Chasing Trillion-Dollar Companies, Founder Ambition, Token Budgets, & Regulatory Capture
In a Nutshell
A handful of frontier AI labs are racing toward recursive self-improvement on a 12-18 month horizon, concentrating talent and compute into an oligopoly while most startups chase hundred-billion-dollar outcomes rather than trillion-dollar ones. Founders must ruthlessly evaluate exit timing, token ROI, and whether they’re on the right side of secular change, because AI compresses normal business cycles by 3-4x and private-market valuations can stay irrational longer than founders can stay productive. Regulatory capture and over-weighting safety over progress risk repeating the nuclear industry’s 40-year stall, which is why lightly-regulated ecosystems in Texas and elsewhere are emerging as counterweights to California’s tightening grip.
These notes were generated by AI and may contain inaccuracies.
70% of France is still nuclear in terms of its power generation. The US is at 18% and hasn't built a reactor in 40 years. A safety lobby in the 1970s essentially killed abundant clean energy for the United States. There are real outcomes where safety has hurt progress. The question is where the spectrum should be on AI regarding safety considerations.
Welcome to No Fires. Discussion focuses on risk management, RSI, how many trillion-dollar companies there can really be, and the ills of regulatory capture. Conviction's low overhead, high signal grant program is accepting applications for 10 exceptional startups building at the frontier. The program runs twice a year with $250,000 in cash on an uncapped note as well as compute and services from partners including OpenAI, Anthropic, Base 10, and others. Previous cohort companies include Cognition, Chai Discovery, Listen Labs, Physical Intelligence, and Flappy Airplanes.
Over the last five years, three companies went from close to zero to a trillion dollars in market cap: Anthropic basically didn't exist five years ago, OpenAI was still quite early with GPT-3 just coming out, and SpaceX was trading at 80-100. This represents a massive inflection in valuations. People now assume there will be a bunch of other trillion-dollar companies formed in three to five years, which is unprecedented in human history. Usually it takes 20 years - SpaceX took since the early 2000s and Google since the 1990s.
The current situation resembles punctuated equilibrium from evolutionary theory, where there are periods of steady state followed by explosions. The history of technology shows waves: a big social wave, a SaaS wave, internet periods including the 1990s internet, social from early 2000s to 2010-2012, SaaS, cloud, big security companies, and crypto. AI will likely have giant breakthroughs in model capability creating another step function. The question is how many more giant companies are coming in the next handful of years, distinct from what happens over 20 years. There are still a lot of hundred-billion-dollar companies to be built, but multi-trillion-dollar companies are harder to reach.
A trillion-dollar market requires $50-100 billion in revenue with good margins. The question becomes where $50-100 billion revenue streams exist for single companies. While there are a dozen-plus companies that are there-ish, how many more will reach this in the next 5 years is the key question. There are tons of companies that can get to $5-10 billion of revenue and become $100 billion companies.
Promising areas include supply chain and energy-type technologies, robotics, and materials. For physical goods companies, the question is whether they have the footprint to get there that fast. The issue is that people are investing against the belief that speed exists, which differs from market size. People are conflating these two things.
Excellent founders are going after niche markets because they're not scared of the labs. There's less head-to-head competition in markets that Harvey, Open Evidence, Decagon, or Sierra entered three to five years ago. For mid to late stage technology markets, people continue investing as if there's velocity to get to a trillion, though some companies will get to $20 billion, some to $100 billion, and some will go to zero. There's a flight to hardware companies with the assumption that labs will never do hardware, American dynamism, and niche applications of AI.
There are companies that should never sell in the near term: Anthropic and OpenAI. Most companies should at least consider selling. There's usually a 12-18 month period where the company is worth the most it will ever be worth. Ben Horowitz wrote about having a pre-planned once-a-year board meeting where the discussion topic is whether to consider exiting in the next six-month period, scheduled in advance so it's not founders or investors pushing for it.
Every year of AI time is like 3-4 years of normal cycle time. Three years in AI is like a decade in normal time. The underlying fact set is changing faster than ever, requiring more frequent double-checking of thinking. Every six months is the suggested frequency for these conversations.
Founders should ask whether they're capturing value as costs fall and capabilities increase. If on the wrong side of secular change and unable to get to the other side, selling may be appropriate. Cursor represents an interesting question about competing on capital, compute access, and whether it's a maximally valuable point in time.
Due to the rapid rise of three trillion-dollar-plus companies in a short timeframe, enormous amounts of venture capital are being returned. People are raising bigger funds and need to put capital somewhere, leading to ongoing rises in valuations over the next year or two. Financing will get easier, not harder. The key question is what founders think is the true likely expected outcome of their company.
At some point, companies will trade at 10x or 15x multiples. Founders should run the math on what their company will be worth, considering dilution and years of work required. The biggest opportunity cost is time - the most productive years of life are on the line. Founders can either walk away with good money to do the next giant thing or roll the dice.
Secondary sales solve some short-term needs but don't create real solutions. People from the 2020-2021 cohort are still running companies five years later that aren't working. The cost is lifetime cost - working on something that will continue to struggle, over-capitalized, with runway for the next 10 years.
Private markets don't have to be rational or right for long periods of time. Being smart about financing ability is equivalent to avoiding margin calls. Founders with technical or structural points of view can get frustrated when investors believe something wrong for extended periods.
People at labs believe there's 6 months or toward the end of the year to be completely done with code, then light RSI by end of next year. Models would be training big chunks of themselves, initially post-training, potentially impacting pre-training. This leads people to believe they have a year and a half of productive work left, working 16-hour days because every week represents 2% of remaining productive time.
A number of very smart and self-aware research scientists have felt there was a knee in the curve on recursive self-improvement or ASI 18 months away, every 18 months for the last 5 years. The extension from code to training code to data pipeline work is easier to believe. The question of gathering data for less verifiable, more complex domains and physical compute accessibility constraints may be more limiting than algorithmic possibility.
Physical compute creates an oligopoly market by creating a ceiling on the rate of progress any single lab can achieve. Assuming compute is roughly par across the ecosystem for big labs, this enforces an oligopoly market up to a point and forces closer competition between players than would otherwise exist.
The belief that RSI is 18 months away suggests there could be a burnout cycle. Some people at major labs have questioned whether they should get married because they don't know what happens in 18 months. Second-order effects are driven by this belief system and potential burnout over time.
As compute becomes the scarce resource, a few dozen researchers drive 80% of results at any given place, representing an interesting human power law.
In any field, there are at most a few dozen people who drive the field forward. This pattern holds across breast cancer research, subfields of mathematics, physics, and the entrepreneurial ecosystem. In AI research, this concentration is becoming more pronounced as compute is increasingly allocated to top performers. Labs have slowed hiring of researchers unless they meet very high standards because the bottleneck is not the researcher cost, but the compute associated with each person.
The concept of return on invested tokens measures ROI when allocating token budgets. This mirrors the historical pattern where internal tools teams at companies were starved for resources because companies preferred having engineers build product rather than internal tools. The death of SaaS is overstated because companies would rather allocate tokens to core product development, margin improvements, or other high-value outcomes rather than spending on SaaS subscriptions.
The shift has moved from "everybody use AI" to measuring spend and moving toward open source, with the next wave being determining which projects and people should receive outsized token allocations based on return on investment.
Minecraft was developed by five to ten people when acquired for billions of dollars by Microsoft, demonstrating that multi-billion dollar single-person companies have already existed. AI has accelerated the ability for individuals to achieve outsized advantages through technology. The question becomes why allocate tokens to people who cannot generate relative outsized returns unless those high performers run out.
Even if displacement of engineers occurs at big tech companies, there are numerous homes for displaced talent. Companies like GE, PG&E, and Hershey's never had the capability to recruit top engineers and would value these individuals even if they were considered mediocre in the context of Google or Meta. This represents a permeation of engineering talent through the enterprise landscape that is likely years away.
Researchers at major labs who have not received outsized token allocations have the opportunity to apply their comparative advantages in domains that will accelerate through AI, including supply chain bottlenecks, bio, and diffusion into other valuable fields. These opportunities are more exciting than concerns about the downfall of mathematics.
The question of how many researchers a top AI lab needs relates to whether current numbers are appropriate. The consideration is whether labs need five times as many people or half as many, with the constraint that researchers must exceed a certain performance bar because the limitation is compute. The best ideas come from a subset of people on average, creating interesting questions about where the N+1 person goes when talent distribution is considered.
An old ex-Googler joke illustrates varying perspectives on contribution concentration: when Google had around 50,000 people, people within search and ads estimated only 20% of people were needed to run the company, while those outside search and ads estimated all 50,000 people were necessary.
Three potential disruptions to current AI development include investors collapsing capex commitments due to market reaction, concerns about debt and returns profiles, and technological disruptions such as alternatives to transformers. Regulatory attempts to restrict model usage or open source to constrain progress pace represent another potential disruption.
California passed a billionaire tax with Democratic Party support. For founders of companies worth $10 billion or more, the law could require forced asset sales. The law appears written broadly enough that once passed, thresholds can be lowered in future years. There is increasing discussion of adding an exit tax in 2028 that would penalize people attempting to leave the state.
The immediate effect is that people attempting to create value in California choose to leave, with some already planning departures by fall. The two negative signs for California are this bill and ballot harvesting initiatives.
Boston was historically the main startup hub counterweight to Silicon Valley in the 1980s-early 1990s, particularly for biotech, despite poor weather. The key factor is having enough smart people aggregated working on common problems, which creates the conditions for renaissances.
Texas is experiencing great technology migration and innovation, particularly around energy, driven by regulatory environment and demand. People are moving from Silicon Valley and other places to experiment in this environment. There is a growing hardware corridor emerging around areas originally centered around SpaceX's location in El Segundo, with both SpaceX and parts of Tesla moving to Texas.
The industry will consume all available compute and power regardless of underlying architectures. Pressure to find more memory or power-efficient architectures becomes more interesting, though catching up to transformers in scale and hardware match remains difficult. People will make bets on alternative architectures as they become more desperate for experimentation.
Whatever new architecture emerges gets copied by labs which then allocate all compute to it. The high-probability outcome is rapid dissemination of any breakthrough, though lower-probability scenarios exist where a team keeps an architecture secret and scales on it exclusively.
Regulatory capture and risk-reward scenarios can create situations where safety burden becomes so high that even dramatically positive outcomes are constrained. The FDA has been criticized for focusing too much on safety and risk while not enough on benefit, creating no risk-reward worth considering.
This same dynamic could emerge in AI labs where safety constraints applied internally create massive advantages for those who push progress forward exponentially. The trade-off involves weighing outcomes like hacked email against better healthcare through AI models sooner.
Every technology has both positive and negative applications. Biotech can create viruses but can also cure cancer. Nuclear power can provide free, cheap, abundant energy or create weapons. France generates 70% of its power from nuclear with no significant accidents. The US generates 18% from nuclear and has not built a reactor in 40 years. Japan generates 25% from nuclear. A safety lobby in the 1970s essentially killed abundant clean energy development in the US.
Significant positive applications exist across human productivity, education, healthcare, daily life benefits, self-driving vehicles, and elderly care. Tech has been successful because it has been lightly regulated, and maintaining this approach preserves optimism, momentum, and progress. Historical examples from biotech and energy show that excessive safety focus can harm development in major industries.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
