Summary of Key Points
Google has expanded its partnerships for its custom AI chips, TPU, to include new players such as Marvell (and potentially AMD), gradually reducing its reliance on Broadcom. The collaboration between Silicon Valley AI companies and chip manufacturers is becoming more diversified, often involving complex stakeholder networks through equity arrangements. Google's TPU shipments are expected to increase by 70%-80% this year, accelerating its efforts to move away from a sole dependence on NVIDIA GPUs. Broadcom faces short-term pressure due to the loss of some orders, while NVIDIA maintains its competitiveness through technological advantages.
1. Google Changes TPU Partners: Marvell Takes Center Stage, Broadcom Faces a Setback
Google previously relied primarily on Broadcom for its custom AI chips (TPUs), but it has now signed a custom agreement with Marvell, covering core components such as AI inference accelerators, storage/network/memory controllers. This news caused Marvell's stock price to rise by nearly 10% and Broadcom's to fall by 4.6%. There are also rumors that Google may invite AMD to participate in the development of its tenth-generation TPU.
Why the Change in Partners?
Google doesn't want to be dependent on a single supplier; having multiple partners helps diversify risks and ensure it obtains chips that better meet its needs, particularly with a focus on efficiency rather than just computational power.
2. The Complex Equity Arrangements in Silicon Valley: Deep Interconnections
This collaboration goes beyond mere chip purchases; it includes equity incentives:
- Marvell has issued warrants to Google, allowing it to buy up to 58.97 million Marvell shares for $206.58 per share (with a potential value of $12.18 billion), but only a portion of the shares will be unlocked if Marvell earns $500 million in custom revenue from Google each year.
- AMD previously granted similar rights to Meta: Meta could buy 10% of AMD's shares for $0.01 per share. NVIDIA invested $100 billion in OpenAI, Google invested $40 billion in Anthropic, and AMD invested $50 billion in Anthropic; in addition, NVIDIA also invested $2 billion in Marvell.
The Purpose?
These arrangements ensure a mutual dependency: chip manufacturers have stable orders, and AI companies have a reliable supply of chips, making it difficult for either party to leave the partnership.
3. Google TPU Shipments on the Rise: A 70%-80% Increase This Year
According to TrendForce Research, Google's shipments of custom ASICs (i.e., TPU) are expected to increase by 70%-80% this year, accounting for about 70% of its AI server sales.
Why the Surge in Demand?
- Google deploys AI servers quickly, and its TPs are updated annually, with the eighth-generation being 2.8 times more powerful than the seventh-generation while also being more energy-efficient.
- Google has a large customer base, using these chips internally and supplying them to companies like Anthropic (with orders in the billions).
- The supply chain is supportive: Broadcom's share of TSMC's advanced CoWoS packaging capacity has increased from 13% last year to 22%, indicating strong demand.
4. Google's Self-Reliance on Custom Chips: Avoiding Dependence on NVIDIA
Google previously relied on NVIDIA GPUs but aims to reduce this dependency:
- NVIDIA's GPUs focus on overall computational power, while Google is more concerned with the efficiency of content generation (e.g., the speed and cost of AI-generated text or images) and the ability to scale clusters for large-scale tasks.
- By developing its own TPs, Google gains control: It can use them internally and sell them to other AI companies like Anthropic, without being at the mercy of others.
5. Broadcom's Challenges and NVIDIA's Response
- Broadcom’s Pressure: After losing some Google orders, Broadcom's AI revenue forecast for the next quarter ($16 billion) fell short of market expectations, and it didn't raise its target for 2027 to $10 billion.
- NVIDIA’s Countermove: NVIDIA has launched the Vera Rubin platform, promising the highest performance per watt and the lowest token costs. Its chips are already in use on platforms like Google Cloud and Microsoft Azure, allowing NVIDIA to maintain its competitive edge through technology.
By breaking down these developments, even those unfamiliar with financial terms can understand Google's strategic moves in the AI chip industry, the interdependencies among Silicon Valley companies, and the underlying dynamics of market competition.