The centralized AI paradigm is dead. That's the provocative thesis from Andrew Trask, an Oxford researcher who previously worked at DeepMind's language modeling team, in a widely-discussed Substack post arguing that networks of smaller AI models have definitively surpassed every frontier AI system on speed, accuracy, and cost. "Networks of neural networks are now faster, cheaper, and more capable than any Frontier AI system," Trask writes flatly. "The game is over." His analysis leans heavily on OpenRouter data—independent benchmarks that he argues reveal what actually happens in practice versus corporate sales pitches.

The Ensembling Revolution Nobody Wanted to Talk About

Trask traces the technical root of this shift back to a dirty secret from academic AI research: model ensembling was so reliably effective at boosting accuracy that it got banned from NeurIPS and other conferences. "If you weighted ensemble models together, you ~always get better accuracy… even if you're ensembling multiple trained versions of the same model," he explains. The reason is elegantly simple—different AI models make different mistakes, and when you combine their outputs through intelligent routing and weighting, those errors tend to cancel out. He describes building similar systems six months ago that reached "the low 50s" on benchmarks where even frontier models struggle. Stanford students are now launching startups around this approach.

Open Source Speed Wins the Cost Race

On cost, Trask identifies a structural advantage for decentralized networks: open source model providers compete purely on delivering fast, cheap results since training is given away free. He points to OpenRouter's independent ratings as evidence that "pound for pound, they're cheaper for the same level of intelligence." The DeepSeek insight about indexing and caching—comparing it to a librarian who walks to a specific chess section rather than reading every page of every book—is accelerating this efficiency by reducing AI inference costs 10-900x per year. "The fastest and lowest cost option is going to be a massive index into the world neurons… not a single blobby network that considers every possible fact in the universe whenever it generates a token," Trask argues.

Skipping Nation-Level AI Entirely

Perhaps most striking is Trask's geopolitical prediction: rather than transitioning from company-level AI (2010-2026) to nation-level AI, humanity will skip straight past national dominance to world-level AI. When the US Government banned Fable-class models, OpenRouter was offering better-than-Fable quality within 24 hours through model combinations. "The network is more than the node," he writes, drawing direct parallels to ARPANET and TCP/IP protocols that linked mainframe computers into something vastly more powerful than any individual machine.

Key Takeaways

  • Ensembling multiple models always wins because different AI systems make different mistakes that cancel out when combined
  • OpenRouter benchmarks show open-source model networks outperforming closed frontier systems at lower cost
  • The DeepSeek indexing breakthrough reduced inference costs 10-900x by enabling targeted retrieval instead of scanning all neurons for every token
  • US Government bans on frontier models have been circumvented within 24 hours via OpenRouter combinations

The Bottom Line

Trask's mainframe analogy is compelling but the transition won't be smooth—incumbent AI companies have "the best advertising/marketing minds in the world" and stock market irrationality could delay the reckoning for years. Still, if you're building anything assuming a single frontier model will remain optimal forever, you're betting against 60 years of computing history. The network always wins.