Memory Now Claims Two-Thirds of AI Chip Costs as HBM Shortages Reshape Industry Economics

High-bandwidth memory has seized control of AI hardware expenses. Once a secondary line item, HBM now represents 63% of total component spending on leading AI chips. That figure stood at 52% just six quarters earlier.

The numbers come from fresh analysis by Epoch AI. Its May 21, 2026 report tracks costs across chips from Nvidia, AMD, Google, and Amazon, weighted by actual production volumes. Total component outlays for these suppliers climbed from roughly $22 billion in 2024 to $52 billion in 2025. HBM alone drove $20 billion of that jump. (Epoch AI)

Logic dies held steady near 13%. Advanced packaging slipped from 19% to 15%. Auxiliary parts such as substrates and power delivery dropped from 15% to 9%. The pattern shows no sign of reversing. Supply tightness and rising prices point to HBM claiming an even bigger slice in 2026.

Memory’s Rising Dominance Forces Hard Choices Across the Supply Chain

Memory makers operate in a tight oligopoly. SK Hynix, Samsung, and Micron control more than 90% of advanced HBM output. All three companies report their 2026 production fully booked. Some contracts stretch into 2027. Prices for HBM3E stacks have climbed 20% in recent months. Spot market swings have reached far higher.

Hyperscalers already feel the pressure. Microsoft lifted its fiscal 2026 capital expenditure forecast to $190 billion. Roughly $25 billion of that increase traces directly to higher component prices. Meta widened its own 2026 guidance by $10 billion, citing the same factor. (The Register)

But the pain spreads further. Consumer electronics makers watch the same inputs. HP reported that memory now accounts for 35% of laptop materials cost, up from 15-18% a quarter earlier. Smartphone bills of materials could rise 15% or more. Conventional DRAM supply has tightened because manufacturers shifted wafer starts toward the far more profitable HBM stacks. One bit of HBM displaces several bits of standard DRAM. (Bloomberg)

And the shift carries environmental weight. Manufacturing HBM demands more materials and energy per gigabyte than ordinary DRAM. Emissions from memory production are climbing faster than those from logic chips. TechInsights analyst Stephen Russell noted the surge will raise semiconductor manufacturing emissions in absolute terms even as individual processes grow more efficient. (Economic Times)

Industry forecasts paint an aggressive picture. Global HBM spending is projected to reach $54.6 billion in 2026, then compound at 42% annually toward $100 billion by 2028. That timeline arrives two years ahead of earlier estimates. HBM could claim 25% of total DRAM wafer production this year and approach 38% before the decade ends. (Global X ETFs)

Chip designers respond in different ways. Nvidia’s Blackwell family stacks more HBM per GPU than its predecessors. Some hyperscalers push custom ASICs that optimize memory access patterns to reduce overall demand. Yet the physics stays stubborn. Large language models require massive key-value caches during inference. Bandwidth hunger grows with model size. No simple substitution exists.

Packaging capacity adds another constraint. TSMC’s CoWoS lines remain oversubscribed even as the relative cost share of packaging declines. The expense of logic dies has not fallen in percentage terms because advanced 3nm and 5nm wafers carry high price tags. Still, memory’s ascent outpaces every other category.

Suppliers race to expand. Micron raised fiscal 2026 capital expenditure to $20 billion to fund a dedicated HBM facility in Idaho, supported by CHIPS Act incentives. SK Hynix plans a $3.87 billion advanced packaging plant in Indiana. Samsung and others pour billions into new capacity. New fabs and packaging lines will not deliver meaningful supply until 2028 at the earliest. Lead times already exceed 52 weeks for some HBM configurations.

Short-term relief looks unlikely. TrendForce analysts warn of DRAM price increases between 50% and 55% in early 2026. Some legacy LPDDR4 parts have grown more expensive than HBM3E because manufacturers deprioritized them. The memory super-cycle shows few signs of peaking.

Investors have taken notice. Micron shares rose more than 340% from the start of 2025 to a January 2026 peak. The three dominant memory producers captured outsized gains as AI infrastructure budgets redirected capital their way. Some projections suggest memory makers could generate $551 billion in AI-related revenue in 2026, twice the expected haul for contract foundries.

Yet risks remain. A slowdown in hyperscaler spending or faster-than-expected gains in memory efficiency could ease pressure. No major analyst expects that outcome before late 2027. Instead, most foresee continued allocation battles, multi-year take-or-pay contracts, and steadily climbing prices.

The bill of materials for a single high-end AI accelerator now tilts heavily toward stacks of HBM3E or its successors. Eight or twelve dies stacked with through-silicon vias. Over 1,200 signal wires routed through each stack. Yields that remain challenging at volume. Each technical improvement in compute nodes only raises the relative importance of feeding those transistors with data at blistering speed.

So the industry finds itself in an unfamiliar position. The memory vendors, long viewed as cyclical commodity players, now sit at the center of the artificial intelligence arms race. Their production decisions set the pace for how many frontier models can train and how quickly inference can scale. Logic designers and packaging specialists must work around them.

Executives at Microsoft, Meta, Google, and Amazon now track HBM wafer starts with the same intensity they once reserved for GPU allocations. Their capital expenditure plans bend to memory availability. Consumer device makers adjust product road maps and pricing. The entire semiconductor value chain feels the reordering.

Epoch AI’s data offers the clearest snapshot yet of this transition. Memory’s share climbed 11 percentage points in under two years. Absolute HBM spending nearly tripled. The trend lines point higher. For an industry built on rapid performance gains, the new constraint is no longer just transistor density. It is the cost and availability of the memory that keeps those transistors fed.

That reality will shape procurement strategies, investment priorities, and competitive positioning for years ahead. The memory wall has not been broken. It has simply moved, and the bill has grown dramatically larger.