Google’s AI Chip Gambit: Marvell Joins Broadcom in TPU Supply Push as Inference Costs Surge

Alphabet Inc.’s Google is negotiating with Marvell Technology Inc. to design two new AI chips—a memory processing unit to pair with its Tensor Processing Units and an inference-optimized TPU. No contract yet. But the move signals Google’s drive to spread risk across suppliers. Days earlier, it extended its Broadcom deal through 2031. The Information broke the story on April 19, 2026.

Inference now dominates AI expenses. Training frontier models happens once, demanding massive compute for weeks. Serving queries? That’s constant, scaling with billions of users. Google handles AI-augmented searches, Gemini chats, Cloud APIs daily. Custom silicon here crushes GPUs on efficiency and price. Ironwood, Google’s seventh-gen TPU, launched this month. Ten times the peak performance of TPU v5p. Scales to 9,216 liquid-cooled chips per superpod—10 megawatts, 42.5 FP8 exaflops. Millions coming in 2026. The Next Web detailed the specs.

Supply Chain Diversification Locks In Flexibility

Broadcom handles high-end TPUs. MediaTek does cost-optimized ‘e’ variants, 20-30% cheaper. TSMC fabs them all. Marvell would design, like MediaTek on Ironwood. Not replacement. Addition. Back in 2023, Google eyed ditching Broadcom via ‘Granite Redux’ with Marvell, projecting billions in savings. Didn’t happen. A Google spokesperson then called Broadcom “an excellent partner.” Now? Multi-vendor setup. Each tackles niches. No single firm dictates terms. Reuters confirmed the talks, citing sources.

Marvell’s custom silicon runs $1.5 billion annually across 18 cloud wins: Amazon’s Trainium, Microsoft’s Maia, Meta’s data unit, Google’s Axion ARM CPU. Fiscal year to February 2026: data center revenue hit $6.1 billion, total $8.2 billion—up 42%. Nvidia dropped $2 billion in March 2026 for NVLink Fusion, blending Marvell chips with its fabric. Marvell grabbed Celestial AI for up to $5.5 billion in December 2025, adding photonics. CEO Matt Murphy eyes 20% custom AI market share, 30% revenue growth in fiscal 2027. Stock up 50% YTD, 30% in April alone. Barclays’ Tom O’Malley upgraded to overweight, $150 target.

Broadcom? Unfazed. Over 70% custom AI accelerator share. AI revenue $8.4 billion last quarter, up 106%. Next: $10.7 billion. Targets $100 billion by 2027. Google-Anthropic ties: $21 billion in 2026, $42 billion in 2027, per Mizuho. Anthropic gets 3.5 gigawatts of next-gen TPU compute from 2027. Shares jumped 6% on the Google extension. Wall Street Journal covered Broadcom’s April 6 announcement.

Market Boom Fuels the Rush

Custom ASICs grow 45% in 2026, hitting $118 billion by 2033—outracing GPUs at 16%. Counterpoint sees Broadcom at 60% by 2027, Marvell 25%. Google serves billions of inferences daily. A few percent off per-query cost? Billions saved yearly. That’s the prize. Development takes years. But direction’s set.

Hyperscalers can’t bet on one supplier. Pricing risk. Supply squeezes. Strategic traps. Google’s four partners—Broadcom, MediaTek, Marvell, TSMC—plus in-house design cover training, inference, cloud compute. Complexity by intent. Marvell wins validation as No. 2 designer. Google gets options for the world’s toughest inference loads. And the chip race accelerates.