4. AI Hardware & Chips

From NVIDIA's dominance to the custom silicon revolution — the hardware powering the AI era.

~75%
NVIDIA market share 2026
$193.7B
NVIDIA DC revenue FY2026
62%
SK Hynix HBM share
3.6M
Blackwell backlog (units)

On This Page

NVIDIA Dominance

GPU Lineup & Roadmap

GenerationChipHBMKey SpecTDPStatus
HopperH10080 GB HBM3~3,958 TFLOPS FP8700WShipping (mature)
HopperH200141 GB HBM3e4.89 TB/s bandwidth700WShipping
BlackwellB200192 GB HBM3e~20 PFLOPS FP4 (5x H100)1000-1200WShipping Q1 2025
BlackwellGB200384 GB (2x B200)Grace CPU + 2x B2002700W totalSold out through mid-2026
RubinR100288 GB HBM43.6 ExaFLOPS FP4 (NVL144)TBDH2 2026 (TSMC 3nm)
FutureRubin UltraHBM4TBDTBD2027
FutureFeynmanNext-genTBDTBD2028

CUDA Moat

NVIDIA-Groq Acquisition

NVIDIA acquired Groq's assets for ~$20 billion in December 2025 (its largest deal ever). Structured as a "non-exclusive licensing agreement." GroqCloud continues operating independently.

AMD Competition

ChipMemoryTDPStatus
MI300X192 GB HBM3750WShipping
MI325X256 GB HBM3e1000WShipping Q4 2024
MI350/MI355XHBM3eTBD2025 (35x inference perf vs MI300)
MI400HBM4TBD2026

Custom Silicon (Hyperscaler Chips)

ChipOwnerPerformanceStatus
TPU Ironwood (v7)Google4.6 PFLOPS FP8/chip, 7.7 TFLOPS/wattShipping 2025. Superpods: 9,216 chips = 42.5 ExaFLOPS at 10MW
TPU Trillium (v6e)Google4.7x over v5eGA, 100k+ chip deployments
Trainium3Amazon/AWS2.52 PFLOPS FP8, 144 GB HBM3e2025 (TSMC 3nm). $10B+ annual run-rate
Maia 200Microsoft3x FP4 over Trainium3 (claimed)TSMC 3nm, 140B+ transistors
MTIA v3-v5MetaUp to 25x compute gains (RISC-V)2026 (TSMC N3, HBM3e). Inference-first, training to follow
BaltraAppleCo-designed with BroadcomMass production H2 2026. DC construction 2027

Custom ASICs: 50-70% lower cost per billion tokens vs H100 for training. Anthropic secured access to up to 1M Google TPU chips.

AI Chip Startups

CompanyApproachStatusValuation/Funding
CerebrasWafer-scale (WSE-3: 4T transistors, 900K cores)Re-filing for Q2 2026 IPO$23B valuation, $1B raised
GroqLPU (deterministic inference)Acquired by NVIDIA for $20BGroqCloud continues independently
SambaNovaReconfigurable dataflow (SN50 chip)Intel acquisition fell through; $350M Series EMulti-year Intel partnership
GraphcoreIPUAcquired by SoftBank (July 2024)Part of SoftBank AI strategy
TenstorrentRISC-V + AI (Jim Keller)Active across DC, auto, robotics, edge$6.9B valuation, $1B+ raised
EtchedTransformer-only ASIC (Sohu)TSMC-supported, HBM3e$5B valuation, $620M raised
d-MatrixDigital in-memory compute (Corsair/Raptor)Raptor: first 3D-stacked DRAM accelerator$2B valuation, $450M raised

Networking: InfiniBand vs Ethernet

TechnologyLatencySpeedBest For
InfiniBand~1-2 µsQuantum-X800: 800Gbps x 144 ports = 115.2 TbpsLarge-scale training. NVIDIA-controlled
NVIDIA Spectrum-XLow800G Ethernet. 760% YoY revenue growth to $1.46BEthernet for AI. Customers: Meta, Oracle
UEC 1.0Near-IBRebuilt from ground up (not just RoCE v2)Open standard. AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, Microsoft

Trend: Ethernet gaining ground. InfiniBand retains edge for largest training runs. By mid-2025, Ethernet leads for many AI back-end deployments.

Memory & Storage

HBM Market Share (Q2 2025)

ManufacturerShareStatus
SK Hynix62%All 2026 supply sold out
Micron21%Meeting only 55-60% of demand
Samsung17%Catching up, ramping HBM4

Supply Crisis

Server OEMs

RankOEMRevenue/ShareNotes
1Dell$12.5B, 20% shareLeading at neoclouds (CoreWeave, Tesla, xAI)
2Supermicro$11.7B, ~9.5%Speed-to-market, liquid cooling expertise
3HPE15% shareEnterprise-focused
4Inspur12% shareDominant in China
5Lenovo11% shareGrowing AI server portfolio

Enterprise AI server market: $245B in 2025, projected $524B by 2030. ODMs (Quanta, Wiwynn, FII) handle bulk of hyperscale volume at ~2-3% margins.

Power per Chip

ChipTDPPerformancePerf/Watt Trend
H100700W~3,958 TFLOPS FP8Baseline
B200 (full-spec)1000-1200W~5x H100 FP4~3-4x improvement
GB200 Superchip2700WMassive training throughputBest perf/watt at scale
Google Ironwood600W4.6 PFLOPS FP87.7 TFLOPS/watt (2x Trillium)
AMD MI300X750WCompetitive with H100Similar

Each generation delivers ~2-3x better perf/watt. GB200 NVL72 rack: 120kW, requires liquid cooling (20 L/min). See Power: Rack Density.

Supply Chain

TSMC Dependency & CoWoS Bottleneck

TSMC manufactures virtually all leading-edge AI chips. CoWoS advanced packaging is the critical bottleneck.

PeriodMonthly CapacityStatus
Late 2024~35,000 wafers/monthFully booked
End 2025~65,000 (doubled)Still sold out (demand up 113% YoY)
End 2026~130,000 (target)NVIDIA has 50%+ booked

Geopolitical Risks

Key Sources