This year we formalized what had been an emerging pattern in our portfolio: an explicit mandate for deep technology — artificial intelligence, biotechnology, advanced materials, the kind of investment that takes a decade to know if you were right. On December 21st, an orbital-class rocket landed itself vertically for the first time — perhaps the clearest demonstration in a decade that "deep tech" is not a marketing phrase. This letter is about why we are betting our next decade on the longest holds we have ever underwritten.
On the Deep Tech Mandate, in Plain Language
Deep tech, as we use the term, refers to companies whose competitive advantage rests on technical work that took years to produce and would take years to replicate. This is in contrast to companies whose advantage rests on go-to-market velocity, network effects, or capital deployment. Both kinds of companies can be excellent investments. Only the first kind benefits structurally from being held by a family fund rather than a decennial fund.
Our deep tech holdings will, on average, take longer to mature, fail more often early, and produce returns that are more concentrated in fewer winners. None of these properties are problems for our structure. All of them are problems for an LP-funded structure. We have, in conversations with peer firms over the past two years, become convinced that the deep-tech category is structurally under-funded relative to its long-term economic value, primarily because the dominant venture structures in the industry are not capable of holding the assets the category produces. We intend to take advantage of this asymmetry.
The mandate is not a category of investments; it is an organizing principle for the next decade of the firm. We will continue to invest in software, internet, and consumer companies as we have. We are, in addition, shifting roughly a third of our total capital allocation toward deep tech over the next three years. The shift is large and it is deliberate. We will write about it in detail in future letters.
On Why Now, and Not Five Years Ago
Three things have changed since 2010 that make this mandate possible. The cost of compute has fallen by an order of magnitude, which has structurally reduced the capital requirements for AI and machine-learning companies to a level that venture capital can productively underwrite. The quality of the technical talent willing to work in startups, particularly in machine learning, has improved meaningfully — partly because the academic labs are no longer the only places where frontier work is being done, and partly because the operator culture of the technology industry has caught up with the scientific seriousness of the underlying problems. And the policy environment around emerging biotechnology has, for the first time in a generation, become predictable enough to underwrite.
None of these are permanent conditions. The compute cost-curve will eventually flatten or reverse; the talent pool will redistribute; the policy environment will shift again. Deep tech investing has always come in windows; we believe one is open now. The companies that should be funded during this window are being founded now. The 2015-2018 vintages will, we predict, be among the most productive deep-tech vintages of the past forty years. We intend to be present for them.
On the Rocket That Landed Last Week
The vertical landing of the Falcon 9's first stage on December 21st is, in our reading, the single most important demonstration of the past five years that decade-long technical bets can produce engineering outcomes the consensus had treated as decades away. SpaceX is not in our portfolio. The investments we made this year that share this property — patient capital deployed against engineering problems considered too hard for venture timelines — are.
The deeper lesson of the landing, beyond the engineering achievement itself, is what it implies about the structure of opportunity in deep tech. The most consequential outcomes are produced by companies whose technical risk is concentrated in a single hard problem and whose execution risk is contained by the discipline of solving that problem. The mistake we and our peers have historically made in deep tech is to fund companies whose technical risk is distributed across multiple problems, where the probability of success on any one of them is acceptable but the joint probability of success on all of them is unacceptably low. We are now structuring our deep-tech diligence around the question of whether the company's risk is concentrated or distributed.
On the Companies We Are Already Two Years Into
Three companies in our portfolio, all founded in 2013, will not produce revenue at material scale until 2018 at the earliest. We knew this when we funded them. We have, this year, increased our positions in all three. The first commercial validations are arriving on schedule — which, in deep tech, is itself a meaningful signal — and the science underlying each has held up against the most rigorous scrutiny we could buy.
The hardest part of supporting deep-tech companies, we have learned, is not the capital. It is the patience required during the years in which the company is producing technical milestones rather than commercial ones. Boards of deep-tech companies require a different kind of discipline from boards of software companies; the metrics that matter cannot be tracked monthly, and the metrics that can be tracked monthly are often misleading. We have spent considerable time this year teaching ourselves and our portfolio CEOs how to govern deep-tech companies under appropriate metrics. The teaching is incomplete; the lessons will accumulate.
A Closing Note
The longest holds we have ever underwritten will not pay back during the term of our current decade. They are intended to define the decade after. We are at peace with the asymmetry. The 2025 letter will, we expect, be reporting on outcomes that the 2015 letter is only able to forecast. We will know then whether the forecast was right.
The Partners
Winzheng Family Investment Fund · December 2015