OpenAI and The Silicon Valley Way
Silicon Valley fell off the path, and must re-embrace it
In business, there are many variables of choice. The Silicon Valley Way, which is an unspoken yet pervasive business methodology, from the beginning has been to simplify business complexity and focus all effort on technical and product innovation. Since the start of modern technology companies, especially those based in California, essentially all of them have followed this formula. Sticking to it has been part of what made Silicon Valley successful. The last decade has seen somewhat of an abandonment of this approach, with victims like OpenAI. But there is an opportunity to get back onto it.
The Silicon Valley Way
There are two main virtues to this approach. The first is to focus all efforts on technology and product. Technologies may be extremely complicated, but in the Silicon Valley Way, that is it. The reason is, true to form, simple to understand. The focus on the total simplicity of business allows for all of the company’s energies and investment to be towards the technology itself—time not spent arguing with the accountant is time spent in the lab. With early companies in Silicon Valley history, like Cisco, there was a maniacal focus on the technical problem.
The second is to force clarity of thinking by avoiding the shroud of complexity. In a recent interview, Doug Leone, the former leader of Sequoia Capital and a veritable dean of Silicon Valley, emphasized the importance of simplicity in the Silicon Valley model. With the best companies, he notes, “it’s never something you scratch your head, you have to go to your [] expert and say, what does this mean? It's never that.” In one crazy example, Paul Graham of Y Combinator invested in Benchling for the dazzlingly simple reason that because biology was important, making their most important software would be valuable.1
This simplicity applies to even the most mundane aspects of the business. Tech companies all use the same lawyers, who use the same form documents from favored institutions like NVCA to create the same legal structures, all of which are incorporated in the same state, Delaware. They use the same rule-of-thumb stock option pool sizes, with standard vesting schedules, and of course all offer equity to their employees (at least, the ones in the corporate offices). They describe their business models with the same acronyms and calculate their financials across the same few common summary statistics. Though this is rarely talked about explicitly, this code is quite rigorously enforced by a lack of funding for startups with the temerity to vary.2 In fact, venture capital as a whole operates basically the same as the days of Arthur Rock.
This is not to say that Silicon Valley avoids business innovation per se, just that those innovations are simple. Free websites supported by ads were novel, Bill Gates transformed software,3 SaaS turned tech on its head, and founder shares turned the Valley “founder-friendly.” In fact, the story of Google is one not only of technical dominance but of brilliant business realizations, like the importance of being the default search engine and auctioning ad space. But these innovations are simple. They involve a straightforward value to the customer and a straightforward exchange of money. Though the act of coming up with these models required creativity, once imagined they are easy to comprehend.
We’re witnessing the suffering caused by adventures in governance now with OpenAI. Like many generational companies, OpenAI’s birth involved changes to its structure born out of concerns that are not crazy. The people who work in AI tend to be true believers and have talked about AI being owned by the world for a long time—even Larry Page once talked about donating true AI to the United Nations, and nonprofits are allowed to own for-profit companies (they’re just usually cafes or bookstores). The idea of creating a corporation that is legally bound to a mission, as nonprofits are, was an attempt to tie the company to the mast of moral obligations when the sirens of profit came calling. This might have worked if not for one problem: nonprofits have no mechanism that forces their boards to behave seriously. There are no shareholder lawsuits, no SEC investigations, not clear metrics for what a “mission” is. And so, OpenAI got what nearly every nonprofit board has, which is a board of misaligned, unserious people.
Though this concept may seem obvious, it is in fact not true for a number of businesses. Finance is a classic counterexample. Most of finance, at its core, comes down to a simple disagreement: two entities look at some type of asset, or collection of assets, and disagree as to its worth, and bet against each other in various ways. The resolution of that disagreement is where the mind-twisters begin, and it shows that finance is just plain complicated. It can require a PhD just to understand some of the valuation methodologies. Something like trading options using Black-Scholes equation is not something the average person can understand and participate in; a Bing ad auction is. Bill Ackman, a billionaire investor who once published a diligence spreadsheet so large it could crash a computer, has called bank balance sheets “inscrutable”—who has ever said that about, say, Cloudflare?
Where a decade of businesses went wrong
In 2011, Marc Andreessen famously said that “software is eating the world.” This mantle, which for a decade was the slogan of his eponymous venture capital firm, was taken up by a wide variety of startups. They took to using software to attempt to create novel, differentiated businesses that rethought existing industries using software. Many of the 2010s-era tech companies were not a total rethink, like Amazon was for bookstores; rather, they had much in common with the industry they sought to replace, like Root Insurance. These businesses were later dubbed “tech-enabled” businesses,
Andreessen’s essay is mostly remembered for its general idea and catchy title, but the actual text is much more nuanced and circumspect. His excitement was driven by the massively decreasing cost of cloud computing and the potential for the internet to allow for scalable distribution, but he is far more nuanced about the ability of those companies to create differentiating value necessary to maintain high valuations. There is also a strong flavor of the rise of the importance of the digital domain and of software commoditizing its complements. In other words, in Andreessen’s original essay, software eating the world is not a given—it is something that occurs for a specific set of reasons. Unfortunately, many tech-enabled businesses did not meet those conditions. There are two main ways these new tech-enabled businesses flew off the rails.
The first is that tech enabled businesses were all about using new technology in old industries with a bet that they would make traditional businesses look more like tech when the opposite turned out to be true. Technology may change the operations of a business, or provide a temporary edge, but if it is ultimately the same service or product, competitors will become wise to the edge and compete the new advantage away, turning a new technique into table stakes that must either be spent expensively acquiring market share before the market structure stabilizes or absorbing differentiation profits until they are commoditized. There are times when technology is so transformative that it fundamentally changes the economics and dynamics of an industry, like with many marketplaces, and so the bigger the transformation the more deceptive it can be. New companies like Uber and Lyft could only exist due to a confluence of technologies—3G mobile, GPS, and touch screens—and so they believed that they were technology companies happening to transform mobility. In fact, so deep was their belief that they both formed self-driving car units, both now disbanded. Yet, like taxi cabs, car drivers remained their biggest expense, and the capital necessary to scale those businesses is much less accessible in a non-ZIRP world. Technology when applied is ultimately a lever for a factor of a business, but when that factor turns out to not be the most important one, that energy may ultimately be misplaced.
The second is that many tech-enabled businesses, particularly in finance, took on balance sheet risk, and the tech companies either did not realize they were doing so or believed that they could create a better underwriting model but, in fact, couldn’t. Much of today’s technology is implicitly assumed to be information technology, and balance-sheet intensive businesses, many information technology companies leverage their edge in the new field of data science to make better statistical decisions. Companies like OpenDoor and Affirm call themselves tech companies, are founded by technologists, and are backed by technology-focused venture capitalists. But they do not actually sell technology. They use technology to make underwriting decisions, for which they deploy capital. Many new so-called technologies are just businesses that were well-timed to use the Internet as a distribution mechanism without having to also build many of the new technologies. And unfortunately, it turned out that these underwriting decisions were either not better, or were poorly applied. And in fact, for long-time experts, many of these mistakes were predictable—for example, Sam Zell observed that WeWork’s business model was not new and had always failed.
Much of the failure comes from a lack of engagement by tech enabled businesses with the underlying industry.4 Take the example of real estate. In the past ten years, Silicon Valley has gone deeper into the real estate game, funding companies that tried to rethink every part of the way we use our physical space. Some, like Airbnb, turned out to be wildly successful; others, like WeWork, became punchlines. Real estate focused startups like to emphasize that real estate is a multi-trillion dollar industry. All of these companies claimed to be leveraging the internet, using technology to get better operations, and even using data science to pick the best buildings. The problem is that in real estate, none of these are where the real edge comes from. The main determinants of the performance of a property, or even a property portfolio requires hyperlocal knowledge and macroeconomic luck that is out of an investor’s control.5 Any such edge would be small and better off sold to many real estate companies as table-stakes SaaS—Appfolio, not Flow.
The funds that manage billions are good at managing their operations, sure, but they are even better at managing their taxes. As a result, successful real estate investors become prodigious producers of paper. A few years ago, WeWork tried to go public for the first time, when Adam Neumann was still in charge, and they suffered from an interesting criticism: it was too complex, and certainly for the tech industry its structure was byzantine. But industries like real estate, or even pharmaceuticals, where master LLCs use spin-off C-corps for individual pharmaceutical assets, complex structures like Up-Cs are not uncommon. Certainly not in real estate, where such optimizations are essential for maximizing returns. Real estate’s privileged position allows it to defer certain tax gains and eliminate others, but only with great structural care and use of the appropriate financial incantations, all delicately and intricately balanced to also navigate countless local considerations. And with millions of real estate owners, there are thousands of them in every Congressional district, so these are not going anywhere.
To that end, why did Silicon Valley focus so much on tech enabled industries over the past decade? The Thielian interpretation is an act of desperation in a world with no new ideas, no technologies that are actually new and worth deploying, an attempt to extract rents in a “tech”-fueled arbitrage like the specious financial machinations of conglomerates of the 1970s. But this is hard to square with an era that produced the foundational research for today’s artificial intelligence, many novel innovations in areas like distributed computing, and the cloud. Mostly, software entrepreneurs heeding the clarion call for software to eat the world have been experimenting, seeing how far their inventions could go, and embracing the utopian view that their magical code could make the world a better place. Modern software was truly novel for many industries 15 years ago, and in a zero-interest rate world, no one was asking whether the productivity improvement was worth the cost of capital.
A fork in the road
Over the past fifteen years, the shift towards tech-enabled business meant that a significant portion of the technology industry moved away from the Silicon Valley Way. But there are signs of a shift back.
The first shot fired in the march back onto the road of the Silicon Valley Way has been the new rise of large language models, convolutional neural networks, and other related technologies. With the past decade’s advances in machine learning, there is no ambiguity as to its value. Machine learning has allowed for useful advances from better translation to image recognition for the blind. Some have calculated that ChatGPT, even with its weird name, is the estimated to be fastest growing consumer application of all time (though it is too early to calculate retention). As Leone would say, you don’t need to call an expert to understand that. To be clear, there are real debates remaining—whether RLHF is the best feedback, whether mere token prediction is enough for many applications, as well as debates as to where value will accrue in the technology stack. But these are technical debates amongst technologists, which is evidence of the vitality of the approach; there is no debate that artificial intelligence would be a breakthrough. It is instructive to compare AI with crypto, though this comparison is admittedly overdone. No one is telling anyone to “have fun staying poor,” grasping for use cases, or running Ponzi schemes—even children understand what AI does. The business models imagined for new AI applications are simple payment for service. And the approach is familiar. It involved a few hundred engineers raising staged venture capital, incentivized with stock options, and working on a focused problem.
There is a groundswell of progress in other areas, too. Many of the most exciting areas of the technology industry today are in areas involving technical breakthroughs. They include new semiconductor architectures, novel battery formulations like lithium air and solid state batteries, large-scale rockets like Starship, quantum computing, a myriad of biotechnology advances like mRNA therapeutics, gene editing, anti-obesity drugs, and bioforges for new manufacturing, and maybe even nuclear fusion. Even Tyler Cowen, who coined the term “Great Stagnation,” predicts it will eventually end and has had to admit we may be emerging from it!
These industries have much in common with semiconductors, which is the original industry the Silicon Valley Way evolved to facilitate. Like semiconductors, they tend to be capital intensive. OpenAI observed in 2018 that the compute expense of machine learning doubles every 3.5 months on average. Cloud companies each spend tens of billions of dollars a year on deploying new servers. Drug development is so expensive, and increasingly so, that in the industry it is jokingly subject to “Eroom’s Law,” which is “Moore’s Law” backwards, to symbolize how costs keep going up. Yet, these new technologies are similar to the historical Silicon Valley technology companies. These technologies all face technical risk and significant upfront development costs, but the resulting industries should all be high-margin with quickly decreasing marginal costs.
In hindsight, it is surprising that zero interest rates did not result in more investment in these types of companies, which suggests that perhaps the conventional wisdom is mistaken. Typically, cash flows are valued at a discount based on the prevailing interest rates. Many traditional Silicon Valley companies involve high upfront costs with cash flows that are far into the future, but with often low market risk. That makes these the types of businesses that one would have expected to thrive in a low-interest rate environment, where they would have functioned as a capital sink for capital searching for a return where cash flows were in the future, but low risk if they passed the technical bar and materialized. However, that is not what occurred. In reality, the zero interest rate phenomenon did not result in an investment in these technologies. Return and risk are inversely correlated, so for low-risk capital to search for high rates of return is to take on more risk. During the era of easy money, low-risk capital flooded to tech-enabled businesses, where there is market risk but low technical risk, even though the pure financial profile (where margins may be much worse when the cash flow streams materialize) is less obviously appealing in a low interest rate world. This suggests that fears that rising interest rates will result in less investment in advanced, capital intensive technologies may be overblown and that in times of low interest rates investors may be willing to take on not just more risk but prefer different types of risks.
Venture investment overall may be down, but companies in AI are raising more than they ever have while industries like biotech have fallen, but not off a cliff. The impact of rising interest rates on true technology may be less than is currently feared. After all, technology is not the same as tech stocks. Venture capital funds are sitting on larger amounts of dry powder than ever, and it is continuing to grow. Though some of this may be illusory, there is still much capital to invest. Before these funds run out, there remains ample opportunity to deploy it in more technologies that comply with the Silicon Valley Way and show returns to continue to raise large funds to deploy in these emerging areas.
Perhaps the greatest risk to Silicon Valley in particular is the separation of design from manufacturing. Apple is famously Designed in California yet manufactured wherever it is cheapest. But moving its manufacturing is no simple task. When it tried to build a Mac Pro, it got “screwed” by a lack of screws—reportedly, the shortage was so bad they had to be brought over in a pick-up truck. It is no surprise that the most advanced electric vehicle batteries that are not science experiments have been designed by Panasonic, Tesla, and CATL—they actually manufacture them. The rapidly decreasing cost of technologies like solar is more generally referred to as Wright’s Law, which is typically explained as “learning by doing.” It is no surprise that many of the best designed drones are made by DJI, which has long been the world’s largest manufacturer of them, nor that many of the best factory robotics systems in the world have been developed by the very factories that use them every day. There is significant evidence that industrial agglomeration effects lead to spillover effects in firm innovation generally. Ensuring the existence of domestic manufacturing, at least for strategic industries, is not just a matter of ensuring supply chains during times of crisis. It is also about maintaining an innovative edge.
As investment decisions are made by the large public institutions, like universities and pension funds, as to continue investing in venture capital, the answer seems to be that greater returns lie ahead by following the Silicon Valley Way and investing in these new, true technology companies. This does not mean that venture capital should not exist for tech-enabled services and SaaS, which have improved people’s lives. But it does mean that those capital allocators should be asking some hard questions. Learning from the past decade, they should be determining whether venture capital partners are truly equipped to evaluate the types of businesses they aim to invest in. They should be asking how they will understand these technologies and better distinguish between funds that invest in true Silicon Valley technologies and those that don’t.
One thing that zero interest rates did do is demolish corporate governance. The revival of this practice will be important for bringing back the Silicon Valley Way. Starting in the mid-2000s, the technology industry became more “founder friendly,” which meant less dilution and keeping founders in positions of power longer. Though this has been good for innovation, it also has a darker side: weakened boards, less accountability for management, more shallow diligence. The main impact is not the rise of fraud but an inability to force startup leadership to focus on achievable goals and true commercial viability. The Silicon Valley Way has always emphasized putting real products in the hands of users and not proliferating useless features or product lines. For the Silicon Valley Way to function, there must be a way to limit these excesses, which will require a revival of better corporate governance. The return of governance, however, must not be a return to the ways of old, which stifled innovators for the benefit of financial investors. Instead, it must be a Governance 2.0, where boards provide true accountability, companies face enough scrutiny to discourage fraud, and founders have the power to see through their instincts and visions.
The Silicon Valley Way is coming back. That said, it is not a given that this future will materialize. The future of technology requires us to make choices and set good incentives. The Silicon Valley Way has been able to keep things simple due to an implicit focus on the end customers and a sense of urgency to make the future happen as soon as possible. The Silicon Valley Way is not a mindset, but it emerges from one. We must emerge from the fog in more ways than one.
Another example is EUV. Extreme ultralight lithography took decades to develop, with doubters that it could even be achieved even a few years before it was—most notably Intel, which arguably was the fatal blow in falling behind TSMC. That said, once the technology itself has been cracked, its utility can be explained to a toddler. It is the only way to make computer chips faster, and since only one company can make them, they sell them to trusted partners for over $100 million apiece.
Complicated technologies may necessitate complicated decisions, including operational decisions; this does not mean a company is not following the Silicon Valley Way. Amazon’s scaled ecommerce business has relied on robots for warehouse operations and machine learning for personalized shopping, but also requires managing over one million blue-collar workers and retail space. Yet, the basic idea—use the Internet to provide a store with infinite selection—is so simple that Jeff Bezos could explain it on local television in 1997. Semiconductor fabs cost tens of billions of dollars to build, and even deciding where to put them is a multi-year endeavor. Yet, ultimately, there is no financial wizardry: there are capital expenses that generate high-margin widgets.
In more ways than one. His recognition of the near-perfect gross margins of software was a major leap because it allowed him to lower prices to target a better point on the price elasticity curve. But even selling software was controversial and an area where he led the charge.
With cryptocurrency, for example, it is instructive that many of the phenomena that caused the industry to spin out of control were novel to technologists but old hat for financiers—Matt Levine, a financial columnist, noted in his opus on crypto that it “constantly reinvented or rediscovered things that finance had been doing for centuries.” With FTX, Levine realized immediately that lending out FTT was like lending one’s own stock as collateral, which as an ex-financier he immediately identified as “messing with very dark magic.” When Terra-Luna blew up, he was yet again on the scene, explaining how the simple, fundamental premise was “insane.” If a banker-columnist is skilled enough to feel that these things are obvious, but a partner at a top venture capital firm isn’t, it raises questions as to who is more equipped to evaluate the businesses being formed in many parts of Silicon Valley.
This also makes real estate an effective way for a large investor to bet on the macroeconomy in the long run with less volatility and generating cash flow along the way. Real estate is excellent at generating cash flows and accessing capital appreciation in tax-efficient ways in large dollar amounts, but not at technology-level rates of return, which allows it so serve the cash flow needs of large institutions at rates of return that would be unacceptable for others.