Bending Spoons: 94% of code written by AI, leverage ratio 2.4x

Luca Ferrari's company is called Bending Spoons, and it is now worth $20 billion. What he does doesn't sound trendy at all: he buys software that others can no longer keep running, fixes it up, makes it generate more cash, and then uses the extra cash plus leverage to buy something bigger. Airtable, AOL, Vimeo, Eventbrite, Komoot, WeTransfer—more than fifty companies in all. And this company now has 94% of its code written by AI, hands over more than 90% of its data analysis to AI, and basically no longer has its own people answer customer service tickets. On one side is extremely high per-person efficiency; on the other is 2.4x leverage. Put these two things together, and a normal person's first reaction is that something is going to blow up.

He specifically picks the things nobody wants to fix anymore

Anish Acharya of a16z opened by saying the company's vibe is like McKinsey's intellectual diversity plus Google's technical feel. Luca explained the business in thirty seconds: buy technology companies with room for improvement, plug them into their own platform—the same underlying technology, the same core team—and turn them into a machine that spits out more cash on a ten-year horizon. If you get it right, you take the extra profit, add leverage, buy a bigger company, and do it again.

This positioning is very unusual. Marc Andreessen has a saying that there is no such thing as stocks in the world; every company is either a bond or an option. Luca said that line is a bit clickbaity, but the direction is right. Apple rode the iPhone, NVIDIA ran into AI—both were inventions, but if you reran that history ten times, eight times it wouldn't happen; someone else would have gotten there first, or it simply wouldn't have worked out. He put it bluntly: radical invention can take people very high, but it is unreliable, and the median outcome is losing everything.

So he has to pick a side. Teams doing radical invention and teams doing extreme optimization have different temperaments, mindsets, and cultures, and a company trying to do both usually does neither well. He stands on the optimization side. His example is Google: thirty years on, search still contributes 80%, and tens of billions of dollars spent on other things in between failed to change that ratio. Anish added that the Transformer paper also came out of Google. Luca said he had even run the numbers: if all the excess money back then had been reinvested in search, shareholder returns might have been better.

Turn 100 of profit into 200, then multiply by three times leverage

The place this model is most likely to crash is debt. Luca's rules are hard: keep leverage below his own 3x ceiling—currently 2.4, and historically closer to 2. Refinance every one to two years, pushing maturities out beyond five years. Hedge 100% of interest rates, so how much interest will be paid in the future is certain. Even under extremely adverse scenarios, ensure all debt can be repaid before maturity.

He volunteered a disclaimer, and put it bluntly: with those companies that eventually blew up, if you had asked them beforehand, every one of them was optimistic about using debt. He treats paranoia as discipline and does not mistake the smooth years of the past few years for ability.

What leverage amplifies—he gave a very concrete piece of arithmetic. A company earns 100 in EBITDA; you grow it to 200, multiply by three times leverage, then go buy the next one. Roll that loop for twenty years, and it could be 100x. He also named Broadcom and TransDigm, saying these two have done exactly this for decades, and without doing so they couldn't have reached today's scale. As for diversification, he warned not to trust it too much: assets you think are uncorrelated may collapse together when they hit the same phenomenon.

AI writes 94% of the code, and people move to designing processes

Two years ago, the share of code written by AI at their company was still 0. The last time he checked, it was 94%. In data analysis—pulling data, cleaning, charting—he estimates at least 90% is done by AI. Customer service is even more thorough: tickets either go to AI or to external vendors, and basically no one in-house answers them one by one anymore. The customer service manager's job now looks more like a product manager's: designing response rules and designing the experience users have when they run into problems. He said the next step is to bring phone and WhatsApp into it as well.

What work remains? Architecture, process design, and building tools for budgets and dashboards. He gave a benchmark: no decent company should now be below 70%, and below that number is waste. Anish picked it up and said everyone should converge toward 100.

They write all their internal tools themselves—more than fifty of them, from A/B testing, marketing automation, and user lifecycle prediction to credential management and recruiting—keeping only general-purpose things like Gmail and Slack bought off the shelf. He said the vendor fees saved don't amount to much; what is really valuable is two things: every tool grows tightly around their own processes, and they are naturally interconnected. After acquiring a company, swapping in the whole set is faster than replacing things one by one.

The year SaaS fell 40%, he didn't panic along with everyone else

A year ago, during that SaaSpocalypse wave, Anish said related companies generally fell 40%. Luca's view is that those valuations never held up in the first place, because they assumed 12% organic growth could last fifteen years. His exact words were sharp: organic growth almost never goes up; it usually goes down, so that set of assumptions simply doesn't make sense. AI just gave everyone an excuse to go back and revise their models, and only after the fall were prices barely reasonable.

He doesn't buy the other narrative—that coding has simply gotten cheaper, so SaaS is doomed. His reasoning is that the technical cost of copying a product was already negligible five years ago relative to the value of "winning this market." For a billion-dollar company, you might have needed only twenty good engineers to clone it back then, but if you couldn't get customers after cloning it, it was all for nothing. He gave a very small example: even if you invest only 0.2 engineer's worth of work hours to build it yourself, the cost is still higher than just paying Airtable directly—a rational person wouldn't do that.

What will really get hurt is the edge. New customers may in the future be chosen by agents, and agents' screening criteria differ from humans', so those who used to choose A now choose B. That will hurt, but it won't make a company dissolve overnight. Anish added a number: software accounts for 8 to 10% of total enterprise spending, and pouring token budgets into rewriting payroll systems has big downside and small upside.

Data is not a warehouse that becomes more valuable the more you accumulate

Many people say that some companies' market caps are already below the value of the data they hold. Luca said he hasn't fully thought this through, but the claim is somewhat misleading. His argument is that the value of incremental data falls extremely fast. If you write a new blog post online today, no one will pay you for it. But if you rewind to the era when models first started scraping data, the first blog post that could be obtained might have been worth a thousand dollars. Now it's worth zero, because it adds nothing incremental to the corpus pool.

Pushing this logic further, the conclusion is not friendly to data moats. In a given niche, the first people to sell the data can make money, and afterward it trends toward zero, because the marginal return on training models with new data is diminishing. He didn't dare make a heavier judgment, saying he isn't qualified to determine whether a company can exceed its own valuation by selling data.

His last startup was shut down by the tax authority

Anish asked him how founders should judge whether to persist or pull out. Luca's answer was that he is probably the least qualified person to answer. Before Bending Spoons, there was a company called Evertale, which he ran for three years, trying to use AI to help people write diaries automatically—too early, users didn't use it, and even less did they pay. They kept pivoting and kept holding on, until only $250,000 to $300,000 was left in the account, most of which was still counting on a tax rebate from the Danish government. Then the government wrote to say that the rebate either didn't hold up or would have to wait twelve months. The company just died that way. He said you could also say red tape made the decision for us.

His advice to the next generation of founders is only three points, all very plain. Aim high, because a person can't do many things in a lifetime, and picking big things has a higher probability of return. Use logic and first principles for decisions; data is input, intuition is also input, but in the end you can't decide just because "I think this is cool." And talent density: if hiring is loose at the start, it's almost impossible to make up for later—this is the point he stressed most.

His efficiency numbers are so good they seem unreal, but throughout the conversation he kept returning to the leverage ceiling, how much debt can still be repaid under extreme scenarios, and how long it takes to make up for that one moment of looseness in hiring. After AI pushes execution costs close to zero, all the hard parts are in capital structure and people.

More hardcore podcast content: Podwise