Work the 500% backwards and the base looks small. More than $100 billion of annualized token sales at the two leading labs today implies something near $17 billion a year ago [2][11], and a repeat of that increase would require the pair to clear $600 billion [13]. That is the trouble with the instrument Warsh handed the market in decrypt.co's account of the keynote [1]: a growth rate coming off a sixfold year decelerates whether or not a single buyer changes behaviour, so the second derivative he says he is watching [4] is close to guaranteed to turn down on the revenue side while the dollar level keeps climbing.
The capex line is the more usable half, or rather the more decomposable one. Roughly 9% over four quarters, the fastest pace since 2021, with more than half of that growth tied to AI buildout [5], puts AI above 4.5 points of the rise and the whole remainder of American business investment below 4.5 [12]. The asymmetry lives in flatness rather than contraction: hold AI-linked capex at today's level and aggregate growth drops under 4.5% [14], which converts the fastest pace since 2021 into an ordinary one without a single announced project being cancelled.
The $100 billion is revenue the labs collect, and the speech did not attempt the other side of that ledger. Warsh asked whether AI will produce a sustained rise in productivity and whether token usage complements or competes with labour, and answered neither [8], while the task force studying those questions has findings that, in his own words, have no bearing on current policy decisions [9]. So promoting AI to a potential factor of production [3], under a section header about preparing for future policy conjunctures [10], changes the Fed's forecasting apparatus well before it changes a rate path, and what the Fed is explicitly not doing is letting the productivity question reach the meeting room. The Fed is not trading on this series, but that does not stop everyone else from using it.
This reads differently depending on which assumption you make. If the hyper-Moore's law framing holds [7], flat dollars buy accelerating capability and the capex series undercounts what is actually being built. If deceleration arrives through power interconnection queues and chip supply rather than through buyers, the print looks identical to a demand break and the tripwire fires on the wrong cause. And if the retirement of secular stagnation [6] proves early, the higher potential growth rate embedded in this framing gets revised back out at the worst moment for anyone who levered to it.
My read is that the second derivative falls for arithmetic reasons and tells the market very little, which is a mildly embarrassing thing to say about a metric a Fed chair chose. What would break it is a print showing four-quarter capex growth above 9% with AI's share still above half. That would mean the acceleration has room, and the derivative is carrying information after all.
