Jun 15–21 | AI Margin Is Product · Harness Moats · Scarcity Pricing

AI margin is now a product decision

The clearest AI signal this week was not model capability. It was cost. Harvey described legal drafting queries that can cost around $20 and large review jobs across 100,000 contracts that can run to roughly $20,000. Its usage has reached around 13 trillion tokens, while Perplexity says revenue has more than tripled since the start of the year and burn is down more than 50%. That is what healthy pull can look like, but it also shows how quickly AI companies move from “can we ship it?” to “can we afford to run it?”

The awkward truth is that token billing may become the new billable hour: disliked, hard to explain, and still difficult to replace when work is variable. The usual SaaS instinct is to hide complexity behind simple pricing, but agentic products break that bargain. Better agents run longer, use more tools, get invoked more often, and compound cost from both sides. The product is no longer just the answer; it is the economic control system around the answer.

Operators should treat model routing, task classification, cost attribution, and billing explanation as core product surfaces. The useful question is not “which model is best?” but “which model is good enough for this workflow, at this latency, with this margin?” Track spend by customer, task, model, and usage pattern. Stress-test the worst customer week, not the average week. If the customer asks why an agent spent what it spent, the answer should be visible in the product, not reconstructed later in finance.

The moat is the harness, not the model

Across Harvey, Perplexity, and Anthropic’s engineering culture, the same point kept appearing in different language: the model is not the product. Harvey’s Legal Agent Bench looks less like a chatbot test and more like a software evaluation: a data room, a partner request, a generated diligence memo, and a rubric that works like unit tests. Fiona Fung described a world where engineers can ship far more code, but the hard problem shifts to specs, tests, monitoring, review, and knowing what good looks like.

This is the less glamorous version of AI progress. A lot of the advantage is scaffolding: tools, context, routing, delegation, evals, audit trails, human review, and workflow design. The provocative read is that many companies are still selling intelligence while the real winners are building operating systems for reliability. Anthropic engineers may be shifting 8x as much code per quarter versus the pre-Claude Code era, but that only creates enterprise value if verification keeps up.

The tactical consequence is to fund the harness before celebrating the demo. Put specs in the repo. Build evals from real customer work, not synthetic chat tasks. Log what the agent touched, why it chose a model, where it failed, and what a human changed. Managers should also use agents as an operating layer: summarising feedback, checking repos, reviewing metrics, and surfacing patterns before the weekly meeting. High agency only scales if high accountability scales with it.

Vertical AI wins by owning workflow, not generic intelligence

Harvey is not becoming valuable because “law” is a label slapped onto a model. It is mapping practice areas, building Shared Spaces for law firms and clients, generating synthetic legal data where real data is scarce, and moving towards task-specific versions for different customer types. Flexport is attacking freight email forwarding, PDFs, ERP updates, and exception handling. Clay’s version of the same pattern is GTM as a programmable workflow rather than a fixed job to automate away.

The contrarian read is that open source and frontier labs may make vertical companies more defensible, not less, if they commoditise the generic layer and force differentiation into private data, process, compliance, and trust. Frontier labs can always launch “legal” or “logistics” features. What is harder is owning the customer’s messy workflow, data boundaries, exception paths, and organisational habits. Enterprise does not stay generic for long once real adoption starts.

For founders, the move is to narrow sooner. Pick the workflow, not just the vertical. Build the data production system if the domain lacks clean training data. Map data residency, conflict risk, auditability, and customer-specific permissions before the sales team scales. The product should manage the process, not merely assist the user. A vertical AI company that cannot explain its workflow advantage is just renting model progress.

Scarcity, not scale, sets the value stack

Aravind Srinivas’s argument that power is the bottleneck to AI reframes infrastructure as a physical execution problem. Chips matter, but so do land, permits, cooling, turbines, grid access, memory, SSDs, and CPU compute. He even argued that 40 out of 100 data centres may not get built because of public resistance. The private aviation discussion made the same point in a different market: aircraft can behave less like depreciating toys and more like scarce inventory when OEM backlogs, pilot supply, maintenance slots, and delivery dates are constrained.

This is a useful correction to software brain. The market often overpays for the visible product and underestimates the constrained layer underneath it. SpaceX’s thin float and first-day options trading were a public-market version of the same mechanism: scarcity plus narrative can distort price before true discovery. In AI infrastructure, the shiny layer is the model. The scarce layer may be power, memory, permits, or operational competence.

Operators should look for the slow-to-replace constraint. If you touch AI infrastructure, inspect power access, local politics, cooling, and supply chain with the same seriousness as model benchmarks. If you buy or sell premium assets, study the operating system beneath the asset: maintenance, downtime, supplier leverage, insurance, and hidden fees. Value accrues where demand can move instantly and supply cannot.

Media is not content. It is legibility with a next action

Lightspeed’s media push, Steve Stoute’s UnitedMasters thesis, and Perplexity’s search-to-agent ambition all point to the same distribution lesson. Claire Zau’s “Hot Startup Rounds” works because it turns insider market knowledge into a repeatable public format: what happened, why it matters, what might change next. Stoute’s line that “everything is advertising” is the cultural version of the same idea. A song can sell glasses; a media format can source founders; an answer engine can become the front door to agentic work.

The lazy critique is that investor media is vanity. The sharper read is that media becomes strategic when it owns both audience formation and the next commercial action. Stoute’s complaint about old labels was not only that they underpaid artists; it was that they failed to own downstream fan relationships, data, merch, tickets, and reactivation. Attention without identity is rented distribution.

Companies should build media as pipeline, not theatre. Pick a format that translates what you uniquely see into language the market can use. Track whether it produces hiring, customers, deal flow, partners, or category authority. If you already have an audience, audit where the next action happens: do you own the email, the CRM, the purchase history, the community, the conversion path? Distribution is not the postscript to product strategy. In many categories, it is the moat.

Outcome pricing is exposing lazy SaaS

Fin/Intercom’s $3.6bn sale to Salesforce stood out because customer support is one of the cleanest examples of software moving from seats to outcomes. Pricing around interventions or resolutions is not a cosmetic change; it forces the vendor to prove value in the unit the customer actually cares about. Meanwhile, Wix and Adobe were discussed as warnings: strong incumbent businesses can still be repriced harshly when the market believes AI weakens the category or the company lacks a credible usage-based wedge.

This is the uncomfortable part for SaaS operators. “Good business” is not enough if the category is being rebuilt. AI makes vendor replacement more plausible because customers can now ask whether a thin internal layer plus models can replace a costly tool, or at least create enough leverage to renegotiate it. Flexport’s interest in replacing or repricing SaaS spend is exactly the kind of buyer behaviour incumbents should fear.

Every software company should run the same audit: does AI improve retention, conversion, support cost, gross margin, or expansion, or is it just a feature demo? If the product can be tied to a business outcome, move the pricing and reporting closer to that outcome. If it cannot, be honest and compete on discipline, depth, or cost. The market is getting less patient with decorative AI and more interested in visible unit economics.

Founder identity becomes an operating system

Kareem Amin’s Clay interview, Ryan Peterson’s Flexport interview, and Steve Stoute’s ownership stories all circled the same uncomfortable idea: strategy often reveals identity. Kareem talks about courage, truth, and justice as operating principles, not poster values. Peterson’s “remote work is white collar fraud” line is deliberately blunt, but beneath it is a claim about standards, density, and execution speed. Stoute’s LeBron story, turning down a $10 million signing bonus, is about understanding that control and upside can matter more than immediate cash.

The contrarian read is that polished mission language is often less predictive than the founder’s real motive structure. Revenge, patriotism, taste, self-respect, post-lack ambition, ownership, intolerance for drift: these are not always investor-deck words, but they shape hiring, pricing, office culture, capital allocation, and willingness to narrow. Real commitment usually shows up as fewer options, not more.

Operators should write down the company’s actual axioms, not the values they wish sounded respectable. What decisions do they change? What must happen in person? Where is the company preserving fake optionality because a hard wedge feels embarrassing? Where is it taking upfront cash at the expense of long-term control? The point is not to copy another founder’s intensity. It is to make the implicit operating system explicit before it starts making decisions for you.