Key Highlights

  • Alphabet, Amazon, Meta, and Microsoft combined are projected to spend close to $725 billion on capital expenditure in 2026, nearly double their 2025 levels.
  • Hardware suppliers are managing the institutional memory of 15 prior industry downturns, not ignoring AI demand.
  • A collective action problem, not a capital shortage, is the structural root of the AI compute bottleneck.
  • US data centre power demand from AI could grow more than 30-fold by 2035, making energy the longest-duration constraint.
  • Amazon is projected to post negative free cash flow in 2026, while Microsoft expects to remain capacity-constrained through the year despite record spending.

Everyone Wants More AI. So Why Is There Not Enough?

Across enterprises, research institutions, and consumer platforms, AI tools are running slower, throwing errors, and throttling access at peak hours. The infrastructure powering artificial intelligence is under serious strain, and the companies building it are struggling to keep pace.

The instinctive explanation is that chipmakers and hardware suppliers are simply not investing enough. But that explanation misses the deeper structural reality — one that is more consequential and considerably harder to resolve than a spending shortfall.

The real story is about rational fear. And it goes back decades.

The Numbers Look Strong. The Problem Is Relative.

Global semiconductor industry capital expenditure reached $166 billion in 2025 and is projected to grow a further 20% to approximately $200 billion in 2026. By any historical benchmark, these are substantial sums. The chip industry is not standing still.

Yet the relative picture tells a different story. The capex-to-semiconductor-market ratio is projected to fall to 19% in 2026, below both the long-run average and the five-year average of 24%. The industry is spending more in absolute dollars while simultaneously becoming more conservative relative to the market it serves. That is not an oversight. It is a deliberate posture shaped by industrial history most outside observers rarely consider.

Fifteen Downturns and Counting

The semiconductor business is one of the most cyclically violent industries in the global economy. The boom-bust cycle is structurally embedded in semiconductors, occurring roughly every three to four years, driven by the inelasticity of supply against more volatile demand. Building a chip factory takes years. Cancelling an order takes an afternoon.

The financial consequences of misjudging this have been severe. The PHLX Semiconductor Sector Index has experienced five major drawdowns since 1996, ranging from 30% during the 2018 trade war to 82% during the dot-com bust, with recovery timelines spanning five months to over six years. Companies that over-built during the dot-com boom spent years writing down idle fabrication assets.

This is the institutional memory every senior semiconductor executive carries into every capital allocation meeting. TSMC's (NYSE:TSM) chief executive recently admitted to feeling "very nervous" about current investment commitments, warning that miscalculation at this scale could constitute a serious setback. Advanced nodes now account for 70 to 80% of planned 2026 capital outlays, meaning each investment decision carries amplified downside exposure if demand softens. That candour reflects hard-earned awareness, not complacency.

The Collective Action Trap

Each major hardware supplier faces an identical dilemma: expand aggressively and risk idle factories if AI demand moderates, or hold back and let competitors absorb the risk first. The rational choice for each individual company produces a collectively inadequate result for the industry as a whole, a classic collective action problem.

Memory manufacturers including Samsung, SK Hynix, and Micron (NASDAQ:MU) are expected to account for 45% of total semiconductor capex in 2026, with SK Hynix and Micron each increasing spending by over 40%. That is meaningful investment. But new wafer fabs still require two to three years from planning to production. Decisions made today determine supply availability in 2028, not now. Supply constraints paradoxically extend the current cycle by pushing deployments out over multiple years, improving near-term supplier pricing power while creating compounding risk if demand softens before new capacity is absorbed.

The Constraint Capital Cannot Quickly Fix

While chip supply dominates the conversation, the more durable constraint may have nothing to do with semiconductors. By 2035, power demand from AI data centres in the United States alone could grow more than 30-fold, reaching 123 gigawatts, up from approximately 4 gigawatts in 2024.

Reaching that scale requires expanding transmission grids, securing generation capacity, and navigating regulatory processes that do not move at technology speed. In constrained markets, obtaining a power connection for a new data centre has become a multi-year permitting exercise. Unlike chip shortages, energy infrastructure constraints are governed by dynamics largely outside the technology industry's direct control — and deserve considerably more strategic attention than they currently receive.

The Buyers Are Not Comfortable Either

Alphabet (NASDAQ:GOOGL), Amazon (NASDAQ:AMZN), Microsoft (NASDAQ:MSFT), and Meta (NASDAQ:META) plan to spend more than $700 billion on capital expenditure in 2026, nearly double what they spent the prior year. The individual positions reveal the pressure beneath that headline figure.

Microsoft set its 2026 capex at $190 billion, with its chief financial officer attributing $25 billion to rising memory chip and component costs, and confirming the company expects to remain capacity-constrained through at least 2026. Amazon held its budget at $200 billion, the largest among peers, while Meta raised its projection to a range topping $145 billion, citing higher component pricing alongside growing competition for land, power, and skilled workers.

The financial strain of sustaining this pace is becoming visible. Amazon is projected to post negative free cash flow of approximately $17 to $28 billion in 2026, Meta faces a potential drop of close to 90% in free cash flow, and Microsoft expects a 28% decline before recovering in 2027. Investors are increasingly concerned that depreciation and operating costs will outpace near-term AI revenue contributions, and markets have already reflected that conditional patience. A moderation in spending by even one of these operators would send significant ripple effects across the entire supply chain.

Capital Allocation Implications

Unlike prior technology cycles, the current rally is not inventory-driven but capex-driven, making it structurally more sensitive to shifts in the investment thesis of a concentrated customer base. Suppliers at full utilisation carry near-term pricing power, but history is consistent: that power is cyclical, not permanent.

Current leaders in AI GPUs, CPUs, and memory may find it challenging to maintain dominant market share as new entrants emerge and workloads shift from training toward inference. Energy infrastructure, advanced chip packaging, and specialised cooling systems, segments where expansion faces physical and regulatory constraints independent of investment volume, may offer more resilient positioning across a full market cycle.

The next true semiconductor supercycle will depend heavily on disciplined capital allocation across the industry. The AI growth story is structurally intact. But the pace of the infrastructure build is governed less by ambition than by the rational caution of an industry that has survived fifteen downturns by knowing when not to over-commit.