Paulo Passoni on why the world is short compute, and how a great company still loses money for its last investors

The J Curve is an English-language podcast and newsletter covering Latin America's technology and venture capital ecosystem, hosted by investor Olga Maslikhova. TJC Debrief is its monthly companion with Paulo Passoni of Valor Capital. This edition opens on Stripe buying OpenRouter and ends in Venezuela, by way of the Excel model Paulo built to show that you can pick a company that goes from nothing to $16 billion of revenue and still lose money on it — and the shortage he thinks nobody in Brasília has noticed, which is not capital or talent but hours of chip time.

What is Stripe actually buying with OpenRouter?

A position in the middle of the money.

Start with the problem OpenRouter solves. For any given task there is a model that does the job well enough at the lowest price. Which model that is changes daily, and it is a different answer for video, for coding, for customer service. The Fortune 100 can track that in-house — they have the teams. Everyone smaller pays someone to track it for them, and there is margin in doing it. That is the orchestrator's job. Paulo puts OpenRouter and Perplexity in it.

Now follow the money. The orchestrator pays for every model it routes to, and that compute sits in Latin America, Europe or the US depending on the provider — so it is making payments all over the world, constantly. Stripe has just bought the company sitting on that flow. That is the bet: if agents end up making most of the payments, and buying intelligence becomes the main money flow, Stripe owns the middle of it.

The obvious caveat: it only works if OpenRouter stays independent and never favours one model over another. Paulo says that neutrality is one of the best things about it.

If routing is not defensible, where does the moat sit?

In what you keep away from the model companies. Routing on its own Paulo calls weak. What is defensible is memory, customer data held on your side, evals built across models, workflows embedded, safety handled once across a single deployment.

Higgsfield is his worked example, and the reason he describes it as an ad-tech company as much as a video orchestrator: the ads side generates performance data on the videos created through it, that data reshapes the orchestration layer, and none of it flows back to the labs. The counter-example is legal. Doing diligence on Harvey and Legora, he found that neither holds memory for its customers — a firm with a house style for drafting clauses still gets a general answer. He expects that to change fast, and thinks whoever holds the memory holds the client, because a law firm whose data trains a rival's answers has a problem.

Why is Harvey already in Latin America?

Because legal AI is a land grab and Harvey got there first. Paulo does not think the product decides this market: Harvey and Legora are both going hard at it, one may be a little better than the other, but not ten times better. So whoever signs a law firm keeps it — switching to something roughly equivalent is not worth the disruption. Get there first, get the client, and they stay.

Which makes early entry rational even before you have won your home market. When to go is a product question: if it travels easily, go as soon as you can, and if it does not, you had better believe you are so far ahead that switching is painless. The destination list is short. A Western applied-AI company goes to the United States, then Europe, and then runs out of options — China is closed, India is its own ecosystem, Africa is too early and likely to go to Chinese applications.

Which turns a commercial question into a geopolitical one, and Paulo says so directly: Latin America is where China will compete with this ecosystem, its companies and governments will have a choice, and an investor sitting between US and European tech firms and the region is doing something closer to foreign policy than to investing.

Why are AI companies reaching revenue scale so much faster?

Because the addressable market is wages. Older software made people more efficient; this software replaces them. Paulo's example is a video shoot — crew, videographer, editors, days of work — now done by one person at a computer. You are not selling a seat. You are selling a production budget.

Two things compound it. The products travel: a video tool is the same in China, Russia, Europe, the US or Brazil. And there is no salesforce — Higgsfield sells the way a mobile games company does, and is partly staffed by people who came out of gaming. CloudWalk does the same in Brazil, where Paulo says the machinery behind InfinitePay's marketing is what outsiders never see.

The obvious caveat: an agentic product still has to be adapted in every market, and that adaptation decides who wins. The end state is deflation in anything that used to need people — not unemployment, he thinks, but different jobs. Agriculture to factories took a century. This takes a decade, and whether political systems can cope is his worry. Maslikhova adds that education cannot move fast either.

How does a great company lose money for its last investors?

The multiple investors pay should decline as a company matures and approaches its IPO. Right now, it often doesn't.

Paulo ran the numbers in Excel. Assume a company grows from zero to $16 billion in revenue over ten years, reaching $400 million by year two — better than almost anything from the last cycle. By the end, it earns a 20% net margin and goes public at 25 to 50 times earnings. That implies an IPO

valuation of five to ten times revenue.

For the economics to work, the revenue multiple has to compress along the way. Paying 100 times revenue at Series A can make sense. Fifty times at Series B can, too. But by Series C and D, the multiple should be closer to 20, falling to around 15 just before the IPO.

On that path, late-stage investors earn annual returns of 20% to 30%. If they keep paying 50 times revenue right up to the listing, their returns fall to 6% or less — and, in some cases, turn negative.

Dilution does not change the conclusion. Once the company is worth $10 billion, each new round typically costs existing shareholders only one to three percentage points.

That leaves the founder making a bad trade:

Paulo points to Elon Musk, whose investors made money at every valuation. That track record is why he can now raise almost any amount of capital.

The exercise began after a trip to San Francisco, where Paulo encountered Series C and D rounds priced at 50 times revenue, with valuations between $10 billion and $30 billion.

For those investments to work, every one of those companies has to become a $100 billion to $300 billion business.

What is Thrive doing that other funds won't?

Two things. Thrive refuses to call itself an early or a late fund, because the boundaries move every year. And it buys into the same company again out of each new fund — which plenty of LPs call a conflict.

Paulo's answer to that: finding an extraordinary company is the hard part, so being told to stop backing one you have already found is backwards. The conflict costs less than the missed deal. He takes the risk seriously — if that company turns out to be a disaster, every fund holding it goes down together — and still thinks repeating on winners wins.

Why is the world short of compute?

Production is not keeping pace with demand — and the price makes that visible.

An hour of access to an NVIDIA chip cost $2.50 last year. Today, it costs $4.50. Paulo says transactions have already cleared at $9, and prices could climb further.

The hyperscalers built their businesses to earn returns of 20% to 25% on invested capital when compute cost $2.50 an hour. In the near term, then, those returns should rise, not fall.

Paulo expects the shortage to last at least three years, and possibly five. Over the next decade, he sees returns on compute settling closer to 15%.

Cheaper open-weight models do not eliminate the constraint. You can replace Anthropic with Kimi, but Kimi still has to run on a chip inside a data centre.

His evidence that the shortage is reaching real companies came from one of his own founders. The company was operating at scale and at the cutting edge. It needed $36 million worth of chips, could barely source them and feared that products already in the market would stop growing.

This is where compute becomes geopolitics. The founder put it plainly: the world is dividing into those who can access compute and those who cannot.

The United States controls much of the supply. American multinationals may therefore capture an enormous share of global markets for the simple reason that they can obtain the hardware. Paulo's aside is that you could explain this dynamic in Brasília and be met with blank faces because the political debate is focused elsewhere.

He believes that blind spot extends across much of the Global South. Its consequence could be even greater technological and economic domination by either the United States or China.

It also changes the nature of Jensen Huang's role. When NVIDIA allocates capital to an applied-AI company in exchange for commitments to buy its chips, Huang begins to look less like a chief executive and more like a national development bank — allocating scarce resources and deciding which companies can become winners.

NVIDIA has also assembled a consortium with major private-equity firms and Goldman Sachs to support the residual value of its chips. The structure turns GPUs into financeable assets, packaged somewhat like collateralised loan obligations.

And leverage, applied to anything, magnifies the outcome.

What should a founder do about it?

Sign the contract that looks too expensive.

The corollary is where the value accrues while the shortage lasts: NVIDIA, memory, and the unglamorous physical layer — buildings, cooling, power — all of which have to be built before a chip earns anything. Paulo notes the challengers, because the profit pool is large enough to attract the best minds in the world: Google's own inference chips, Anthropic building its own, and Etched, the Harvard dropouts' company, which shipped its first cluster to Jane Street. He calls that healthy. It is capitalism.

Why did only fourteen Latin American companies raise a Series B in seven months?

Partly a vacuum, mostly exits. The vacuum is technological: companies built with the new tools have not matured to Series B yet, and investors reserving capital for what they know is coming will skip a cohort of perfectly good businesses built the old way. Some of those still get funded, especially the ones becoming AI companies quickly.

But the real answer is that nobody can get out. Finding a company at $10 million of revenue and watching it reach $50 million is worth nothing if there is no exit at the end of it, and Paulo counts himself among the people who have made that mistake. Which drives the whole market toward fewer, larger bets — companies you can argue will be publicly traded, because the alternative is selling to a local incumbent.

Valor's last two positions follow the logic exactly: more CloudWalk, bought as secondary, and Plata. Both massive, both liquid on his reading, and both — because there is so little competition for Latin American growth deals — priced reasonably.

What does a company have to look like to be growth-investable now?

Paulo describes a simple staircase. From a $20 million revenue base, he wants 5× growth in a year — ideally 10×. Below 5×, he is not interested. From $100 million, he wants 2× to 3×. From $500 million, at least a doubling to $1 billion. Beyond $1 billion, he is content with 30% to 50% annual growth for five years, which still compounds into several billion dollars.

Margins depend on the business. A global orchestrator may reach 40% to 50%, but not the 70% to 80% of classic software. Some Latin American companies can operate at 50% to 70%.

On unit economics, he looks for two things. First, an honestly calculated LTV-to-CAC ratio: 5× to 10× is healthy; 1× is not. The common mistake is a CAC figure that quietly excludes part of the acquisition cost.

Second, he looks at the absolute capital required — not just the ratio. The era of burning billions to build the next Uber is over. Higgsfield reached $700 million in annualised revenue on less than $200 million of capital. Both Higgsfield and ElevenLabs have shown that they can stop burning cash and continue growing.

The metric that may matter most is not a growth rate at all: revenue per employee.

Above everything sits talent. Paulo's advice to founders is to stop asking what valuation investors will pay and start asking whether they can attract the mathematicians and physicists capable of building the product. Without that talent density, nothing else follows.

What would Paulo not back today?

A company built on a hole in the legislation, however fast it is growing, because the business can disappear the morning the hole closes. And, more broadly, software — he is unsure which software companies will be worth anything, notes multiples have derated to around three times revenue. Maslikhova offers the Airtable acquisition as the illustration; he agrees.

The caveat is that some of them will adapt and use AI to their advantage. His doubt is not about the strategy but the staffing: the companies that most need to move usually cannot attract the people who would move them.

Are Brazil's big seed rounds building the pipeline?

Only if revenue keeps pace with valuation. The higher the seed or Series A price, the more revenue a company must add in absolute terms. Decade, for example, raised an $85 million seed round before generating revenue. At that entry price, it would need to reach at least $200 million within 18 months.

Valuation multiples tend to originate in San Francisco and spread outward at a discount. That discount creates an attractive model: companies from Eastern Europe, Sweden, Kazakhstan or Argentina can keep more stable, often higher-quality and considerably cheaper technical talent at home while selling globally from day one.

It is the early Israeli model — and arguably the best risk-reward available.

Why does Brazil raise more structured credit than equity?

Because the instrument can manage risk better than a bank — and the regulator designed it well.

A FIDC is a Brazilian receivables fund regulated by the CVM. Paulo considers its guardrails significantly stronger than those of the American equivalent.

They are stronger precisely because Brazil is a riskier market — which he sees as a feature, not a caveat.

The structural advantage is duration. A bank funds long-term loans with deposits that can be withdrawn, creating a mismatch. In a FIDC, capital is locked up for a term that matches the underlying loans. It is better asset-liability management inside a vehicle often dismissed as “emerging-market credit.”

The mechanics explain the returns. A lender begins with its own equity, building a record of actual losses. It can then raise a FIDC against that history, typically with a 70% advance rate while absorbing the first 30% of losses.

As the asset class proves itself, the structure tightens. A mezzanine tranche appears in the middle, and the equity tranche beneath it becomes much smaller. Returns on that equity can then become enormous: Paulo estimates that one Mercado Pago FIDC may generate more than 400% or 500% annually on its equity tranche.

Where does vertical lending go wrong?

The risk lies with the founder, not the instrument.

Vertical lending is one of Valor’s favourite theses, despite its unpopularity with many venture firms. The model appears in import finance, education, agriculture, cars and many other sectors. The danger is temperamental.

A founder focused on maximising the revenue curve is unlikely to underwrite conservatively. The consequences emerge later.

The recurring mistake is underestimating reserves. Regulators tell banks how much capital to hold. A FIDC founder receives no such instruction and may assume that the equity tranche is the entire buffer. It is not.

When losses rise, the founder must inject enough equity to protect the mezzanine and senior tranches — and that capital must already be available. Failing to make those investors whole can permanently close access to the credit market, destroying the business’s long-term value.

The alternative is an emergency call to shareholders: the underwriting was wrong, and the company needs more money. That means raising venture equity at the moment of maximum dilution.

Paulo’s prescription is to grow slightly slower and hold what may look like excess cash. It is not excess; it is operating capital. Because no regulator sets the reserve level, the company must regulate itself.

He considers CloudWalk the safer version of the model, largely because of its distribution: seven million small merchants who are expensive to acquire but inexpensive to serve.

Why Venezuela?

Because it combines a compressed economic base with unusually resilient talent.

GDP per capita has fallen so far that normalisation alone could double or triple it over a decade. That does not require a return to 1995 levels — only enough recovery to create a substantial investment opportunity.

One company positioned for that recovery is Cashea. The model is Kaspi, the Kazakh company that began in fintech, expanded into e-commerce and became part of the country’s operating infrastructure.

The perception of political risk has also shifted. US intervention under the Trump administration recast Venezuela, almost overnight, as moving away from China and Russia and towards the United States. The expectation is that this realignment could survive a change in administration.

But the core asset is talent. Years of instability have produced founders who are difficult to intimidate.

Venezuelans are highly entrepreneurial and hardworking, with a resilience comparable to Argentines. Even so, the opportunity is unlikely to support a Venezuela-only fund. Instead, regional investors — including Kaszek and Valor — are adding the country to broader Latin American strategies.

Why does the US suddenly care about Latin America?

Minerals, energy and migration, in a decoupling from China. Maslikhova puts it that this is the first time the US government has cared about the region; his answer is that the reasons are unromantic: self-sufficiency requires inputs that sit in Latin America and have to flow north rather than east, and migration is not an irrelevant consideration. Africa is the cautionary example, and the difference is proximity.

Paulo's warning is about method rather than motive. In the 1960s and 70s the US worried about Soviet alignment and sponsored military regimes that looked fine and turned out corrupt and inept. So the how matters — and within the region there will be winners and losers, because what decides a country's outcome is not Washington but whether its own leaders find a model that works while staying aligned.

COMPANIES - AI AND INFRASTRUCTURE

⠀COMPANIES — LATIN AMERICA

⠀PEOPLE AND FIRMS

⠀IDEAS AND FRAMEWORKS

A monthly companion to The J Curve, in which Olga Maslikhova and Paulo Passoni take apart the month's news at the intersection of Latin America and global markets. Where the interview series is built around one person's history, Debrief is built around what has happened since the last one — and it is where the archive's standing positions get revisited, revised and occasionally reversed on the record.

Managing partner at Valor Capital Group, the cross-border firm investing between the United States and Latin America. Previously a managing partner at SoftBank's Latin America fund and an investor at Third Point. The positions he discusses here are CloudWalk, bought through secondary, Plata, and Status at the early stage.

TJC Debrief — what he would do with $500 million, and why an AI margin is not a SaaS margin · Paulo Passoni, Valor Capital — why speed is the last moat, and the CloudWalk–Ramp gap · Brian Requarth, Latitud — the anatomy of a fundraise

🎙 The J Curve  is where LATAM's boldest founders & investors come to talk real strategy, opportunity and leadership.