There is a concept in venture capital that sounds simple and is almost never executed correctly: invest before the market. Every serious early-stage investor claims to do it. Very few actually do. The gap between the claim and the practice is not a matter of will — it is a matter of temporal courage. Investing ten years before the market does not mean being slightly early. It means being so early that the market does not yet have language for what you are backing. It means writing a check at a moment when the consensus view is that your investment thesis is either naive or delusional.
This essay is about the actual mathematics and psychology of early timing in deep technology investing: why the early bets generate the highest returns, how to distinguish "too early" from "correctly early," and what the empirical record of the last fifteen years tells us about the relationship between investment timing and fund performance.
The Temporal Structure of Venture Returns
To understand why early timing produces superior returns, you need to understand how technology adoption curves interact with valuation multiples. Technology adoption follows an S-curve: slow initial adoption, rapid acceleration through the growth phase, plateau at maturity. Valuation multiples at any point in this curve roughly track the market's confidence in the total addressable market and the timeline to realising it.
At the very beginning of the adoption curve — what Geoffrey Moore calls the "chasm" period, before the technology crosses into mainstream commercial use — valuation multiples are extremely low. The market has not priced in the growth phase because the growth phase has not started. An investor who enters at this point gets the asset at the pre-adoption price and participates in the re-rating that happens as the market realizes the technology is real.
The mathematics of this dynamic are what make early-stage deep technology investing so attractive in structural terms. Consider a technology company whose total enterprise value at maturity is $10 billion. If you invest at seed stage, when the company's valuation is $20 million, you are getting the asset at 0.2% of its mature value. Even accounting for dilution from future financing rounds, a well-structured seed investment in a company that reaches $10 billion in enterprise value can generate 500— or more on invested capital.
Now consider the same company at Series B, after it has demonstrated commercial traction and the market has begun to price in the growth story. The valuation might be $300 million. The return to full maturity is approximately 33—. Both investments are profitable. But the seed investor has generated 15 times the return per dollar invested for essentially the same company.
The only difference is timing. The seed investor accepted uncertainty about whether the company would succeed in exchange for a dramatically lower entry price. The Series B investor bought certainty at a premium. The mathematical advantage of the earlier entry is so large that it persists even after you discount for the higher failure rate at seed stage.
Mistral: The Price of Scientific Conviction
Mistral AI provides one of the most vivid recent illustrations of the early-timing premium. Founded in April 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix — all former researchers from DeepMind and Meta AI — Mistral closed its seed round of €105 million at a valuation of approximately €240 million within weeks of founding. At the time, the conventional wisdom in AI was that large language model development was a game only for trillion-dollar tech companies. The compute requirements, the data requirements, and the engineering team requirements were all seen as prohibitive barriers to entry for new entrants.
Mistral's founding team had a specific technical thesis: that the efficiency of language model architectures was dramatically improvable, and that a small team with the right insights could build highly competitive models at a fraction of the compute cost that OpenAI and Google were spending. Within six months, they released Mistral 7B — a model that benchmarked comparably to much larger models from established players, demonstrating that their efficiency hypothesis was correct.
By June 2024, Mistral had raised $415 million in its Series B at a valuation of $6 billion — from a seed valuation of approximately €240 million, a return of roughly 25— in fourteen months for the earliest investors. The investors who backed Mistral at seed were not making a bet that was obvious in April 2023. They were making a bet on a specific technical thesis held by a specific team, before any of the empirical results existed that would later validate that thesis. Their early entry price reflected genuine uncertainty; their subsequent return reflected the resolution of that uncertainty in the most favorable possible way.
Cohere and the Enterprise AI Timing Thesis
Cohere presents a different but equally instructive case. Founded in 2019 by Aidan Gomez, Ivan Zhang, and Nick Frosst — alumni of the Google Brain team that produced the original Transformer paper — Cohere was built on a specific commercial thesis: that large enterprises would need dedicated, customisable language model infrastructure that they could deploy within their own security perimeters, separate from the consumer-facing AI products that OpenAI and Google were building.
In 2019 and 2020, this thesis was considered premature. GPT-3 had not yet demonstrated that language models had commercial utility at enterprise scale. The market for enterprise AI infrastructure did not yet exist in any meaningful sense. Cohere's seed investors were backing a thesis about what enterprise AI would look like three to five years in the future, not what it looked like at the time of investment.
By 2024, Cohere had raised approximately $270 million in its Series C at a valuation of $5 billion. The enterprise AI market that Cohere had predicted in 2019 had materialised precisely as the founders had anticipated. Financial services firms, healthcare companies, and large manufacturing enterprises were paying significant premiums for AI infrastructure that could be deployed on-premises or in private cloud environments, with the customisation and security controls that consumer-facing AI products could not provide. Cohere's earliest investors had been right about the market and right about the timing — a combination that produced returns that no late-stage entry could have matched.
The Physics of Being Early: What "Ten Years Before" Actually Means
The phrase "ten years before the market" is often misunderstood to mean "ten years before the technology is useful." What it actually means is ten years before the technology is conventionally recognised as useful — before the market has developed the analytical frameworks to properly value it, before institutional capital has been allocated to the sector, before the McKinsey reports have been written and the Gartner curves have been published.
There is a systematic gap between when a technology first becomes technically capable of addressing a commercially important problem and when institutional investors begin to allocate capital to it at scale. This gap exists for several reasons: large institutional investors require established track records before allocating to new categories; investment frameworks are typically backward-looking; and the humans who manage capital tend to require social proof before committing to genuinely contrarian positions.
The consequence is that the best entry points in deep technology investing — the points at which the expected return per unit of risk is highest — consistently occur in the period before institutional consensus forms. These are the moments when a technology has produced enough scientific evidence to make a credible case for commercial viability, but when the broader investment market has not yet caught up to the evidence.
Identifying these moments requires a different kind of analytical work than conventional due diligence. It requires being close enough to the research frontier to know when a scientific result is genuinely significant, being intellectually honest enough to distinguish "this is interesting" from "this changes everything," and being commercially experienced enough to trace the path from laboratory result to market application. Very few investment teams have all three capabilities simultaneously.
The Anatomy of a Correctly Timed Early Investment: Figure AI
Figure AI, founded in 2022 by Brett Adcock and a team of robotics engineers from Apple, Tesla, Boston Dynamics, and Google, represents a masterclass in identifying the moment when a long-developing technology crosses the commercial viability threshold. Humanoid robotics had been a research topic since the 1970s. Honda's ASIMO, Boston Dynamics' Atlas, and dozens of academic research platforms had demonstrated increasingly capable robots over decades. But none of them had gotten close to the commercial viability required for deployment in actual industrial settings.
What changed in 2021 and 2022 was a combination of factors: advances in simulation-to-real transfer (the ability to train robotic control policies in simulation and deploy them reliably in the physical world), improvements in transformer-based perception models that could process visual and tactile information in real time, and reductions in the cost of the actuator and sensor hardware that physical robots require. These advances, taken together, crossed a threshold that the Figure team correctly identified as commercially significant.
Investors who understood the research trajectory — who knew what had changed technically in the preceding three years — could see what the Figure team saw. Investors who were evaluating the opportunity purely on historical track record (decades of robotics companies that had failed to commercialise) could not. The early investors in Figure, who got in at valuations that were a fraction of the $2.6 billion at which the company raised its Series B in 2024, were rewarded for understanding the technology well enough to recognise that this time was genuinely different.
When "Too Early" Destroys the Investment Thesis
Correctly early and incorrectly early look identical at the time of investment. The difference only becomes clear in retrospect, which is why the discipline of evaluating investment timing is one of the hardest skills in venture capital to develop and the easiest to fool yourself about.
The characteristic of "too early" investments is that the technology requires a fundamental scientific advance — not just engineering progress or cost reduction — before it becomes commercially viable. Cold fusion has been "ten years away" since the 1950s because it requires solving a basic physics problem that current understanding does not have a clear path to solving. Many fusion energy companies have failed not because they were bad companies but because the underlying science was not ready.
The characteristic of "correctly early" investments is that the technology is already scientifically established and the barriers to commercial viability are engineering and cost barriers — problems that respond to capital investment and human ingenuity rather than requiring new scientific breakthroughs. LLMs were "correctly early" in 2019 because the scientific basis (transformers, attention mechanisms, scaling laws) was established; what remained was engineering work and compute cost reduction. Humanoid robotics was "correctly early" in 2022 because the underlying science of control theory, computer vision, and machine learning was established; what remained was systems integration and cost reduction.
The analytical question for any early-stage deep technology investment is therefore: what barriers remain between the current state of the technology and commercial deployment, and are those barriers of a type that capital and engineering talent can systematically reduce? If the answer is yes, the investment may be "correctly early." If the answer requires a fundamental scientific advance that no one knows how to make, the investment is likely "too early" in the sense that matters commercially.
Physical Intelligence and the Frontier of Generalist Robotics
Physical Intelligence (Pi), founded in 2023 by a team including Sergey Levine from UC Berkeley and Chelsea Finn from Stanford, raised $400 million at a $2.4 billion valuation in 2024 — before the company had a commercial product. The investment thesis was not about a product that existed; it was about a scientific approach (foundation models for physical manipulation, analogous to how GPT-4 is a foundation model for language) that the founders believed would produce the first truly generalist robotic control system.
The investors who backed Physical Intelligence at this stage — primarily Jeff Bezos, Khosla Ventures, and Tiger Global — were betting on a research paradigm, not a product. They were making the judgment that the foundation model approach to robotics would prove as transformative as the foundation model approach to language, and that this team, with its unique combination of academic depth and commercial experience, was best positioned to execute on that paradigm.
Whether that bet proves correct is not yet determined. But the return profile if it does prove correct is extraordinary — and the investment was made at a moment when the conventional market had not yet formed a view on whether the paradigm would work. That is precisely the window that early-timing investing aims to capture.
Constructing a Portfolio Around Temporal Advantage
At Estes Capital, our portfolio construction philosophy explicitly incorporates investment timing as a variable. We aim to identify technology categories where we believe we are currently in the pre-consensus window — where the scientific evidence is sufficient to support a credible commercial thesis, but where the mainstream investment market has not yet allocated capital at scale.
Currently, we believe that window is open most clearly in three areas: AI infrastructure (specifically the infrastructure required for deploying large models reliably and securely at enterprise scale), advanced energy technology (particularly electrochemical manufacturing for clean hydrogen and next-generation nuclear), and biological manufacturing (the application of programmable biology to production of materials, pharmaceuticals, and chemicals that are currently derived from fossil fuel feedstocks).
In each of these categories, the science is established — the fundamental feasibility is not in question. What remains is the engineering and cost trajectory, and the pattern of that trajectory is clear enough from recent results to support investment conviction. The mainstream investment market is beginning to form views in each of these categories, but we believe we are still in the early stages of institutional capital formation — still in the period where the temporal advantage of early entry produces the most compelling return profile.
The hardest part of this strategy is not identifying the right categories. It is maintaining conviction during the period between entry and consensus, when the investment looks wrong from the outside, when the product does not yet exist, and when the rational short-term view is that you have made a mistake. The investors who generated the extraordinary returns on Anthropic, Mistral, and Cohere were not smarter than everyone else. They were more patient, more willing to hold conviction in the face of short-term uncertainty, and more rigorous in their original analysis.
If you are building a deep technology company in a category where the market does not yet exist, and you need a partner who can provide both early-stage capital and the long-term conviction to hold through the pre-consensus period, we are the fund you are looking for. Let's talk.
