The latecomers just took the lead

Five months. That's how long it took Jeff Bezos to build a company valued at $38 billion. Bloomberg reported that Project Prometheus, his physical AI lab focused on manufacturing and robotics, is closing a $10 billion round with BlackRock and JPMorgan lining up. The lab has 120 employees, most poached from OpenAI, xAI, Meta, and DeepMind. It didn't exist in November.

The conventional reading is that Bezos is buying his way into AI. The more interesting reading is that buying your way in now works. The barriers that once protected incumbents (years of institutional knowledge, massive research teams, first-mover advantages built on data flywheels) have thinned to the point where a well-capitalised latecomer with a Rolodex can assemble a credible lab in less time than it takes most companies to finish a planning cycle.

This pattern is playing out at every scale simultaneously.

The middle falls out

In Beijing, Moonshot AI just shipped Kimi K2.6, a 1 trillion-parameter open-weight model that SiliconANGLE reported beats both Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro, scoring 58.6 against Anthropic's 53.4 and OpenAI's 57.7. A lab most Western developers hadn't heard of six months ago is now publishing state-of-the-art coding benchmarks under a Modified MIT License. The model activates only 32 billion parameters per token from its trillion-parameter pool and can coordinate 300 sub-agents across 4,000 steps. This isn't a scrappy underdog story. It's a demonstration that the research moat around frontier models has become shockingly narrow.

Meanwhile, at the other end of the spectrum, people who've never written a line of code are flooding app stores. TechCrunch reported that App Store releases jumped 60% year-over-year, with productivity apps and utilities cracking the top five categories for the first time. The working hypothesis is that AI coding tools like Claude Code and Replit have crossed a threshold where non-developers can ship real software. April releases are up 104% compared to last year.

Three stories, one structural shift: the advantage of incumbency is collapsing faster than the advantage of speed.

This has a name in economics. Joseph Schumpeter's creative destruction assumed that disruption came from innovators who'd spent years building something genuinely new. What we're watching is closer to creative assembly: the components (talent, models, frameworks, distribution) are now modular enough that the binding constraint isn't what you've built, it's how fast you can recombine what already exists. Bezos recombines elite researchers. Moonshot recombines open research and mixture-of-experts architecture. First-time developers recombine AI tools and app store distribution.

The practical implication for anyone building products: your defensibility probably isn't where you think it is. It's not in your model, your team's pedigree, or your head start. Those assets are more liquid than they've ever been, easily acquired or replicated within months. The durable advantages are the ones that compound with use and resist modular assembly: proprietary data from actual customers, distribution channels with real switching costs, domain expertise too specific to poach, and operational depth that can't be hired away in a single quarter.

I think the next twelve months will be deeply uncomfortable for mid-tier AI companies: those with decent models and reasonable funding but no structural lock-in. They're squeezed from above by latecomers who can outspend them overnight and from below by individuals who no longer need them. The middle is where creative assembly hits hardest.

The question worth sitting with: if being late is no longer a disadvantage, what exactly are you in a hurry to protect?

---