Etzioni on AI: Startup speed is not a strategy
AI startups should run experiments fast but make big commitments slowly, Oren Etzioni argues, because speed alone can't build a defensible business when every competitor can build quickly too. He goes through the five usual arguments for speed, says only one holds up and makes the case for proprietary data and relationships as a startup's real advantage. Read More

The world of AI startups has fetishized “speed” and whipped itself into a frenzy. Sure, the feature that took a quarter to build back in 2023 takes a mere AI-fueled weekend in 2026, but many AI startups are running on FOMO, caffeine and tokens without a clear strategy.
A startup is a marathon, and a team that attempts to sprint the whole way will burn out before the finish line.
Five arguments for speed, and what’s wrong with four of them
Sam Altman’s Startup Playbook puts the speed creed in one line: “I have never, not once, seen a slow-moving founder be really successful.” I was singing from the same hymnbook earlier this year. In my column on ten commandments for AI startups I wrote that “velocity is the new moat.” Then everyone got fast, and I stopped singing.
Here are the five arguments usually made for that creed.
The window is closing. Each model release opens a gap that the frontier labs close within a release cycle or so. But if the labs will close it anyway, speed merely buys you a nicer funeral. The argument supports picking a different direction, not running faster in this one.
Competitors are everywhere. AI makes building cheap, so clones show up in weeks. True, and that’s the reason velocity isn’t a moat. When building is cheap, defensibility matters more than being first. Peter Thiel’s last-mover argument from Zero to One is stronger today than when he wrote it. In his words: “It’s much better to be the last mover—that is, to make the last great development in a specific market and enjoy years or even decades of monopoly profits.”
The data flywheel. Early users generate data that improves the product, so get there first. Except that each new model generation resets the flywheel, and data that taught a fine-tuned model something in 2024 is worth little once a frontier model does the task out of the box.
The data that matters is data the labs can’t easily get: clinical records, industrial sensor readings, proprietary transactions. Access to proprietary data often comes through trust and partnerships, which take time, and moving too fast can cost you the trust required to obtain that precious data.
Momentum attracts capital and talent. That’s true. However, capital raised on growth hype may be planting the seeds of its own demise. A high valuation can box you into the growth story that justified it. Startups that raise at a saner pace retain the option to change direction.
Learning speed. Fast iteration finds product-market fit sooner. This one holds, but there are two kinds of speed.
Fast iteration means cheap experiments: a prototype, a pilot, a feature shipped to 20 users and pulled if it flops. Fast scaling means hiring ahead of revenue, signing multi-year contracts, buying compute and raising at a valuation you then have to grow into.
Iteration should be as fast as you can make it, because each cycle is cheap to undo. Scaling should come more slowly, because each step is a major commitment that’s hard to unwind, and scaling too early is one of the most expensive mistakes a startup can make: a 2011 study of more than 3,200 high-growth technology startups blamed premature scaling for about 74% of internet startup failures.
Jeff Bezos drew a similar distinction in his 2015 letter to Amazon shareholders. A few decisions are one-way doors, irreversible once you walk through, and deserve deliberation. Most decisions are two-way doors, and those should be made fast by a small group. A startup’s experiments are two-way doors. Its big commitments, a large raise, a long contract, a hiring spree, are not.
Judgment and defensibility: what speed can’t buy
At AI House, Sri Chandrasekar, one of our managing directors, puts it this way: “We don’t hand founders a thesis. We expect them to find the right problem, and that takes judgment, not speed. Building is nearly free now, so run every experiment fast and manually, and only build the machine once something works.”
Finding the right problem takes iteration, acumen and judgment. Speed helps with the first and does little for the other two. A hamster on a wheel at 200 mph has plenty of speed and no chance of progress. Even Altman’s playbook says Y Combinator looks for “businesses that get more powerful with scale and that are difficult to copy.”
The secret weapon
Prof. Herbert Simon, the Nobel laureate who helped invent AI at Carnegie Mellon, advised his graduate students to pick important problems for which they had a secret weapon, because they would win only “by having access to some knowledge or research methods the others do not have.”
Swap students for founders and research methods for data and relationships, and the advice holds. A startup’s secret weapon is an asset you hold that a competitor can’t acquire by moving faster: a dataset nobody else can assemble, a relationship that took years to nurture.
Farecast, the airfare prediction company I founded in 2003, developed a secret weapon by gathering a massive amount of labeled data. By 2007 we were mining more than 175 billion airfares, and every prediction we made was later scored against what the fare actually did, so the labeled set grew every day, and it grew only by waiting for fares to play out.
A startup that signs a data agreement with a hospital network has the same kind of weapon. So does a payments company whose fraud labels come from its own transaction stream. In each case the data is the asset a faster rival can’t copy.
What to do instead
Laurence Peter, of Peter Principle fame, said it decades ago: “If you don’t know where you are going, you will probably end up somewhere else.”
So pick a direction where frontier labs won’t go and clearly articulate why. Run experiments fast but make commitments slowly. Jim Collins, the Good to Great author, called it firing bullets before cannonballs: cheap shots to find the range, then the one big shot once you have it.
Chase the data and the relationships the labs can’t buy and accept that trust is the slow road to both. Raise at a pace that leaves you the option to pivot when the next model lands. And name your secret weapon in one sentence.
To repurpose an old joke: two bulls stand on a hill above a herd. The young one says, let’s charge down there. The old one retorts: charge and they scatter. Walk, and they’ll be waiting when we arrive.
In AI right now, everyone is charging. The founders who end up with the prize will be the ones who are more judicious.