Franklin Templeton Says AI CapEx Is Still in the Early Innings as Skepticism Hits

A 10% semiconductor pullback is testing the AI capex story

The immediate question is not whether AI capital spending is real, but whether this pullback starts a deeper skepticism loop. The Philadelphia Semiconductor Index plunged 10% last week, the kind of sharp move that can push a sector from enthusiasm straight into proof mode. That is why Franklin Templeton's stance stands out. Katrina Dudley still describes the build-out as a "decade-long cycle", and she says the spending trend looks durable at least through 2026, through 2027, and well into 2028.

The market divide is clear. Bears argue that AI infrastructure spending may outrun measurable returns. Dudley's counter is that this looks more like early-cycle volatility than a broken thesis. For investors, though, the key watchpoint is perception: if hyperscaler spending stops being viewed as demand and starts being viewed as excess, the rerating can quickly move the other way.

For now, the bull case still rests on a large spending base. Hyperscaler commitments of $750 billion in capital expenditures for 2026 support that view, and SOXX has already begun to rebound after the selloff. The recent dip was enough to shake confidence, but not yet enough to break the capex narrative. If earnings start to lag that spending, repricing could accelerate. If they do not, this pullback may look more like a test than a turning point.

Dudley's "early innings" argument is about duration, not linearity

Dudley's "early innings" framing does not say spending will rise in a straight line. It says the overall investment cycle is still unfolding and still large enough to absorb volatility. That is the more useful read, because the deeper debate is not whether AI needs infrastructure, but how big that build-out will ultimately be and how quickly it will be renewed.

What is still being estimated

Goldman Sachs' analysis makes clear that the scale of AI infrastructure investment is not a single fixed number. It is highly sensitive to assumptions around The economic useful life of AI silicon, The cost and complexity of next-generation data centers, The chip and architecture mix, and Elongation from power, labor, and equipment bottlenecks. Current estimates of the final scale are therefore more conditional than they usually appear.

That matters because the bear case does not require AI demand to collapse. It only requires slower hardware refresh cycles, longer payback periods, or bottlenecks that change the timing and efficiency of the build-out. If AI chips remain economically productive longer than expected, cumulative spending could still be enormous and yet fall short of the most bullish market expectations.

Record earnings have not fully calmed investors

That tension showed up again last week. Samsung and SK Hynix pointed to big numbers and deals amid record earnings, but Bloomberg also stressed that investor skepticism is dragging stocks lower. That is the practical meaning of "early innings": the long-term case can still be intact while markets remain demanding in the short term.