If they surveyed 3K “ai” startups the likely answer is some tiny fraction have any genuinely unique IP and the others at best are likely using ML libraries and services from other parties. Again it will probably devolve in to a discussion around the boundaries for what IP is actually “AI”, but yes, it’s mostly sizzle with no steak in the market.

AI is whatever hasn't been done yet. - Douglas Hofstadter (1980)

A view perhaps exists that "true AI IP" is perhaps only algorithmic. However, let's be frank: it's rare that a new algorithm is the best, fastest, cheapest, only, or most appropriate solution for a problem.

Remembering that most inventions (AI or otherwise) aren't fundamentally enabling new ("green field" / "blue ocean") areas (they're just competing with existing approaches which are more costly, slower, less convenient, less reliable, less available, or less accurate than they might be - eg. after the invention) we should perhaps expect applications of existing algorithms to new datasets or domains to be the most numerous 'AI' applications in the marketplace.