Anthropic’s Multi-Billion Dollar Compute Play
Anthropic is rapidly transforming from a cloud buyer into a long-horizon infrastructure planner. The company has reportedly signed a USD 1.8 billion (approx. RM8.28 billion) compute deal with Akamai Technologies to handle rising demand for its Claude AI, particularly for long-running agents and automation-heavy workloads that stress AI inference scaling. Akamai’s distributed GPU and edge architecture is positioned to support low-latency inference and agent tasks, providing an additional supply lane beyond traditional hyperscale clouds. In parallel, Anthropic may pay Google USD 200 billion (approx. RM920 billion) over five years for cloud computing commitments and chip access, according to reporting cited in the source material. While the exact terms remain unconfirmed, the arrangement builds on an existing TPU capacity expansion scheduled to start coming online in 2027. Together, these AI compute infrastructure deals show Anthropic racing to secure training and inference capacity years in advance, turning compute access into a strategic moat.
How Capacity Lock-Ins Are Redrawing AI Market Dynamics
Anthropic’s long-term agreements with Google and Akamai highlight a structural shift in the AI ecosystem: major labs are reserving future compute years ahead to avoid shortages that could derail product roadmaps. By pre-booking cloud computing commitments and specialized chips, frontier model developers can plan larger training runs and more intensive enterprise inference traffic without waiting in cloud queues or scrambling for GPUs. For infrastructure providers, a customer willing to commit this far in advance de-risks investments in data centers, chip procurement, and power upgrades. That alignment encourages faster buildout cycles and bigger clusters tailored to specific workloads. But it also narrows the spot market for everyone else. As top labs soak up large blocks of capacity, smaller AI builders may face higher prices, longer wait times, and weaker negotiating power, reinforcing a tiered market where access to compute becomes as decisive as model quality.
SAP’s Frontier AI Lab Bet on Tabular Foundation Models
While Anthropic secures generic compute scale, SAP is investing directly in domain-specific frontier AI labs. SAP has entered into a definitive agreement to acquire Prior Labs, a specialist in Tabular Foundation Models (TFMs), and plans to invest more than €1 billion over the next four years to grow it into a globally leading frontier AI lab focused on structured business data. Prior Labs will continue to operate as an independent entity, giving it autonomy while benefiting from SAP’s resources and enterprise reach. TFMs are designed for tables, numbers, and statistics—areas where large language models often struggle. They can power predictions on payment delays, supplier risk, upsell opportunities, and customer churn, unlocking AI value in core business systems. By pairing Prior Labs’ research team and benchmarks with SAP’s SAP-RPT-1 groundwork and customer data environment, this deal signals a push to own the infrastructure and expertise for mission-critical, structured-data AI rather than relying solely on generic cloud models.
The Emerging AI Compute Arms Race and Its Consequences
Taken together, Anthropic’s cloud-scale deals and SAP’s investment in Prior Labs illustrate an AI compute arms race with several layers. At the top, labs like Anthropic secure enormous AI compute infrastructure deals for both training and AI inference scaling, turning capacity into a strategic asset. At the application layer, enterprise players such as SAP build frontier AI labs tuned to their data and customers, tightening vertical integration between infrastructure, models, and business workflows. The consequences cut in two directions. On one hand, these investments could accelerate innovation and stabilize access for large customers, enabling more powerful, reliable AI services. On the other, they risk concentrating bargaining power and capacity among a small cluster of well-financed firms. As more providers lock in multi-year cloud computing commitments, independent developers and smaller enterprises may find that the real bottleneck in AI is no longer algorithms, but who controls the underlying compute pipelines.