Scarcity of specialised computing hardware shapes what AI products can offer, and the effects reach users in forms that rarely mention hardware at all. The chain from a supply constraint to a changed feature is short and traceable.
Where the constraint actually sits
The bottleneck is not the chip design or the software. It is manufacturing capacity for advanced packaging and high-bandwidth memory, both of which are concentrated in very few facilities.
Expanding that capacity means building fabrication plants, which takes years and enormous capital, and the decision to build is made against demand forecasts that may be wrong.
Power and cooling form a second constraint. A datacentre needs a grid connection sized for the load, and utility interconnection queues are long in most regions.
Neither constraint responds to demand quickly. A surge in orders does not produce more chips this quarter or next.
So capacity is effectively fixed on the timescale over which products are planned.
Why training and serving compete
The same accelerators train new models and answer user requests, and every organisation must split its fleet between the two.
Training a frontier model occupies a large cluster continuously for months. That hardware is unavailable to users for the duration.
Serving demand, meanwhile, is not optional. Requests arrive whether or not a training run is in progress, and degrading response times drives users away.
The split is therefore a strategic decision made repeatedly. Investing in the next model costs current capacity; protecting current capacity delays the next model.
Announcements about roadmap slippage often trace back to this allocation rather than to any research difficulty.
How rate limits encode scarcity
A rate limit is the most direct expression of capacity in a product. It caps how much of a shared resource any one customer can consume.
Limits are usually tiered, with higher allowances for larger commitments, which allocates scarce capacity towards revenue.
Sudden tightening of limits on an existing plan almost always signals a capacity problem rather than a policy change, whatever the accompanying message says.
Providers also shift users to smaller models under load, which preserves availability at the cost of quality and is frequently invisible to the user.
Reading limit changes as capacity signals gives a more accurate picture of a provider's position than its public statements do.
Why smaller models attract the investment
When capacity is scarce, efficiency gains have the same effect as new hardware, and they arrive far faster.
Techniques that shrink a model or reduce the arithmetic per request free capacity across the entire fleet at once, which is why work on distillation and quantisation intensifies during shortages.
A smaller model that is nearly as good on most requests is commercially superior to a larger one that is slightly better and several times more expensive to run.
This is why product ranges have widened, with fast cheap models handling routine work and expensive models reserved for hard requests.
Scarcity pushed the field towards efficiency sooner than performance ambitions alone would have.
How pricing follows capacity
Prices per unit of usage have fallen substantially over recent years, driven by efficiency work and competition.
Underneath that trend, individual providers raise effective prices during tight periods by adjusting which model serves a tier, lowering included allowances, or introducing surcharges for peak usage.
Flat-rate consumer subscriptions are exposed here, because heavy users can cost more than they pay, and the usual response is a usage cap rather than a price rise.
Enterprise contracts increasingly include reserved capacity, which is a customer paying for hardware to sit ready rather than for requests served.
The emergence of reserved capacity as a product is itself evidence of how binding the constraint has been.
What gets cut first
Features with high compute cost per user and unclear revenue are the first casualties, and video generation sits squarely in that category.
Free tiers shrink next, because they consume capacity that could serve paying customers, though they are rarely removed outright since they feed the funnel.
Long context handling gets quietly limited, since memory consumption grows quickly with input length and long requests occupy hardware disproportionately.
Research access programmes and generous trial allowances contract, which is less visible but slows external work on the models.
None of these decisions is announced as a capacity measure, so the pattern is only apparent across several providers at once.
Why datacentre timelines stay slow
Announced construction plans imply capacity arriving in years, not months, and each project depends on land, power, water and permits before any hardware is installed.
Grid connections are the common delay. Utilities plan generation and transmission on decade timescales and cannot absorb large new loads at short notice.
Local opposition around water use and electricity prices has slowed projects in several regions, adding uncertainty to schedules that were already long.
Some operators have responded by siting facilities near existing generation or by contracting for dedicated supply, which shortens one delay while adding others.
The consequence for product planning is that capacity relief is always several quarters away, and interim measures must carry the load.
How teams design around scarcity
Applications built with capacity in mind route requests by difficulty, sending routine work to a small model and escalating only what needs the larger one.
Caching repeated results, batching requests and shortening prompts all reduce consumption without changing what the user sees.
Teams also build fallbacks across providers, so a capacity incident at one supplier degrades quality rather than halting the product.
The discipline this imposes tends to survive after the shortage eases, because the same techniques reduce cost and latency permanently.
Scarcity has been an unusually effective teacher of engineering restraint in a field that otherwise rewards scale.