Nvidia's Blackwell GPUs Go Live
Akamai is deploying thousands of Nvidia's new Blackwell GPUs, creating a massive, distributed AI platform. The move establishes Blackwell as the new industry standard for large-scale AI training and inference, setting the benchmark for Apple's own cloud and on-device hardware ambitions.
Nvidia's Blackwell architecture represents a monumental leap in raw compute power, with a single GPU packing 208 billion transistors. This is achieved by linking two massive dies with a 10 TB/s interconnect, effectively creating one unified, giant GPU. The performance jump is staggering, offering up to 20 petaflops of AI performance on a single chip and a 2.5x boost over the previous Hopper generation. The new architecture introduces 5th-generation tensor cores that support new, lower-precision FP4 and FP6 data formats. This capability, combined with a second-generation Transformer Engine, allows Blackwell to dramatically increase throughput for large language model (LLM) inference, a critical factor for real-time AI applications. For context, a rack of 72 Blackwell GPUs can achieve 1,440 petaflops and is designed for trillion-parameter models. Akamai's deployment shifts the focus from centralized training to distributed inference, placing thousands of Blackwell GPUs across its 4,400+ global locations. This strategy directly tackles the latency barrier that hinders large-scale AI adoption, aiming to bring computation closer to end-users and reduce delays by up to 2.5 times compared to traditional hyperscalers. The move is a clear indicator that the industry's economic and infrastructure challenges are shifting from training models to the cost of running them in production. This new performance benchmark directly challenges Apple's own silicon ambitions. While Apple is pushing on-device AI with its neural engines, it is also expanding its backend infrastructure, with plans to mass-produce its own AI server chips, reportedly codenamed "Baltra," in the second half of 2026. This investment in custom silicon for its data centers signals a long-term strategy to control its entire AI hardware and software stack. On the manufacturing front, Apple is accelerating its US-based production, though the focus is currently on Houston, Texas, not Fremont. The Houston facility is slated for mass production of advanced AI servers and, for the first time, the Mac mini in 2026. As part of a larger $600 billion US investment, Apple also plans to purchase over 100 million advanced chips from TSMC's new Arizona factory in 2026. The competition for talent to design and build this next generation of hardware remains fierce. Bay Area tech giants like Nvidia are leveraging massive stock packages that have created "golden handcuffs," making it incredibly difficult to lure away experienced engineers. One Nvidia employee's $420,000 equity package in 2023 was valued at nearly $2 million by late 2025, a powerful incentive to stay put. Meanwhile, the regulatory landscape is tightening. The US government is now considering per-customer caps on the number of advanced AI accelerators sold to China, potentially limiting any single Chinese firm to 75,000 of Nvidia's last-generation H200 chips. This follows earlier restrictions and highlights the increasing geopolitical complexities impacting the semiconductor supply chain.
