Skip to main content

Nexus Expert Research

Data Center Chip Demand Heading Into 2027: What’s Driving It and What’s Capping It

Data center chip demand is growing because hyperscalers plan roughly $720–740 billion of 2026 capital spending, a large share of it on AI servers and networking. Demand now spans GPUs, custom ASICs, HBM memory, advanced packaging and switch silicon. Power, CoWoS packaging and HBM supply, not buyer appetite, are the binding constraints heading into 2027.

The five layers of data center chip demand

Data center chip demand is the total value and volume of semiconductors bought to build, expand and refresh data centers, including accelerators, CPUs, memory, networking silicon and power chips. AI infrastructure accounts for most of the growth in data center chip demand.

Nexus Expert Research splits data center chip demand into five layers that can each be tested on their own. The model is called the Nexus 5-Layer Chip Demand Stack:

  • Compute: GPUs, custom ASICs (chips designed for one customer’s workloads) and CPUs.
  • Memory: HBM (high-bandwidth memory, DRAM stacked next to the accelerator), server DRAM and data center SSDs.
  • Packaging: advanced packaging such as TSMC’s CoWoS, which joins accelerator dies and HBM stacks into one package.
  • Networking: switch ASICs, NICs (network interface cards) and the optics that link thousands of accelerators.
  • Power and cooling: grid connections, power delivery and liquid cooling, which decide whether installed chips can run.

The tightest layer caps shipments at every other layer. A buyer can order GPUs, but they ship only as fast as HBM and CoWoS allow. They earn revenue only once a powered, cooled rack is ready.

How big is the data center chip market?

IDC forecasts data center semiconductor revenue of $477.1 billion in 2026, rising to $843.2 billion by 2030, when data center chips would be nearly half of the total semiconductor market. Gartner, using a different definition, expects the AI data center ecosystem to account for 36.5% of 2026 semiconductor revenue and more than 53% by 2030.

Source (publication date)What is measured2026 figureLonger-range figure
IDC (April 29, 2026)Data center semiconductor revenue$477.1 billion$843.2 billion by 2030
IDC (April 29, 2026)Total semiconductor revenue$1.29 trillion (+52.8%)$1.75 trillion by 2030
Gartner (August 24, 2026)Total semiconductor revenue$1.6 trillion (+92% from $809 billion)$1.9 trillion in 2027
Gartner (August 24, 2026)AI data center share of semiconductor revenue36.5%More than 53% by 2030
Omdia (August 28, 2025)AI data center chips (GPUs and accelerators)$207 billion (2025 figure)$286 billion by 2030

The forecasts diverge mainly over memory pricing. Gartner’s August 2026 forecast puts 2026 memory revenue at $837 billion, against $594.7 billion in IDC’s April 2026 forecast. Repricing of memory between April and August explains most of the gap between the two totals.

The Omdia forecast is now out of date. Nvidia alone booked $89.0 billion of data center revenue in the three months to July 26, 2026, so Omdia’s $286 billion figure for 2030 sits well below current spending. When sizing the data center chip market, use the newest forecast and record its publication date and definition.

How much are hyperscalers spending on AI infrastructure?

The four largest US hyperscalers (Amazon, Alphabet, Microsoft and Meta) guided to roughly $720–740 billion of combined 2026 capital expenditure as of late October 2026. Amazon leads at about $220 billion. Capex also covers buildings and power, but Alphabet says about 65% of its Q3 2026 technical infrastructure investment went to servers and networking equipment.

Company2026 capex guidanceQ3 2026 capex2027 signal
AmazonAbout $220 billion (cash capex)$75.9 billionExpects to spend “meaningfully more” in 2027
Alphabet$195–205 billion (unchanged)$51.4 billionExpects capex to “increase significantly” in 2027
MicrosoftAbout $170 billion for calendar 2026 (including finance leases)$45.2 billion (fiscal Q1 2027)Over $50 billion expected next quarter
Meta$135–145 billion (including finance lease principal)$34.9 billionNo formal 2027 capex figure given

The four hyperscalers spent about $207 billion between them in Q3 2026. Microsoft trimmed its calendar-year figure from about $175 billion to about $170 billion because some third-party capacity deliveries were delayed. That reflects a supply constraint, not a demand cut. Microsoft expects capex to accelerate in the second half of its fiscal year, “primarily from CPUs, inclusive of higher memory and storage costs and GPUs.”

Capex is not the same as chip spending. The companies define it differently: Amazon reports cash capex, while Microsoft and Meta include finance leases. The totals also include land, buildings and power equipment.

Which chips power AI data centers, and who makes them?

AI data centers use five main chip types: GPUs and other accelerators for computation, custom ASICs such as Google TPUs and AWS Trainium, host CPUs, high-bandwidth memory (HBM) stacked beside each accelerator, and networking silicon including switch ASICs, NICs and optical DSPs. Nvidia, AMD, Broadcom, SK hynix, Samsung, Micron and TSMC are the main suppliers.

Chip categoryJob inside an AI data centerMain suppliersDemand signal
GPU / merchant acceleratorParallel math for training and inferenceNvidia, AMDNvidia data center revenue $89.0B in Q2 FY2027 (+117%); AMD data center $6.7B in Q2 2026 (+107%)
Custom ASIC (XPU)Accelerator built for one hyperscalerBroadcom (design partner), Google TPU, AWS TrainiumBroadcom AI semiconductor revenue $16.7B in fiscal Q3 2026 (+221%)
CPUHost node, orchestration, agent and RL workloadsIntel, AMD, Arm-based designs (Graviton, Axion, Cobalt, Grace/Vera)Late-2025 server CPU demand inflection reported by Intel and SemiAnalysis
HBMMemory stacked next to the acceleratorSK hynix, Samsung, MicronSK hynix held 50% of HBM revenue in Q2 2026 (Counterpoint)
Networking siliconSwitch ASICs, NICs, opticsNvidia, Broadcom and othersAI back-end switch sales passed front-end in 2Q 2026 (Dell’Oro)
Foundry and packagingMakes and packages leading-edge diesTSMC, plus OSATs such as ASE and SPILHigh-performance computing was 68% of TSMC’s Q3 2026 revenue

GPUs and merchant accelerators: Nvidia and AMD

Nvidia remains the largest supplier of data center accelerators by a wide margin. For the quarter ended July 26, 2026, Nvidia reported revenue of $96.2 billion, including $89.0 billion from data center. It guided the next quarter to $108.0 billion, plus or minus 2%, assuming no data center compute revenue from China.

AMD is Nvidia’s main merchant competitor. AMD’s data center segment revenue was $6.7 billion in Q2 2026, up 107% year over year. Meta and AMD announced plans to deploy up to 6 gigawatts of AMD Instinct GPUs. AMD expects its data center segment revenue to more than double in 2027, a company projection rather than a result.

Custom ASICs: Google TPU, AWS Trainium and Broadcom XPUs

Custom ASICs are the fastest-growing compute category in AI data centers, but merchant GPUs still make up most accelerator revenue. Broadcom’s AI semiconductor revenue, which covers custom accelerators and AI networking, reached $16.7 billion in fiscal Q3 2026, up 221%. Broadcom guided fiscal Q4 to $21.7 billion and says it has six XPU customers.

Hyperscalers are also starting to sell their in-house chips to others. Alphabet called TPU systems sales “a meaningful contributor” to Google Cloud revenue, which rose 103% to $30.7 billion in Q3 2026. Amazon plans to offer Trainium outside AWS from late 2027.

Custom silicon concentrates revenue in a handful of buyers. Broadcom estimates its top five end customers accounted for about 55% of net revenue in fiscal Q3 2026.

Why CPUs are back in AI data centers

Server CPU demand rose again in late 2025 because AI agents and reinforcement learning run large amounts of CPU-bound code alongside GPUs. SemiAnalysis reported in February 2026 that Intel saw an unexpected rise in data center CPU demand in late 2025 and was moving wafer supply from PCs to server CPUs. Microsoft’s October 2026 guidance lists CPUs first among its capex drivers.

Why is HBM the tightest part of the AI chip supply chain?

HBM is the tightest part of the AI chip supply chain because every AI accelerator needs several HBM stacks, only three companies make it at scale, and most capacity was committed in advance. IDC calls HBM “the primary constraint in the AI accelerator supply chain” and expects no meaningful new supply before late 2026 at the earliest.

  • Supplier concentration: According to Counterpoint Research, SK hynix held 50% of HBM revenue in Q2 2026, down from 64% a year earlier. Samsung ranked second.
  • Forward contracting: Micron says agreements cover the vast majority of its calendar 2027 HBM bit supply, at significantly higher year-over-year prices.
  • Market size: BofA estimates the 2026 HBM market at $54.6 billion, up 58%, as cited in SK hynix’s 2026 market outlook.
  • Spillover: IDC forecasts 2026 DRAM revenue of $418.6 billion, up 177%. Part of the reason is that HBM uses more wafer area per bit, which squeezes supply of conventional DRAM.

Read HBM revenue figures with care. Gartner’s Shrish Pant says “pricing expansion is accelerating growth in 2026,” which means much of the growth in memory revenue comes from higher prices, not higher unit volumes.

How does CoWoS advanced packaging limit AI chip output?

CoWoS (Chip-on-Wafer-on-Substrate) is TSMC’s 2.5D packaging technology that places accelerator dies and HBM stacks side by side on a silicon interposer, and AI chip output cannot exceed available CoWoS capacity. TSMC has said CoWoS capacity will grow at a compound annual rate above 80% from 2022 to 2027. Even so, capacity still falls short of demand.

Estimate (source)2026 CoWoS capacityLater outlook
Economic Daily News, via TrendForce (June 2026)120,000–140,000 wafers per monthSupply-demand gap narrowing from about 20% to about 10% by end-2026
Goldman Sachs, via TrendForce (October 2026)1.275 million wafers for the year (about 106,000 per month)2.73 million wafers in 2027 (+114%)
Commercial Times analysts, via TrendForce (October 2026)More than 130,000 wafers per month200,000 per month targeted in 2027

The CoWoS estimates disagree by about 25%, so no single figure should be treated as firm. Monthly capacity numbers come from analysts and supply-chain media, not from TSMC disclosures. On October 27, 2026, TrendForce, citing Commercial Times supply-chain sources, reported that TSMC is “reportedly weighing price hikes of up to 20%” on CoWoS, a move TSMC has not confirmed. DigiTimes, as reported by Wccftech in December 2025, said “NVIDIA has booked 800,000 to 850,000 wafers for 2026,” more than half of TSMC’s CoWoS capacity, and a Morgan Stanley estimate puts Nvidia at about 60%.

Chart comparing analyst estimates of TSMC CoWoS advanced packaging capacity for 2026 and 2027

Why is networking silicon now a major share of AI data center spending?

Networking silicon now takes a major share of AI data center spending because training and inference clusters link thousands of accelerators that exchange data constantly. Dell’Oro Group reported that in 2Q 2026, switch sales for AI back-end networks overtook front-end network switch sales for the first time.

  • Scale: Dell’Oro forecasts that cumulative AI back-end switch spending will approach $1 trillion over 2026–2030, with scale-up networking accounting for more than half the market by 2030.
  • Technology: Ethernet made up about two-thirds of data center switch sales in AI clusters in 1Q 2026.
  • Speed: Most AI back-end switch ports now run at 800 Gbps. Dell’Oro expects 1.6 Tbps by 2027 and 3.2 Tbps by 2030.
  • Supply: Dell’Oro expects the market to stay “supply-constrained rather than demand-constrained” for the whole five-year forecast.

How do power and cooling constrain data center semiconductor demand?

Power and cooling limit data center semiconductor demand by setting how many chips can be switched on and when. Chips can be deployed only as fast as data centers secure grid connections and liquid-cooled halls. In 2026, Microsoft, Alphabet and Amazon all said they were short of capacity.

  • Global electricity: In its Base Case, the IEA projects data center electricity use more than doubling, from about 415 TWh in 2024 to around 945 TWh by 2030.
  • US share: Lawrence Berkeley National Laboratory found that data centers used about 4.4% of US electricity in 2023 and projected 6.7% to 12% by 2028.
  • Load growth: Grid Strategies (November 2025) links about 90 GW of the 166 GW of forecast US peak-load growth to 2030 to data centers. It also says that data center figure is probably overstated by about 25 GW.
  • Queues: ERCOT had more than 438,000 MW of large-load connection requests by mid-2026, nearly 90% of them from data centers.
  • Capacity markets: PJM’s 2028/2029 capacity auction (July 2026) came in 6,831 MW short of its reliability requirement.
  • Rack density: Schneider Electric reports that average rack density rose from about 16 kW in 2025 to 27 kW in 2026, while only one in five operators say they can support 50–70 kW racks. Nvidia’s Vera Rubin NVL72 rack is 100% liquid-cooled.

Power shortages usually delay chip deployments rather than cancel them. Microsoft CFO Amy Hood said in April 2026 that Microsoft expects “to remain constrained at least through 2026.” Alphabet started renting third-party compute at the end of Q3 2026, citing a “supply-constrained environment.”

Training vs inference: how does the workload mix change chip demand?

GPU demand in data centers is now driven more by inference than training. Deloitte’s 2026 TMT Predictions (November 18, 2025) projected that inference would account for “roughly two-thirds of AI compute in 2026,” up from a third in 2023 and half in 2025. Training still drives the largest single clusters, but inference runs continuously and scales with every user, agent and token served.

FactorTraining workloadsInference workloads
Demand patternLarge, lumpy cluster build-outsContinuous, grows with usage
Chips favouredTop-end GPUs, scale-up networkingGPUs plus custom ASICs, CPUs for agents, high memory bandwidth
Memory pressureHigh (model size)High and rising (context length, concurrent users)
LocationConcentrated mega-campusesMore distributed, closer to users

The move toward inference spreads demand across more suppliers. IDC notes that “inference demand compounds on itself,” because each new model generation raises inference volume. At CES 2026, Lenovo said it expects AI spending to shift to 80% inference and 20% training. Both figures are forecasts, not measurements.

Will data center chip demand keep rising through 2027?

Data center chip demand is very likely to keep rising through 2027, because Alphabet and Amazon have both guided to higher 2027 spending and suppliers report sold-out capacity. The main risk is the pace of growth. Free cash flow at several hyperscalers turned negative in 2026, and power availability may delay deployments.

Signals pointing up:

  • Alphabet expects 2027 capex to “increase significantly,” and Amazon expects to spend “meaningfully more.”
  • TSMC guided Q4 2026 revenue to $52.3–54.3 billion. It also said its forecast of mid-to-high-50s percent annual AI accelerator revenue growth from 2024 to 2029 could prove conservative.
  • Micron has contracted most of its calendar 2027 HBM supply.

Risks that could slow growth:

  • Free cash flow in Q3 2026 was negative $9.1 billion at Alphabet and negative $2.1 billion at Meta.
  • Amazon’s trailing twelve-month free cash flow was negative $42.4 billion.
  • Omdia expected AI’s share of data center investment to peak in 2026.
  • IDC expects memory pricing to normalise, though it should stay structurally above pre-AI levels.

The outlook is uneven across chip types and regions. Memory revenue can fall even while unit demand rises, and Nvidia’s guidance excludes China.

Expert calls vs analyst forecasts: what operators see before the data shows it

Expert calls and analyst forecasts answer different questions. Analyst forecasts size the market. Expert calls explain why the market is moving and often surface allocation, delay and design changes months before those changes appear in published data. The strongest research uses both.

DimensionAnalyst forecasts (IDC, Gartner, Omdia)Expert calls (operators and buyers)
Time lagPublished periodically; IDC’s and Gartner’s 2026 totals, published four months apart, differ by $310 billionSame-week view from people placing orders
GranularityMarket totals and segmentsSite, vendor and allocation detail
Best useSizing, benchmarking, board decksTesting assumptions, timing, competitive moves
Main weaknessDefinitions differ; revisions can be largeSmall samples; individual bias
CompliancePublicMust avoid material non-public information

The Ground-Signal Checklist: six demand signals experts see first

The Ground-Signal Checklist is Nexus Expert Research’s list of six data center chip demand signals that operators usually see before forecasts reflect them.

  • Energization dates: when utilities will actually power new halls.
  • HBM and CoWoS allocation changes: which buyers gained or lost packaging and memory slots.
  • Rack density and cooling orders: liquid-cooling retrofits and power-delivery upgrades.
  • Third-party capacity contracts: leases with neoclouds and colocation providers, which both Microsoft and Alphabet cited in October 2026.
  • GPU vs ASIC mix in procurement: how new clusters divide between merchant and custom silicon.
  • Component lead times: Accuris tracking showed semiconductor lead times reached 40 weeks in March 2026.

Original example: A team sizing 2027 liquid-cooling demand could start from Schneider Electric’s rack-density figures, then ask three colocation power directors when customer racks above 100 kW will actually be energized. The gap between their answers and the published forecast is the research finding.

How to validate a chip demand forecast with expert calls

Choose one layer of the 5-Layer Chip Demand Stack to test, such as HBM allocation or power energization.

List the roles that see that layer: hyperscaler capacity-planning directors, OSAT packaging managers, memory account managers, utility interconnection engineers.

Interview both buyers and suppliers, so each side’s bias is checked by the other.

Ask about timing, allocation and direction, never about confidential prices or unreleased financials.

Compare answers with the latest public guidance, such as earnings-call capex figures and analyst forecasts.

Repeat after each earnings cycle, because hyperscalers update capex guidance every quarter.

Frequently asked questions about data center chip demand

Is Nvidia still the largest data center chip supplier?
Yes. Nvidia’s data center revenue was $89.0 billion in the quarter ended July 26, 2026. In comparable 2026 quarters, AMD’s data center segment brought in $6.7 billion and Broadcom’s AI semiconductor revenue was $16.7 billion.

What is the difference between a GPU and a custom ASIC?
A GPU is a general-purpose parallel processor sold to many customers. A custom ASIC, such as a Google TPU or AWS Trainium, is designed for one company’s workloads. It gives up flexibility in exchange for better cost and power efficiency on those workloads.

Does China matter for data center chip demand forecasts?
China matters mainly as a source of uncertainty. Nvidia’s outlook for the third quarter of fiscal 2027 assumes no data center compute revenue from China, so any change to export rules would shift forecasts.

Why are AI data centers raising memory prices for PCs and phones?
HBM and server DRAM take wafer capacity that would otherwise make conventional memory. IDC forecasts that mobile semiconductor revenue will fall to $89.8 billion in 2026, partly because memory now makes up a larger share of device bills of materials.

How often do hyperscalers update their capex guidance?
Amazon, Alphabet, Microsoft and Meta update capex guidance with each quarterly earnings report. In 2026, the updates came in late January or early February, late April, late July and late October.

Are expert calls on chip demand compliant?
Expert calls are compliant when experts are screened and briefed not to share confidential information, and when questions stay on industry trends rather than material non-public information.

Next step: test your chip demand forecast with Nexus Expert Research

The quickest way to stress-test a data center chip demand forecast is to ask three timing and allocation questions of operators who see the relevant layer every week. Nexus Expert Research recruits those operators from scratch for each engagement, including Big 4 hyperscaler directors, typically in under 24 hours.

Naveed Saqib

Muhammad Naveed Saqib is a content strategist at Nexus Expert Research, where he writes on the expert network industry, market research, and business intelligence for professional audiences. He focuses on turning complex, research-heavy topics into clear, well-sourced content that readers can actually trust and act on.

Write a comment

Your email address will not be published. Required fields are marked *