Efficiency as strategy means supply stays tight longer than capex implies

Semiconductor Engineering reported that efficiency is key for the next two to five years, and that the biggest companies have likely locked in much of what they need. On Amazon's July 30th earnings call, CEO Andy Jassy said demand is strong into 2028. The largest players, like NVIDIA and Google, have massive capital and strong supply chain management, according to Semiconductor Engineering. Recently, all the major hyperscalers announced earnings and increased capex plans, partly due to rising costs, Semiconductor Engineering noted.

I read that as a supply signal, not a demand celebration. When the largest hyperscalers frame efficiency as the plan for two to five years, they are telling you that supply constraints will persist longer than the capex headlines suggest. This is not a temporary squeeze that resolves in 2027. It is an admission that the bottleneck will outlast the buildout cycle, and that operators who secured firm capacity positions early will extract value from those who did not. The operators with scale and early procurement have already locked their positions. The efficiency talk is about making those positions stretch.

Announced capex is intent. Firm supply positions are execution. The gap between the two determines who pays spot premiums and who operates from strength.

Efficiency talk is scarcity management

When hyperscalers describe efficiency as the strategy, they are describing scarcity management. Semiconductor Engineering stated that the number one strategy for coping with supply chain bottlenecks is efficiency — get the most results out of what you have and what you can get. That phrasing is critical: out of what you have and what you can get. It assumes that getting more is constrained, and that the constraint is binding enough to make efficiency the primary lever for the next several years.

This is not a temporary condition. A two-to-five-year efficiency horizon means that the supply side will not catch up to demand within the current capex cycle. Power is the binding constraint. Interconnection queues, substation lead times, cooling retrofits, and chip allocation will remain binding longer than the market is pricing. Operators who secured firm positions early — through advance power purchase agreements, early interconnection deposits, or long-term chip commitments — will operate from leverage. Late entrants will pay spot premiums or wait.

Semiconductor Engineering also noted that if you are lucky, you may find companies that acquired more resources than they can use. That line is telling. The market is already bifurcating into operators with excess capacity and operators scrambling for access. Excess capacity is described as a matter of luck, not market availability. Supply is tight and unevenly distributed. The operators with excess are the ones who locked early and locked large. The operators hunting for that excess are the ones who did not.

Capex obscures the execution timeline

Capex announcements signal intent, but they obscure the execution timeline. Semiconductor Engineering reported that recently, all the major hyperscalers announced earnings and increased capex plans, partly due to rising costs. Rising costs are a symptom of constrained supply, not a sign that supply is loosening. When capex increases but costs rise faster, operators are paying more for the same or less capacity. The bottleneck is tightening, not easing.

Capex cycles overshoot. They overshoot on dollars but undershoot on delivered capacity. The useful question is not how much capex was announced, but how much energized capacity will be online by year-end 2027, and who controls it. Announced megawatts are not energized megawatts. The gap between the two is where value migrates. Operators who locked firm interconnection positions and energization schedules early will convert capex into capacity on schedule. Operators who announced capex but lack firm supply positions will face delays, cost overruns, and spot-market exposure.

Semiconductor Engineering stated that the biggest companies have likely locked in much of what they need. That is the key phrase. Likely locked in means that the firms with scale and early-mover advantage have already secured their positions. The remaining supply is either spoken for or available only at spot premiums. The efficiency strategy is about making those locked positions stretch, not about waiting for supply to loosen. It is about maximizing utilization of what is already secured, because adding incremental capacity is constrained by lead times, permitting, and utility schedules that extend well beyond the current earnings cycle.

Who wins and who waits

The operators who win are the ones who locked capacity early and can now optimize utilization while others scramble. Semiconductor Engineering reported that the largest players, like NVIDIA and Google, have massive capital and strong supply chain management. Capital is necessary but not sufficient. Supply chain management — meaning early procurement, firm commitments, and locked-in delivery schedules — is what converts capital into capacity. The firms with strong supply chain management are the ones who placed orders, secured interconnection, and locked power agreements before the bottleneck became consensus.

The operators who wait are the ones who either underestimated the duration of the bottleneck or lacked the capital and conviction to lock capacity early. They now face a market where efficiency is the strategy because incremental supply is constrained. They will either pay spot premiums for available capacity, accept longer lead times, or partner with operators who have excess. Semiconductor Engineering noted that if you are lucky, you may find companies that acquired more resources than they can use. The secondary market for capacity — leases, partnerships, or outright sales — will become a meaningful channel, and the operators with excess will extract value from those without.

Meta's recent announcement that it will sell or rent excess AI computing power, including a potential ten billion dollar deal with Anthropic, is a concrete example. Operators with excess capacity are monetizing it. Operators without are paying for access. The fact that such deals are being structured now, in 2026, confirms that the bottleneck is real and that the secondary market is already active. The operators who locked capacity early are now in a position to monetize it, either by using it themselves or by leasing it to others at a premium.

The Case for Faster Supply Relief

The counterargument is that efficiency gains may be larger and faster than I expect, allowing constrained supply to stretch further than the current bottleneck implies. If software optimization, better utilization, or architectural improvements deliver step-function gains in output per watt or output per chip, then the supply constraint may ease without proportional capacity additions. In that scenario, the premium on early capacity lockup would diminish, and late entrants would have time to catch up without paying punitive spot rates.

Semiconductor Engineering asked, but what happens when supply eventually catches up? That question is valid. If supply does catch up within the two-to-five-year window, then the efficiency strategy becomes a bridge rather than a permanent condition, and the operators who paid spot premiums or waited may not face lasting disadvantage. If efficiency gains are genuinely large, then the bottleneck may resolve sooner than the hyperscalers are signaling, and the value of early capacity lockup may erode.

The risk to my thesis is that I am underestimating the pace of efficiency improvement and overestimating the persistence of the bottleneck. If utilization rates improve faster than I expect, or if software optimization reduces demand growth per unit of infrastructure, then the supply constraint may loosen earlier than the two-to-five-year horizon suggests. In that case, the operators who locked capacity early would still benefit, but the gap between early movers and late entrants would narrow, and the secondary market for capacity would soften.

The bottleneck migrates

The bottleneck migrates, and the efficiency strategy reflects that migration. Semiconductor Engineering stated that efficiency is key for the next two to five years, which implies that the bottleneck will shift across layers — from chips to power to cooling to memory and networking — but will not disappear. The operators who can navigate each layer as it becomes binding are the ones who will maintain capacity growth. The operators who cannot will face delays at each transition.

Power is the binding constraint today. Interconnection queues and substation lead times gate AI capacity harder than chip supply does. But as power constraints ease in some regions, the bottleneck will migrate to cooling, to networking, or to memory bandwidth. The efficiency strategy is about optimizing across all layers simultaneously, because the constraint is systemic rather than isolated. The operators with firm supply positions across multiple layers — power, cooling, silicon, and network capacity — are the ones who will convert efficiency into sustained growth. The operators with gaps in any layer will face binding constraints that limit their ability to scale.

Semiconductor Engineering reported that demand is strong into 2028, according to Amazon CEO Andy Jassy. That demand signal is useful, but it does not guarantee that supply will keep pace. The efficiency strategy is the hyperscalers' admission that supply will not keep pace, and that the gap between demand and supply will persist longer than the current capex cycle. The operators who prepared for that gap by locking capacity early are the ones who will capture the demand. The operators who did not will face a market where efficiency is necessary but not sufficient, and where incremental capacity is available only at a premium or with long lead times.

Utilization Prints to Track

I am watching disclosed utilization rates or efficiency metrics in Q3 and Q4 2026 earnings calls that quantify how much demand growth operators absorbed without proportional capacity additions. If hyperscalers report sustained revenue growth with flat or declining capex-to-revenue ratios, it confirms that efficiency is stretching constrained supply and that the bottleneck remains binding. If utilization rates remain high or increase, it signals that operators are running close to capacity limits and that incremental demand will require new supply, not just better optimization.

I am watching spot pricing or lead times for power, cooling equipment, or rack space in secondary markets through year-end 2026. If spot prices rise or lead times extend, it confirms that supply is tight and that late entrants are paying premiums. If spot prices soften or lead times compress, it suggests that supply is catching up faster than I expect and that the efficiency strategy may be a shorter bridge than the two-to-five-year horizon implies.

I am watching announcements of capacity sales, leases, or partnerships by players like Meta or xAI that confirm excess inventory exists and is being monetized rather than held. If such deals proliferate and pricing details emerge, it will provide a market-clearing price for constrained capacity and confirm that the secondary market is active. If deals are sparse or pricing remains opaque, it suggests that excess capacity is limited and that the operators who locked early are using it themselves rather than monetizing it, which would confirm that the bottleneck is tighter than the market is pricing.