The AI Stack Isn't Finished, So Stop Building as Though It Is

Platforming, in the sense this argument uses it, has little to do with which vendor wins. It is the arrival of a shared technological, industrial, and institutional architecture stable enough that independent firms can build on top without fearing the ground moves under them ^[1]. The revenue that makes the current moment look like a maturing ecosystem measures a different thing. A reported $2 billion run rate at Cursor and $1.2 billion in recurring revenue at Salesforce's Agentforce show how fast buyers will try the tools ^[5][7], not how settled the foundation beneath them is ^[1].

That distinction is the whole of the investment problem. Adoption has run faster than it did for the personal computer or the internet at the same stage ^[13], and roughly 2.4 billion people now touch a generative AI product each month ^[4]. But the essay's own standard is not adoption. It is whether AI becomes a general-purpose technology on the order of electricity or the internal combustion engine ^[10], and by that measure the complementary innovation and organizational rework such a shift requires is still in its infancy ^[11]. The binding constraint is not immature models or the usual adoption friction; those are real but not what holds things back ^[16].

Two findings from the underlying research make the posture concrete. Where AI is reliable remains jagged and task-by-task rather than uniform ^[14], and the deep productivity gains of a general-purpose technology historically arrive only after a lag, following the intangible investment that a J-curve describes ^[15]. Both argue against committing capital to a single architecture on the assumption it is finished.

So the counsel is to spend on things that survive a change in the stack. Learn faster than you commit, put money into complements rather than raw model capability, and build organizational capability that no shared model can hand your competitor as easily as it hands it to you ^[3]. Electricity, the steam engine, and the internet each transformed economies only after successive waves of complementary innovation built around them ^[12]; the firms that captured value were rarely the ones that bet earliest on a particular configuration.

The uncomfortable part is that many buyers are already staffed and provisioned as though the platform exists. Harvey is embedded across large law firms and Shopify treats AI use as a baseline expectation ^[8][9]. Treating a well-marketed layer as permanent architecture is exactly the commitment the evidence says to defer.