Context

OpenAI announced it would cut off access to its models inside Cursor, the AI coding tool, after Elon Musk's SpaceX (which also owns xAI) acquired Cursor. NLW uses that move, and the pattern it fits, to work through what the next phase of frontier-lab competition means for anyone building on top of AI. The pattern is not new: Anthropic cut off Windsurf in 2025 when OpenAI was about to acquire it, blocked OpenAI and xAI from its API, and changed its subscriptions to exclude third-party tools. The through-line for product people is that dependence on a single AI vendor is now a real strategic risk, and the question of control has moved beyond the model to the harness (the software layer that wraps and orchestrates models) as well.

The Big Idea

Frontier labs now routinely cut off tools and harnesses owned by their competitors, so building resiliently on AI means not depending on any single vendor for either your models or your harness. The control conversation has moved from just the model layer to the harness layer too, and it is about resilience and sovereignty, not only cost.

The Cursor cutoff makes it concrete: if the models and the harness you rely on can both be pulled out from under you because of a fight between vendors, you need the ability to route around any one of them.

Key Insights

Labs now cut off rivals' tools

The core pattern. Being a customer of a frontier lab no longer guarantees access if you get entangled with a competitor.

• The examples: OpenAI cut Cursor off (SpaceX/xAI owns it); Anthropic cut Windsurf off in mid-2025 when OpenAI was acquiring it; Anthropic blocked OpenAI's and xAI's API access, and changed subscriptions to stop covering third-party tools like OpenClaw and Hermes. As one observer summed it up, "not your weights, not your product."
• Why it matters: whether a given cutoff is pettiness or rational (protecting against a competitor harvesting training data), the effect on you is identical. Your access to a model or tool can disappear over a conflict you are not party to.
• The counter-example that proves the point: Anthropic said it would keep supporting Cursor, but commenters noted Anthropic now pays SpaceX heavily for compute, so its restraint tracks its own interests, not customer loyalty.

Sovereignty moved to the harness layer

The most important reframe. Enterprises already knew to think about model independence (open-weights policies, routing tasks to the cheapest capable model). This episode's point is that the same control question now applies to the harness, the software that wraps models, orchestrates tools, and runs your workflows.

• Why: the Cursor cutoff hit at the harness level, not just the model level. If your workflows are welded to a harness a lab controls, you inherit that lab's fights and policies.
• The response: decouple your workflows from any single provider's models and harness. NLW predicts a lot more discourse around "open harnesses" (he cites DeepSeek releasing an open, plugin-based harness), not just open models, because enterprises now need optionality at both layers.

The Reverse Information Paradox

A strategic concept from Microsoft's Satya Nadella, and the deepest idea in the episode. You pay for intelligence twice: once with money, and again with the proprietary knowledge you must reveal to make the model useful.

• The mechanism: models learn from "exhaust", the prompts you write, the tools your agents use, and especially the corrections you make when the model is wrong. Each correction becomes institutional know-how the vendor absorbs. Over time the information asymmetry skews: the seller learns more about you while you learn little about what they are learning.
• The prescription: distribute the learning infrastructure so each firm controls its own learning loop, rather than feeding its most valuable knowledge into a black box only the vendor can see. This is why Microsoft's newer models are built to be customized and post-trained by the customer.
• For PMs: treat your usage data and corrections as an asset you may be giving away. Building on a closed vendor can quietly transfer your hardest-won domain knowledge to a company that could become your competitor.

Cheaper tokens unlock outsized demand

A token-economics insight (Aaron Levie's framing of Jevons Paradox) with direct planning implications. Every task an enterprise could automate is either ROI-positive or not based on the cost of the tokens to do it.

• The dynamic: as tokens get cheaper at a given capability level, more tasks cross into ROI-positive territory (process every contract, read every log, watch every data stream), so consumption rises disproportionately. Levie's estimate: even a 50% drop in token price can drive a 5x increase in tokens for these workloads.
• The minimum-feasibility threshold (Brandon Gleklen's point): many tasks need some minimum quantity of tokens to actually be done well. If the price does not let you hit that minimum, the task simply is not attempted. A price drop that crosses the threshold green-lights use cases that were previously non-viable.
• The evidence: after OpenAI cut prices, OpenRouter usage of one model rose 5.6x and another 13.8x, and about a third of users stayed even after prices returned to full.
• For PMs: falling token prices do not just save money on existing features; they open entirely new use cases. Watch for the price point that makes a currently-infeasible workflow suddenly worth automating.

Labs now compete on efficiency

A shift in the basis of competition. Frontier labs are no longer racing only on raw capability but on the frontier of efficiency (cost and tokens per task). OpenAI's aggressive price cuts (one model down 80%, another 20%) were aimed at driving usage on third-party platforms, and signal that being cheapest-per-task is now a competitive lever alongside being smartest. For anyone choosing models, efficiency is becoming as much a differentiator as capability, and it moves fast.

Tokens are not a value proxy

A sharp measurement caution. When Cursor said OpenAI models were only about 5% of its traffic, OpenAI's PM argued that token share understates value, because more efficient models use far fewer tokens to accomplish the same task, so a small token share can represent a much larger share of value created. The general lesson: token or usage volume is a misleading proxy for value or revenue, since a more efficient model does more per token. Do not judge a model's or feature's importance by raw token counts alone.

Mental Models & Frameworks

The Reverse Information Paradox

Nadella's model for the hidden cost of closed AI: you pay once in money and again in revealed proprietary knowledge, and the information asymmetry compounds in the vendor's favor over time as it learns from your prompts, tool use, and corrections. Use it as a lens on any AI build: ask what institutional knowledge you are feeding a vendor, whether that vendor could become a competitor, and whether you should own the learning loop (via customizable or open-weights models) instead of exporting your edge.

The model-and-harness portfolio

A resilience model: treat both your models and your harness as a portfolio you route across, not a single vendor you marry. Keep the ability to send lighter, cheaper, or more sensitive tasks to open-weights models, and keep your workflows decoupled from any one harness (increasingly, an open harness) so a vendor conflict cannot strand you. The goal is optionality at both layers, so no single company controls your access.

Jevons Paradox for tokens

The counterintuitive economics that cheaper tokens increase total token spend rather than decrease it, because lower per-task cost pulls a flood of previously-uneconomical tasks into the ROI-positive zone. Pair it with the minimum-feasibility threshold (a task needs enough tokens to be done at all, so a price drop can flip it from impossible to worth doing). Use both to forecast that your AI costs and your automatable surface will grow together as prices fall, not shrink.

Decision Principles

Principle: Own your tools, keep relationships direct

• When: you depend on AI models or tools that are routed through another company's layer.
• Why: if your access is mediated by a third party (a harness owned by one lab, models accessed through a competitor's product), you are exposed to conflicts between those companies. Owning your tools and keeping your relationships with the models and products you rely on direct, not routed through layers, insulates you from their fights.

Principle: Have an open-weights and open-harness policy

• When: setting enterprise AI strategy.
• Why: you do not need to abandon closed frontier models (most use cases have not even moved yet), but you need the capability to integrate and route certain tasks to lighter, cheaper, more controllable open-weights models, and increasingly to open harnesses. Like any capability it takes time to build, so start now rather than when a cutoff forces you.

Trade-offs & Nuance

Closed frontier power vs control

Closed frontier models remain the most capable, and NLW is clear he does not see enterprises abandoning them. The trade-off is that using them means less control and feeding your knowledge into a system you do not own (the Reverse Information Paradox). The sophisticated posture is not all-or-nothing: keep using frontier models where their capability wins, while building the ability to route lighter or more sensitive tasks to models and harnesses you control. A fully closed setup maximizes capability but concentrates risk; a mixed setup trades some peak capability for resilience and sovereignty.

Practical Application

Build a model and harness routing capability

• Do: stand up the ability to route tasks across multiple models (frontier and open-weights) by matching task difficulty to model capability and cost, and decouple your workflows from any single harness so you are not locked to one vendor's tool.
• Why it works: it turns a vendor cutoff or price change from an existential event into a routing change, and it captures the cost savings of using cheaper models where they suffice. This takes time to build, so treat it as a capability to start on now.

Protect your learning loop

Audit what proprietary knowledge you are handing to AI vendors through prompts, agent tool use, and especially your corrections when the model is wrong. Decide which of that is core IP, and for those areas favor customizable or open-weights models you can post-train on your own data, so your institutional know-how compounds for you rather than for a potential competitor.

Reprice your infeasible use cases

Keep a list of automation ideas you shelved because the token cost made them ROI-negative, and revisit it whenever model prices drop. A price cut that crosses a task's minimum-feasibility threshold can suddenly make a previously-impossible workflow (processing every contract, reading every log) worth building.

Questions to Consider

• If a frontier lab cut off the models or the tool our product depends on tomorrow because of a fight with another vendor, how badly would we be stranded, and could we route around it?
• Is our control strategy limited to models (open-weights routing), or have we also thought about the harness layer that wraps and orchestrates those models?
• What proprietary knowledge are we feeding an AI vendor through our prompts, agent actions, and corrections, and could that vendor use it to compete with us (the Reverse Information Paradox)?
• Which automation ideas did we shelve as too expensive per token, and would a plausible drop in model prices cross the threshold that makes them worth building?
• Are we judging a model's or feature's importance by raw token or usage volume, when a more efficient model can create more value with far fewer tokens?

Bottom Line

Frontier labs now cut off tools and harnesses owned by their rivals as a matter of course, so the next phase of AI competition makes single-vendor dependence a real risk. The move for anyone building on AI is to own the layers you can and keep optionality across both models and harnesses (open-weights and open harnesses included), protect your own learning loop rather than feeding your edge into a black box, and plan for cheaper tokens to expand what you automate, not just cut your bill.

Concepts to Explore

Jevons Paradox

The economic observation that making a resource more efficient (cheaper to use) can increase total consumption of it rather than reduce it, because the lower cost unlocks far more uses. Applied to AI, cheaper tokens pull a wave of previously-uneconomical tasks into being worth automating, so total token spend rises even as per-task cost falls. Worth understanding for anyone forecasting AI cost or capacity, because it inverts the intuition that price cuts shrink spend.

Notable Quotes

"Not your weights, not your product." (Yuchen Jin, on frontier labs cutting off competitors' tools)

"You essentially pay for intelligence twice, once with money and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful." (Satya Nadella, the Reverse Information Paradox)