The Product Desk

The most expensive component was the one nobody specified

Around try five, the agent said the target was impossible, and it was right. Claude Code had diagnosed a 2.2-second artificial latency ceiling that capped the achievable Lighthouse score near 89, against a requested target of 100. The Haiku evaluator — a second, cheaper model asked to decide whether the main agent was finished — sent it back 14 more times and spent $0.67 doing active harm. An LLM-as-judge stop condition outranked a correct termination.

The fix is cheap and specific. A deterministic budget check sits above the judge in the control flow, and an agent that declares a goal unreachable terminates after one confirmation instead of fourteen rejections.

The verifier is the optimization target, which makes it product's job

a16z's more durable claim is not about dollars. The verifier does not merely end the loop; it defines what the loop treats as progress. An incomplete signal teaches an agent to get better at passing the check without getting better at the user's task. On SpecBench, frontier agents routinely passed visible tests while failing held-out tests that exercised the same features in combination. One produced a 2,900-line "compiler" that simply memorized the test inputs.

Whoever writes the acceptance criteria for an agentic feature is writing its optimization target. "Keep working until the tests pass" is verifiable and still only a proxy.

Two items land on the launch checklist: a held-out evaluation the agent never sees, and combinatorial testing that exercises features together instead of one at a time.

Three sources, three different fixes

The reporting converges on the diagnosis and splits on the remedy. The remedies turn out to be complementary rather than competing.

• Instrument it. Returns on iteration are logarithmic. One web-agent benchmark moved from 38.8% to 43.2% success going from 1 to 10 samples, then bought 0.2 points for double the tokens at 20. Cost per iteration and progress per dollar have to be visible while a run is still going, split between main agent and evaluator.
• Shorten it. Enterprise IT analysis argues reliability falls sharply as autonomous step count rises, and that interrun variance opens a wide gap between a best-attempt demo and behavior users can depend on. Past roughly five autonomous steps, the work becomes checkpointed sub-workflows, and acceptance moves from best run to consistency across ten identical runs.
• Make failures durable. The most-cited open-source "self-improving" agent banks successes as reusable skill files and has no automatic path to learn from a failure that appears only in production. It caps runs at 90 turns and never converts a bad run into a permanent regression case.

A study of nearly 150,000 real agent actions in IT operations sits underneath all three. The agents that improved were handling low-risk tasks with human analysts iteratively shaping behavior. That is the observed pattern. Autonomy plus headcount savings is the deck's pattern, it runs ahead of the published evidence, and an exec will quote it back anyway.

What this does to packaging

Nobody, human or loop, knows at kickoff whether a task needs 20 iterations or 500. Flat seat pricing on an agentic feature transfers that entire variance onto gross margin, and selling "more compute" as a premium tier charges users more for output that can be worse past the plateau. The defensible shape is tiers pinned to sample budgets at the knee of the returns curve, with cost per successful task reported as the unit metric rather than usage.

Caveat before this reaches a board deck: the web-agent benchmark is unnamed, the reasoning-degradation studies are uncited, and a16z discloses it has not independently verified third-party information. Reproduce the returns curve on the actual workload before it prices anything.

DeepSeek's hike is a capacity story, not a greed story

Separate what was pitched from what is being done. The pitch was accessible AI: CEO Liang Wenfeng had previously told employees and investors that large profits would not be a priority. What is being done is a price increase, and the detail that makes it legible is the fundraise around it. DeepSeek generated an estimated $400–500M in annualized revenue at rock-bottom prices, and is reportedly raising 50B yuan at a 500B yuan (about $74B) valuation, roughly 150x revenue, per The Information. Reversing the accessible-AI line mid-fundraise, while demand for V4-Flash overwhelms serving capacity, points at one cause: capacity is the constraint and the subsidy is ending.

The opposite move landed the same week. Meta shipped Muse Code in preview with a contributor tier Alexandr Wang claims is more than 10x cheaper than Claude and Codex, framed as loss-leading against data-center capex. Google was reportedly willing to pay $1.5bn for a 35-person coding startup, about $43M per employee. Prices are moving up and down in the same seven days. Cents-per-token is not a plannable input, and it is not a defensible axis.

The hardware layer says the same thing with different numbers

Rental economics already turned. H100 hourly rates reportedly fell from nearly $8 in early 2024 to about $1.70 by late 2025, then climbed back toward $2.35 by March 2026. A collapse and a rebound inside two quarters. Underneath that, The Information reports (Qianer Liu and Phoebe Liu, Aug 6) that Nvidia is testing Rubin Ultra variants with less memory than it publicly announced, because memory rather than compute is now the binding shortage.

That distinction is a feature-scope question, not only an infra one. Compute determines how fast a token comes out; memory determines how many users can hold a long conversation at once, and what each one costs you. If the flagship 2027 part ships lighter on memory than the vendor quoted against, the constraint lands on the roadmap rather than the datacenter.

Where the sources disagree, and why it changes your hedge

Three explanations are live at once, and each implies a different move. The provider-side reading says a subsidy ended, so the hedge is a qualified fallback model per workload. The financing-side reading says compute pricing behaves like a commodity market, because much of the hardware sits in separate entities with lenders holding claims on the chips, so the hedge is contract language plus counterparty questions: who owns the hardware serving us, who can claim it on default, and is this provider's credit supported by a third party. The supply-side reading says memory is scarce, so the hedge is a funded memory-efficiency epic. Meanwhile AWS plans roughly $220B of 2026 capex and still says it will not meet all demand through 2027, which makes availability rather than price the negotiating axis.

All three hedges are cheap. All three are painful to start late. The sequencing that wins the room puts the sensitivity model first, because it converts an argument about direction into a named list of features that break, and a named list is what gets an epic funded.

Discipline note: every figure here is reported rather than filed — the rental rates come from third-party reports that may not match your contract or region, and Nvidia's language is hedged at "testing." Validate against your own invoices before any of it drives a pricing page.

Three leaders, three lock-in primitives a team can ship

A shopper stands at the Costco shelf and does not compare prices. She already paid the membership fee, so the comparison happened months ago, in a different mood. That is the mechanism, and each Q2 leader runs a version of it. Amazon bundles entitlements users will not re-buy elsewhere. Costco charges the fee and lets a prepaid sunk cost do the persuading at the shelf. McDonald's runs 220 million enrolled loyalty profiles that feed pricing decisions, which is accumulated user state and history that is expensive to recreate somewhere else.

This is a roadmap finding rather than a marketing one, because the split only shows up in deterministic transaction data. Here is what teams tell themselves users prefer, and here is what repeat spend says: the two diverge the moment a real switching cost exists. A retention backlog prioritized from NPS, preference surveys or interview quotes has been overfunding sentiment features and underfunding structure. Rewards are near saturation — 76% of women and 68% of men already belong to a program — which is why the leaders moved to structure.

The same test, run over a century

A separate analysis of seven companies past 100 years old reaches the same split from the opposite direction, and the return column carries the argument. Asset-light fee capture clusters at the top. Marriott is up more than 10,000% since its 1993 IPO collecting management fees while independent owners hold the buildings. McDonald's is up more than 6,000% since 1990 on rent and franchise fees. Friction-based retention sits at the bottom. Bank of America's stated moat is that switching banks is a pain, plus regulation, and it is tied for the weakest return in the cohort.

Disney is the sharper warning. It owns Mickey Mouse, Cinderella and Star Wars, it has pricing power across every line, and the stock has been roughly flat for a decade. Demand was permanent. The distribution channel was not. Customer affection got read as structural defensibility, which is the expensive version of this mistake.

Friction moats do not decay gracefully. They fail all at once, the quarter someone productizes migration.

Caveat: those return figures come from third-party charts rather than primary filings, with no valuation inputs anywhere. Use the taxonomy, not the numbers.

The same measurement error, now on acquisition

A 1,000-question study of Google AI Overviews in the legal category found at least one named provider in 59% of all answers and 92% of commercial queries, almost always woven into the guidance rather than presented as a list. Behind each answer, Google ran an average of 6.4 hidden fan-out searches, including reputation checks. No single firm dominated the results.

The unit of visibility is the fan-out sub-question, not the keyword, which means comparison pages, docs and objection-handling content are what actually gets retrieved. Third-party reputation is a ranking input, so review volume and credible independent mentions need a named owner rather than living as a marketing side quest. The surface is also unowned, and the last several times a discovery channel moved, showing up early mattered more than being good at it.

The forcing function is a single number. Track named-provider rate monthly against the 59% and 92% benchmarks. It is the cheapest early-warning metric available on a channel that has no page two.