Alphabet Raises 2026 Capex to $195–205B, Boosting AI Chip Suppliers

Alphabet announced during its Q2 earnings report that it would increase its 2026 capital expenditure budget to between $195 billion（約31兆円） and $205 billion（約33兆円） for building out AI data centers, up from previous guidance of between $180 billion（約29兆円） and $190 billion（約30兆円）. The company also plans to significantly increase capex in the year after that. This spending will flow through a specific ecosystem of AI infrastructure suppliers, each playing a distinct role in Alphabet's custom-chip strategy.

Broadcom stands at the center of this ecosystem as the co-developer of Alphabet's tensor processing units (TPUs), which are designed for specific AI tasks. Broadcom manages the physical design and supply chain for these custom application-specific integrated circuits (ASICs) and records the revenue when they are delivered. The company also operates a large networking business that will benefit from Alphabet's data-center infrastructure spending. Celestica, Alphabet's main hardware integration partner, manufactures custom printed circuit boards and builds the physical server enclosures for Alphabet's systems; it also connects server trays into racks and provides network switches. Alphabet is believed to be Celestica's largest customer, representing about 28% of the company's total revenue in Q1, with the top three hyperscaler customers making up more than 50% of Celestica's revenue overall.

Memory and optical components round out the supply chain. Samsung is the primary supplier of high-bandwidth memory (HBM)—a specialized form of DRAM packaged with chips to optimize performance—and supplies more than 60% of Alphabet's HBM needs for its TPUs. SK Hynix supplies the remainder of Alphabet's HBM and is also the main HBM supplier for Nvidia's GPUs. Micron, though not a major Alphabet HBM supplier, likely provides other solutions such as high-density enterprise solid storage devices and DDR5 server DRAM. Lumentum benefits from Alphabet's use of direct optical circuit switches in its TPU architecture, which keep data in light form to reduce power and latency; Lumentum is one of only two global suppliers capable of producing the high-speed 200-gigabit laser chips that generate these light signals at scale. Taiwan Semiconductor Manufacturing, which has a near monopoly on advanced logic chip manufacturing, produces the vast majority of Alphabet's TPUs (despite the company's reported 3 million TPU order from Intel) and also provides the chip-on-wafer-on-substrate advanced packaging that bonds HBM to the TPU die. Nvidia also benefits from Alphabet's capex increase, as the company continues to buy Nvidia GPUs for research, custom kernels, sparse models, and other tasks where GPUs outperform TPUs.

Alphabet's increased capex guidance reflects the company's commitment to building out proprietary AI infrastructure, moving away from reliance on off-the-shelf components toward custom silicon and integrated systems. The shift to custom tensor processing units (TPUs) with direct optical circuit switches—which keep data in light form to reduce power consumption and latency—signals a fundamental architectural difference from traditional data-center design. This strategy creates a distinct ecosystem of suppliers optimized for Alphabet's specific needs: Broadcom handles the physical design and supply chain of the TPUs themselves; Celestica integrates the hardware; specialized memory makers (Samsung and SK Hynix) provide the high-bandwidth memory that bonds to the chips; and Taiwan Semiconductor Manufacturing provides advanced packaging and foundry services. The article also notes that while Alphabet has reportedly ordered 3 million TPUs from Intel, Taiwan Semiconductor Manufacturing still manufactures the vast majority of Alphabet's TPUs, underscoring the Taiwanese foundry's critical role. Because Alphabet's capex increase likely signals broader hyperscaler demand—the article suggests other companies may follow suit—these suppliers stand to see sustained revenue growth beyond Alphabet alone.

No comments yet. Be the first to share your thoughts!