Frontier Overhangs (Stratechery…

1. Key Themes

"Pacing the Frontier" conveniently solves several of Anthropic's business problems, not just safety concerns

Thompson's core thesis is that Dario Amodei's safety-motivated call to slow AI development happens to also address five distinct commercial "overhangs" facing frontier labs.

"'Pacing the Frontier' is framed as — and I believe motivated by — a concern about safety, but it also happens to address several distinct problems faced by the frontier labs. These problems take the form of overhangs that have formed by virtue of how rapidly models are improving; slowing model improvement would reduce the overhangs."

Model capability is becoming "good enough," which shifts value toward harnesses and products, not raw model performance

Once performance crosses a threshold, competitive advantage moves from proprietary integration to modular, customer-driven differentiation — exactly the pattern Christensen described.

"This doesn't, in and of itself, suggest that new model capabilities aren't desired; it does, however, suggest that current model capabilities are 'good enough' for customers to not do whatever is necessary to get access to the cutting edge, which reduces the value of the cutting edge to its proprietors... Pure capability no longer translates directly into a moat."

Frontier labs face a structural race to own end-user touchpoints before their capability edge erodes

Thompson revisits his "economic imperative" thesis: labs need to become the interface, not just the backend, or risk being commoditized.

"It has long been clear to me that the frontier labs have the economic imperative to move closer to the user. If you own the user touchpoint, then you have meaningful lock-in... it's in the frontier labs' long-term interest to not simply be a commodity input into software but to simply replace software outright."

High AI prices reflect a supply-constrained "price umbrella," not sustainable margin advantage

Current pricing power is temporary and self-inflicted — a byproduct of compute misallocation between inference and training/R&D.

"OpenAI and Anthropic charge high prices because they can: there is so much demand for their products at their current price levels that they can barely keep up... If they slowed down progress they could re-allocate computing and fully capture the market."

Capital markets, not compute or demand, are the real near-term risk to the AI buildout

Thompson flags that the shift from free cash flow to debt and now exotic funding vehicles (insurance floats, pension funds) increases systemic fragility.

"It's one thing to spend all of your free cash flow; it's another thing to tap the debt markets. And, beyond that, it's a completely new nerve-racking thing to bring safety-seeking assets to bear. AI better deliver before it's too late."

2. Contrarian Perspectives

Pushing the frontier faster is safer than pacing it, at least for cybersecurity

Thompson argues the real, present-day safety risk (autonomous cyberattacks) is asymmetric in favor of attackers, meaning slowing frontier progress helps bad actors more than it protects anyone.

"This reality actually makes the case for pushing the frontier, not pacing it. The capability overhang is entirely on the offensive side: it's defenders that need models that are not only good enough to mount a defense, but to do so in an entirely automated way... In other words, when it comes to the tangible safety risk that exists today — bad actors using aligned LLMs to attack infrastructure — pacing the frontier actually increases the window in which bad things can happen."

Chinese model competition is overstated — it's a pricing illusion, not a cost advantage

Against the consensus narrative that cheap Chinese models threaten Western labs' economics, Thompson argues the "cheapness" is an artifact of supply-demand imbalance among the leaders, not superior unit economics.

"I highly doubt that Chinese models are cheaper to serve on a marginal cost basis, they just seem cheaper because Anthropic and OpenAI are so supply constrained that they are charging far more than they would if there were sufficient supply to meet the demand for intelligence."

AI "safety" advocacy is motivated more by competitive positioning than pure philosophy

Thompson suggests the timing of Amodei's "pace the frontier" argument — arriving just as OpenAI regained the capability lead — reveals a competitive, not purely ethical, motive.

"I'm not sure it's a coincidence the demand to pace the frontier came when the frontier was, for the first time in a while, set by OpenAI... Anthropic is fine with Anthropic being in the lead; anyone else requires government intervention."

3. Companies Identified

Anthropic — Frontier AI lab (Claude models), subject of Amodei's "pace the frontier" argument. Why mentioned: Central case study for the thesis that safety rhetoric aligns with business self-interest (data retention, pricing, competitive positioning).

"Anthropic has told its backers it will be profitable this quarter, as it moves to allay investor concerns about the aggressive cash burn of frontier AI companies ahead of its blockbuster initial public offering... Anthropic's gross margins are above 80 per cent before accounting for revenue shared with distribution partners, including Amazon, and the cost of training its models."

OpenAI — Frontier AI lab (ChatGPT/Codex), competitor to Anthropic. Why mentioned: Framed as the trigger for Anthropic's "pace the frontier" push once it regained model leadership.

"I'm not sure it's a coincidence the demand to pace the frontier came when the frontier was, for the first time in a while, set by OpenAI."

Microsoft — Enterprise software/cloud giant, maker of Copilot/E7. Why mentioned: Case study of harness-model modularity undermining the "harness lock-in" thesis, per Satya Nadella's on-record comments.

"We're using the same harness that we use in GitHub and the same thing in security, too... we will have GPT, we will have Anthropic in there and any open weight model... it took a while for Nadella's claims to reflect shipping reality, but sure enough you can now choose a model for Copilot Cowork."

Meta — Builder of the Muse personal agent product and Muse Spark model. Why mentioned: Presented as a bearish signal for frontier labs — proof that "good enough" (not state-of-the-art) models can produce sticky, defensible products.

"Muse is, by a significant margin, the best and most approachable personal agent product I have tried... Muse Spark 1.3, the most advanced Meta model, is still not state-of-the-art, and that is the bearish signal: it is good enough for a very good personal agent product, and critically, a personal agent is much stickier than a chatbot."

Nvidia — GPU/compute infrastructure provider. Why mentioned: Used to illustrate capital/funding risk in the AI buildout as financing shifts from cash flow to debt and exotic capital sources.

"Right now, however, is the danger zone, as hyperscalers blow through the debt markets and Google at least starts to tap equity... AI better deliver before it's too late."

Google — Hyperscaler. Why mentioned: Cited as an example of hyperscalers moving beyond free cash flow into equity markets to fund AI infrastructure.

"Right now, however, is the danger zone, as hyperscalers blow through the debt markets and Google at least starts to tap equity."

Fable — Enterprise AI product tied to Anthropic's data policies. Why mentioned: Case study showing customers rejecting a data-retention requirement, disproving the idea that pure capability guarantees adoption.

"Fable wasn't good enough: customers pushed back, and Fable usage stayed relatively low; when Fable 5.1 was released, the Anthropic-gets-to-keep-your-data provision was gone."

4. People Identified

Dario Amodei — CEO of Anthropic. Why mentioned: Author of "We Must Pace the Frontier" and central target of Thompson's critique linking safety rhetoric to business strategy.

"I'm referring, of course, to the question of Effective Altruism and the extent to which it is intermingled with Anthropic and CEO Dario Amodei's insistence that We Must Pace the Frontier."

Satya Nadella — CEO of Microsoft. Why mentioned: Quoted from a Stratechery interview confirming Microsoft's multi-model harness strategy, undercutting the model/harness integration moat thesis.

"We're using the same harness that we use in GitHub and the same thing in security, too... we will have GPT, we will have Anthropic in there and any open weight model."

Clayton Christensen — Late business theorist, author of The Innovator's Solution. Why mentioned: His modularity/integration framework is used to explain why "good enough" model performance shifts competitive advantage away from proprietary integration.

"Companies can introduce new products faster because they can upgrade individual subsystems without having to redesign everything... firms have the slack to trade away some performance with these customers because functionality is more than good enough."

Sam Altman (implied context) — CEO of OpenAI. Why mentioned: Referenced as the original reason for Anthropic's founding and the ongoing rivalry underlying the "pace the frontier" debate.

"Anthropic exists because Amodei and his cofounders didn't trust Sam Altman and OpenAI."

5. Operating Insights

Don't assume model-harness integration is a durable moat — as multi-model harnesses (like Microsoft's Copilot Cowork) mature, the ability to swap underlying models undercuts lock-in strategies built purely around proprietary integration; product/UX and data portability matter more.

"It will be up to end users to determine how competitive Microsoft's harness is with Claude's, but clearly the harness and the model can be different things."

Build stickiness through user data/context accumulation, not raw capability — Meta's Muse shows a personal-agent product with a merely "good enough" model can create switching costs superior to a state-of-the-art chatbot, because of accumulated personal context.

"Once you have put all of your information into a personal agent and actually incorporated it into your day-to-day life, it is much more of a challenge to change to something else."

Watch for "good enough" thresholds where non-performance factors (privacy, cost, convenience) start driving purchase decisions — this is when modular/commoditized offerings gain ground against integrated incumbents, per Christensen's disruption framework; operators should track when customers start trading off capability for other attributes.

"This is evidence of Christensen's theory in action: customers demonstrated the willingness to base their model-choice decision on something other than pure performance, namely, data retention policies."

6. Overlooked Insights

The offense/defense asymmetry in automated cyberattacks is a structural, not incidental, problem — the math strongly favors attackers regardless of AI safety measures, meaning enterprises will eventually be forced to trust autonomous defensive agents, a significant but under-discussed operational shift for security-focused companies.

"The attack only needs to work once for the entire endeavor to have a positive payoff... Truly effective defense will mean truly trusting agents to act autonomously, but most companies won't do that until they are forced to by regular and unremitting hacks by fully autonomous attackers."

Anthropic's "profitability" claims exclude the single largest cost category (training), which functions like disguised depreciation — a subtle but critical caveat for investors evaluating the IPO narrative.

"The bigger issue is the exclusion of training costs, which are massive (and which, in a true depiction of gross margins, would count as depreciation)... Anthropic would be fine if only they didn't have to pay for training!"