AI infrastructure & semiconductors
AI infrastructure and semiconductors is the silicon, systems, networking and cloud capacity layer that trains and serves frontier models. Nvidia's data centre revenue ran at a $110B+ annual rate by late 2024, the company shipped roughly 3.5M Hopper GPUs into 2024, and Blackwell ramp through 2025 took data centre revenue past $35B in a single quarter. The category covers training and inference accelerators (Nvidia, AMD, Cerebras, Groq, SambaNova and Tenstorrent), hyperscaler in-house silicon (Google TPU, AWS Trainium and Inferentia, Microsoft Maia, Meta MTIA), GPU clouds (CoreWeave, Lambda, Crusoe, Together AI and Nebius) and the networking, memory and packaging stack underneath.
The sector spans training accelerators, inference accelerators, hyperscaler in-house silicon, GPU cloud platforms, AI networking and interconnect, high-bandwidth memory and advanced packaging, AI data centre capacity and power, and edge and on-device AI silicon.
Revenue comes from GPU and accelerator hardware sales, GPU cloud rental priced per hour per GPU, foundry wafer revenue tied to advanced nodes, licensing of IP cores and architectures, networking and switching hardware, and long-dated power purchase and colocation contracts underpinning new AI data centres.
AI infrastructure & semiconductors is part of AI & ML.
$136B
Global market size
423
Public companies
Key VC investors
Key strategic buyers
How AI infrastructure & semiconductors companies monetize?
AI infrastructure & semiconductors companies monetize through accelerator hardware sales, GPU cloud rental and long-dated capacity contracts.
Accelerator hardware sales
Direct hardware revenue from GPUs, AI accelerators and full rack systems. Nvidia DGX and HGX, AMD Instinct, Cerebras CS-3 and Groq LPU systems ship via OEM, ODM, hyperscaler and direct enterprise channels.
GPU cloud rental
Per-GPU per-hour pricing on H100, H200, B200 and GB200 capacity. CoreWeave, Lambda, Crusoe and Nebius sell this directly; AWS, Azure, GCP and Oracle Cloud price it inside broader commitments.
Foundry wafer revenue
Wafer fabrication revenue on advanced nodes (3nm and 2nm). TSMC captures the majority of advanced AI silicon manufacturing; Samsung Foundry and Intel Foundry Services compete with smaller share.
IP licensing and royalties
Licensing of CPU, GPU, NPU and DSP IP cores plus per-chip royalties. Arm licenses to virtually every smartphone and a growing share of data centre silicon; Synopsys and Cadence license EDA tools and IP blocks.
Networking and switching hardware
Switches, NICs, optical interconnect and DPUs for AI clusters. Nvidia (NVLink, Spectrum-X, Quantum and BlueField), Broadcom (Tomahawk and Jericho), Arista and Marvell lead AI networking.
Long-dated capacity contracts
Multi-year power, colocation and capacity contracts underpinning new AI data centres. Stargate (OpenAI, Oracle, SoftBank and MGX, $500B announced January 2025), Microsoft-CoreWeave, Anthropic-AWS Project Rainier and Meta-Hyperion are reference examples.
AI infrastructure & semiconductors valuations in May 2026
Public AI infrastructure & semiconductors comps trade at 7.9x EV/Revenue. Median revenue multiple across AI infrastructure & semiconductors M&A deals was 3.2x in the last 12 months. Median revenue multiple across AI infrastructure & semiconductors VC rounds was 29x in the last 12 months.
7.9x
Median EV/Revenue as of May 2026 for public AI infrastructure & semiconductors companies
24x
NVIDIA is the highest valued public AI infrastructure & semiconductors company based on EV/Revenue (excluding outliers)
3.2x
Median EV/Revenue across AI infrastructure & semiconductors M&A deals in the last 12 months
29x
Median EV/Revenue across AI infrastructure & semiconductors VC rounds in the last 12 months
AI infrastructure & semiconductors market segments
Major AI infrastructure & semiconductors segments include training accelerators, GPU cloud platforms and high-bandwidth memory & packaging.
Training accelerators
Silicon designed for training frontier models. Nvidia H100, H200, B200 and GB200 dominate the category with sustained majority share; AMD Instinct MI300X and MI325X are the principal merchant challenger with anchor wins at Microsoft and Oracle; Cerebras CS-3 (wafer-scale) competes on specific workloads; Google TPU v5p and TPU Trillium anchor hyperscaler-native training.
Inference accelerators
Silicon optimised for low-latency model serving. Nvidia continues to dominate on broad workloads; Groq LPU (founded by Jonathan Ross), SambaNova SN40L, Tenstorrent and Etched (transformer-specific ASICs) compete on latency and cost; AWS Inferentia 2 and Google TPU serve hyperscaler workloads; on-device NPUs from Apple and Qualcomm anchor the edge tier.
Hyperscaler in-house silicon
Custom silicon designed in-house by hyperscalers to reduce Nvidia dependency. Google TPU is the most mature (TPU v5p and Trillium shipped at scale); AWS Trainium 2 and Inferentia 2 ramping; Microsoft Maia and Cobalt; Meta MTIA.
GPU cloud platforms
Pure-play providers renting Nvidia and AMD GPU capacity by the hour. CoreWeave (NASDAQ: CRWV; IPO March 2025) leads scale; Lambda, Crusoe, Nebius (Yandex spin), Together AI and Voltage Park serve the mid-market.
AI networking and interconnect
Switching, NICs and optical interconnect linking GPUs into training clusters. Nvidia NVLink and Spectrum-X, Broadcom Tomahawk and Jericho, Arista 7800R, and Marvell pluggable optics anchor the category.
High-bandwidth memory and packaging
HBM3e and HBM4 stacks plus advanced CoWoS packaging underpin every frontier accelerator. SK Hynix leads HBM share with Samsung and Micron competing; TSMC CoWoS capacity is the principal industry bottleneck for AI silicon supply and is sold out 12-18 months forward.
AI data centre capacity and power
Power-dense colocation, hyperscale build-out, grid interconnects and PPA-backed greenfield campuses. Equinix, Digital Realty, Vantage, QTS (Blackstone) and Compass build colocation; Stargate ($500B announced January 2025 by OpenAI, Oracle, SoftBank and MGX) anchors greenfield US capacity.
Edge and on-device AI silicon
NPUs and accelerators for laptops, smartphones, wearables and embedded devices. Apple Neural Engine, Qualcomm Hexagon, Intel Lunar Lake NPU and AMD XDNA define PC-class silicon; Arm and MediaTek license edge AI IP into mobile.
Fractional CFO, financial modelling and deal advice for AI infrastructure & semiconductors companies
See how Flow helps AI infrastructure & semiconductors founders.
We speak founders' language and have great operational understanding of AI infrastructure & semiconductors businesses.
Book an intro call - we'll look under the hood and recommend concrete next steps.
Fractional CFO
For founders who want to improve their FP&A functions, build an investor-ready financial model, and prepare for the next VC round.
Capital raising
For bootstrapped and already-VC-backed entrepreneurs who are looking to raise late stage venture or growth capital.
M&A
For category-leading technology companies who are exploring exit alternatives with either financial or strategic acquirers.
Key AI infrastructure & semiconductors KPIs to track
Data centre revenue, gross margin, GPU shipments, utilisation and power capacity under contract are the metrics investors track in AI infrastructure & semiconductors.
| KPI | Definition |
|---|---|
| Data centre revenue | Hardware and silicon revenue tied to AI training and inference. Nvidia data centre segment runs above a $110B annual rate; the headline gauge of AI capex. |
| Gross margin | Hardware gross margin. Nvidia data centre at 75-78%; AMD MI300 series at 50-55%; specialty AI silicon often below; commodity GPU cloud at 55-70%. |
| GPU shipments / installed base | Volume of accelerators shipped and active in customer clusters. Nvidia shipped roughly 3.5M Hopper GPUs into 2024; the cleanest read on training capacity available globally. |
| Utilisation | GPU and cluster utilisation. The principal economic signal for GPU clouds; healthy operators run sustained utilisation above 75%. |
| Power capacity (MW) under contract | Megawatts of power contracted for AI data centres. The new binding constraint on growth; tracked by hyperscalers, colocation operators and GPU clouds. |
| Backlog / RPO | Remaining performance obligations on multi-year contracts. CoreWeave reported over $30B in RPO at IPO; Nvidia and AMD report supply commitments running 12-18 months forward. |
| Capex intensity | Capex as a share of revenue. Hyperscalers running capex above 40% of revenue on AI build-out; GPU clouds running above 200% in growth years. |
| Yield and packaging capacity | TSMC CoWoS wafer yield and packaging capacity. The hard constraint on advanced AI silicon supply; tracked by every analyst covering the category. |
Main AI infrastructure & semiconductors players globally
The most active AI infrastructure & semiconductors companies and category leaders globally.
| Company | HQ | Overview |
|---|---|---|
Nvidia nvidia.com |  Santa Clara | Dominant supplier of AI training and inference accelerators (NASDAQ: NVDA). Data centre segment passed a $110B annual rate by late 2024; market cap crossed $3T in 2024 and remained the largest or second-largest listed company globally through 2025. CUDA software stack remains the principal moat. |
AMD amd.com | Santa Clara | Principal Nvidia challenger in data centre accelerators (NASDAQ: AMD). MI300X and MI325X ramped through 2024-25 with anchor commitments from Microsoft, Meta, Oracle and OpenAI; data centre AI revenue passed $5B in 2024 and ran well above $10B in 2025. |
TSMC tsmc.com |  Hsinchu | World's largest semiconductor foundry (TWSE: 2330; NYSE: TSM). Fabricates the majority of advanced AI silicon (Nvidia, AMD, Apple, Google TPU and Broadcom) on 5nm, 4nm, 3nm and 2nm nodes. CoWoS advanced packaging capacity is the principal industry bottleneck. |
Broadcom broadcom.com | Palo Alto | AI networking silicon, custom ASICs and the largest enterprise software business after VMware (acquired November 2023 for $69B). Designs Google TPU and Meta MTIA accelerators; market cap above $1T through 2025. |
Cerebras Systems cerebras.net | Sunnyvale | Wafer-scale AI accelerator manufacturer. CS-3 system built around the WSE-3 chip; filed for IPO in 2024 (delayed pending CFIUS review of G42 investment); principal customer concentration in G42 (UAE). |
Groq groq.com | Mountain View | Founded by Jonathan Ross (Google TPU originator); designs the LPU inference-optimised accelerator. Raised $640M Series D in 2024 at $2.8B valuation led by BlackRock; offers token-priced API competing with Nvidia GPU clouds on inference latency. |
SK Hynix skhynix.com |  Icheon | World leader in HBM3e high-bandwidth memory (KOSPI: 000660). Single largest HBM supplier into Nvidia Hopper and Blackwell; HBM revenue grew several-fold across 2023-25 as AI capex ramped. |
CoreWeave coreweave.com | Livingston | GPU cloud (NASDAQ: CRWV; IPO March 2025). Largest pure-play GPU cloud globally; over $15B revenue run-rate by 2025 with anchor customers Microsoft, Meta, OpenAI and Mistral; reported over $30B in remaining performance obligations at IPO. |
Lambda lambdalabs.com | San Francisco | GPU cloud and on-premise AI infrastructure. Raised $480M Series D in early 2024 at $1.5B valuation; principal Nvidia partner for mid-market AI training and inference. |
Crusoe crusoe.ai | Denver | Stranded-energy-powered AI cloud and data centre operator. Raised $600M Series D in late 2024 at $2.8B valuation; builds gas-flare-powered data centres in West Texas oilfields and is delivering capacity into Stargate. |
CFO-as-a-service without the overhead
Fractional CFO services priced for startups - senior operator, no permanent headcount.
Key AI infrastructure & semiconductors market trends
The Stargate $500B build-out, power as the binding constraint and HBM & CoWoS as the supply chokepoint are reshaping AI infrastructure & semiconductors right now.
Stargate and the $500B build-out
OpenAI, Oracle, SoftBank and MGX announced the Stargate joint venture in January 2025 to deploy up to $500B into US AI data centres over four years. The first 1.2GW Abilene, Texas site is under construction with Crusoe; the programme has reset the scale of single-site AI infrastructure announcements.
Hyperscaler in-house silicon
Google TPU v5p and Trillium, AWS Trainium 2 and Inferentia 2, Microsoft Maia and Cobalt, and Meta MTIA are all in serious production deployment. Hyperscalers are not exiting Nvidia, but are routing inference and selected training workloads to in-house silicon to compress unit economics.
Power as the binding constraint
GPU supply has eased relative to 2023; power, transmission, substation interconnect and cooling have replaced it as the binding constraint on AI capacity growth. Microsoft signed a 20-year PPA with Constellation to restart Three Mile Island Unit 1 in September 2024; Amazon, Google, Oracle and Meta have all announced nuclear deals through 2024-25.
Export controls reshaping global silicon
US export controls expanded in October 2023 and again in December 2024 and January 2025 restrict advanced AI silicon to roughly 40 countries. Nvidia H20 and B20 (China-compliant variants) shipped through 2024-25 before the H20 ban in April 2025; Huawei Ascend 910C ramped to fill the gap inside China; Cambricon and Biren are growing fast on domestic substitution; the divergence between Western and Chinese AI silicon is now structural.
HBM and CoWoS as the supply chokepoint
SK Hynix, Samsung, Micron and CXMT HBM capacity plus TSMC CoWoS advanced packaging are sold out 12-18 months forward. Nvidia, AMD, Broadcom and Google each made multi-billion-dollar long-term commitments in 2024-25 to secure allocation; HBM4 ramp in 2026 will be the next inflection.
GPU cloud public market debut
CoreWeave's IPO in March 2025 marked the first scaled public-market test of the GPU cloud model. Trading dynamics, customer concentration disclosures (Microsoft anchor) and capex intensity have shaped how Nebius, Lambda, Crusoe and Together AI are pricing follow-on rounds. Multiple peers expected to file in 2026.
Similar verticals to AI infrastructure & semiconductors
Explore niches like AI software and AI consulting.
Explore other sectors
We know tech inside & out.
We live and breath tech - true understanding of how startups operate is fundamental at what we do.
Recent insights across AI infrastructure & semiconductors and beyond
Talk to us
Schedule a call to get a health check on your business and see how we could help.
Fractional CFO
- Fractional CFO for Software
- Fractional CFO for AI & ML
- Fractional CFO for Fintech
- Fractional CFO for Consumer internet
- Fractional CFO for Digital media
- Fractional CFO for E-commerce & marketplaces
- Fractional CFO for Consumer products
- Fractional CFO for Mobility
- Fractional CFO for Digital health
- Fractional CFO for Industrial technology
- Fractional CFO for Digital infrastructure
- Fractional CFO for IT services
Stages
Countries
- UK Fractional CFO
- Ireland Fractional CFO
- France Fractional CFO
- Germany Fractional CFO
- Spain Fractional CFO
- Portugal Fractional CFO
- Italy Fractional CFO
- Netherlands Fractional CFO
- Belgium Fractional CFO
- Switzerland Fractional CFO
- Austria Fractional CFO
- Denmark Fractional CFO
- Sweden Fractional CFO
- Norway Fractional CFO
- Finland Fractional CFO
- Poland Fractional CFO
- Estonia Fractional CFO
- US Fractional CFO
- Canada Fractional CFO
- Mexico Fractional CFO
- Brazil Fractional CFO
- UAE Fractional CFO
- Australia Fractional CFO
Cities
- London Fractional CFO
- Dublin Fractional CFO
- Paris Fractional CFO
- Berlin Fractional CFO
- Madrid Fractional CFO
- Lisbon Fractional CFO
- Milan Fractional CFO
- Amsterdam Fractional CFO
- Brussels Fractional CFO
- Zurich Fractional CFO
- Vienna Fractional CFO
- Copenhagen Fractional CFO
- Stockholm Fractional CFO
- Oslo Fractional CFO
- Helsinki Fractional CFO
- Warsaw Fractional CFO
- Tallinn Fractional CFO
- New York Fractional CFO
- Toronto Fractional CFO
- Mexico City Fractional CFO
- São Paulo Fractional CFO
- Dubai Fractional CFO
- Sydney Fractional CFO