The Philadelphia Semiconductor Index has now fallen for seven consecutive sessions, a losing streak long enough to make even eternal optimists check their assumptions. The proximate anxiety is familiar: cheaper, more accessible artificial intelligence, of the sort Meta is enabling by releasing open-source model weights, might mean less money flowing into silicon.

The counterargument arrives courtesy of a nineteenth-century economist. The Jevons Paradox holds that when a resource becomes cheaper to use, consumption of it rises rather than falls. Applied to semiconductors, the theory suggests that open weights and efficient models should expand the boundaries of AI applications — and with them, the number of chips required to train and run those applications.

The market has so far declined to be convinced. Seven down days say as much.

Why cheaper AI might sell more chips

Meta's decision to publish model weights, rather than keep them locked behind an API, is the kind of move that reshapes developer behaviour at scale. Any team can now take a frontier-class model, modify it, and deploy it on hardware of their choosing. That choosing is the part chipmakers care about.

If the Jevons logic holds, each incremental deployment adds inference workloads — and inference, unlike a single headline training run, is a continuous drain on accelerators, memory and networking. Cheap software, in this reading, is not a threat to hardware demand. It is the demand.

The historical comparison is apt. Coal consumption rose after James Watt's engine made coal use more efficient, which is precisely the observation William Stanley Jevons recorded in 1865. Whether graphics processors follow coal's trajectory is the question now hanging over the sector.

What the losing streak actually reflects

Seven sessions of decline do not constitute a verdict on AI demand, but they do indicate that investors no longer treat accelerator spending as an article of faith. The earlier phase of the rally rewarded any company within shouting distance of a datacentre. The current phase appears to demand evidence.

That evidence will arrive in the usual forms: capital expenditure guidance from the large cloud operators, order books at the foundries, and memory pricing. Until then, the index is trading on sentiment, and sentiment has turned cautious.

There is an irony worth noting. The same open-source strategy that Meta presents as generosity doubles as a demand-generation machine for the entire compute stack. If developers build on free weights, someone still pays for the electricity and the silicon beneath them.

Can the paradox deliver a rebound?

The bullish case is straightforward. Broader application of AI, enabled by freely available weights, means more total computation, more accelerators, more advanced packaging and more high-bandwidth memory. The bear case is equally simple: efficiency gains may outpace adoption, leaving unit demand flat while prices fall.

History leans toward the bulls, but history does not buy semiconductors on margin. The seventh consecutive decline suggests the market wants numbers, not nineteenth-century analogies.

The test ahead is empirical. If Meta's open weights translate into measurable growth in deployed AI workloads over the coming quarters, the Jevons case strengthens and the sector's rebound case with it. If not, the index will have been right to be skeptical, and no paradox will save it.