Jun 08–14 | AI Needs Control Planes · Liquidity Is Design · Proven User Loops

AI value is moving from model choice to control planes

The AI debate is still framed as a horse race between OpenAI, Anthropic, Google and open source. That is becoming the wrong unit of analysis. Nebius, Factory and the wider model-war discussion all point in the same direction: the scarce product layer is not simply the model, but the system that decides which model to call, when to call it, how to measure the result, and how to stop the bill becoming absurd.

This matters because enterprise AI adoption is entering its hangover phase. The first phase was board pressure. The second was token maxing: give everyone access, celebrate usage, assume productivity will show up later. The third is the CIO discovering that hundreds of thousands of dollars a month are going on trivial prompts, duplicated context and frontier models doing work an open model could handle. Factory’s claim that 80–90% of tasks can run on open models is less a prediction than a budget memo.

The contrarian read is that cheaper AI is not bearish for spend. It is probably bullish for consumption. Nebius argues open source expands demand by making more workflows economical, while managed inference can cut costs by up to 70% through caching, distillation, speculative decoding, orchestration and deployment tuning. Lower unit cost does not mean lower total spend if it unlocks more use cases.

The operator move is to stop treating AI access as a perk and start treating it as resource allocation. Classify tasks by criticality: frontier model, open model, no model. Put expensive reasoning behind routing rules, evals and budget controls. The company that wins will not be the one with the longest model menu; it will be the one with the best control plane.

Data loops are the new software moat

Alex Sacerdote’s line that “data is the new code” is easy to nod along with and harder to operationalise. The stronger version is this: the moat is not the model, the interface or even the workflow on day one. It is the repeatable loop that collects data, labels it, improves the system, deploys it, observes the result and feeds that back into the next version.

That is why Scale’s story is more useful than another abstract AI market map. Its advantage was not owning a magic model; it was owning a dirty, unavoidable bottleneck in the stack: collection, annotation, training, deployment, monitoring and feedback. Factory makes a similar point from the engineering side. The future engineer is less a person typing code and more a person designing the scaffolding, constraints, docs, tests and execution environments that allow agents to work safely.

The uncomfortable implication is that AI adoption is less blocked by capability than by behaviour change and data discipline. The first large wave of automation may hit repetitive digital work before physical robotics because emails, tickets, spreadsheets, codebases and CRMs already produce the data exhaust models need. The hardest problems are often not the ones that look hardest to humans; they are the ones without clean digital feedback.

Founders should audit their data exhaust with the same seriousness they apply to pricing or pipeline. Which workflows generate proprietary signal? Where is feedback captured? Who labels quality? How often does the system improve? A demo is not a moat. A closed loop around a high-value repetitive workflow might be.

Private-company liquidity has become an operating system

SpaceX and OpenAI are not just capital-markets stories. They show that elite companies can stay private for far longer while still needing many of the functions public markets used to provide: liquidity, price discovery, employee trust, acquisition currency and massive financing capacity. The question is no longer simply “when do we IPO?” It is “how do we manage private-company time?”

The SpaceX coverage was useful because it separated company quality from IPO theatre. A fixed-price IPO at a reported $1.77 trillion valuation and $75 billion offering may not produce the tidy first-day pop bankers normally engineer. But the pop is a weak proxy for the quality of the asset. SpaceX could trade awkwardly on day one and still be one of the defining capital allocation stories of the decade.

137 Ventures’ SpaceX position makes the deeper point. The firm kept buying across roughly two dozen investments over 16 years because the private-company lifecycle itself had changed. Secondaries reportedly exceeded $240 billion and reached 31% of venture volume, while SpaceX tenders moved from roughly annual to twice-yearly. That is not financial plumbing; it is company design. Employees need to buy houses, pay loans and manage life without forcing the company public too early.

Operators should treat liquidity policy as retention hygiene, not an afterthought for the CFO. Decide who can sell, when, how former employees are treated and how tenders fit the company’s long-duration story. Public markets are becoming optionality, not a trophy. Private liquidity is becoming part of the trust contract.

Great companies need truth mechanisms more than alignment

Ed Catmull’s Pixar story and Brian Singerman’s Founders Fund philosophy come from different worlds, but they converge on the same operating principle: great organisations are built to tell themselves the truth before the market does. Catmull watched admired companies scale, polish their processes and then make obvious-in-hindsight mistakes. Singerman strips venture down to founder judgement: can this person actually do the thing better than anyone else?

Pixar’s brain trust is the cleaner management lesson. It gave blunt feedback but had no authority to impose fixes. That separation matters. Most companies either avoid candour to preserve harmony or turn feedback into hierarchy. Pixar tried to make critique wide and decision rights narrow, so the work could be attacked without the person being diminished and without every comment becoming a command.

The contrarian read is that alignment can become a liability. If everyone agrees too quickly, the organisation may be protecting status, process or ego rather than the customer experience. Catmull’s line that “quality is the best business plan” is not a slogan about taste. It is a statement about compounding trust. Singerman’s 98%+ founder-quality framing says the same thing from the capital side: polish, portfolio theory and process matter less than whether the person in front of you has authentic judgement.

The practical move is to install truth mechanisms deliberately. Run “where are we blind?” reviews. Separate critique from decision rights. In hiring and investing, ask questions that cannot be answered with borrowed startup language. If a team only produces neat updates, it is probably hiding the ugly first drafts where the real work happens.

Product winners begin with proven behaviour, not founder imagination

Mark Pincus’s “Proven Better New” framework is a useful antidote to originality theatre. Start with a behaviour already proven in the market, make it obviously better, then add one new wrinkle. Words With Friends was not a grand reinvention of social gaming; it was Scrabble plus the social graph. The discipline is that “better” must be obvious to the target user, while “new” should be small enough not to ask for too much belief.

This is more ambitious than it sounds. Founders often want a product to carry the full weight of their category thesis on day one. Pincus argues the opposite: the first version should be humiliatingly small because smallness is how you falsify quickly. “Kill hope before hope kills you” is not cynicism. It is a launch operating system. Hope appears when a team has built too much to admit it still has not learned the core thing.

The sharper implication is that distribution is not a post-launch growth layer. It is part of the product’s design. FarmVille’s pre-access and paid keys worked because scarcity, anticipation and monetisation were built into the loop. Zynga’s focus on day-365 retention and its ASN metric show the same bias: durable behaviour beats top-of-funnel theatre.

Every product review should force three questions before roadmap debate begins. What exact behaviour is already proven? What improvement would make ten out of ten target users say yes? What is the smallest new element we can test without smuggling hope back into the plan? In an era where AI makes building cheaper, the risk is not building too slowly. It is building too much before the market has taught you anything.

Lean companies are becoming the default, not the phase

Several episodes circled the same operating shift: capacity is decoupling from headcount. Lovable was discussed at $500 million ARR with 146 employees. Cursor’s revenue-per-head was framed as extraordinary. Uber cutting 23% of HR sits in the same pattern. AI-native companies are making it harder to defend bloat as a sign of scale.

The useful version of this idea is not “fire everyone and use AI”. It is that the winning unit is increasingly the load-bearing individual: someone who can own an outcome across product, systems, GTM and execution. Factory’s claim that engineers and salespeople sit together is not office-layout trivia. It reflects a view that the customer journey from first impression to tenth renewal is one product surface, not a baton pass between departments.

The contrarian implication is that small and fast may stop being a startup stage and become a durable operating model. Traditional scaling added layers because coordination costs rose with ambition. AI and better tooling change part of that equation. They make high-agency people more productive, but they also expose low-leverage coordination work faster.

Operators should measure revenue per head, but not as a vanity metric. Use it to ask where the organisation is adding judgement and where it is adding motion. Hire fewer people who can carry more context. Put product, GTM and engineering closer together. Spend on tooling, documentation, recovery and decision quality where it makes strong people stronger. The new org chart is not flatter because flat is fashionable; it is flatter because leverage is concentrating.