Google's TPU Merchant Play Validates Its AI Infrastructure Moat — Anthropic Orders 1 Million Units

The magnitude of Google's AI commitment is staggering, positioning it at a critical inflection point in the technological adoption curve. Chief Technologist Amin Vahdat has framed the long-term bet as potentially reaching $1.9 trillion in infrastructure spending over the next ten years. This decade-long projection is built on a current annual capital expenditure rate of $175 billion to $185 billion, a level that could persist for several years. Viewed through the lens of the S-curve, Google is not just riding the growth phase-it is actively building the rails for the entire next paradigm.

This is an industry-wide sprint, not a solo race. The five largest US cloud and AI infrastructure providers-Microsoft, Alphabet, Amazon, Meta, and Oracle-are collectively committing to spend between $660 billion and $690 billion on capital expenditure in 2026. That figure represents a near doubling of 2025 levels, signaling a decisive shift from early experimentation to massive, coordinated infrastructure build-out. The sheer scale of this "infrastructure sprint" creates a powerful feedback loop: the deployment of compute power accelerates AI model development, which in turn drives even greater demand for infrastructure.

Google's own journey exemplifies this inflection. Its Tensor Processing Unit (TPU) program has crossed a fundamental threshold, moving from internal use to becoming a merchant silicon vendor. The production forecasts for its seventh-generation Ironwood TPU illustrate this exponential ramp: after a jump from roughly 2 million to 4 million units in 2026, the forecast climbs to 5 million in 2027 and a projected 7 million units by 2028. This isn't just scaling-it's a transition from a proprietary compute layer to a foundational infrastructure product. The company is now selling the very chips that power its own AI ambitions, a move that could redefine its competitive position and revenue model as the AI paradigm matures.

The Infrastructure Layer Thesis: Building the Fundamental Rails

Google's strategic pivot is clear: it is no longer just a cloud customer or a model developer. It is positioning itself as the fundamental infrastructure layer for the AI paradigm. This means building the physical and technological rails that every other player must ride. The company's move to become a merchant silicon vendor is the clearest signal of this ambition. Its proprietary Tensor Processing Units (TPUs), once the exclusive compute engine for its own services, are now a product for sale. The production forecasts for its seventh-generation Ironwood TPU-jumping from 2 million to 4 million units in 2026 and targeting 7 million by 2028-show this is not a side project but a core, scaling business. This transition from internal use to external sales is the hallmark of a company that has crossed the S-curve inflection into a foundational role.

This approach contrasts sharply with its closest rivals. Microsoft's strategy is built on a powerful partnership with Nvidia, relying on the latter's dominant GPU technology to power its Azure cloud. While effective, this creates a dependency on a third-party supplier and a shared supply chain. Amazon, meanwhile, leverages its unparalleled scale in logistics and its existing AWS cloud footprint to drive its infrastructure build-out. Its model is one of massive, integrated deployment. Google's path is distinct: it is vertically integrating the compute silicon itself, controlling both the hardware design and the cloud platform where it is sold. This gives it a unique advantage in total cost of ownership, as evidenced by claims that its all-in TCO runs 44% below a comparable Nvidia GB200 server.

Yet, the most critical rail for any infrastructure layer is power. The industry is hitting a fundamental bottleneck where public electrical grids cannot deliver the required capacity. This is forcing a strategic shift toward on-site generation, and Google is acting decisively. The company has secured a $5 billion partnership with Brookfield to develop on-site power generation for its data centers. This isn't just about securing electricity; it's about controlling a non-negotiable input for exponential growth. By locking in dedicated power supply, Google is mitigating a primary operational risk that could otherwise stall its entire infrastructure sprint. This deep partnership with an energy provider creates a tangible moat, ensuring its compute capacity can be deployed without waiting years for grid interconnection.

The bottom line is that Google is constructing a multi-layered moat. It controls the compute silicon, it owns the cloud platform for distribution, and it is securing the power supply. This integrated approach aims to capture the value at the infrastructure layer, where the exponential growth of AI demand will be most concentrated. While competitors play different games, Google is building the field.

The Energy Bottleneck: A Non-Linear Constraint

The primary risk to the AI infrastructure S-curve is no longer capital or compute-it is physical saturation. Public electrical grids are hitting a hard ceiling, unable to deliver the concentrated power required by modern AI-optimized racks. These racks demand between 30 kW to over 100 kW, a massive jump from the 5-15 kW typical of traditional workloads. This exponential increase in power density has overwhelmed local grid capacity, turning what was once a future concern into an immediate, structural bottleneck. The result is a strategic pivot away from grid dependence and toward a new paradigm: off-grid data centers powered by their own energy sources.

This shift redefines the investment and operational model. Instead of simply building servers, companies must now build entire energy systems. The industry response has evolved from efficiency initiatives to securing power independence. Major deals, like Google's $5 billion partnership with Brookfield for on-site generation, exemplify this new reality. The core business risk has changed from software lag to grid interconnection delays, which can now exceed three years. This forces a fundamental rethinking of capital allocation, where a significant portion of the $175-$185 billion annual capex is redirected toward energy infrastructure.

The scale of this demand is staggering. Global AI-driven data center power consumption is set to grow by nearly 126 GW annually through 2028. That figure is almost as large as Canada's entire annual power demand. The bottom line is that the infrastructure sprint is now a race for power. Companies that control the energy supply chain-through partnerships, on-site generation, or vertical integration-will have a decisive advantage in deploying compute at the pace required by the AI paradigm. The bottleneck is physical, and the solution requires a new class of infrastructure investment.

Financial Impact, Valuation, and the Adoption Curve

The disconnect between massive infrastructure investment and the current revenue streams is the central tension for Google's AI bet. The industry is spending nearly $700 billion on AI infrastructure this year, a sprint that dwarfs the growth of the pure-play AI vendors it is building for. While companies like OpenAI and Anthropic are scaling rapidly-with Anthropic's revenue run rate surpassing $9 billion in early 2026-their combined revenues remain a fraction of the infrastructure investment being deployed on their behalf. This creates a classic lag: the capital expenditure is being front-loaded to build the rails, while the commercial adoption of the AI services that will ultimately fill them is still accelerating. The sustainability of this capex depends entirely on whether the adoption curve for AI services can steepen fast enough to justify the exponential build-out.

Google's stock performance reflects this market sensitivity. The shares have delivered a strong rolling annual return of 82%, a powerful rally that has rewarded the long-term vision. Yet, that momentum has stalled recently, with the stock underperforming over the last 20 days. This choppiness highlights how investors are weighing the massive future bets against near-term profitability. The valuation metrics are elevated, with a forward P/E near 29, which prices in significant future growth. The market is not discounting the capex; it is demanding proof that the revenue from that capex will materialize on the required timeline.

The commercial traction of Google's TPU merchant business is the critical early indicator. The company's transition to selling its proprietary silicon is the linchpin of its infrastructure layer thesis. Here, the validation from a major AI developer is paramount. Anthropic's commitment to 1 million TPUs (1-plus GW of compute) is a pivotal, real-world endorsement. It moves the TPU program from a promising internal project to a validated external product. This deal provides a tangible revenue anchor for the capital expenditure, offering a clearer path to recoup the investment in silicon design and manufacturing. It is the first concrete sign that Google's infrastructure build-out is not just a cost center but a scalable, merchantable product.

The bottom line is that Google is betting on a future where its infrastructure spend directly fuels its own revenue. The stock's recent pullback is a reminder that exponential growth narratives require exponential execution. The TPU deal with Anthropic is a crucial validation point, but the company must continue to convert its massive capital allocation into commercial success across its cloud and silicon businesses. The adoption curve for AI services must accelerate to match the S-curve of infrastructure investment.

Catalysts, Risks, and What to Watch

The thesis of Google's infrastructure build-out hinges on a few critical forward-looking events. The company's massive capital expenditure is only de-risked if it can execute on its energy partnerships and secure the power needed to run its servers. The commercial success of its TPU merchant business is the clearest validation that its silicon strategy is creating value. And the timeline for grid saturation will force the industry to accelerate its pivot to off-grid solutions, a shift where Google is already positioned.

The execution of Google's own energy projects is paramount. The company's $5 billion partnership with Brookfield to develop on-site power generation is not a theoretical hedge; it is a direct investment to control a non-negotiable input. The catalyst here is the successful deployment of these projects. If Google can bring its own power online faster than grid interconnection delays, it will demonstrate a tangible operational advantage. This execution will be the first real-world test of whether its integrated energy strategy can de-risk its capital expenditure and ensure its compute capacity is deployed on schedule. Any delay or cost overrun in these energy projects would directly threaten the timeline and economics of its entire infrastructure sprint.

The timeline for grid saturation acts as a powerful external catalyst. Developers expect power constraints to become acute by 2027–2028. This looming deadline is a non-linear constraint that will force faster adoption of off-grid solutions across the industry. For Google, this creates a window of opportunity. As the clock ticks down, the strategic value of its energy partnerships and on-site generation will become more apparent. The company that can demonstrate it has secured power independent of the grid will be best positioned to capture the next wave of AI compute demand. The risk is that if grid capacity is restored or new supply comes online faster than expected, the urgency-and thus the premium-of off-grid solutions could diminish.

Finally, the commercial traction of the TPU merchant business is the key indicator of demand for Google's infrastructure layer. The company's transition to selling its proprietary silicon is the linchpin of its vertical integration strategy. The commitment from Anthropic to purchase 1 million TPUs is a pivotal early validation. The next catalyst is whether this deal can be replicated and scaled. Investors should watch for announcements of new commercial contracts, particularly with other major AI developers. The bottom line is that Google must convert its massive capex into a scalable, merchantable product. The TPU deal with Anthropic provides a revenue anchor, but sustained growth will require a broader customer base. This commercial success will determine if the company's infrastructure build-out is a cost center or a profitable engine for future growth.