someone else ran the company, and he defined what it built. Once again he was the hardware guy rather than the businessman, though the hardware had changed in a way that would have surprised his younger self.
On March 27, 2009, Bechtolsheim published a short essay on Arista's corporate blog called "The Silicon Choice for Cloud Networking." Its subject was technical: whether switches should be built on proprietary custom chips, known as ASICs, or on standard "merchant silicon" bought from chip companies. Underneath the technical question was a confession.
— Andy Bechtolsheim, "The Silicon Choice for Cloud Networking," Arista Networks blog, March 27, 2009In my own career, I have led the development of several generations of very high volume network switch silicon. However, even I could not design better silicon switch chips than what is available now on the merchant market.
It is an unusual sentence for a legendary engineer to write. He had built the Catalyst switch chips at Cisco, and he was declaring that the competitive advantage he had spent a decade constructing had evaporated. He laid out the reasons with his usual precision. In the 1990s a switch chip could be built for a few hundred thousand dollars, and at finer geometries a full custom chip now cost tens of millions. Switch architecture had stabilized so far that differentiation at the silicon level was no longer possible. Customers had grown wary of being locked into proprietary features in what he estimated as a $16 billion LAN switching market. He compared the moment to the one that had pushed him out of Sun. In the 1980s and 1990s a dozen CPU architectures flourished, and the cheap, ubiquitous x86 swallowed most of the volume. "It may not have had all the capabilities of RISC," he wrote, "but it delivered the best cost-performance." The same fork, he argued, now ran through networking. "It will take a while for this to play out, but the outcome is inevitable. The history of silicon technology has demonstrated that, over the long term, economics always wins."
Over the following years he kept extending the argument. In 2011 he wrote that a single merchant chip could replace a chassis full of proprietary ones: "Fewer chips means faster speeds, lower latency, less power consumption, lower cost, and higher reliability." He told Light Reading in 2012 that with an advanced chip costing around $20 million to design, "if you can't sell 100,000 chips, it's hardly worth doing." By 2018 he was describing Arista, a company that sells physical boxes, as something else: "over 90 percent of our R&D headcount is in software… we are a software company, not really a hardware company."
The hardware guy now ran a software company. The man who had designed the Megapenny's boards by hand now argued that nobody should design their own boards anymore, at least not the part that used to matter. The constant was not the medium but the method. Find the commodity, ride its volume curve, and put the differentiation in the layer where differentiation still lives.
He was skeptical of fashions, sometimes bluntly. At a Tech Field Day session in March 2012, at the height of excitement over OpenFlow and software-defined networking, he observed that OpenFlow had no significant production deployments. He grouped it with TRILL and FCoE, standards announced years earlier that customers had largely ignored. His explanation was that customers have existing networks and dislike architectures that will not interoperate with them. He favored ECMP, equal-cost multipath routing, because it was mature, it was reliable, and it already worked. In September 2017 he forecast that 400-gigabit Ethernet ports would reach parity with 100-gigabit ports around 2021. He preferred the 100-gigabit-per-lane approach of an industry consortium called the 100G Lambda MSA, which Arista belonged to, over the slower and less practical IEEE proposals. In October 2018 he explained why 400G mattered in a single blog sentence: "Because the easiest way to go faster is to go faster."
When Timothy Prickett Morgan of The Next Platform interviewed him in March 2022 about co-packaged optics, an approach that moves optical components onto the same package as the switch chip, Prickett Morgan described the experience as being "like being able to ask Shakespeare to tell you what he really meant in Hamlet." Bechtolsheim cut through the hype at once. Co-packaging was not a new kind of optics, he said, and not a new wavelength or a new interface. It was a packaging technique, and it existed for one reason. "The goal is always power reduction. In other words, if there was no power reduction, nobody would talk about co-packaged optics." He then broke the projected 20 percent saving into four buckets of power draw and explained why some of them would not shrink at all.
The market rewarded the method. In March 2015 he reported, with an engineer's flat satisfaction, that a Crehan Research report showed Arista gaining share in data center switching. By 2023 he held about 14 percent of the company directly or through his family trust. His estimated net worth was $7 billion in September 2018, about $18 billion in March 2024, and $28.9 billion in 2025. That made him, by Bloomberg's and Forbes's reckoning, roughly the sixty-eighth richest person in the world. The cause of the climb was the thing he had first simulated on a PDP-11 in 1974. Artificial intelligence turned out to be, among other things, an enormous problem of moving data between machines without melting the building. By 2025 he was traveling the conference circuit with keynotes on AI "centers of data" and explaining on podcasts why packet loss and traffic scheduling now determined how fast models trained. Arista's website described his job as responsibility for "advanced AI, silicon, and optics initiatives."
The Man in Sandals
He goes by Andy. On Arista's blog, the bylines on a single page alternate between "Andreas Bechtolsheim" and "Andy Bechtolsheim," as if the company could never quite settle the question. The baronial von shows up in German biographies and on the Forbes list, and almost nowhere else.
In 2002 Chris Yeh, then running a founders' group for the SDForum, was riding a shuttle bus to the organization's Visionary Awards. He sat beside a friendly man in a conservative suit with a pleasant German accent. They talked about blogging and emerging technologies, and the man "listened attentively and appreciatively" to Yeh's opinions. At the end of the ride Yeh introduced himself. "I'm Andy," the man said. As they stepped off the bus, he pointed down at his feet. He was wearing sandals with the suit. "I can't stand those dress shoes," he said. Only later did Yeh realize that his seatmate was one of the honorees. (Yeh spent the rest of that night with Jonathan Abrams, who told him about a new project called Friendster. Silicon Valley rarely wastes an anecdote.)
The sandals capture a whole manner: formal on the outside and ruthlessly practical underneath. He does whatever the occasion demands except where the demand is uncomfortable and pointless, and then he quietly declines. He shows the same disposition toward technology. He wears the industry standard on the outside and puts his own architecture underneath.
In May 2012 he returned to Stanford to give his Engineering Hero lecture, titled "The Process of Innovation." The startup chronicler Hervé Lebret got up at 4 a.m. for it. Bechtolsheim argued that innovation was not a matter of R&D budgets. Apple spent less on R&D than its peers, he noted, and did not really study its customers. He said innovation was not even about marketing. It was about bringing a needed product to customers. He listed the ways new products fail: too early, too late, too difficult, not relevant, too expensive. He offered a theory of why start-ups win. In Yeh's notes it reads: "A startup succeeds because of focus. It is spending 100% of effort on a problem that the big companies have overlooked." Lebret recorded a remark on failure that was half sociology and half memoir from a man who had left a country where it is hard to start over. In Silicon Valley, Bechtolsheim said in Lebret's paraphrase, not trying is the risk. Elsewhere, if you fail, you may have to hide for thirty years.
He did not hide, and he did not fully assimilate either. He has lived most of his life in the United States and never applied for American citizenship. He remains a German national. In November 2025 the German American Business Association of Northern California chose him as the inaugural guest of a series called "Living Legends." Nearly two hundred people came. So did Germany's consul general in San Francisco, Theo Kidess, and a group of computer engineering students from the German International School of Silicon Valley, whose questions, the organizers reported, "added incredible energy to the evening." He told them that a decade of extraordinary technological acceleration was beginning, and that they should learn fast and make things happen rather than wait for the perfect moment.
He was a German boy from a house with no neighbors, talking to German children in California about the network that now connects every house.
The Trust of a Longtime Contact
The SEC's complaint was precise in the way that legal documents are and that engineering documents should be. Bechtolsheim had learned of the impending deal "through his and Arista Networks's longstanding relationship" with the other company. That company "consulted with Bechtolsheim concerning the potential acquisition." He knew, or was "reckless in not knowing," that the information was material and nonpublic. "Immediately after learning this information," he allegedly traded. The price of an Acacia option before the announcement was the price of not knowing. Afterward it was the price of knowing.
— Joseph Sansone, Chief, SEC Market Abuse Unit, March 26, 2024We allege that Bechtolsheim, while serving as the chairman of a publicly traded company, abused the trust of a longtime business contact who had shared highly sensitive information about an imminent corporate acquisition.
The word in the SEC's statement that stays with you is trust. Bechtolsheim's career had run on a particular kind of trust, the confidence that his judgment about technology was better than other people's. Executives call him for that reason. Founders take his checks, even uncashable ones, for that reason. A senior executive of Tech Company A presumably called that afternoon for that reason too. The SEC's allegation was that the trust had been spent.
Arista moved carefully. The company told Fortune that his role had shifted to a nonexecutive one in November 2023, and the complaint noted that he had resigned as chairman and chief development officer that December, three months before the charges became public. "While the SEC announcement did not involve any trading in Arista securities," a spokesperson said, "Arista takes compliance to the company's code of conduct and insider trading policy seriously." He stayed on as chief architect. Asked after the court approved the settlement in June 2024, Arista declined to say anything beyond that.
Cisco completed its acquisition of Acacia in March 2021 for $4.5 billion. Acacia makes optics. By 2022 Bechtolsheim was explaining to The Next Platform exactly which watts in an optical module could be saved and which could not.
None of this has an explanation, at least not one available to an outsider working from the public record. Neither the complaint, the settlement nor anything Bechtolsheim has said in public explains why a man with billions, whose penalty would come to the equivalent of $84, would risk his name for $415,726 split between two accounts that were not his. It may have been carelessness, or something worse, or something smaller and stranger than either. The record does not say. It says only that a man who had spent his life planning ahead, writing the SKU numbers into the letter to Santa so that nothing would be left to chance, acted within minutes of hearing something he should not have used.
He had grown up in a house from which you could not see another house, and he spent fifty years building the machines that let every house see every other one. The most valuable thing those machines carry is information, which travels faster than it can be judged. He knew this better than almost anyone. He wrote it into a blog post in 2018: the easiest way to go faster is to go faster.
The tape recorder was a reel-to-reel model, and it belonged to his father, a village schoolteacher with a baron's name who lived until 2018. The boy had taken it apart on the floor of the farmhouse while his father watched, worried, from somewhere nearby. Then the boy put it back together. "And it still worked," Bechtolsheim told the interviewer, sixty years later.
Part IIThe Playbook
The principles below come from a single pattern repeated across five decades. Bechtolsheim sees a cost curve bending, builds the machine that sits on the right side of the bend, and lets someone else run the company that sells it. Some of these lessons are transferable to almost any builder or investor. One of them, the last, is a warning. Each is drawn from the evidence in Part I rather than from the mythology that has grown around him.
The principles
- 1.Build the tool you need, then notice who else needs it.
- 2.Assemble from commodity parts; compete on architecture.
- 3.When licensees can't see the category, become the company.
- 4.Leave over a cost curve, not a quarrel.
- 5.Be willing to obsolete your own expertise in public.
- 6.Write the check before the paperwork exists.
- 7.Back people serially, not deals singly.
- 8.Treat project selection as the main event.
- 9.Follow the watts.
- 10.Prefer the standard that already interoperates.
- 11.Separate the architect from the operator.
- 12.Build a wall around your information edge.
Principle 1
Build the tool you need, then notice who else needs it
The SUN workstation started as a personal CAD machine. Baskett's "3M" specification described the computer a Stanford logic designer would want on his own desk. Two decades later Bechtolsheim framed his first Google check in the same terms, as an attempt "to solve my own problem." In both cases the product was validated before any market research, because the first customer was also the designer and a demanding one.
The discipline is in the second half of the sentence. Many engineers build the tool they need, and few notice when the tool turns out to be a category. Bechtolsheim noticed because other people kept building things with his boards that he had not designed. There was a terminal server, a router that became Cisco's, a laser printer controller and the graphics prototype that became Silicon Graphics. Those derivative uses showed that the board was general-purpose. Watching them was a form of market research.
Builder-users have a structural advantage in judging quality, because they have a baseline. A user who has suffered with bad search knows good search in one query.
Tactic: Keep a running list of the tools you built for yourself, and watch for anyone who repurposes them without being asked, since that is your strongest early signal of a market.
Principle 2
Assemble from commodity parts; compete on architecture
Sun's founding design was a merchant microprocessor (the Motorola 68000), a standard bus (Multibus), a licensed operating system (Berkeley Unix) and a bitmapped display. Within its first year Sun swapped Bechtolsheim's own Ethernet board for a commercial one from 3Com. Twenty-seven years later he argued the same thing about switches, that merchant silicon would beat custom ASICs because "economics always wins."
The insight is that originality and proprietary components are different things. The value of the SUN workstation was never in any single chip. It was in the integration, in putting networking, graphics and Unix together in one affordable box. Arista applied the same logic a generation later. It bought chips from the best of "at least five silicon companies" and put 90 percent of its R&D headcount into the software that made those chips behave like a coherent network.
⚙
The Commodity Thesis, Twice
The same argument, made about computers in the 1980s and about switches in 2009.
| Dimension | Sun, 1982 | Arista, 2009 onward |
|---|---|---|
| Core processor | Motorola 68000, off the shelf | Merchant switch silicon from multiple vendors |
| Interconnect standard | Multibus and Ethernet | Standard Ethernet, ECMP routing |
| Differentiation layer | System integration plus Berkeley Unix | EOS software; over 90% of R&D headcount in software |
| Economic trigger | VLSI made a $10,000 "Megapenny" workstation possible | Custom chip costs rose from hundreds of thousands to tens of millions |
| What the incumbents defended | Proprietary minicomputers | Vertically integrated custom ASICs and customer lock-in |
Tactic: List every component in your product that you build in-house, and for each one ask whether a vendor shipping in a hundred times your volume will be better at it within three years; if the answer is yes, move your engineers up the stack now.
Principle 3
When licensees can't see the category, become the company
Bechtolsheim's first business model was the one that suited his temperament. He licensed the design, let manufacturers build it and collected royalties. It had worked at sixteen with the Intel 8008 controller. It failed with the workstation because the licensees "had all the pieces, but they didn't see the market opportunity."
The lesson is that licensing transfers a design and does not transfer the vision behind it. A company that sees your design as a component will sell it as a component. If the value lies in a new category, and the SUN workstation was a new category, someone has to build a company around that category's whole logic, from the hardware through the software to the customer it serves.
Bechtolsheim did not make that leap alone. Khosla's phone call, "Let's just start a company," supplied the business conviction he said he lacked. The engineer's contribution was being frustrated enough to say yes.
Tactic: If you have licensed your technology and your partners keep shipping it as a feature rather than a product, take that as evidence that you need to build the product yourself, and find a business-minded co-founder before your frustration curdles into resignation.
Principle 4
Leave over a cost curve, not a quarrel
In 1995 Bechtolsheim saw x86 cost-performance "edging up" on SPARC while Sun was putting "all the wood behind one arrowhead." He did not stay to fight the strategy. He left, built switches for a decade, and returned in 2004 when Sun bought his x86 server company. The reunion went smoothly in large part because nothing had been broken.
The tactical point is that disagreements over economic curves resolve themselves with time. If you are right about the curve, the market will eventually force your former colleagues to agree, and you want to be standing there with the product when it does. A public fight only makes the eventual acquisition awkward.
↻
The Long Loop
Bechtolsheim's career as a series of exits and returns, most of them friendly.
1977
Declines to follow Justin Rattner's Intel internship to Oregon; enrolls at Stanford.
1982
Co-founds Sun Microsystems on February 24 with badge #1.
1995
Leaves Sun, worried about x86 cost-performance; co-founds Granite Systems with David Cheriton.
1996
Cisco acquires Granite for $220 million; he joins Cisco.
2001
Co-founds Kealia with Cheriton to build AMD Opteron servers.
2004
Sun acquires Kealia; he returns as senior VP and chief architect.
2008
Leaves Sun to become chairman and CDO of Arista Networks.
2023
Resigns as chairman and CDO; remains chief architect.
Tactic: When you disagree with your company's bet on a technology curve, write down your thesis with dates and numbers, leave cleanly, and build the thing yourself, so that you are positioned as the acquisition rather than the critic.
Principle 5
Be willing to obsolete your own expertise in public
"Even I could not design better silicon switch chips than what is available now on the merchant market." Bechtolsheim wrote that sentence in 2009 about a skill he had spent the late 1990s perfecting at Cisco. Most experts defend the value of their expertise long after the market has stopped paying for it. He announced its depreciation on a corporate blog and built a company on the far side of the write-down.
The move earns credibility in a way a self-serving argument cannot. When a former master of custom ASICs says custom ASICs are finished, customers listen. The confession was also good strategy. By conceding the silicon layer, Arista freed itself to compete on software, where incumbents were weakest.
Tactic: Once a year, name the skill you are proudest of and ask what it would take for a commodity to beat it; if that day has come, say so publicly before a competitor says it for you.
Principle 6
Write the check before the paperwork exists
The Google check was made out to a company that did not legally exist. The VMware seed, as Diane Greene has told it, followed a meeting of about an hour. The Mythic investment came within a week of an introduction. None of this was impulsive in the usual sense. Bechtolsheim was judging technology in domains where he had decades of first-hand expertise, so the judgment itself took very little time.
Speed signals conviction in a way a term sheet cannot, and it can also force structure into being. Bechtolsheim's uncashable check motivated Page and Brin to incorporate. A fast yes from a credible expert works like an accelerant. It tells founders and every later investor that the person most qualified to judge has already judged.
The caveat matters. Speed is only an advantage when the judgment behind it is earned. The Claria investment, made in a domain far from his expertise, ended when the company stopped operating in 2008.
Tactic: Set a rule that in your area of real technical depth you will decide within one meeting, and outside it you will not invest at all.
Principle 7
Back people serially, not deals singly
Bechtolsheim invested in every start-up his former Sun colleague George Haber founded: CompCore, GigaPixel, Mobilygen and CrestaTech, along with Moovweb. He co-founded three companies with David Cheriton and co-invested with him in Google. The pattern is a portfolio organized around a few people whose judgment he trusts rather than around market theses.
◆
A Partial Angel Ledger
Selected investments documented in the public record, with known outcomes.
| Company | Category | Known outcome | Status |
|---|---|---|---|
| Google (1998) | Search | First check, $100,000, before incorporation | Scale |
| VMware | Virtualization | Seed investor | Scale |
| CompCore (Haber) | Chips | Acquired by Zoran | Scale |
| GigaPixel (Haber) | Graphics | Acquired by 3Dfx | Scale |
| Mobilygen (Haber) | Video chips | Acquired by Maxim, 2008 | Scale |
| Tapulous | iPhone games | Acquired by Disney, 2010 | Scale |
| Claria | Online advertising | Ceased operating, 2008 | Risk |
| Mythic (2018) | Analog AI chips | Invested within a week of introduction | Try |
Serial backing cuts the cost of due diligence to almost nothing, because the diligence on the person was done years ago at a lab bench or in an engineering review. It also builds loyalty. A founder who knows the check will come again thinks bigger.
Tactic: Identify the three to five people whose technical judgment you have watched up close for years, and commit in advance to backing whatever they start next.
Principle 8
Treat project selection as the main event
"The most important thing is to pick the right project," Bechtolsheim told Stanford in 2012. His list of why products fail (too early, too late, too difficult, not relevant, too expensive) is mostly a list of selection errors, not execution errors. Only "too difficult" is about building. The rest are about choosing.
His own career reads as a record of good selections followed by ordinary, competent execution. He picked the workstation as VLSI made it affordable, Gigabit Ethernet as network backbones congested, x86 servers as Opteron arrived, and merchant-silicon switches as custom chip costs exploded. Each time the hard part was the choice and its timing. The engineering was difficult but tractable.
He also gave a reason why start-ups can win selections that large companies lose: "A startup succeeds because of focus. It is spending 100% of effort on a problem that the big companies have overlooked."
Tactic: Before committing to a project, run it through all five failure modes and write a one-line rebuttal for each; if you cannot rebut "too early" with a specific enabling technology that exists today, wait.
Principle 9
Follow the watts
"The goal is always power reduction," Bechtolsheim said of co-packaged optics in 2022. "If there was no power reduction, nobody would talk about co-packaged optics." The Sun Fire servers of 2005 were designed around data center heat and power. His 2011 case for merchant silicon put "less power consumption" on the same line as speed and cost. In 2018 he argued that analog AI chips would win on power because the brain is "remarkably power efficient, and it is analog."
Power is the honest metric. Marketing can inflate speeds and specifications, but a watt is a watt, and in a data center every watt costs money twice, once to deliver and once to cool. Following the watts filters out hype. A technology that does not reduce power at the same performance, or raise performance at the same power, is usually a solution looking for a problem.
Tactic: When evaluating any infrastructure technology, ask for its performance-per-watt against the incumbent, and treat a missing or vague answer as disqualifying.
Principle 10
Prefer the standard that already interoperates
At the peak of the OpenFlow excitement in 2012, Bechtolsheim noted that it had no significant production deployments. He favored ECMP, which was "mature, reliable," and needed no new and unproven architectures. His reasoning was about customers: they have existing networks and will not adopt systems that cannot talk to them. In 2017 he backed the 100G-per-lane consortium standard over slower IEEE proposals on grounds of practicality and forward compatibility.
This looks like conservatism, and it is not quite that. It is a theory of adoption. New technologies spread when they slot into what already exists. "The easiest way to go faster is to go faster" sounds like a tautology, but it is a strategy: stay on the established path and push its speed rather than build a new path that nobody can use yet.
Tactic: For any new standard you are tempted to adopt, count its production deployments rather than its announcements, and default to the incumbent until the count is meaningful.
Principle 11
Separate the architect from the operator
Bechtolsheim has run companies only briefly. He was CEO of Granite for about a year before Cisco bought it. At Sun, McNealy ran the operation. At Arista, Jayshree Ullal ran it, and the two shared the 2015 EY award. Bechtolsheim called himself "really more the engineer, the hardware guy, than the businessman," and he built his career around that self-knowledge rather than against it.
Founders who know they are architects often feel they must become CEOs anyway, because that is what founders are supposed to do. Bechtolsheim's record argues for the opposite. Find an operator whose strengths complement yours, take a title like chief development officer or chief architect that keeps you on the product, and give up the parts of the job that you would do badly.
Tactic: Write down the three parts of running a company you most dislike, and if they include managing people, finance and sales, recruit an operating partner before you raise your first round.
Principle 12
Build a wall around your information edge
Bechtolsheim's entire value as an investor and adviser comes from the flow of information toward him. Executives consult him, founders pitch him, and partners share roadmaps under NDA. According to the SEC, that flow is exactly what made the July 8, 2019 trades possible. A senior executive called him for advice and, in doing so, disclosed an imminent deal. The settlement cost him his chairmanship, a five-year bar from public company leadership and the phrase "abused the trust" in a federal regulator's press release.
The lesson is not moral, or not only moral. It is structural. The more consequential your network becomes, the more material nonpublic information reaches you without your asking, and the more your reputation depends on visible, mechanical barriers between what you hear and what you trade. Good intentions do not protect you here. What protects you is procedure: pre-cleared trading plans, no discretionary trading in accounts you influence, and a habit of treating every confidential conversation as a reason to do nothing.
A $923,740 fine is trivial for a billionaire. The other costs are not trivial for anyone.
Tactic: If people bring you confidential information as a matter of course, put every account you can influence, including relatives' accounts, under written pre-clearance rules before you need them.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Andy Bechtolsheim, Oral History, Computer History Museum, 2015There wasn't much interaction with other people, but that did not bother me since I was mostly interested in understanding how things worked.
— Andy Bechtolsheim, Stanford Engineering Hero lecture, 2012 (as recorded by Chris Yeh)Being first is not important (Google, Facebook), being the first to solve a relevant problem is all-important.
— Andy Bechtolsheim, Stanford Engineering Hero lecture, 2012Innovation is the never-ending search for better solutions.
— Andy Bechtolsheim, "The Easiest Way to go Faster is to go Faster," Arista Networks blog, October 23, 2018Why 400G Ethernet? In one sentence, because the easiest way to go faster is to go faster.
— Andy Bechtolsheim, interview with The Next Platform, March 2022The goal is always power reduction. In other words, if there was no power reduction, nobody would talk about co-packaged optics.
Maxims
- Plan ahead to the SKU. Specify exactly what you want, down to the part number, so that luck has less work to do.
- The category beats the component. Licensees with all the pieces can still miss the product; someone has to see the whole.
- Economics always wins. Over a long enough horizon, the cheaper architecture with adequate capability swallows the better one.
- Concede the layer you can't win. When commodity parts match your best work, move your engineers to where differentiation still lives.
- Leave over curves, not quarrels. If you're right about the economics, your old company will come to you, so make sure the door can still open.
- Decide fast where you are deep. A one-hour yes is earned by decades of knowledge, and outside that domain the same speed is just risk.
- Selection is most of success. Products fail for being too early, too late, too hard, irrelevant or too expensive, and four of those five are choosing errors.
- Follow the watts. Power is the metric hype cannot inflate.
- Count deployments, not announcements. A standard matters when customers run it in production.
- Your information edge needs a wall. The network that makes you valuable also brings you knowledge you must not use.
