The rapid acceleration of Alphabet's custom silicon roadmap presents a structural reassessment of competitive dynamics in AI hardware. At its core, Google's shift toward externally commercializing Tensor Processing Units via partners like Broadcom reveals a company executing vertical integration across the entire AI stack—from chip design through data center construction, power generation, and model co-optimization. For NVIDIA, this evolution poses both immediate tailwinds and formidable long-term headwinds.
The Architecture of Google's Custom Silicon Dominance
Google has deployed its 7th-generation Ironwood TPU at scale, with over 100,000 units now operational 24. The Ironwood architecture scales to 9,216-chip liquid-cooled pods, delivering 4,614 TFLOPs FP8 per chip and packing 192 GiB of HBM with 7,380 GB/s bandwidth per chip 12,16,25. Manufactured on TSMC's 3nm process, Ironwood represents the kind of foundational productive asset that, in an earlier industrial age, would have been compared to a steel mill—a concentrated locus of computational capacity that can be leveraged across the entire value chain.
The roadmap ahead is equally telling. Google's 8th-generation TPU, split into training (8t) and inference (8i) variants, promises up to 2x better performance-per-watt versus Ironwood 11, with the inference variant (8i) entering production in Q3 2026 32. But perhaps more revealing is Google's architectural decision for the next phase: the company has selected MediaTek—not Broadcom or Qualcomm—to design its TPUv9, codenamed "Triggerfish" 6,27. This choice signals deliberate diversification of the custom silicon supply chain, reducing single-vendor dependency and introducing competitive pressure across the design ecosystem. Google is further splitting next-generation AI hardware production across three distinct partners: Broadcom for training accelerators, MediaTek for inference silicon, and Marvell for memory integration 26. This is the logic of an industrial trust—controlling the critical nodes of the value chain while maintaining multiple suppliers to preserve optionality and competitive discipline.
The Paradox of Massive GPU Commitments
Yet this narrative of custom silicon dominance coexists with a seemingly contradictory reality: Google remains an extraordinarily large NVIDIA customer. Alphabet's aggregate GPU commitment across various contracts may reach 435,000 units 35—a volume that underscores the sheer scale of AI compute demand in this era. More striking still, Google has entered into a GPU supply agreement with SpaceX valued at approximately $920 million per month (roughly 92 billion yen), with a ramp-up period concluding in September 2026 and a contractual termination right if full access is not provided by that deadline 20,21,22,35. Google Cloud has simultaneously expanded its hardware compatibility matrix to include NVIDIA Hopper GPUs and RTX PRO 6000 Blackwell GPUs 18, while Confidential Space now supports NVIDIA Hopper compute 5,10,18.
This apparent contradiction resolves when one recognizes that demand for AI compute has vastly outstripped supply across all accelerator types. GPU utilization across Alphabet, Meta, and Amazon infrastructure is effectively 100% 40—a condition that rarely persists in mature competitive markets. In NVIDIA's case, this means that even as Google builds custom silicon capacity at scale, the total addressable market for GPU compute remains insatiable. The two technologies coexist not because GPU and TPU are substitutes in some theoretical sense, but because aggregate demand has grown faster than supply for both.
This demand overflow has created a visible architecture of rationing. Alphabet's 2026 compute capacity is fully allocated, with internal researchers queued behind paying Google Cloud customers for TPU access 35. Google has reportedly limited Meta's access to Gemini AI capacity due to compute demand exceeding available supply—a claim corroborated by multiple sources 3,4,8,13,14,15. In such an environment, NVIDIA's GPU supply becomes a complement rather than a substitute: when TPU capacity is exhausted, customers migrate to NVIDIA GPUs, and both accelerator makers prosper from sheer scarcity.
The Ecosystem as a Competitive Moat
The strategic significance of Google's TPU commercialization emerges most clearly in the third-party customer landscape. Anthropic, the Claude-maker now backed by Salesforce and Amazon, has expanded its Google Cloud TPU usage to up to 1 million units and over 1 GW of capacity 1,16, with an additional 1–2 GW contracted for in June 2026 9. Anthropic's total compute commitment as of June 2026 is approximately 10.8 GW 30. Google and Broadcom will jointly provide multiple gigawatts of next-generation TPU capacity to Anthropic starting in 2027 37. Similarly, Apple is utilizing Google Cloud TPUs for large-scale AI model pretraining and Private Cloud Compute workloads 18,19,23,33.
This ecosystem expansion is the move of a company that has discovered a new revenue stream while simultaneously entrenching its infrastructure moat. Each new customer locked into TPU-optimized workloads becomes less price-sensitive to NVIDIA GPU offerings, because workload-specific performance asymmetries create real switching costs. Models optimized for NVIDIA Hopper-class GPUs may underperform on TPUs, and vice versa 16. Google TPUs and Amazon Trainium chips demonstrate superior performance-per-watt compared to general-purpose GPUs on optimized workloads 31. While OpenAI's Jalapeño inference chip claims comparable performance-per-watt to both NVIDIA Blackwell and Google TPUs 29, the proliferation of workload-specific accelerators reinforces the broader industry insight: no single architecture will dominate all workloads, and fragmentation becomes a feature that preserves competitive diversity while eroding any single vendor's pricing power.
The Broadcom Relationship and Its Limits
Broadcom's role in this ecosystem warrants careful attention. The company's long-term agreement with Google for TPU production runs through 2031, with cumulative revenue potentially exceeding $500 billion 32. Broadcom's TPU shipments are forecast to reach 7 million units annually by 2028 38—a volume that would make the company a substantial beneficiary of Google's vertical integration strategy. Yet the introduction of MediaTek as the designer of TPUv9 introduces a subtle but meaningful shift in the power dynamic. Broadcom remains positioned as a foundry and integrator, but Google has deliberately introduced a competitor into the custom silicon design process. This is industrial discipline: maintain multiple suppliers, prevent any single vendor from becoming indispensable, and preserve the optionality to adjust supply chain composition.
Broadcom itself is diversifying to hedge this concentration risk. The company is simultaneously expanding its custom accelerator partnerships with Meta (MTIA) and other hyperscalers 17,28,41, recognizing that a deep dependence on Google's TPU roadmap alone would leave Broadcom structurally vulnerable should Google accelerate vertical integration further.
Energy as the Binding Constraint
The most underappreciated challenge in Google's AI infrastructure expansion is not silicon supply—it is energy procurement. Google's AI buildout is projected to contribute a 37% increase in electricity consumption in 2025 36, following a pattern in which total electricity demand has risen over 250% since 2019 7,11. Scope 3 supply chain emissions grew 25% year-over-year 39, reflecting the energy intensity of GPU and TPU manufacturing, deployment, and operation.
This energy bottleneck is reshaping Google's strategic calculus. The company is pursuing nuclear energy for its Alabama data center 2, partnering with Westinghouse on AI applications for AP1000 reactor construction 11, and investing in TAE Technologies for fusion research 11. These moves are not peripheral to Google's AI strategy—they are central. The primary constraint limiting Google Cloud's expansion is not the availability of chip design talent or fabrication capacity; it is the physical scarcity of reliable, carbon-competitive power 34. Data center siting faces institutional and social frictions that slow approval and construction timelines 34, meaning that even where grid capacity exists, regulatory and permitting delays impose real costs on scaling.
For NVIDIA, this energy constraint represents an implicit cap on the pace of AI infrastructure deployment, regardless of GPU supply. If Google—the company with the strongest balance sheet, the deepest technical talent, and the most urgent motive to build—finds itself energy-constrained, then the broader industry faces genuine limits on compute expansion. This scenario would dampen NVIDIA's growth trajectory not through direct competitive loss to custom silicon, but through a slowdown in the total addressable market.
Strategic Implications
The evidence assembled here suggests that NVIDIA's position is bifurcating between near-term strength and medium-term vulnerability. In the immediate term—the next 18 to 24 months—NVIDIA remains indispensable. Google's $920 million per month SpaceX GPU commitment, the 100% utilization rate across large-scale infrastructure, and the queuing of compute jobs all confirm that NVIDIA's accelerators are selling into an unlimited immediate market. The scarcity is still the primary economic driver.
However, the trajectory is toward structural erosion. Google's custom TPU roadmap is not a peripheral bet or a boutique internal optimization project. It is the centerpiece of the company's competitive strategy in AI: the way Alphabet preserves margin, deepens lock-in with cloud customers, and ensures that its own AI development remains cost-advantaged relative to rivals. The split of next-generation TPU design across Broadcom, MediaTek, and Marvell demonstrates that Google has moved from TPU-as-internal-tool to TPU-as-platform, architected to scale and support third-party ecosystems. Each Anthropic customer, each Apple deployment, each enterprise on-premises TPU installation represents a shrinkage in NVIDIA's addressable market.
Workload-specific performance fragmentation 16 preserves NVIDIA's CUDA moat for GPU-native applications, but this moat is eroding at the margins as the pool of TPU-optimized models expands. The heterogeneous accelerator landscape that is emerging—TPUs, Trainium, Gaudi, Blackwell, and proprietary inference chips—suggests that the era of GPU ubiquity in AI is ending. NVIDIA will remain a critical node in the stack, but it will be one node among several, with pricing power constrained by viable alternatives tailored to specific workloads.
The energy constraint introduces a final wrinkle. If power availability becomes the binding constraint on data center expansion (as the evidence suggests for Google), then the total AI compute market may grow more slowly than historical GPU demand curves would predict. This would compress NVIDIA's growth rates not through direct displacement but through a moderation in the pace of infrastructure buildout across the industry.
For investors assessing NVIDIA's long-term positioning, the critical question is not whether custom silicon will someday rival NVIDIA GPUs—it clearly will. The question is the speed of this transition and the width of NVIDIA's sustainable margin during it. The evidence here suggests a 5- to 10-year window during which NVIDIA remains highly profitable but faces accelerating competitive encroachment, followed by a more mature competitive landscape in which NVIDIA captures a declining share of AI accelerator spending. This is not a case for bearishness in the short run; it is a case for recognition that inflection points, once they are visible, tend to move quickly.
