Apple is developing its first self-designed AI server chip, Baltra, with Broadcom. The chip is planned for mass production in the second half of 2026 and deployment in Apple’s own data centers in 2027. It uses TSMC’s 3nm N3E process and a Chiplet architecture, targeting AI inference workloads. The move reduces Apple’s reliance on Nvidia and supports Apple Intelligence cloud infrastructure.
Apple and Broadcom signed a multi-billion-dollar custom ASIC deal in December 2024 and renewed it in July 2026, extending the partnership to 2031. Total deal value is estimated at over $30 billion. With Apple joining, Google, Microsoft, Amazon, Meta, OpenAI, and Apple are all now competing in ASIC servers.
The push toward custom silicon is driven by cost and performance. Nvidia’s high-end AI accelerators sell for about $30,000 each**, hold nearly **90% market share**, and carry roughly **75% margins**. JPMorgan estimates Google’s new **TPU Ironwood**, built with Broadcom, offers FP8 performance close to Nvidia’s B200 and B300 at an estimated **$13,000. Cloud providers’ in-house chip purchase costs are about 40% of Nvidia/AMD GPU retail prices, with rental costs at 60–70% of GPU rentals. As AI shifts from training to inference, ASIC advantages grow. About 150 companies are developing over 200 AI semiconductor designs.
Major ASIC products are already commercializing. Google TPU share rose from 71% in 2023 to 74% in 2024, with TPUs expected to account for nearly 78% of Google’s AI server shipments in 2026. Microsoft Maia 200 uses TSMC 3nm, supports up to 6,144 chips via Ethernet, and delivers about 30% better performance at equal cost. Amazon Trainium2 shipments grew over 200% in 2024, with another doubling in 2025; Trainium3 launches in December 2025 with mass production expected in Q1 2026. Meta MTIA moves to Samsung foundry for its third generation in a project worth about $7.3 billion, targeting a 5GW data center by 2030. OpenAI is also working with Broadcom on the Jalapeño inference chip.
Capex is surging. TrendForce says the top eight cloud providers’ combined capex exceeded $420 billion in 2025**, up **61% year over year**, and could top **$710 billion in 2026. Global AI server shipments are expected to grow over 28% in 2026. GPUs still lead at 69.7%, but ASIC share rises to 27.8%. Custom ASIC server shipment growth could reach 44.6%, about three times GPU server growth.
JPMorgan forecasts ASIC shipments of 6.8 million units in 2026, or 42% of AI chip shipments, rising to 12.5 million in 2027, or 53%, surpassing GPU shipments of 10.9 million. Morgan Stanley sees the AI ASIC market growing from $12 billion in 2024 to $30 billion in 2027. Marvell projects growth from $6.6 billion in 2023 to $55.4 billion in 2028, a 53% CAGR. Broadcom sees a $60–90 billion custom AI chip market by 2027.
Broadcom holds 80–85% of the high-end ASIC market and expects AI revenue of $56 billion in FY2026**, potentially over **$100 billion in FY2027. TSMC is the main manufacturing beneficiary. Nvidia is responding with a $3.5 billion investment in MediaTek and ecosystem moves.
For cloud providers, ASICs complement rather than replace GPUs: training uses general-purpose GPUs, while inference shifts to custom ASICs. Apple’s entry deepens the AI infrastructure rebuild around cost and energy efficiency. The second half of 2026 will be key, with Baltra production, Google Ironwood ramp, and Amazon Trainium3 capacity growth clarifying the ASIC vs. GPU landscape.
ICgoodFind: Leading cloud providers are all-in on AI ASICs, and custom compute is rising fast. ASIC and GPU share competition is intensifying, creating new opportunities across chip design and advanced foundry supply chains.
