Your New Favorite VC Metric Is Two Guys and an FDA Letter

A founder pitched me on Monday using a slide I have now seen four times this quarter. It was a revenue-per-employee chart. Salesforce was on it at roughly $530K per head. Atlassian at $700K. Cursor sat on the right side of the chart at $6.7M. Medvi, tiny and highlighted, floated above at around $200M.

The founder tapped the slide. "This is the new power law," he said. "You are either above the Cursor line or you are not investable."

I nodded, because that is what you do in pitches. Then I closed the Zoom, opened a second browser tab, and pulled up MEDVi LLC, warning letter number 721455, dated February 20, 2026. FDA Office of Compliance. Misbranding. Falsely implied FDA approval. Falsely represented Medvi as the compounder of products it did not compound.

The best proof point of the biggest new VC thesis of 2026 has an open FDA file.

Agentic leverage is real, and the math is obscene

Here's the thing. I don't think the thesis is wrong. I think the math is honest.

Cursor (Anysphere) is staring down a $60B valuation round with fifty employees and $2B in ARR. That is roughly $6.7M in revenue per human head, which is something like 10x the enterprise-SaaS baseline. More than half the Fortune 500 uses the product. Corporate customers generate 60% of revenue. This is not a trick of headcount accounting. It is a genuine throughput step-change, and if you came up through engineering, you recognize the shape: it's the difference between a batch-processing system and a streaming one. Same input, same customer, radically different operational envelope.

Sequoia, which just closed a fresh $7B AI fund this week, has formalized this into underwriting. Their partner Julien Bek put it plainly: for every dollar spent on software, six go to services. Agentic leverage is the wedge into those six dollars. Pace closed a $10M Series A from Sequoia to automate insurance BPO. Auctor took $20M to build the agentic OS for software implementation. In both deals, the line item the agent is eating is not a software license. It is a pool of human contractors billing hourly. That's the actual TAM unlock.

If you have ever read a distributed systems paper, you know exactly what Sequoia is doing. They added a throughput-per-node metric to the cost model. Frankly, it was overdue.

Here is where I start to squint. When "output per human employee" becomes the metric your investor scores you on, the rational founder response is to keep humans off the cap table. Not just to stay lean. To never staff up in the first place.

That is fine for a developer-tools company, where the product is the moat and the customer expects self-serve. That is less fine for a regulated business, where the humans you don't hire are compliance officers, medical review boards, and customer service reps with judgment about edge cases. Those humans are expensive precisely because they absorb liability. Take them out, and the liability doesn't disappear. It just shifts somewhere downstream, usually to the customer or the regulator.

Which brings me to Medvi.

The FDA letter inside the poster child

The New York Times published a profile on April 2 framing Medvi as the potential proof of Sam Altman and Dario Amodei's predicted one-person billion-dollar company. $401M in 2025 revenue. Two employees. $20K of starting capital. GLP-1 telehealth, built on a stack of Claude, ChatGPT, Grok, ElevenLabs, Midjourney, and Runway. 250,000 customers. A 16.2% net margin. The kind of numbers that make a seed partner's spreadsheet crash.

Six weeks earlier, the FDA had sent them a warning letter citing misbranding. The website copy suggested FDA approval of compounded products when there was none. It implied Medvi was the compounder when it was not. Drug Discovery & Development later catalogued more than 5,000 active Medvi-related ads on Meta, with several running under apparent fictitious personas with fabricated medical titles. An industry-wide action hit 30 telehealth companies for the same category of violation in the same window.

Yeah. Ouch.

Here is the honest read: the incredible revenue number is partially explained by the warning letter. When you replace the compliance function, the sales function, and the medical review function with agents, the unit economics light up in exactly the way Sequoia's model loves. And when the company generates $3M a day with two humans, there is nobody left in the org chart to apply the brakes.

I don't want to single out Medvi. Maybe they resolve the FDA action. Maybe the ads are nuanced. But as a datapoint in the agentic-leverage thesis, it is impossible to un-see. The metric and the liability structure point in opposite directions, and the metric is winning the pitch decks.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more

What a liability-weighted version of the metric looks like

If you have ever run throughput benchmarks on a production system, you already know the trap. Nominal throughput always goes up when you disable the expensive per-request checks. The graph looks great. The graph also lies, which you learn at 3 AM when the incident postmortem comes in.

Revenue per human employee is a real, legible metric. For agentic companies, it is currently measuring two things at once, and nobody is separating them well. It measures (1) genuine software-eating-services productivity, which is the Cursor story, and (2) the willingness to operate without the humans who regulators expect to see, which is part of the Medvi story. Both are priced the same right now. They should not be.

Here is the fix. It is boring. It will not trend on Twitter.

Add a second column next to every agentic-leverage company's revenue-per-employee number. Label it "regulated roles removed from the org chart." Compliance officers. Medical review boards. Licensed professionals tied to the product. Customer-service escalation with discretion over refunds or harm. Moderation. Underwriters. If the agent is doing the job, fine, but the diligence should model the blast radius when the agent is wrong and the audit shows up. That blast radius has a real name in finance: contingent liability. Enterprise SaaS comps already price it implicitly, through customer concentration and churn covenants. We are not doing that for agentic companies yet. We should.

Then add a third column: plausible regulator-attention scenarios. What is the worst five-business-day news cycle look like for this business? What happens to the revenue curve if the primary sales channel (Meta ads, Google ads, affiliate networks, App Store rankings) disqualifies the category? For Medvi, that scenario is live right now. For Cursor, the regulatory surface is mostly IP ingestion and enterprise data handling, which is a totally different blast radius. Same metric on the pitch deck. Radically different downside distributions.

If I were underwriting agentic-leverage deals at scale, this is the slide I would build. It is less sexy than $6.7M per head. It is the slide that tells me which of these companies survives the first regulator letter.

The boring answer to the hottest thesis of 2026

Here is where I land. The agentic-leverage thesis is not fake. Cursor is a genuinely excellent business. Sequoia's new $7B AI fund is going to generate returns that make a lot of people look slow. But "output per human," priced naively, is a metric that cannot tell you it is lying, because the variable that would give it away (humans removed from the org chart) is exactly the variable the metric is hiding.

The first rule of production engineering is that the best metric is the one that tells you when the system is wrong. Right now the VC industry is marking up agentic companies on a metric with no such property. Eventually, an FDA letter, or a class action, or an ad-platform deplatforming event, or a regulator inquiry will do the telling. It will happen to a name everyone in the room recognizes. When it does, the pricing will reset, and the firms that priced the contingent liability first will look smart, and the firms that priced the hype first will pretend they saw it coming.

I would rather be boring on the way in than right on the way out.

— SWEdonym