AI Compute Infrastructure Enters the Physical Layer Breakthrough Phase: TPU Foundry Shifts and Domestic Optical Communications Reshape the Global Semiconductor Supply Chain

AI Compute Infrastructure Enters the “Physical-Layer Breakthrough Phase”: TPU Foundry Deployment and Domestication of Optical Interconnects Reshape the Global Semiconductor Value Chain

When Google quietly placed an order with Intel for over 3 million custom-designed TPUs (scheduled for delivery in 2028), and when Amazon signed a multi-billion-dollar optical fiber manufacturing agreement with Corning to bolster domestic U.S. optical interconnect capabilities, a long-underestimated inflection point has clearly emerged: The competitive focus of the AI industry is rapidly shifting downward—from the “software narrative” of algorithms and models—to the “physical-layer infrastructure” composed of chips, optical modules, advanced packaging, and specialty materials. This is not hype; it represents a fundamental reallocation of capital expenditure. According to Goldman Sachs’ latest estimates, global AI-related hardware capex will reach a record $580 billion between 2024 and 2026—with nearly 65% concentrated on the physical deployment of data center infrastructure. The market reaction triggered by this chain of events—Intel surging 11% in a single day, Corning rising nearly 10% pre-market, and Wolfspeed jumping 13% on its SiC aviation collaboration—is no isolated rally. Rather, it signals a clear, medium-to-long-term reassessment of underlying industry logic.

Foundry Model Breakthrough: x86 Giants Enter the Core AI Chip Supply Chain

Google’s decision to entrust Intel with TPU manufacturing carries strategic significance far beyond a single order. Historically, TPUs—as Google’s in-house AI accelerators—have been fabricated exclusively by pure-play foundries such as TSMC. Although Intel has pursued its IDM 2.0 strategy, its foundry services (IFS) have long faced dual challenges: client trust deficits and process maturity gaps. This collaboration signifies three pivotal developments:

First, Intel has now validated its advanced capabilities—including Gate-All-Around (GAA) transistor architecture, EMIB 2.5D packaging, and Foveros Direct 3D stacking—earning recognition from leading cloud providers for both cutting-edge process technology and heterogeneous integration.

Second, custom TPU foundry engagement is, at its core, a deep “IP + manufacturing + system-level co-development” partnership. Google will open up portions of its microarchitecture instruction set and memory subsystem design specifications, elevating Intel from a mere “foundry partner” to a “co-definer” of the silicon.

Third, the scale of 3 million units—estimated at 256 GB HBM3 bandwidth and 1,000 TOPS INT8 performance per chip, yielding aggregate compute capacity exceeding 300 ExaOPS—foreshadows the widespread adoption of “thousand-GPU liquid-cooled clusters” in 2027–2028. This era will impose entirely new physical constraints on thermal management, power delivery, and interconnect design.

Notably, this move resonates technologically with Intel’s recent launches of the “Arrow Lake” client CPU and the “Granite Rapids” server CPU—all three sharing the Intel 18A process node and a unified chiplet interconnect standard (UCIe). In essence, Intel is building a full-stack physical-layer foundation spanning AI training (TPUs), inference (Xeon CPUs + GPUs), and endpoints (PCs/edge devices). The market’s 11% one-day surge in Intel shares reflects collective pricing of its IDM 2.0 strategy transitioning from “technical feasibility” to “commercial closure.”

Domestication of Optical Interconnects: A Paradigm Shift from “Bandwidth Anxiety” to “Manufacturing Sovereignty”

If TPU foundry deployment resolves the generation of AI compute, then Amazon’s multi-billion-dollar fiber agreement with Corning directly targets the critical bottleneck in AI compute distribution: optical interconnects. In today’s top-tier AI clusters, inter-GPU communication bandwidth demands have reached several terabytes per second—while traditional copper cables suffer severe signal attenuation beyond ~3 meters. Silicon photonics and high-density multimode fiber represent the only viable solutions. Yet their core materials (e.g., Corning’s SMF-28® Ultra fiber, low-loss grating arrays) and packaging equipment remain heavily dependent on overseas supply chains. A 2023 U.S. Department of Commerce report revealed that domestic optical interconnect component self-sufficiency stands below 35%, with pronounced gaps—especially in InP laser chips and silicon photonic modulators required for 800G/1.6T optical modules.

The pivotal breakthrough embedded in this agreement lies in its unprecedented vertical integration depth: Corning will not only supply optical fiber preforms but also co-build domestic U.S. optical module packaging and test facilities with Amazon—and jointly develop next-generation “bend-insensitive fiber” optimized for ultra-dense cabling inside liquid-cooled racks. This initiative directly activates two critical upstream value chains: (i) specialty glass and rare-earth-doped materials (e.g., purification of erbium and ytterbium); and (ii) midstream opto-chip packaging and test equipment (e.g., ASM Pacific’s die bonders, Veeco’s MOCVD systems). Lumentum’s 4.2% pre-market gain and II-VI’s (now Coherent) concurrent strength confirm that market consensus is crystallizing around a valuation premium for “manufacturing sovereignty” in optical communications.

Hardware Stack Revaluation: A Three-Dimensional Resonance Across Equipment, Materials, and Advanced Packaging

This broad-based hardware-sector rally is no coincidence. The Nasdaq Semiconductor Index rebounded 6.1% in a single day—with equipment makers (Applied Materials +7.6%, ASML ADR +6.3%), materials suppliers (Wolfspeed +13%), and OSATs (Amkor, JCET-related stocks) all rising in unison—revealing the complete capex transmission pathway:

• Equipment: AI chips demand higher precision in EUV multi-patterning and atomic layer deposition (ALD). Applied Materials’ Producer platform orders are already booked through Q2 2025.
• Materials: SiC substrates are seeing rapid penetration into AI power management (e.g., GPU voltage regulator modules—VRMs). Wolfspeed’s collaboration with GE Aviation serves to validate high-voltage, high-current reliability—a capability directly transferable to AI server 48V power architectures.
• Packaging: TSMC’s CoWoS capacity utilization remains persistently above 120%, spurring ASE and SPIL to accelerate ABF substrate line expansions—while Intel’s EMIB technology emerges as a compelling alternative gaining traction.

Critically, easing inflation expectations—New York Fed’s May one-year inflation forecast dropped to 3.46%—provide a macro foundation for tech stock valuation recovery. As real-rate pressure recedes, investors grow more willing to pay premiums for capital expenditures with high certainty. This is precisely the core logic behind semiconductor equipment and materials outperforming software names (e.g., Microsoft and Adobe declined).

Physical-Layer Infrastructure: A Structural Investment Theme That Transcends Cycles

Looking back, the true cyclical survivors following the dot-com bust were physical-layer infrastructure vendors: Cisco (routers), Emulex (HBAs), Seagate (hard drives). Today’s AI wave follows a similar pattern: As large-model parameter counts approach physical limits (e.g., training energy consumption exceeding 100 megawatts), industry evolution inevitably converges toward more efficient transistors, lower-loss optical pathways, and more compact 3D packaging. Google’s TPU foundry deal and Amazon’s fiber agreement are concrete declarations of this convergence.

For investors, it is imperative to discard the outdated “AI = software” paradigm and adopt the new framework: “AI = silicon-based physical systems.” The hardware stack revaluation is not a short-term trading theme—it is a decade-scale structural opportunity driven jointly by Moore’s Law deceleration, AI’s power-wall constraints, and geopolitical supply-chain restructuring. When 3 million TPUs light up on Intel’s fabs, and when Corning’s optical fibers are drawn into cable on American soil, we are witnessing not merely order fulfillment—but the foundational ceremony of a new infrastructure era.