Chasing Trillion-Dollar Companies, Founder Ambition, Token Budgets, & Regulatory Capture
Summary
- Elad argues the last five years were a trillion-dollar anomaly, with Anthropic, OpenAI, and SpaceX roughly making the leap from close to zero to $1 trillion. That does not establish a new cadence: a trillion-dollar company likely needs $50 billion-$100 billion of revenue with good margins, and he can identify only one unnamed contender that might reach the mark within three to five years.
- Sarah’s pushback is that investors still underestimate AI market expansion by valuing Harvey or Abridge per lawyer or doctor instead of asking what outcome-based pricing unlocks. Coding already shows consumption and delivered value potentially going “a hundred X from here”; Elad agrees on the opportunity but insists investors are conflating eventual market size with the speed required to build physical capacity and revenue.
- Elad sees a troubling flight from ambition among some of the best new founders: fear of the neo-labs is pushing them toward niche AI, hardware, or supposedly lab-proof markets. Labs will naturally absorb certain products, but not all of them; avoiding both categories sacrifices opportunities to compete through product and distribution.
- Most companies should at least consider selling, and many have a 12-to-18-month maximum-value window, Elad says, even though companies such as Anthropic and OpenAI should not sell in the near term. Boards should revisit exits every six months because “every year of AI time is like three to four years of normal cycle time,” while founders model dilution, a probable roughly 10x, perhaps 15x, eventual multiple, and the irreplaceable cost of spending five or six productive years trapped in a company that no longer works.
- Sarah reports manic expectations among people at the labs: coding may be effectively solved in roughly six months or by year-end, followed by “light RSI” around the end of the next year. She accepts that models can help improve training but disputes confidence in the clock: some scientists have forecast an 18-month recursive-self-improvement inflection “every eighteen months for the last five years,” while data and physical compute remain credible bottlenecks.
- Compute scarcity is creating both an oligopoly and a human power law: a few dozen researchers may drive roughly 80% of results, so labs increasingly allocate scarce compute by “return on invested tokens.” That logic also makes the “death of SaaS” look overstated—enterprises may reserve tokens for core products and major margin gains instead of rebuilding inexpensive software.
- A radically better architecture might emerge, but Sarah expects the industry to consume all available compute and power regardless; Elad’s high-probability outcome is that breakthroughs get copied by the incumbent labs. Policy may move the map faster: they discuss California tax proposals driving departures and Texas attracting an energy-and-hardware ecosystem because experimentation is easier.
- Elad’s broad warning is that safety can become regulatory capture: a high compliance burden can protect labs already advancing internally at exponential speed. His comparison is nuclear power—about 70% of French generation versus 18% in the US and 25% in Japan—where he believes excessive safety politics suppressed abundant energy; Sarah counters that reactors are being built now, though Elad replies, “We’re not making much.”
Deep dive
1. The trillion-dollar burst is an anomaly, not a cadence
Elad’s baseline is a five-year valuation anomaly: Anthropic barely existed, OpenAI was still early around GPT-3, and SpaceX traded at “eighty, a hundred, something like that.” Three companies then moved roughly from near zero to $1 trillion, compressing journeys that historically took 15-20 years.
His model is “punctuated equilibrium”: technology produces a “Cambridge explosion,” consolidates, then waits for another breakthrough. Social, SaaS, cloud, security, crypto, and successive internet cycles followed that rhythm; AI can have further waves, but some consolidators have already emerged.
Sarah’s disagreement is about imagination, not arithmetic. Investors may understand intellectually that AI sells services value yet still price Harvey or Abridge per seat, lawyer, or doctor instead of modeling outcome-based revenue; coding offers visible evidence that consumption and value can expand “a hundred X from here.”
Elad’s distinction survives the pushback: plenty of businesses can reach $5 billion-$10 billion of revenue and become $100 billion companies, but $1 trillion likely demands $50 billion-$100 billion with good margins. Energy and robotics may support that scale eventually; physical footprint makes reaching it within three to five years another matter.
2. Fear of the labs is shrinking ambition—and changing exit math
Elad sees good founders retreating into niches because neo-labs might enter their markets: hardware, “American dynamism,” narrow applications, or functions an inference cloud should provide. Some markets will be eaten, but others can be won through product and distribution; his concern is the trend among the best founders, not the median one.
Sarah shares the disappointment that founders are becoming less ambitious, while noting their portfolios contain companies challenging central lab premises. The disagreement is one of degree: not every founder is meek, but Elad believes enough of the newest exceptional founders now are.
For exits, Elad separates companies that should “never, ever sell” near term—Anthropic and OpenAI—from the majority, which should consider selling and may have a 12-to-18-month peak-value window. His governance fix is a pre-scheduled, non-emotional board discussion every six months, because three years of AI change now resembles a decade.
Sarah would ask whether the company captures value as costs fall and capabilities rise, and whether competing through capital and compute—her Cursor example—might be a maximally valuable point in time. She also says financing structure should match the thesis horizon and that a secondary can solve short-term needs without solving the underlying problem. Elad expects rising valuations for one or two years but says founders should ignore investor, press, and Twitter narratives and run the math: future dilution, a probable roughly 10x, perhaps 15x, multiple, expected outcome, and years of work. Private markets can remain irrational long enough to create margin-call-like risk.
3. Recursive self-improvement is credible; its countdown is not
Sarah reports manic expectations among people at the labs: code may become a solved problem in roughly six months or by year-end, followed by “light RSI” at the end of the following year, with models training major portions of themselves—probably post-training first, with pre-training possibly following.
That belief produces the brutal calculation Sarah describes: some people conclude they should work 16 hours daily because every week represents about 2% of their remaining productive career. Elad knows people at one major lab who have wondered whether to get married before the world changes; Sarah calls the psychology “kind of tragic,” akin to deciding how to spend the last two years of one’s life.
Sarah believes self-improving training is a natural extension of progress in code and math, particularly for training code and data pipelines. Her hedge is the timeline: an 18-month inflection has been predicted repeatedly for five years, while collecting data in less-verifiable domains and securing physical compute may constrain progress more than algorithms do.
4. Token budgets will expose the power law in human contribution
Physical compute imposes a ceiling that can enforce an oligopoly: if capacity is distributed roughly pro rata among major labs, no player can accelerate without limit. It keeps competitors closer together until the constraint lifts, even if their underlying research quality differs.
Elad says a few dozen researchers may generate roughly 80% of a lab’s results. Some labs have therefore slowed researcher hiring unless candidates clear an extremely high bar—not because salaries are prohibitive, but because each hire consumes scarce compute that could be allocated to stronger ideas.
His emerging metric is return on invested tokens, or ROIT. Enterprises are moving from “everybody use AI and do whatever you want” toward measured spending, more open-source deployment, and eventually explicit decisions about which people and projects deserve outsized token budgets.
That allocation problem is why Elad thinks the “death of SaaS” is overstated: firms may prefer tokens on core products or large margin gains over replacing cheap software. Minecraft—built by roughly five or ten people and sold for billions—shows extreme leverage already existed; AI has accelerated it, while displaced engineers could diffuse into GE, PG&E, Hershey’s, and other enterprises. Elad says that displacement, if it happens, is likely many years away.
5. Better architectures may diffuse faster than they disrupt
Sarah’s potential shock list includes investors withdrawing from CapEx because markets reject the debt-and-return profile, restrictions on existing or open models, and alternatives to transformers. Yet she expects all available compute and power to be consumed regardless of architecture, making memory and power efficiency more valuable without changing the industry’s direction.
Catching transformers at scale and matching their hardware ecosystem remains difficult. Elad’s high-probability outcome is that any breakthrough gets copied by compute-rich labs; the lower-probability path requires a neo-lab to keep its architecture secret long enough to scale decisively before an employee carries the knowledge elsewhere.
Policy is already reshaping geography. Sarah calls the discussed California billionaire-tax proposal the fastest way to chase out value creators; Elad mentions possible broader implementation and talk of an exit tax in ’28. They stop short of predicting instant Miami migration, arguing that ecosystems instead self-assemble around critical mass.
Texas is their clearest specimen: Sarah sees energy experimentation pulling in strong technologists, while Elad adds an emerging hardware corridor linked to SpaceX and, he thinks, part of Tesla’s move. His interpretation is categorical: these shifts are driven by regulatory differences, not by a simple ranking of where people prefer to live.
6. Safety rules can protect incumbents while suppressing upside
Asked whether labs could use compute access to control verticals, Elad pivots to regulatory capture through safety. If compliance blocks outsiders while incumbents keep progressing internally—and one AI year equals three or four normal years—a single year of protected development becomes an enormous competitive advantage.
Elad relays, without endorsing it, a drug developer’s view that regulators can evaluate risk without equally weighing benefit, thereby slowing development. Nuclear supplies the sharper example: France generates about 70% of its power from it, versus 18% in the US and 25% in Japan; Elad blames the 1970s safety lobby for four decades without US reactor construction. Sarah objects, “We’re making them now”; he answers, “We’re not making much.”
Elad still supports “proper safeguards,” but argues historical regulation went too far in energy, medicine, and biotech. The upside at stake spans productivity, education, healthcare, self-driving, and elder care; his closing call is to keep technology lightly regulated enough that society does not “lose optimism, lose momentum, lose progress.”