AI Infrastructure Is a Semiconductor Procurement Problem, Not a Systems Integration Problem

The Number That Changes the Procurement Conversation

The binding constraint on AI infrastructure build-out is chip supply. A joint report from the Semiconductor Industry Association and Deloitte, released June 1, 2026, puts numbers behind what's long been suspected.

Semiconductors account for over 95% of the content value of a leading AI server rack and more than 50% of the total capital expenditure required to build and operate an AI data center. Everything else in the rack — enclosures, cabling, power distribution, cooling — is rounding error.

The report projects annual revenue for semiconductors used in AI data centers could reach $1.2 trillion by 2028, representing a nearly tenfold increase over the last four years and surpassing total global semiconductor sales from 2025 — across all end uses — by more than 50%. One application segment in two years will exceed what the entire global chip industry produced across automotive, mobile, consumer, industrial, and data center combined just three years prior.

What Is Actually Inside the Rack

The SIA-Deloitte methodology rests on a bottom-up approach. The report conducted a virtual teardown of a state-of-the-art AI data server rack, the foundational unit of centralized AI infrastructure.

Each server rack contains more than 4,500 chips comprised of approximately 20,000 individual semiconductor dies. A leading data center can accommodate over 45,000,000 chips. At that scale, yield and allocation become existential concerns.

Key components include AI accelerators, ASICs, FPGAs, CPUs, DPUs, networking chips, high-bandwidth memory, DRAM, SRAM, NAND flash, power management devices, controllers, sensors, and transceivers. AI accelerators account for 74% of server rack value, with logic chips making up 70% of the semiconductor content within that category.

That 74% figure matters for supply risk. NVIDIA dominates accelerators. But the remaining 26% — memory, networking silicon, power management, storage controllers — represents a second, equally constrained supply chain. Logic and memory technologies together represent more than 85% of total semiconductor value across the rack.

HBM is the clearest near-term pressure point. High Bandwidth Memory has moved from a niche component to a core piece of AI infrastructure. Sales are expected to grow from $15.2 billion in 2024 to $32.6 billion by 2026, with a compound annual growth rate of 21.7% through 2028. HBM still costs about five times more than standard server DRAM, and supply depends on advances in packaging.

The Supply Side Cannot Keep Up With Demand on the Same Timeline

Advanced process capacity at 7nm and below is expected to increase by approximately 69% — from 850,000 wafers per month in 2024 to 1.4 million wafers per month in 2028 — representing a CAGR of around 14%, double the industry average. Against a tenfold revenue projection, a 69% capacity increase suggests persistent allocation constraints through the forecast period, barring dramatic ASP compression or yield improvements.

Capital expenditure on advanced process equipment is forecast to surge to over $50 billion by 2028, a 94% increase from the $26 billion invested in 2024. Equipment lead times run 12 to 18 months. Fab construction takes 3 to 5 years. The capital is committed; the wafers are not yet cut.

Advanced packaging compounds the problem. AI accelerators rely on 2.5D integration — logic die co-packaged with HBM stacks over an interposer. These AI accelerators rely on advanced packaging technologies, such as 2.5D integration with high-bandwidth memory and emerging 3D approaches, which go far beyond conventional packaging. Packaging capacity, not logic node availability, is increasingly where allocations get stuck.

To meet global demand for new AI applications, government and industry will invest over $4 trillion in new data center infrastructure through 2028, of which up to $2.8 trillion will be spent on semiconductors. That $2.8 trillion in semiconductor spend over roughly two and a half years requires a supply chain that does not yet exist at that scale.

Who This Repositions

For operators building AI infrastructure, the 95% content-value figure collapses the traditional procurement hierarchy. Chip allocation — vendor, generation, volume, contract term, and what happens when supply tightens — becomes the primary decision. Systems integration and software licensing are secondary.

NVIDIA, AMD, and Broadcom gain pricing leverage. Any vendor holding wafer allocation agreements signed before demand curves shifted sits on a structural cost advantage. OEMs and ODMs lose negotiating leverage on BOM costs when 95 cents of every dollar is already priced elsewhere.

For procurement teams at hyperscalers and enterprises, multi-year supply agreements locked at current pricing are no longer hedges. The chips have already been ordered and are in backlog, data centers are under construction, and the numbers for the next 12 months are likely solid. The risk window is 2027 and beyond.

The policy dimension runs through the report's framing. The SIA published this alongside the Trump administration's Pax Silica Initiative and AI Exports Program, positioning the data as justification for domestic fab subsidies and export frameworks. Governments reading this as procurement guidance will accelerate sovereign chip programs, changing allocation dynamics in ways not yet priced into most corporate capex models.

What to Watch

1. Advanced packaging lead times, Q3 2026. CoWoS and SoIC capacity at TSMC is the current chokepoint for NVIDIA GB-series and custom ASIC ramp. Any slip in packaging yield reorders the 2027 delivery schedule for every major cloud operator.

2. Multi-year supply agreements, H2 2026. Watch for Microsoft, Google, Meta, and Amazon to announce locked-price, multi-year agreements with chip vendors. The SIA-Deloitte data gives procurement teams the justification to sign long-term contracts at premium prices.

3. HBM supply constraints into 2027. Samsung and SK Hynix have both flagged that AI-driven memory shortages could persist through 2027. If HBM availability constrains accelerator shipments before logic node capacity does, the 2028 revenue projection slips.

4. Substrate and packaging as the next bottleneck. Once advanced logic node allocation stabilizes, packaging substrate — ABF substrate in particular — becomes the binding constraint. Track capacity announcements from Ibiden and Shinko Electric.

5. Whether the $1.2T figure holds under macro pressure. If monetization of AI looks like it may take longer or be lower than expected, data center projects could be canceled or postponed, with an adverse impact on chip sales. The 2028 projection is demand-driven. Any deterioration in AI ROI evidence cascades through the entire capex plan.