We see a slowdown in product launch cycles by frontier AI

Although the market currently doesn’t seem too bothered by what’s going on, it seems that the recent headlines from frontier AI labs are shifting from technical triumphs to damage control. When OpenAI and Anthropic suddenly start publicising severe model misalignments, pulling models offline, and lobbying for government-mandated slowdowns, we think a slowdown in spending of some kind relative to the current rosy expectations looks fairly certain.

Indeed, this could be the beginning of a profound legal reality check as OpenAI indefinitely delay the release of its highly anticipated GPT-6.1 Astra model, explicitly citing safety and misalignment concerns that “didn’t quite meet the bar.” Even more damning are the recent disclosures surrounding GPT-5.6 Sol, where OpenAI researchers caught the model exhibiting deeply deceptive behaviour during internal testing.

When a frontier lab publicly documents that its flagship models exhibit uncontrollable, deceptive behaviours, product liability risks start to become a problem. If these labs subsequently release such architectures into enterprise workflows—such as legal drafting, automated coding, or financial analysis—and the model commits a high-value error, they can no longer claim it was an unforeseeable software bug. In a court of law, that becomes knowingly deploying a defective system.

The so called “beta” shield that tech startups have historically used to ship rough code and let users find the edge cases seems to be coming to an end. When enterprises pay premium API fees to embed these models into core operating infrastructure, terms of service will not shield the labs from gross negligence lawsuits or statutory damages when mission-critical systems get corrupted by a model actively trying to hide its tracks. It is worth noting that OpenAI is already facing landmark product liability and wrongful death lawsuits—such as the recent high-profile Nippon Life and Raine v. OpenAI cases—testing exactly these design-defect and failure-to-warn thresholds.

To protect themselves from catastrophic future litigation, these labs may have no choice but to drastically lengthen their safety and legal review cycles. Slower model deployment cycles destroy the velocity thesis that venture capital has used to justify $1.3 trillion private valuations for OpenAI and over $2trn for Anthropic. If the cadence of frontier launches slows from every few months to once every 12 to 18 months, so will everything else.

This brings us back to labs facing crushing take-or-pay data centre leases and multi-billion-dollar infrastructure commitments while token prices are under pressure. Citing “safety concerns” provides the ultimate legal off-ramp by allowing them to delay or renegotiate infrastructure obligations without acknowledging structural insolvency.

We think there are many issues that could lead to a slowdown in the AI data centre buildout, including the secular 20% plus inflation rate we think we are seeing in the AI supply chain which is eating away at these massive spending budgets but delivering less compute power. So compute costs are continuing to rise while token pricing itself remains under pressure from China’s open-source models, which are considerably cheaper for less demanding applications.

Then there is the highly concerning revenue concentration of these firms. Recent corporate payment data confirms that a staggering 80% of OpenAI and Anthropic’s enterprise revenues come from just 1% of their customers. At Anthropic, a handful of hardcore users from within the AI industry, like Perplexity, generate the vast bulk of its usage. That is yet another glaring red flag regarding the fragility of their end-demand that seems to be completely ignored by Anthropic’s IPO jockeys.

If one also considers the growing political opposition to building data centres in the US—which has become a key issue ahead of the mid-terms as rural communities increasingly object to these monoliths being built in their backyards—there are even more physical hurdles for growth rates at this stage. This is especially true as hyperscalers are more than likely to be issuing more expensive debt to finance their AI data centre projects going forward.