Nvidia Vera Rubin Demand, September 2026: Anthropic Locks In 2.5 GW at $40 Billion Per-Gigawatt Revenue

Anthropic's reported lock-in of 2.5 gigawatts of Nvidia capacity, anchoring a per-gigawatt annual revenue rate of roughly $40 billion, arrives alongside an unresolved HBM4 supply chain, a record $150 billion buyback, and a maturing CUDA alternative taking shape in Beijing.

The most telling metric in Nvidia's September 2026 news cycle was not a quarterly earnings figure but a capacity unit: Anthropic reportedly locked in 2.5 gigawatts of Nvidia infrastructure, including the Vera Rubin GPU generation, anchoring a per-gigawatt annual revenue rate for Nvidia of roughly $40 billion. At $228.38 per share and fiscal 2026 revenue of $215.94 billion, the company's valuation already reflects the assumption that this pricing density is structural rather than transitory. The commitment is notable not just for its scale but for its product generation: the lock-in covers Nvidia's Vera Rubin GPU family.

That commitment is consistent with a broader demand acceleration visible across September's coverage. CoreWeave reached production scale with Nvidia's Vera Rubin NVL72 GPU clusters in September 2026, naming Cognition as an early live customer. On the hyperscale side, Elon Musk reportedly disclosed a month-by-month expansion roadmap for xAI's Colossus 2 program targeting more than 1.2 million Nvidia GPUs by year-end. Separately, Nvidia is said to be pursuing its own infrastructure positions — reports cited arrangements tied to a Hut 8 data-center campus in Texas and substantially larger Ohio commitments — though the terms and ownership structure of those arrangements remain unconfirmed.

Nvidia's own financial posture reflects the confidence that demand picture implies. Capital expenditure reached $6.04 billion in fiscal 2026, an increase of 86.7% from $3.24 billion the prior year, yet against revenue of $215.94 billion that yields a capex intensity of just 3% — a lean ratio made possible by a fabless model in which the heaviest investments sit at foundry and memory suppliers rather than on Nvidia's own balance sheet. The company reinforced that posture by announcing a $150 billion share buyback in September 2026, its largest ever.

The supply chain on which that trajectory depends is, however, where the most concentrated vulnerability resides. Reports indicate Nvidia is pressing SK Hynix, Samsung, and Micron to deliver 16-layer HBM4 memory at volume in the fourth quarter of 2026, with supply contracts not yet confirmed. A patent-infringement petition by Netlist, seeking a US import ban on Micron memory chips incorporated into Nvidia AI systems, adds a legal dimension to that dependency. Nvidia has also reportedly issued a two-year deadline for its packaging suppliers to complete glass-substrate technology development, a transition that industry observers suggest could roughly double HBM packaging costs by 2027. The memory and substrate roadmap supporting Vera Rubin and successor products is technically ambitious and contractually incomplete.

The competitive and geopolitical environment has grown commensurately more complex. In China, Huawei has claimed that its domestic AI chip sales now exceed Nvidia's in that market — a figure that is not independently verifiable but that coincides with the joint release by DeepSeek and Huawei of TileLang, an open-source programming language for the Ascend chip family explicitly designed to enable development outside Nvidia's CUDA ecosystem. The practical depth and scalability of that alternative software stack remain to be demonstrated at production scale. At the policy level, Nvidia and AMD were reported to have lobbied the Trump administration in September to ease proposed AI chip export restrictions in a pending defense bill, while separate reports indicated that Chinese regulators were considering permitting companies including Alibaba and ByteDance to purchase specific Nvidia GPU models. The scope and durability of any such access expansion remain unverified.

Three developments will be most informative over the near term. Whether Anthropic's 2.5-gigawatt lock-in translates into contracted delivery timelines will determine how firmly the $40-billion per-gigawatt revenue metric is embedded in Nvidia's forward revenue visibility. Whether HBM4 supply agreements are signed — and on what schedule — will indicate whether the Vera Rubin production ramp proceeds without the memory constraints that shaped earlier product transitions. And whether US-China chip policy stabilizes into a durable framework will clarify the commercial potential of a market that simultaneously represents a meaningful revenue opportunity and a concentrated regulatory variable. The buyback and the capacity commitments both express confidence that these uncertainties resolve in Nvidia's favor; tracking each signal against that implied confidence is the work of the next two quarters.