Astro Teller: The $1B Bet No CEO Will Back, Moonshots 3x Cheaper in 16 yrs, and Clean Water at 1¢/L
Summary
At X, a moonshot is a testable conjunction of a huge problem, a science-fiction-sounding solution, and a breakthrough technology that offers at least “a tiny chance” of connecting the two. The necessary temperament is equal parts audacity and humility: pursue the unlikely, but begin by admitting, “It probably won’t work,” then test the hardest assumption as cheaply as possible.
X operates less like a conventional R&D lab than a tightly filtered venture portfolio. Teller says up to 200 ideas a year make it far enough to receive code names, yet only two graduate five or six years later—a stated 2% hit rate. Across 16 years, approximately 2,000 code-named projects produced 35–50 graduates, with Google Brain illustrating how an approximately 18-person team at graduation can underpin an enormous outcome.
A central corporate obstacle is governance, because executives routinely reject bets whose expected value they readily understand. Teller contrasts $1 million guaranteed with a 1% chance at $1 billion—the latter has 10 times the expected utility, yet support disappears when employees ask whether their CEO or board would actually tolerate it. His verdict: “You don’t need a lecture on innovation. You need a new manager.” He endorses keeping the core accountable for roughly 10% profit while placing 10X bets in an edge organization reporting directly to the CEO.
X says the cost of producing a graduate has fallen roughly threefold over 16 years, or, in Teller’s uncertain estimate, perhaps 10–20% annually, through better practice, earlier graduation, and AI. Alex challenged whether this merely reflects a richer economy; Teller rejected the comparison as confusing “the cost to paint a house with the value of the house.” The return can still be extraordinary because a Waymo-sized winner covers both itself and the discarded experiments.
AI is already a teammate across X, but Teller regards “we use AI” as an implementation detail equivalent to saying “we use electricity.” On the current clean-water project, salaries still exceed the AI bill, though Teller could not say by how much. He hopes AI will shorten the path from crazy idea to evidence-based de-risking and permit even greater audacity. He does not expect 100% automation within a decade: choosing socially worthwhile problems, winning community acceptance, and building distribution will require humans for “at least the next decade or two.”
X’s kill discipline is built around solving the monkey before constructing the pedestal. Teams must attack the assumption that could invalidate the project, because the easy pedestal creates visible progress while concealing that the monkey may never recite Shakespeare. Teller calls himself the “crown prince of failure” because moonshots are learning processes: “You learn nothing when you’re right. You only learn things when you’re wrong.”
Scientific discovery is not the investable endpoint; industrialization is the remaining 95% of the work. A single flake of room-temperature superconductor might win a Nobel Prize, Teller argues, but commercial value requires thousands of tons per day, suitable ductility, and manufacturability. The same techno-economic threshold governs clean water: 10 cents per liter is insufficient when the world-changing target is approximately one cent.
Deep dive
1. A moonshot begins as a falsifiable story, not an ambitious slogan
Teller’s three-part definition starts with a named, consequential problem; without one, the work is “arguably an academic exercise.” It then requires a science-fiction-sounding product that would resolve that problem and a breakthrough technology offering at least “a prayer” of making the product real.
X calls that conjunction a “moonshot story hypothesis”—testable, but emphatically not a promise of success. The explorer needs audacity to “suspend your disbelief for non-stupid reasons” and equal humility to acknowledge from day one that the journey is unlikely to work.
Early rejection often comes from first-principles techno-economics: enumerate the bill of materials, raw weight, achievable price, and customer willingness to pay. Teller’s conclusion is commercial as well as moral: “Purpose and profit can support each other,” while something structurally loss-making probably will not change the world at scale.
The funnel makes that skepticism concrete. Teller says up to 200 ideas annually make it far enough to receive code names and that X graduates about two after five or six years; over 16 years, approximately 2,000 code-named projects yielded somewhere between 35 and 50 graduates.
2. Google Brain shows why apparent overnight successes require long incubation
Google Brain began at X roughly 15½ years ago, when neural networks were “absolutely dead” and academics including Andrew Ng and Yann LeCun argued that scale might revive them. Ng and Jeff Dean asked whether industrializing neural networks—making them tens of thousands of times larger—could produce the missing breakthrough.
Teller identifies TPUs and the transformer—the “T in ChatGPT”—as coming from Google Brain. X’s role is to create “seed crystals” that can later reshape the world; his broader framing is that an apparent overnight success is typically 15–20 years in the making.
The unsolved portfolio remains grounded in hard thresholds. Nearly three billion people are water-stressed and lack clean drinking water, but X stopped approaches projected at 10 cents per liter because transformative clean water requires roughly one cent all-in; it also keeps revisiting education and the broader challenge of shifting energy across both time and location.
New materials excite Teller, but discovery is “the first 5% of a business.” One flake of room-temperature superconductor could earn a Nobel Prize; the moonshot is manufacturing thousands of tons daily, with properties such as sufficient ductility for winding wires.
3. Efficient moonshots require protected governance and unusual culture
Teller partly accepts the great-stagnation framing: after the 1960s and NASA, society “lost the explorer spirit.” His economic diagnosis is that during World War II and the Cold War, people cared less about return on investment; when return mattered more, audacious projects became harder to justify.
His “dirty little secret” is that moonshots are easy if efficiency is irrelevant: assemble “delusionally optimistic” people and pour money on them. X’s real project is systematizing radical innovation so its expected return makes sustained funding rational.
Teller’s choice experiment exposes the governance gap: $1 million guaranteed versus a 1% chance at $1 billion. Almost everyone chooses the mathematically superior second option, but hands fall when he asks whether their manager, CEO, or board would genuinely support it; the “choice Bs” therefore need a sequestered organization where their mess is tolerated.
His organizational prescription is to keep the core accountable for roughly 10% profit year after year while placing 10X—1,000%—bets in a protected edge organization reporting directly to the CEO. Peter compares this structure with Lockheed’s Skunk Works and Steve Jobs moving the Mac team off campus.
An audience question asked why X does not produce 100 baby moonshot companies a year. Teller said the limiting factor is culture engineering—intellectual honesty, humility, teamwork, and long-term thinking can be made adaptive among a few hundred people, but he does not know how to preserve that microcosm among thousands.
4. AI lowers experimentation costs without replacing the human mandate
Teller resists making AI the headline: pitching “we use AI” is like pitching “we use electricity.” Every moonshot, he says, combines people and agents, but the obsession remains the actual problem—clean water, grid storage, education, the electric grid, or preventing trillions of dollars a year from reaching landfills.
Teller hopes AI will shorten the time from a crazy idea to evidence that it is no longer crazy, while allowing X to raise its audacity and accomplish more in less time. The basic moonshot-factory structure, he says, should remain the same.
On X’s current clean-water effort, the salary bill remains larger than the AI bill, though Teller admitted he did not know the ratio. Automation will expand wherever it produces real benefit, but “spinning up an agent” is not synonymous with creating value.
Nor does he expect a fully automated moonshot factory within the next decade. Humans still choose problems and ensure solutions are socially acceptable, distributable, and welcomed by communities; those functions will require humans for “at least the next decade or two.”
X constrains ambition with a commercial clock: move from crazy to demonstrably not crazy in less than a decade, then reach an enduring business in less than another decade. Teller would hear a Mars proposal, but his test would remain: how does it become a durable business in less than two decades for Alphabet, X’s investor?
5. Killing projects early is the factory’s central compounding advantage
Teller’s monkey-and-pedestal analogy targets performative progress. If the mission is teaching a monkey atop a 10-foot pedestal to recite Shakespeare, teams instinctively build the easy pedestal; X insists they train the monkey first, because the pedestal can always be built afterward.
The point is learning speed, not failure for its own sake. Teller wants teams optimizing X’s portfolio, not protecting a beloved teleporter for five years after it becomes a zombie: “You learn nothing when you’re right. You only learn things when you’re wrong.”
One current experiment has produced a result that “would be a problem for physics” if correct. Five months in, the team is working “ferociously hard” to discover how it is fooling itself—a confidential, cold-fusion-like anomaly that leaves Teller both “deeply uncomfortable and deeply excited.”
His asymmetric rule is deliberately severe: a false positive can consume many years and many tens of millions of dollars, while rejecting a genuine moonshot costs X zero if problems and possible solutions are effectively inexhaustible. Hence the counterintuitive conclusion: X probably still does not say no quickly enough.