Across four days and more than twenty sessions, 500 Global's VC Unlocked: AI Edition brought together experienced investors and operators to pressure-test how venture capital works when the underlying technology resets every few weeks.

Put two dozen veteran investors in a room for four days and you'd expect them to converge. They did — on exactly one thing. The AI model itself is no longer a moat. "Code is disposable," as one put it, and a clever technical lead lasts maybe a week before it's copied.

Then the fighting started. Because once you admit the model is free, you're stuck with the harder question: is anything built on top of it actually defensible? or is the whole stack just a race to see who gets cloned slowest? Here are the five arguments that split the room.

1. If the model is a commodity, what's actually worth owning?

This was the argument underneath all the others, and the room broke cleanly in two. One camp said that there is no durable software moat left at all. Jeremiah Owyang, a General Partner at Blitzscaling Ventures, built his entire investment thesis on it: because code is a commodity, you underwrite the flywheel instead. Things like proprietary real-time data, community, distribution, and network effects, i.e. the things that compound and can't be cloned in a weekend. On a panel about the future of AI and Venture, Seb Barriga, GP of Milemark Capital said, "I don't care about LLMs at all, they're commoditized." Lu Zhang, Founding Partner of Fusion Fund also falls into that camp, “Companies will say ‘my AI is better, my model is better’. But that may not be the case two months later, and not even two weeks later.”

The other camp didn’t dispute that models commoditize, but they argued the tech moat still exists in places the model can't reach. Helen Liang, Founder and Managing Partner of FoundersX Ventures, locates it in the capital-intensive lower layers of the stack: energy, advanced chips, and scarce deep-domain data. Jake Saper, a General Partner at Emergence Capital, locates it in the guaranteed outcome — a foundation lab, as he likes to say, "is not going to be a fund administrator." Investor Miguel Paredes locates it in proprietary data plus domain knowledge plus switching cost.

So the real question was not whether the model is a moat, since everyone agreed it isn't. It's whether anything sitting on top of it is genuinely non-copyable, or merely slower to copy. The working test that emerged across the week: figure out where the asset a competitor structurally cannot reach actually sits — data, distribution, outcome ownership, or expensive infrastructure — and treat any code-level edge as temporary by default.

2. Are AI gross margins structurally broken, or just early?

A sharper, more technical fight broke out over economics — specifically gross margin. Traditional software keeps 77 cents of every revenue dollar after the cost of delivery. The fight was over whether AI ever can.

Victor Chang, VP of ServiceNow Ventures, presented a "FOMO-free" diligence framework arguing that an AI company may only ever have a path to roughly 50 percent margins, because compute is the new COGS (cost of goods sold) — and many AI companies are running negative gross margins today. Jai Das, Founding Partner of Sapphire Ventures, shared a cautionary tale: a high-profile AI coding company that reportedly sought a sale because it "couldn't get margin under control."

The optimists countered that intelligence is collapsing in price. Google engineer Ajinkya Rasam showed that the cost of a million tokens has fallen by roughly a thousandfold, and Saper argued that margins improve as models get cheaper and spend shifts to open source. The best operators are already posting high-60s margins through deliberate "token-maxxing," automatically routing easy work to cheaper models.

Both can be right at once, and seeing why reframes the whole metric. Jevons' paradox says that as the unit price of compute falls, total consumption rises to fill the gap. So even as per-token costs collapse, the total compute bills will continue to climb or stay the same. That means the number for investors to watch was never cost per token — it's cost per successful outcome, including all the human cleanup a demo never shows you. The practical takeaway: today's negative margins are a red flag, unless the founder can show the cost curve bending in their favor.

3. Is AI the cloud curve, or the smartphone supercycle?

The spiciest moment of the week came in the closing capstone exercise, when one participant laid out a genuinely contrarian bear case: “AI adoption might flatten into commodity infrastructure the way cloud computing did”. Prices fall, durable efficiency gains never materialize, users churn after a few months, and the technology "just becomes a feature". Which, the participant added, "kills all of us here with our little funds."

Jai Das took almost the opposite view. His position was: stop talking about bubbles, because bubbles are how transformative technology actually gets built. By his estimates, we're at "inning zero or one," with more than a decade of enterprise adoption still ahead and several trillion-dollar companies yet to be founded.

What makes this debate useful is that it's falsifiable. The two sides predict observably different futures. Watch retention. If users churn after roughly three months and the efficiency improvements never translate to real metrics, the cloud-commodity bear case wins. If value persists at genuinely early adoption levels, the supercycle case wins. The consensus in the room was that it will take about three years of data to tell the signal from the noise — but in the meantime investors know exactly which numbers to watch.

4. Should a fund automate investment decisions, or only the busywork?

The investors turned the same scrutiny on their own jobs. Salil Deshpande, General Partner and Founder of Uncorrelated Ventures, runs his firm as a solo GP with a fleet of roughly eighteen AI agents — but he deliberately keeps sourcing and investment decisions human. The agents handle the systematic, repeatable work; one of them is careful to introduce itself as his assistant rather than impersonating him. Renata Quintini, Co-Founder and Managing Director of Renegade Partners, draws the line just as hard: "AI doesn't decide anything for us." Access to information, she argues, has been democratized; but the quality of judgment has not.

The participants then ran their own experiment during hackathon. In a diligence screening, an AI agent independently selected 9 of 60 startup finalists and only 3 overlapped with the humans' picks. The group flagged this divergence and sealed it as a 24-month bet to watch. Some in the room were openly more excited about handing a system the data and letting it carry more of the decision-making, noting that at least one major fund already uses AI to score companies and gates deals based on the score.

The consensus was that AI should widen a firm’s coverage and dispatch the "obvious no's." For now, most funds are keeping the actual investment and follow-on decisions firmly human.

5. Is "Wait" a real investment verdict, or a cop-out?

The smallest debate was the most revealing about how AI investors today actually think. In a live investment-committee simulation, a facilitator flagged that "Wait seems like a cop-out answer," and in the debrief, he warned that consensus voting tends to average strong-yeses and strong-nos into a mushy middle. However, the best investments tend to come from strong disagreement.

But the program's own deal-memo rubric explicitly lists "wait — see them next round" as a legitimate verdict, and Quintini's "decide before the decision" protocol formalizes it: pre-commit, in advance, exactly what a company must come back to you and prove before you'll buy.

The difference comes down to one thing. A "wait," used to avoid taking a position is a cop-out. A "wait" anchored to a pre-committed, falsifiable condition is a discipline. The entire difference is whether you've specified, ahead of time, what would change your mind.

The thread that ties it all together

The overall consensus during the program (that intelligence is commoditizing) turned out to be the boring part. Every real debate was about durability: once the model is effectively free, what survives, and how would you know if you were wrong about it?

This is the actual lesson. The investors who sounded most credible were the ones who could name the specific evidence that would prove them wrong — the churn number, the margin curve, the retention cliff. It's the discipline 500’s Tony Wang and Amjad Ahmad pressed on the cohort from Day 1 and the capstone returned to on Day 4. In a market where the technology resets every few weeks, that may be the only edge that doesn't commoditize: not having the contrarian take, but committing to how you'll find out your take is wrong.