Jun 1–7 | Moats Move to Control · Token Budgets · Data Access

The AI moat is moving to the control plane

A striking number of this week’s conversations landed on the same uncomfortable point: if the model keeps improving, the thin application layer gets less defensible. Mercor’s Brendan Foody put it most bluntly — “the model is the product” — while Legora, Merge, Fivetran and Exa all described their products less as clever wrappers and more as systems for deciding what models can see, do, cost and prove.

That is the real shift. The durable layer is no longer “we added an LLM to workflow X”. It is evals, routing, permissions, auditability, retrieval, data access and domain-specific primitives. Legora’s legal product depends on legal workflows, enterprise guardrails and optimal model routing. Merge’s Gateway is a policy and cost-control layer. Fivetran is arguing that the data platform becomes the agent’s context substrate. These are not separate categories so much as pieces of a new enterprise operating system.

The contrarian read is that the most defensible AI startups may look less like apps and more like control planes for work. A beautiful interface is not enough if a frontier model can reproduce the workflow in six months. The company with the moat is the one that owns the rules, memory, permissions and feedback loops that make model use reliable inside a real organisation.

For operators, this is a useful audit. If your AI product disappeared tomorrow, would the customer lose a governed system of work or just a nicer way to prompt a model? If you do not own routing, evals, access rights, cost controls or workflow-specific data, you may be building a feature. The product question has changed from “what can the model do?” to “who controls how the model is allowed to work?”

Tokens are becoming COGS, not experimentation spend

The token conversation has moved from engineering enthusiasm to CFO discipline. Merge described “token-maxxing” bills that shock finance teams. Uber reportedly put a cap around $1,500 per engineer per month. Mercor is already spending more on tokens for internal agents than on employee headcount. Anthropic, Google and Cognition were discussed less like normal software stories and more like signs that AI turns software into a capital-hungry infrastructure business.

The lazy answer is to celebrate usage. The better answer is allocation. Legora prefers performance over latency, then cost, because a lawyer waiting for a better answer is different from a chatbot returning “hello”. Exa claims better retrieval can cut spend by around 20x in some cases because the model sees only the context it needs. Dara Khosrowshahi made the same point from Uber’s side: use expensive frontier models to explore, then route scaled workloads to cheaper or open-source systems.

The provocative implication is that “more tokens” is a bad management signal. Token spend is not proof of AI maturity any more than cloud waste was proof of software sophistication. The best AI companies will be the best allocators of compute: premium inference where judgement matters, cheap models where repetition dominates, retrieval where context can replace brute force, and human review where errors are expensive.

This needs to become an operating dashboard, not an innovation budget. Track token spend by function and workflow. Define which work deserves frontier models, which should run on smaller systems, and where AI spend is actually replacing labour rather than adding another line item. The coming finance question is not “are we using AI?” It is “why did this task deserve expensive inference?”

Agents need data rights before they need more intelligence

The strongest data-access argument came from George Fraser at Fivetran: the old reason to centralise data was reporting; the new reason is agent context. An agent cannot act intelligently if it sees only one app, stale exports or a vendor-approved slice of customer data. Exa’s Will Bryk made the same point from the retrieval side: agents do not want ten links; they want everything relevant, filtered and fresh enough to act on.

This makes vendor lockdown a strategic issue, not a technical annoyance. Fraser called out SAP-style restrictions on agent access and pointed to Open Data Infrastructure as a way to score vendors on egress charges, complete-copy access and terms-of-use restrictions. He also argued that “data gravity” is often architecture debt wearing a clever name: with change-data-capture, you do not need nightly full copies of everything. You move changes, continuously.

The less obvious read is that vendors restricting access may accelerate their own disintermediation. If customers cannot use the incumbent’s data in agentic workflows, they will build workarounds, centralise more aggressively, or buy from AI-native competitors that treat data access as a default. A locked product may protect today’s UI while weakening tomorrow’s system of action.

Operators should treat data rights as procurement language. Put continuous access to a full copy of your own data into vendor contracts. Review API limits, export terms, egress fees and agent restrictions before they become blockers. Retrieval quality should sit next to product quality and sales productivity as a business metric, because the agent with incomplete context is not an intelligent worker; it is a confident intern with missing files.

Enterprise AI wins by orchestrating messy work, not demoing chat

Happy Robot’s story is useful because it breaks the “voice AI” category. The company started in logistics, but the real product is not a more human-sounding caller. It is a coordination layer across phone calls, emails, airline websites, CRMs, ERPs, transport systems, Snowflake, documents and Notion. Its “Twin” layer exists because the valuable context lives across systems and channels, not inside one neat SaaS object.

That same pattern shows up in Legora and Merge. Legora is building specialised review bots, security review, incident agents and SRE agents around legal workflows. Merge’s Agent Handler and Gateway sit between agents and the tools they call. Happy Robot runs campaigns like 20,000 to 50,000 daily outreach attempts for duty collection, but the important bit is not the scale of calls; it is the ability to preserve context, ask permission, trigger the right system and avoid exposing sensitive variables like maximum buy price to the model.

The contrarian read is that enterprise AI is mostly deterministic workflow wrapped around probabilistic language. The hard part is not making the agent sound human. It is deciding what the model should not see, which actions require approval, how context moves across systems, and where every step is logged for audit. The risk is less “model thinking” than “model acting” with the wrong permissions.

A practical starting point is to map coordination cost, not departments. Where do people chase status, reconcile channels, copy data, ask for approvals or wait on another function? Those are better AI wedges than generic “knowledge work”. But before agents touch production systems, build the governance layer: identity, permissions, tool access, sensitive-field blocking, logs and offboarding. Trustworthy orchestration will beat impressive demos.

Distribution is a systems problem, not a marketing department

Dara Khosrowshahi described Uber in a way that reverses the usual consumer-company story. Uber is supply-led: drivers, couriers, restaurants, merchants and AV partners come first, then demand fills in. That explains why Uber can be useful to autonomous vehicle companies without owning the AV stack. If multiple AV winners emerge, Uber’s edge is demand, marketplace density and partner distribution. Its AVs are reportedly 30%+ busier on the Uber network than first-party AVs outside it.

Spotify’s Gustav Söderström offered a consumer version of the same idea. Spotify did not win podcasts, audiobooks and music by building separate perfect apps. It chose one surface, freemium, personalisation and ubiquity, then absorbed the organisational complexity internally. Its weekly three-hour E-team meeting with roughly 14 leaders is not tidy management theory; it is an attempt to stop the org chart leaking into the product.

Balaji and Steven Glinert made the geopolitical version explicit: power belongs to whoever can coordinate the larger coalition. Their supply-chain discussion was noisy in places, but the operator point is clean. You do not have resilience because you have a supplier list. You have resilience when you can see the graph, identify tier-two and tier-three dependencies, and build “allied weight” through partners who can absorb shocks.

For founders and GTM leaders, the question is not simply “how do we get more demand?” It is “what scarce side of the system makes the strategy work?” In a marketplace, that may be supply. In consumer software, it may be one coherent distribution surface. In hard-tech or regulated markets, it may be partner trust and supplier visibility. Good product theory does not survive a weak coalition.

Deep-tech GTM starts with time compression

Impulse Space is a useful reminder that hard-tech categories often sound wrong at first. “Space tug” is less interesting than what Eric Romo and Tom Mueller are really selling: time compression and design freedom after launch. Helios aims to move payloads to GEO in roughly eight hours rather than six to ten months through transfer orbit. That is not just faster transport; it can pull forward revenue, reduce operating burn and change how customers design satellites upstream.

The company also shows disciplined market selection. Romo described the original commercial orbital-transfer idea as financially ugly: limited customer count, roughly 20% gross margins and heavy working-capital needs because launches had to be bought ahead of demand. The product did something real, but the business model was too thin. The defence opportunity around Mira became clearer only after a first flight helped Space Force reframe the mission: rapid manoeuvre, rendezvous and characterisation in contested space.

The contrarian read is that vertical integration here is not founder machismo. It is a learning-rate strategy. Mueller keeps propulsion, machine shop, tanks, assembly and testing close because outsourced bottlenecks slow iteration and create fragile dependencies. In hard tech, cycle time can be the product. A machine shop or test stand is not overhead if it shortens the path from failure to version three.

The operator test is harsher than TAM. Ask whether the product changes customer economics on time, margin and working capital at once. Ask whether it changes the customer’s own design decisions before purchase. And when considering insourcing, ignore ideology: bring in the parts that constrain learning speed, quality variance or supply resilience. Everything else can stay outside.

Consumer AI should increase agency, not surrender

The consumer thread this week pushed against the voice-only, black-box automation story. Steven Sinofsky argued that AI costs will push more compute back onto local devices, because paying per token forever starts to look like earlier scarcity cycles in CPU and GPU history. Tony Fadell was sceptical that screens disappear soon; maps, visual context, correction loops and trust still need a display. Spotify’s AI thesis was similar in spirit: users should be able to steer the algorithm, not merely be optimised by it.

That matters because the fashionable AI promise is often “the machine does it for you”. The better consumer wedge may be “the machine lets you control it precisely”. Gustav’s framing of AI at Spotify is not more engagement for its own sake, but the user telling the system: this is who you think I am; here is who I actually am. Fadell’s warning against cognitive surrender lands in the same place. When generic output becomes cheap, taste, judgement and correction become more valuable.

The contrarian implication is that the screen may be under-rated. Voice will matter, local agents will matter, but many high-trust consumer products still need visible state, editable choices and fast recovery from mistakes. A hidden optimiser is convenient until the user wants to understand or override it. Backward compatibility can be a tax, as Sinofsky argued, but user control is not legacy baggage.

For product teams, design AI as steerability before automation. Use local execution where privacy, latency or cost make it valuable. Expose controls in plain language. Let people correct defaults, inspect recommendations and recover from bad actions. The next great consumer AI product may not be the one that removes the interface; it may be the one that makes the interface finally listen.