신경망 시장(2027-2040년)

Neural networks have become the organising principle of the global computing industry. Every frontier language model, every image and video generator, every autonomous-driving perception stack and most modern speech, vision and recommendation systems is a neural network. The hardware that trains and runs them, the models themselves and the software used to deploy them now make up one of the fastest-growing and most heavily capitalised markets in technology. The scale of investment is unprecedented. OpenAI closed a US$122 billion round at an US$852 billion valuation in April 2026, and Anthropic raised US$65 billion at a US$965 billion valuation in May 2026. Mistral AI raised €3 billion, AMI Labs US$1.03 billion and SambaNova US$1 billion. Cerebras raised US$5.55 billion in the largest US IPO of 2026. NVIDIA licensed Groq’s inference technology for about US$20 billion. Consolidation is reshaping edge AI and photonics, through onsemi–Synaptics (about US$7 billion), Marvell–Celestial AI (US$3.25 billion), ADI–Alif (US$1.35 billion) and Microchip–Hailo.

NVIDIA remains the reference platform for data-centre accelerators. Custom chips from Google, Amazon, Meta and OpenAI are gaining ground, as are inference specialists such as Cerebras, Groq, Etched and Rebellions, and China’s domestic GPU makers.

Five shifts define the outlook:

• Inference overtakes training as the dominant workload. Reasoning models and agentic AI multiply the compute used per task.
• Architectures diversify. Mixture-of-experts transformers dominate today. State-space, liquid, diffusion and other post-transformer designs from Liquid AI, Inception, Pathway and others are gaining traction.
• Neural networks move into the physical world. World models and robot foundation models are carried into vehicles, robots and devices.
• New computing substrates begin to commercialise. These include neuromorphic, photonic, analog, thermodynamic and biological computing.
• Value migrates towards models, software and Asia. Sovereign AI programmes in Korea, Japan, India, the Gulf and Europe build national capacity.

There are also serious constraints:

• grid power;
• HBM memory and advanced packaging capacity;
• export controls dividing the US and Chinese markets;
• EU AI Act obligations;
• the gap between infrastructure spending and end-user revenue.

How companies handle memory bandwidth, power, custom silicon and inference economics will decide who captures value through 2040.

The Global Neural Networks Market 2027–2040 is an analysis of the hardware, models, software and applications behind modern AI. It contains more than 160 tables and 28 figures. Detailed forecasts to 2040 are broken down by:

• market segment;
• hardware type;
• training vs. inference;
• cloud vs. edge;
• architecture family;
• end-use sector;
• region.

Bull, base and bear scenarios are included, and the report reconciles its figures with third-party estimates. The report goes beyond market sizing to assess how close each company is to delivering working commercial products. Every one of the 194 profiled companies has a full one-page profile covering:

• company overview;
• neural network technology and products;
• dated developments for 2025–2026;
• market position and outlook.

Each profile also carries a commercial readiness rating on a five-stage scale: Commercial at scale, Commercial – early, Integrating, Transitioning or Development. Alongside it are an estimated time to a working commercial product or model and a statement of the readiness evidence. This lets investors, strategists and procurement teams see at a glance which companies are shipping today, which are piloting with customers, which are in transition after acquisitions or pivots, and which remain years from market. Summary tables and a chart break readiness down across seven company groups. They show that established chipmakers and leading model developers are overwhelmingly commercial, while neuromorphic, photonic, analog and thermodynamic computing developers are mostly still integrating with customers or in development.

The report covers:

• data-centre GPUs and custom ASICs;
• challenger and Chinese accelerators;
• edge neural processors;
• neuromorphic, photonic, analog and biological computing;
• HBM memory, packaging and interconnect;
• foundation and open-weight models;
• post-transformer architectures;
• token economics;
• applications in automotive, healthcare, robotics, defence and finance.

It also includes regional analysis, a full record of funding, M&A and IPOs, regulation and a 2027–2040 technology roadmap.

Report contents include:

• Global market forecasts 2025–2040 by segment, hardware type, workload, deployment, architecture, end use and region, with bull, base and bear scenarios
• Data-centre accelerators: NVIDIA and AMD roadmaps, hyperscaler custom chips, challengers, China’s chip ecosystem, Korean and Japanese developers
• Edge neural processors for smartphones, AI PCs, vehicles, industrial systems and robots, plus edge AI consolidation
• Neuromorphic, photonic, analog and in-memory, thermodynamic, quantum and biological computing
• HBM roadmap, the memory wall, advanced packaging and optical interconnect
• Architectures: CNNs, transformers, mixture of experts, state-space models, liquid, graph and topological networks, spiking networks, world models, quantum and biological networks
• Foundation model developers, valuations, open-weight models, token economics, MLOps and AI memory infrastructure
• Applications: cloud, enterprise and agentic AI, consumer devices, automotive, healthcare, robotics and physical AI, finance, defence, telecoms, science
• Regional analysis: North America, China, South Korea, Japan, Taiwan, Europe, UK, India, Middle East, Russia and Australia
• Funding rounds, M&A, IPOs and strategic compute partnerships for 2024–2026
• Regulation: EU AI Act, US export controls, UK, China, ISO/IEC 42001, FDA–EMA principles, litigation
• Risks: power, memory supply, geopolitics, capital intensity, safety
• Technology commercialisation timeline and ten predictions to 2040
• Commercial readiness assessment of 194 companies, with stage, time to product and evidence
• 194 one-page company profiles including 1X Technologies, Agibot (Zhiyuan Robot), Aidoc, Akhetonics, Alchip Technologies, Aleph Alpha, Alibaba Group (Qwen, T-Head), Alif Semiconductor, Ambarella, AMD, AMI Labs, Ampere Computing, Analog Devices, Anduril Industries, Anthropic, Arago, Arlequin AI, Arm Holdings, Aspinity, Astera Labs, Axelera AI, Axera Semiconductor, Ayar Labs, Baidu (incl. Kunlunxin), Baseten, Biren Technology, Black Forest Labs, Black Sesame Technologies, Blaize, Blumind, BrainChip, Broadcom, ByteDance, Cambricon Technologies, Cartesia, Celestial AI, Cerebras Systems, ChipAgents, Chipmind, CogniFiber, Cohere, CoreWeave, Cortical Labs, Crusoe, d-Matrix, Darktrace, Databricks, DeepSeek, DeepX, EdgeCortix, ElevenLabs, EnCharge AI, Enflame Technology, Engramme, Etched, Extropic, Ferroelectric Memory Company (FMC), Field AI, Figure AI, FinalSpark, Fireworks AI, Fractile and more.....

Table of Contents

1 EXECUTIVE SUMMARY

• 1.1 Report scope and market definition
• 1.2 Key findings
• 1.3 Market size and forecast 2025–2040
• 1.4 Growth drivers and restraints
• 1.5 Technology inflection points 2027–2040
• 1.6 Competitive landscape summary
• 1.7 Investment and deal activity 2025–2026
• 1.8 Strategic outlook and recommendations

2 INTRODUCTION

• 2.1 Defining neural networks
• 2.2 Historical development
• 2.3 Taxonomy of neural network architectures
• 2.4 Training and inference
  • 2.4.1 Hardware requirements of training and inference workloads
• 2.5 Scaling laws and compute demand
  • 2.5.1 The cost-per-token curve
• 2.6 The neural network value chain
• 2.7 Research methodology
  • 2.7.1 Approach
  • 2.7.2 Reconciliation with related Future Markets reports

3 NEURAL NETWORK ARCHITECTURES AND ALGORITHMS

• 3.1 Multilayer perceptrons and feedforward networks
• 3.2 Convolutional neural networks
  • 3.2.1 3D CNNs in medical imaging: JLK
• 3.3 Recurrent networks and LSTMs
• 3.4 Transformers and attention
  • 3.4.1 Self-attention and its cost
  • 3.4.2 Efficiency techniques
• 3.5 Mixture-of-experts models
• 3.6 Post-transformer architectures
  • 3.6.1 Why replace attention
  • 3.6.2 State-space models and hybrids
  • 3.6.3 Venture-backed alternatives
  • 3.6.4 Diffusion language models
• 3.7 Liquid neural networks and continuous-time models
  • 3.7.1 Liquid AI
• 3.8 Graph and topological neural networks
  • 3.8.1 Arlequin AI
• 3.9 Generative architectures: GANs, VAEs, diffusion and world models
  • 3.9.1 Diffusion models
  • 3.9.2 World models
  • 3.9.3 World foundation models as training infrastructure for physical AI
• 3.10 Spiking neural networks
  • 3.10.1 Digital, analog and hybrid spiking implementations
  • 3.10.2 Photonic spiking and reservoir networks
• 3.11 Physics-informed networks, neural operators and KANs
• 3.12 Quantum neural networks
  • 3.12.1 Neural networks for quantum key distribution: WiMi
  • 3.12.2 Hybrid spiking-quantum networks: MicroCloud Hologram DSQ-Net
  • 3.12.3 Wider quantum machine learning context
• 3.13 Biological neural networks and organoid intelligence
  • 3.13.1 Cortical Labs
  • 3.13.2 FinalSpark and other developers
• 3.14 Memory-augmented neural networks
  • 3.14.1 Engramme and Large Memory Models
• 3.15 Model compression, small language models and on-device AI
  • 3.15.1 Distillation
  • 3.15.2 Quantisation, pruning and sparsity
  • 3.15.3 On-device AI
  • 3.15.4 Precision, efficiency and edge compilation toolchains
  • 3.15.5 Precision and sparsity roadmap, 2026–2028
• 3.16 Architecture adoption outlook

4 NEURAL NETWORK HARDWARE

• 4.1 Hardware landscape
  • 4.1.1 Energy efficiency: TOPS/W tiers and benchmarks
• 4.2 GPUs
  • 4.2.1 NVIDIA
  • 4.2.2 AMD
  • 4.2.3 Intel and other GPU suppliers
  • 4.2.4 GPU unit shipments, average selling prices and generational mix
• 4.3 Hyperscaler custom ASICs
  • 4.3.1 Google TPU
  • 4.3.2 AWS Trainium and Inferentia
  • 4.3.3 Microsoft Maia
  • 4.3.4 Meta MTIA
  • 4.3.5 OpenAI and Broadcom
  • 4.3.6 Broadcom and Marvell
  • 4.3.7 Custom accelerator volumes and the design-services layer
• 4.4 Challenger accelerator companies
  • 4.4.1 Efficiency and total cost of ownership: merchant GPUs, custom and start-up ASICs
  • 4.4.2 Wafer-scale yield economics
  • 4.4.3 Development cost as a barrier to entry
• 4.5 China’s accelerator ecosystem
  • 4.5.1 China’s processor market, domestic unit shares and the memory gap
  • 4.5.2 Capability gap by workload
• 4.6 South Korean and Japanese accelerator developers
  • 4.6.1 South Korea
  • 4.6.2 Japan
• 4.7 Edge neural processing units
  • 4.7.1 Edge NPU architectures and dataflow
  • 4.7.2 Edge accelerator performance and efficiency benchmarks
  • 4.7.3 Process technology and the cost of edge NN silicon
  • 4.7.4 Mobile NPU efficiency and ultra-low-power audio and sensing
  • 4.7.5 Robot compute platforms and the two-tier control hierarchy
• 4.8 Open-source neural network hardware
  • 4.8.1 RISC-V in production edge NN silicon
• 4.9 Neuromorphic processors
  • 4.9.1 Neuromorphic drivers, restraints and near-term outlook
• 4.10 Photonic neural networks and optical interconnect
  • 4.10.1 Optical neural network architectures
  • 4.10.2 Performance, efficiency and the precision problem
  • 4.10.3 Optical switching and light sources for AI fabrics
• 4.11 Analog, in-memory and thermodynamic computing
  • 4.11.1 The data-movement problem in numbers
  • 4.11.2 Memory technologies for compute-in-memory
  • 4.11.3 Precision limits and the analog roadmap to
  • 4.11.4 Processing-in-memory and compute-in-memory products from memory suppliers
  • 4.11.5 Emerging non-volatile memory as synaptic devices
• 4.12 Memory, packaging and interconnect
  • 4.12.1 High-bandwidth memory
  • 4.12.2 Advanced packaging
  • 4.12.3 Interconnect
  • 4.12.4 Rack bill of materials, HBM content and packaging capacity
  • 4.12.5 Energy per bit and the case for optics
  • 4.12.6 The memory wall in numbers
  • 4.12.7 HBM unit and revenue outlook
  • 4.12.8 Inference memory hierarchy and the KV cache
• 4.13 Hardware market forecasts
  • 4.13.1 Data-center neural network accelerators
  • 4.13.2 Edge neural processing units
  • 4.13.3 Neuromorphic and novel neural network hardware
  • 4.13.4 Comparison with Future Markets’ low-power AI semiconductor forecast
  • 4.13.5 Related-report context: optical computing forecasts

5 SOFTWARE, MODELS AND PLATFORMS

• 5.1 Frameworks and compilers
  • 5.1.1 The framework layer
  • 5.1.2 Compilers, kernels and the CUDA moat
  • 5.1.3 Inference engines and on-device runtimes
• 5.2 Foundation model developers
  • 5.2.1 Capital concentration at the frontier
  • 5.2.2 Europe, China, Japan and Russia
  • 5.2.3 Architecture challengers and new labs
  • 5.2.4 Implications
• 5.3 Open-weight vs. proprietary models
  • 5.3.1 The shifting balance
  • 5.3.2 Who uses open weights and why
• 5.4 Inference economics and token demand
  • 5.4.1 From training to inference
  • 5.4.2 Token volumes
  • 5.4.3 Reasoning models and test-time compute
  • 5.4.4 Falling unit costs and rising total spend
  • 5.4.5 Cost per token by model class and hardware
  • 5.4.6 Memory footprint of inference workload classes
• 5.5 Platforms, MLOps and inference serving
  • 5.5.1 Market structure
  • 5.5.2 Consolidation and valuations
  • 5.5.3 Inference-serving economics
• 5.6 Training data and labelling
  • 5.6.1 From labelling to expert data
  • 5.6.2 Scale AI, Surge AI and the 2025 realignment
  • 5.6.3 Crowdsourced platforms
  • 5.6.4 Synthetic data
• 5.7 Memory and agent infrastructure
  • 5.7.1 Why memory matters
  • 5.7.2 Engramme
  • 5.7.3 Mem0, Zep, Letta and LangMem
  • 5.7.4 Outlook
• 5.8 Software and model forecasts

6 MARKETS AND APPLICATIONS

• 6.1 Hyperscale cloud and AI labs
  • 6.1.1 Capital expenditure as the primary demand signal
  • 6.1.2 Neoclouds, Oracle and Stargate
  • 6.1.3 AI labs as buyers
  • 6.1.4 Processor demand by data-center facility type
  • 6.1.5 Sovereign AI build-outs
  • 6.1.6 Chinese hyperscaler capital expenditure
• 6.2 Enterprise and agentic AI
  • 6.2.1 From copilots to agents
  • 6.2.2 Platforms and leading suppliers
• 6.3 Consumer devices
  • 6.3.1 The on-device shift
  • 6.3.2 Small models and edge silicon
    • 6.3.2.1 Smartphone and AI PC NN processor benchmarks
• 6.4 Automotive and autonomous mobility
  • 6.4.1 End-to-end neural driving
  • 6.4.2 Compute platforms
  • 6.4.3 Open-source silicon and efficient models
  • 6.4.4 Compute requirements by level of autonomy
  • 6.4.5 Robotaxi operating data and in-vehicle compute content
• 6.5 Healthcare and life sciences
  • 6.5.1 Medical imaging
  • 6.5.2 JLK and 3D convolutional networks
  • 6.5.3 Drug discovery and regulation
• 6.6 Industrial, robotics and physical AI
  • 6.6.1 Robot foundation models
  • 6.6.2 Humanoids
  • 6.6.3 Industrial applications
  • 6.6.4 On-board neural network compute for humanoid robots
  • 6.6.5 Vision-language-action model lineage and specifications
  • 6.6.6 Training paradigms, sim-to-real transfer and remaining technical gaps
  • 6.6.7 Fleet learning and orchestration
  • 6.6.8 Humanoid unit and price forecasts
• 6.7 Financial services
  • 6.7.1 Established and emerging uses
• 6.8 Defence, security and government
  • 6.8.1 A funding surge
  • 6.8.2 Arlequin AI and topological networks
  • 6.8.3 Government and intelligence
• 6.9 Telecommunications and quantum-secure communications
  • 6.9.1 AI-RAN and network operations
  • 6.9.2 Neural networks in quantum key distribution
• 6.10 Science and materials discovery
  • 6.10.1 Structural biology
  • 6.10.2 Materials informatics
  • 6.10.3 Weather and climate
  • 6.10.4 Neural operators and scientific machine learning
• 6.11 End-use market forecasts
  • 6.11.1 Reconciliation with Future Markets’ physical AI forecast

7 GLOBAL MARKET FORECASTS 2025–2040

• 7.1 Forecast approach
• 7.2 Total market
• 7.3 By segment
• 7.4 Hardware vs. software
• 7.5 Training vs. inference
• 7.6 Cloud vs. edge
  • 7.6.1 Comparison with Future Markets’ edge AI chip forecast
• 7.7 By architecture family
• 7.8 By region
• 7.9 Scenario analysis

8 REGIONAL ANALYSIS

• 8.1 North America
  • 8.1.1 Hyperscaler capital expenditure
  • 8.1.2 Stargate and gigawatt-scale campuses
  • 8.1.3 Industrial policy: CHIPS Act and the Intel stake
  • 8.1.4 Export controls and revenue share
  • 8.1.5 Capital markets and challengers
• 8.2 China
  • 8.2.1 Domestic accelerators
  • 8.2.2 The AI chip IPO wave
  • 8.2.3 Models: DeepSeek and the open-weight ecosystem
  • 8.2.4 Procurement policy
  • 8.2.5 Quantum-neural and holographic computing companies
• 8.3 South Korea
  • 8.3.1 Memory: SK hynix and Samsung
  • 8.3.2 GPUs and national infrastructure
  • 8.3.3 Neural processing units: Rebellions and FuriosaAI
  • 8.3.4 Models and applications
• 8.4 Japan
  • 8.4.1 Rapidus and leading-edge manufacturing
  • 8.4.2 Edge AI chips and the JST programme
  • 8.4.3 Model developers
  • 8.4.4 SoftBank
  • 8.4.5 Physical AI as national industrial policy
• 8.5 Taiwan
• 8.6 Europe
  • 8.6.1 Mistral AI and European compute
  • 8.6.2 Start-ups and novel architectures
  • 8.6.3 EU policy: Chips Act 2.0 and the AI Act
• 8.7 United Kingdom
  • 8.7.1 Compute: AIRR and Isambard-AI
  • 8.7.2 Policy and legislation
  • 8.7.3 Companies
• 8.8 India, Middle East, Russia, Australia and rest of world
  • 8.8.1 India
  • 8.8.2 Saudi Arabia and the UAE
  • 8.8.3 Russia
  • 8.8.4 Australia and Singapore

9 COMPETITIVE LANDSCAPE, INVESTMENT AND M&A

• 9.1 Market structure
  • 9.1.1 Data-center accelerator concentration
  • 9.1.2 The rise of custom silicon
  • 9.1.3 China as a separate market
  • 9.1.4 Concentration elsewhere in the stack
• 9.2 Venture funding
  • 9.2.1 Scale and concentration of capital
  • 9.2.2 Hardware and infrastructure rounds
  • 9.2.3 Edge, neuromorphic and novel computing
  • 9.2.4 Post-transformer and memory start-ups
  • 9.2.5 Assessment
  • 9.2.6 Physical AI and robot foundation model funding
• 9.3 M&A and licensing
  • 9.3.1 Deal patterns
  • 9.3.2 Deals that did not happen
• 9.4 IPOs and public markets
  • 9.4.1 The United States
  • 9.4.2 China’s domestic GPU listings
  • 9.4.3 Pipeline and implications
• 9.5 Strategic compute partnerships
  • 9.5.1 Gigawatt-denominated deals
  • 9.5.2 Backlog, sovereign deals and the edge
  • 9.5.3 Circularity and risk
• 9.6 Start-up landscape
  • 9.6.1 Layers of the landscape
  • 9.6.2 Dynamics across the layers

10 REGULATION, STANDARDS AND ETHICS

• 10.1 EU AI Act
  • 10.1.1 The Digital Omnibus
  • 10.1.2 Market implications
• 10.2 US export controls
  • 10.2.1 From prohibition to tightening, 2022–2024
  • 10.2.2 The H20 episode and revenue sharing,
  • 10.2.3 The H200/MI325X rule and 25% duty,
• 10.3 United Kingdom
• 10.4 China
• 10.5 Standards and certification
  • 10.5.1 ISO/IEC
  • 10.5.2 Other standards and frameworks
• 10.6 Sector regulation: healthcare and drug development
• 10.7 Liability, copyright and litigation
  • 10.7.1 Product liability and design
  • 10.7.2 Copyright

11 MARKET CHALLENGES AND RISKS

• 11.1 Energy and power
  • 11.1.1 Rack density, cooling and the AI share of data-center load
  • 11.1.2 Cooling overhead and the system-level efficiency gap
  • 11.1.3 Embodied carbon and resource intensity
• 11.2 Memory and packaging supply constraints
  • 11.2.1 Memory requirements by accelerator generation
  • 11.2.2 Memory cost per AI server and the 2029–2030 correction risk
• 11.3 Geopolitics and supply-chain concentration
  • 11.3.1 Choke points in the neural network hardware supply chain
• 11.4 Capital intensity and valuation risk
  • 11.4.1 Capital intensity ratios and the shift from capex to opex
• 11.5 Talent, data and intellectual property
  • 11.5.1 Chip design capacity and AI-driven EDA
• 11.6 Safety, security and reliability
• 11.7 Risk assessment

12 FUTURE OUTLOOK 2027–2040

• 12.1 Technology roadmap
  • 12.1.1 Three waves of commercialisation
  • 12.1.2 Readiness and commercialisation timeline
  • 12.1.3 Architecture and hardware co-evolution
  • 12.1.4 Energy-efficiency roadmap
  • 12.1.5 Photonic neural network roadmap
  • 12.1.6 Physical AI readiness by application
  • 12.1.7 HBM and memory technology roadmap to
• 12.2 Ten predictions for
• 12.3 Opportunity assessment
• 12.4 Wildcards and discontinuities
  • 12.4.1 Upside discontinuities
  • 12.4.2 Downside discontinuities
  • 12.4.3 Unconventional discontinuities
  • 12.4.4 Physical AI scenario variables

13 COMPANY PROFILES (195 company profiles)

14 APPENDICES

• 14.1 Detailed year-by-year forecasts
• 14.2 Glossary

15 REFERENCES