Why The AI Slowdown Is Already Here

Deployment friction, not benchmark scores, is the real sign model-to-model gains are slowing even as AI capex keeps climbing toward $500 billion.

Trace Cohen
By Trace Cohen · The VC Read
Founder & Editor-in-chief, Value Add VC
2 min read
Why The AI Slowdown Is Already Here: key figures
ShareXLinkedInEmail

THE RUNDOWN

1

Deployment friction -- stalled enterprise pilots, slow API rollouts after a model announcement -- is a more honest capability signal right now than benchmark scores, which keep getting gamed.

2

2026 AI capex is still tracking toward roughly $500 billion even as perceived model-to-model gains slow, meaning capital commitment and capability progress have become disconnected signals.

3

Founders betting their roadmap on the next model generation finally fixing a reliability problem should treat that as a bet, not a certainty, given hallucination rates haven't meaningfully improved all year.

4

Public-market scrutiny arriving before the spending curve bends is the real catalyst to watch -- several frontier labs will answer to quarterly earnings before the capex-vs-capability gap forces a correction.

The VC Read

Value Add VC analysis

For investors, the edge is moving to deployment speed and reliability engineering rather than raw model access, because the model-access moat is the one most likely to evaporate next.

Analysis

The AI slowdown The Register's Rupert Goodwins just wrote about is real, yet many startup valuations are still priced for the version of 2026 where every model release was strictly better than the last one. That stopped being true months ago. Frontier labs are shipping models that have to be kept on tighter guardrails for "uncontrollable" behavior -- Goodwins' phrase, and the right one -- and nobody in the field has actually fixed hallucinations, they've just gotten better at hiding them behind more confident-sounding output.

The tell isn't benchmark scores, which keep climbing because the benchmarks keep getting gamed. It's deployment friction: how many enterprise pilots quietly stall at the "we need a human to check every output anyway" stage, and how long it takes labs to go from announcing a frontier model to actually shipping the API access that makes it useful. That gap has been widening all year, not shrinking, and it's the clearest evidence that "bigger model, better results" stopped being a reliable formula sometime around mid-2026.

“For the labs themselves, slower perceived progress is exactly the environment where capital discipline questions get asked that didn't get asked in 2024 or 2025.”

For founders building on top of frontier APIs, this means the real risk isn't that your model provider gets leapfrogged -- it's that the next model you're counting on to finally fix your reliability problem doesn't, because the problem was never really a capability gap. For the labs themselves, slower perceived progress is exactly the environment where capital discipline questions get asked that didn't get asked in 2024 or 2025.

Room for disagreement: the capex numbers don't show anyone actually slowing down -- Big Tech's 2026 AI infrastructure spend is still tracking toward roughly $500 billion, and OpenAI, Anthropic and Google are all still raising or deploying at a pace that assumes the next model generation justifies it. A genuine slowdown in perceived capability doesn't mean a slowdown in capital commitment, and by the time the spending curve actually bends, several of these companies will already be public and answering to quarterly earnings calls instead of venture boards. That lag is exactly why this is a bet worth having a position on either side of.

ShareXLinkedInEmail

More from the Pulse archive: Artificial Intelligence coverage → · All topics

Key Sources

1 source

THE WIRE in your inbox— Tech, startup & VC news with The VC Read, a few times a week. Free to subscribe, no spam.

No spam, unsubscribe anytime. Sent via Substack · Privacy · or subscribe on Substack ↗