The industrial logic of AI is now unmistakable: the master resource of this age is compute, and the firms that command its most efficient, large-scale production will dictate the terms of the next economic era. Just as the steel barons of the late 19th century secured their empires by integrating ore, transport, and mills, today’s AI lords are racing to lock in gigawatt-scale data centers, proprietary accelerators, and the loyalty of a few frontier labs whose appetite for raw computation is reshaping the infrastructure landscape. In this contest, Alphabet has placed a decisive bet on becoming the Bessemer process of AI workloads, using its custom TPU silicon and deep partnership with Anthropic to cement a position in the foundational layer. Yet the same forces that have made Anthropic a demand engine for Google Cloud are also fueling a multi-cloud scramble that draws in rivals from AWS to SpaceX, and forcing Alphabet to confront the fragility of its concentrated exposure.
The Demand Engine: Anthropic’s Compute Hunger and Alphabet’s TPU Mill
Anthropic’s need for compute has grown so voracious that it has effectively become a primary driver of hyperscale capacity planning. The numbers betray an industrial mobilization on a scale not seen since the railway booms: Anthropic expects to spend $20 billion on server leasing in 2026 alone 22, and its server spending has already tripled year-over-year 22. For Alphabet, this translates into commitments of breathtaking magnitude. The company has secured agreements with Google and Broadcom totaling 3.5 gigawatts of TPU capacity 1,2,3,12,14, with further reporting pointing to 5 gigawatts of Google infrastructure coming online from 2027 5,12,22. A Broadcom-mediated supply line will deliver 3.5 GW of next-generation TPUs starting that same year 11, while Google’s own capacity scaling—including a joint venture with Blackstone—targets 500 megawatts by 2027 10,13,36 and expansion beyond that milestone planned 30. These are not short-term experiments; they are long-term contracts of approximately five years 18 that have been public since at least October 2024 10. The partnership is further operationalized through integrated solutions like Google Cloud Rapid Cache, which enables Anthropic to achieve read throughput of up to 2.5 TB/s 19.
Such scale, however, begets dependency. By one estimate, approximately 40% of Google Cloud’s demand is generated by Anthropic 6, and the two leading AI labs—OpenAI and Anthropic—combined account for roughly half of all hyperscale cloud bookings 28. Within Amazon, Anthropic alone represents nearly 50% of the cloud order book 9, and a similar concentration is implied for Google. This is the double-edged sword of anchor tenancy: while the revenue streams underwrite capacity expansion, they also create a single-point vulnerability. Alphabet has already felt the strain, as internal capacity constraints emerge from the dual demands of Anthropic and Meta 11.
Multi-Cloud Chess: From AWS to Azure to SpaceX
Anthropic, for its part, has no intention of staking its future on any single provider. Its strategy is a classic hedging of industrial supply: secure as much capacity as possible, across as many architectures and vendors as the market can bear, and then optimize workloads to each. The firm has secured up to 5 GW of new capacity from Amazon Web Services—with roughly 1 GW expected online by end-2026 4,7,8,12,23—and maintains a 1 GW deal with Microsoft Azure 17. Its hardware portfolio spans NVIDIA GPUs, AWS Trainium chips, and Google TPUs 12, with engineers actively matching training and inference jobs to the most suitable architectures 33. This multi-cloud posture 22 is a deliberate erosion of any single provider’s bargaining power, and it prevents Alphabet from charging monopoly rents on its TPU advantage.
More disruptive still is the emergence of SpaceX as a compute lessor. No longer merely a launch company, SpaceX has become a supplier of raw GPU capacity, leasing compute capacity from its Colossus 1 and Colossus 2 superclusters 23,27,37,38, including over 220,000 NVIDIA GPUs from Colossus 1 24. The terms are revealing: the leases are structured as short-term, 180-day agreements 15 with a 90-day cancellation provision 15, initially conceived as a stop-gap to address Anthropic’s acute compute shortages 26. Yet the payments—$1.25 billion per month through 2029 20,27—are of an order that signals permanence. SpaceX’s own IPO prospectus envisions this as a long-term line of business 27, and its terrestrial superclusters are only the beginning.
The New Pick-and-Shovel King? SpaceX’s Terrestrial and Orbital Ambitions
If the current landscape is defined by a few hyperscale foundries, the next act may be written by the entity that controls the most radical expansion of the compute frontier. SpaceX is openly pursuing orbital AI infrastructure, citing the energy constraints of terrestrial data centers and the promise of continuous space-based solar power 21,25. While orbital data centers remain experimental, they represent a potential paradigm shift that could undercut the value of ground-based investments over time 16,29. In this vision, the old constraints of land, power, and cooling are eclipsed by a platform that can scale without the same geographic bottlenecks. For Alphabet—committed to a joint venture with Blackstone for 500 MW by 2027—the question becomes: does the terrestrial strategy lock in a durable cost advantage, or does it risk becoming a stranded asset as the industry moves to orbit?
The broader industry is already thinking in gigawatt campusses. Projects like Stargate (5 GW) and Meta’s Hyperion campus (5 GW) 8,31 illustrate that the AI buildout is entering a phase where the unit of capacity is not the data center but the power plant. Total AI compute capacity could reach 35 GW by 2028 34, and eventually a hundred-gigawatt-plus installed base 35. In this race, the decisive advantage will not belong to the most advanced model builder, but to the operator that can deliver the cheapest, most abundant, and most reliable compute. SpaceX’s growing inventory of NVIDIA H100 GPUs 32 and its unconventional financing give it a profile that is part industrialist, part disruptor—a modern-day steel magnate emerging from the aerospace sector.
Alphabet’s Strategic Imperatives: Integration, Capacity, and the Blackstone Venture
For Alphabet, the path forward demands a clear-eyed recognition of where its true strengths lie. Its bet on custom TPU silicon has created a productive asset that few rivals can replicate, and the Anthropic relationship validates the roadmap. But concentration is the enemy of resilience. The company must accelerate its capacity expansion—both through the Broadcom pipeline and the Blackstone venture—while simultaneously diversifying its tenant base to reduce the asymmetric risk of a single customer. The Blackstone vehicle is a step toward hybridizing infrastructure development, but the timeline (2027 and beyond) may leave a window for AWS to capture more of Anthropic’s marginal workloads with its nearer-term 1 GW delivery.
More fundamentally, Alphabet must monitor the orbital play not as a novelty but as a potential strategic disruption that could redraw the cost curves for the entire industry. The firm that controls the cheapest compute will control the AI future; if that firm turns out to be SpaceX, then every hypercaler’s terrestrial foundation becomes a legacy asset. Alphabet would be wise to explore partnerships, capacity contracts, or even acquisition of orbital compute startups to ensure it is not blindsided. In an era where the hundred-gigawatt frontier is within sight, the only unforgivable sin is to be left without a mill.
Implications for the AI Empire
The unfolding multi-cloud expansion of Anthropic is more than a procurement story; it is a signal of how bargaining power is shifting in the AI stack. The frontier labs are the new steel customers—indispensable, but also capable of playing multiple suppliers against one another. For Alphabet, the immediate revenue tailwinds are real, but the long-term contest will be won by the entity that most ruthlessly drives down the cost of compute and integrates the most critical layers of the stack. If Google can maintain its TPU advantage and scale capacity faster than rivals, it will secure a platform moat that endures. If not, it risks becoming a high-cost foundry in a world that is rapidly commoditizing the means of computation. The next five years will reveal whether Alphabet’s integration play is the equivalent of Carnegie’s Homestead Works—or a sprawling, underutilized mill in a market that has moved on.
