Broadcom AI Chips Surge as Custom Silicon Challenges NVIDIA

Quick Answer

Broadcom AI revenue $4.2B in Q2 2026 (+60% YoY), $20B+ for FY2026. Customers: Google TPU, Meta MTIA, Microsoft Maia, ByteDance. Custom silicon captures 30-50% cost savings vs NVIDIA. Broadcom also designs Tomahawk/Jericho networking for AI data centers (Broadcom Q2 2026 earnings, 2026).

Data last verified September 2026 from Broadcom Q2 2026 earnings and 10-Q SEC filing.

Broadcom AI revenue growth

Source: Broadcom Q1-Q2 2026 earnings reports.

Google TPU partnership with Broadcom

Broadcom's most important AI partnership is with Google for the Tensor Processing Unit (TPU). Key TPU versions: (1) TPU v1-v4 — earlier generations, (2) TPU v5e (2023) — efficient inference, (3) TPU v5p (2023) — performance for training, (4) TPU v6 (Trillium, 2024) — 2x perf per watt, (5) TPU v7 (Ironwood, 2026) — next-gen for Gemini 2.0. Google is reportedly one of Broadcom's largest customers, accounting for approximately 25-30% of Broadcom's AI revenue (Broadcom, Google, 2026).

Meta MTIA partnership with Broadcom

Meta's Training and Inference Accelerator (MTIA) is designed in partnership with Broadcom. MTIA versions: (1) MTIA v1 (2023) — first-gen inference accelerator, (2) MTIA v2 (2024) — 2x performance, used for Meta's Reels and Feed ranking, (3) MTIA v3 (planned 2026) — third-gen, focus on training workloads. Meta has stated that MTIA is now deployed at scale for inference and will be expanded to training. Broadcom is Meta's primary design partner for MTIA (Meta, 2026).

Microsoft Maia partnership with Broadcom

Microsoft's Maia AI accelerator is also designed in partnership with Broadcom. Maia versions: (1) Maia 100 (2023) — first-gen AI accelerator, deployed for Azure OpenAI inference, (2) Maia 200 (planned 2026) — second-gen, focus on training and inference. Microsoft is using Maia to reduce dependence on NVIDIA for Azure AI workloads, particularly Azure OpenAI Service. Broadcom is Microsoft's primary design partner for Maia (Microsoft, 2026).

Custom AI chip market

The custom AI chip market is growing rapidly: (1) 2024 — ~$15B, (2) 2025 — ~$30B, (3) 2026 — ~$50B (projected), (4) 2027 — ~$80B (projected). The growth is driven by: (1) Hyperscaler demand for cheaper AI compute, (2) Performance optimization for specific workloads, (3) Supply diversification from NVIDIA. Market share leaders: Broadcom (design partner for major hyperscalers), Marvell (custom chips for AWS Trainium), Alchip (Asian hyperscalers), GlobalUnichip (TSMC partner) (Gartner, 2026).

Broadcom vs NVIDIA

Source: company earnings, Gartner (2026).

Why custom silicon is winning for hyperscalers

• Cost: Custom chips can be 30-50% cheaper than NVIDIA GPUs for equivalent AI workloads.
• Performance optimization: Custom chips are optimized for specific AI workloads (training vs inference, specific model architectures).
• Supply diversification: Hyperscalers want to reduce dependence on NVIDIA for supply risk.
• Power efficiency: Custom chips can be more power-efficient than GPUs for specific workloads.
• Vertical integration: Hyperscalers want to own their AI hardware stack for differentiation.

Broadcom's networking chip business

Broadcom's networking chip business is a critical enabler of AI data centers. Broadcom designs: (1) Tomahawk 6 — 102.4Tbps Ethernet switches for AI data center fabrics, (2) Tomahawk 5 — 51.2Tbps switches, (3) Jericho 3 — deep-buffer switches for large AI clusters, (4) Silicon Photonics — optical interconnect for high-bandwidth AI networks, (5) NICs — 400G and 800G Ethernet adapters for AI servers. Broadcom networking chips are used in nearly every major AI data center (Broadcom, 2026).

Broadcom's stock and valuation

Broadcom (AVGO) is one of the best-performing AI stocks in 2026: (1) Stock price up 70%+ YTD, (2) Market cap $1.2T+ (now a top 10 US company), (3) Forward P/E ~40x (premium due to AI growth), (4) Stock split 10-for-1 in 2024, (5) Dividend yield ~1.2%. Investors see Broadcom as a less expensive way to invest in the AI custom silicon trend than NVIDIA (Bloomberg, 2026).

What this means for the AI industry

Broadcom's surging AI revenue signals: (1) Custom silicon is a major trend in AI, (2) Hyperscalers want to reduce dependence on NVIDIA, (3) Broadcom is the largest beneficiary of the custom silicon trend, (4) NVIDIA will face increasing competition from custom chips in the hyperscaler segment, (5) The AI chip market is large enough to support multiple winners, not just NVIDIA. Broadcom's success is a positive signal for the broader AI infrastructure buildout (Broadcom, 2026).

Other custom AI chip makers

• Marvell Technology: Custom chips for AWS Trainium, networking chips for AI data centers.
• Alchip Technologies: Custom chips for Asian hyperscalers (Alibaba, Tencent, Baidu).
• GlobalUnichip: TSMC partner for custom AI chip design.
• Tenstorrent (Jim Keller): Open-source RISC-V AI accelerator.
• Cerebras Systems: Wafer-scale AI accelerator for training.
• SambaNova Systems: Reconfigurable dataflow AI accelerator.
• Groq: LPU (Language Processing Unit) for inference.
• Etched.ai: Transformer-specific ASIC.

Resources and next steps

Follow Broadcom investor relations at investors.broadcom.com. For AI chip industry analysis, follow Gartner, IDC, and Forrester. The next Broadcom earnings are expected December 2026. For AI infrastructure investment trends, follow Bank of America, Goldman Sachs, and Morgan Stanley. For Broadcom customer news, follow Google, Meta, and Microsoft quarterly earnings.