Allocation, Queues, and Reading the Market
When capacity is fixed and demand exceeds it, something has to give, and what gives is access. The scarce packaging and substrate slots get divided among buyers by commercial arrangement rather than by spot availability. Buyers who committed early, who signed long-term agreements, who prepaid, who built priority relationships with the foundry and the substrate suppliers, get their allocation. Buyers who did not, wait. This is the mechanism by which the top-of-stack constraint becomes a market outcome: not everyone experiences the shortage equally, because allocation decides who feels it. Two companies with equally good chip designs can have entirely different fortunes based purely on who secured packaging slots earlier.
Queues are the visible symptom of this allocation regime. When you read that a particular accelerator has a long lead time, that the wait stretches out over many months, you are reading the length of the packaging and substrate queue, not a shortage of logic dies. The lead time is a direct readout of how far demand exceeds the binding downstream capacity. It lengthens when a new generation increases per-unit capacity consumption, and it shortens only as downstream capacity slowly catches up. The queue is the constraint made legible, and learning to read it as a packaging signal rather than a generic demand signal is a core skill.
Allocation also explains behavior that looks strange from the outside. A company can announce a powerful new design and still ship very few of them for a long time, because the design's existence is a fab-and-design achievement while its shipment rate is a packaging-and-substrate achievement, and the two are decoupled. A company can hold logic dies in inventory and still be unable to ship product, because it lacks the downstream capacity or the matched memory to finish them. These are not contradictions; they are what the inverted bottleneck looks like from the buyer's seat. Dies are necessary but no longer sufficient, and the sufficiency lives downstream where the capacity is allocated.
A worked example makes the readout concrete. Suppose two rival buyers each unveil a strong new accelerator in the same season. One committed to packaging and substrate slots two years earlier and locked matched memory supply; the other did not. The first ships in volume near launch; the second announces impressively and then trickles product out over many quarters, its dies effectively stranded for want of downstream capacity. Nothing about the second design was inferior. The gap is entirely an allocation gap, decided long before either product existed. Read that way, a divergence in shipment rates between two capable competitors is not a verdict on their engineering; it is a record of who reserved the scarce downstream lane first.
The allocation regime feeds back into strategy in ways worth naming. Because downstream capacity must be committed years ahead, the largest buyers behave less like customers placing orders and more like partners co-investing in capacity, underwriting substrate expansion, prepaying for packaging, securing memory supply through long agreements. This deepens the concentration: the buyers with the most capital secure the most allocation, which lets them ship the most product, which funds the next round of commitments. Scale compounds. The structure of the supply chain thus shapes the structure of the market above it, tilting advantage toward the players who can commit earliest and largest against a downstream chain that rewards exactly that.
For reading the market in practice, this suggests watching a specific set of signals rather than the headline wafer and design news. Watch the foundry's packaging capacity guidance and new-facility announcements. Watch the substrate suppliers' expansion decisions and their willingness to commit capital. Watch memory capacity, since it is a co-equal downstream constraint. Watch lead times as a live readout of the demand-to-capacity gap. And watch the allocation commitments, who is prepaying whom, because those reveal who has secured the scarce slots ahead of the shortage becoming public. These signals sit downstream of the design news everyone else is watching, which is precisely why they carry information the headline coverage misses.
The closing synthesis returns to the thesis and hardens it into a discipline. A GPU generation causes every downstream consequence: it sets the capacity each unit consumes, it sets the queue, it sets who gets allocated, and it sets the finished-unit supply of the industry. The design at the top is the cause, and everything below, packaging load, substrate demand, memory demand, lead times, allocation fights, is consequence. To reason well about this market you invert the usual attention: instead of asking how many chips can be designed or how many wafers can be printed, ask how much downstream capacity the design will consume and whether that capacity was committed in time. Answer that, and the downstream stops being a surprise and becomes a prediction. That inversion, from wafers to packaging, from design to consequence, is the entire discipline this course exists to install.