JJB

Founder & CEO @MegawattFinance · Building startups since 2010 · In Web3 since 2016 · Shipping AI products · Where technology, capital, and the grid converge

[JJB @jacopoburiollo]

Inventory is a good first case for this because it comes with the two things on-chain credit usually lacks: a verifiable quantity and a market price someone else sets.
A lender can check what exists and what it is worth without trusting the borrower's own valuation. Most physical assets fail one or both of those tests, which is why the borrowers who arrive first tend to be the ones whose collateral was already legible.

[JJB @jacopoburiollo]

Half of running a physical asset is deciding what to spend money on now to avoid spending more later.

Nobody gets credit for the failure that did not happen, which is why the discipline is rare and the returns are not evenly distributed.

[JJB @jacopoburiollo]

Most bad forecasting in infrastructure comes from reading an announcement as a commitment, then being surprised when a thing that was never certain fails to happen.

[JJB @jacopoburiollo]

The compaction habit has a second benefit: writing the summary forces you to decide what actually mattered, which is a step people otherwise skip. Half of what accumulates in a long working session turns out to be process rather than substance, and only becomes visible when you have to compress it into something a fresh context can use. The discipline is closer to writing a handover than to managing a tool.

[JJB @jacopoburiollo]

The tell I have learned to watch for is whether anything changes on their side between meetings. Free enthusiasm but a document, an internal owner, a date, a small commitment. Meetings without any of those are a process running on someone else's calendar, and the founder is the only participant paying for the time.

[JJB @jacopoburiollo]

The most useful people in my network are the ones who have told me I was wrong about something specific.

Almost none of them were trying to be helpful at the time.

They just knew the thing better than I did and could not let it go.

[JJB @jacopoburiollo]

Most of that list runs into the same wall: there is no counterfactual. A founder takes one path and never sees what the other would have produced. Which is why survivor stories keep getting read as causation, and why the interesting cases, the companies that did everything right and still died, mostly go unrecorded.
AI could help with the part nobody wants to do, which is reading the failures. Far more gets written about companies that worked, and the ones that did not usually leave no account at all. Anything that reconstructs what happened from fragments widens the sample.

[JJB @jacopoburiollo]

The thing that surprises people about financing energy assets is how little of the risk is technical.

The panels work. The batteries work.

What decides the return is whether someone will pay a reasonable price for the output a decade or two out, and nobody has a model for that.

[JJB @jacopoburiollo]

This is the argument for staying small longer than the plan says you should. An organisation built for the old productivity range has approval steps, handoffs, and coordination layers that exist to manage the limits of human throughput. Add AI to that, and you get output the system cannot absorb. Start without those layers, and there is nothing to unwind. Which means the constraint is not adoption; it is that most companies have to dismantle something before the gains show up.

[JJB @jacopoburiollo]

The evaluation problem you describe generalises past search. Anywhere a model's output depends on a retrieval layer you cannot see, you are trusting two systems and can only assess one. For anything where being wrong is expensive, that opacity is the thing that decides how much of the output you can actually rely on, and it is usually undocumented.

[JJB @jacopoburiollo]

The same problem shows up outside benchmarking. In any domain where the model is now competent, the comparison you actually need is against what a good human would have produced on the same task. Almost nobody has that reference available in the moment. So you end up evaluating output against your own expectations, which drift upward as the tools improve. Losing the human baseline in benchmarks and losing it in daily use are the same failure at different scales.

[JJB @jacopoburiollo]

Worth pairing that chart with the wallet-level data on who actually bought. Traced buyer by buyer, the $30B came almost entirely from crypto-native holders, protocol treasuries, DAOs, funds, with a small number of wallets holding most of it. Which makes the growth real and the composition narrower than the headline suggests. The Caracas-to-Calcutta phase is still ahead rather than underway, and the constraint on it looks less like access than like whether an allocator can defend the position to someone who was not in the room.

[JJB @jacopoburiollo]

Most RWA infrastructure is built for the moment an asset gets issued.

Very little of it is built for the moment someone has to enforce a claim against the thing underneath.

That gap will decide which of these products survive a bad year.

[JJB @jacopoburiollo]

The correlation argument holds until liquidation, and that is where it gets tested. Uncorrelated collateral is only useful if it can be sold at a knowable price during the same stress event that made it valuable, and for real assets, the sale itself is slow and thinly bid precisely when everything else is falling. So the diversification is real on the balance sheet and can disappear at the moment it matters most. Shorter-duration fixed income arriving first makes sense for exactly that reason.

[JJB @jacopoburiollo]

Forty percent of corporate renewable PPAs from one sector is the number that reframes it. That is not procurement; it is a buyer large enough to decide which projects get built. Which changes what a developer optimises for, since a counterparty with that much credit and that much urgency can contract for output others cannot.

[JJB @jacopoburiollo]

The compliance-review framing is the useful part, because it explains why the gap persists even where demand exists. An allocator needs to defend a decision to someone who was not in the room, which means the asset has to arrive with independently verifiable risk data attached. For anything illiquid, that data has to be produced deliberately, by someone accountable for it, and continuously rather than at issuance. Which makes the missing piece less about tokenization standards and more about who does the ongoing work of establishing what the thing is worth.

[JJB @jacopoburiollo]

People picture infrastructure risk as something dramatic.

A storm, a failure, a fire. Those are the risks that are already insured and priced.

What actually decides the return is whether the thing sells its output at a decent price for twenty years.

[JJB @jacopoburiollo]

Supplier insolvency is the risk that sits quietly inside every long-dated warranty. A twenty-year performance guarantee is only worth the balance sheet standing behind it, and battery manufacturing has thin margins and heavy capex, so that balance sheet is often the weakest link in the chain. Worth pricing the counterparty, not just the warranty terms.

[JJB @jacopoburiollo]

Using these daily in an unbounded field, the thing I notice is that verification still happens, it just runs on a multi-year lag. The model is often good at a judgment call and nobody can tell for years whether it was right. Makes it genuinely hard to know whether capability improved or my confidence did.

[JJB @jacopoburiollo]

Fourteen years from 1% to over half. What follows is harder, because the same success compresses what each new plant earns in exactly the hours it generates. California is further into that than anywhere in Europe, which makes it the best available preview of what high-penetration markets do next.