AT&T’s AI rollout targets OpenAI-driven growth, drawing a pro-compute reaction from NVIDIA leadership

AT&T is indicating that its artificial intelligence deployments will rely more heavily on models from OpenAI over time, and the direction of travel is being noticed by major AI infrastructure providers. In a report published by Yahoo Finance, AT&T Chief Data and AI Officer Andy Markus said that OpenAI models already power about 25% of the telecom’s total AI usage, with a longer-term goal of reaching roughly 70% to 80%.

Markus’s remarks place emphasis on AI “usage” rather than a one-time project. The framing suggests AT&T is building ongoing workflows where AI models are invoked repeatedly, across operations and customer-facing applications. While the report does not specify which particular AI use cases drive the 25% figure, it indicates that the company sees a substantial portion of its AI activity as already model-driven rather than rule-based.

The Yahoo Finance piece also connects the AT&T comments to an industry debate about what some executives call a “token apocalypse.” Tokens are the building blocks of text and other language inputs and outputs processed by large language models. The concern behind the term is that demand for AI-generated content could lead to runaway consumption of model inputs and outputs, raising costs and straining infrastructure if each request requires too many tokens. Markus’s outlook, described as moving toward higher reliance on OpenAI models, suggests AT&T expects those costs to remain manageable or to be offset by improvements in model efficiency, scaling, or pricing.

NVIDIA, whose business is heavily tied to accelerating AI training and inference, was cited in the same context. Yahoo Finance reported that NVIDIA’s CEO was “cheering” the same trend, implying approval of continued growth in AI workloads even as the “token apocalypse” concern circulates among market participants.

The connection between carrier AI strategy and GPU supply chains is straightforward at a high level: more AI inference calls generally mean more demand for compute and the software stack that runs models efficiently. NVIDIA’s position in the ecosystem comes primarily through its data center graphics processing units and related AI platforms, which many companies use to serve models at scale. However, the report summarized by Yahoo Finance does not provide specifics from NVIDIA about what the company expects in terms of token volume, pricing, or customer mix.

NVIDIA does run public channels for company and product updates, but the only research link available for this story is the NVIDIA newsroom homepage. Without additional primary detail quoted directly in the Yahoo Finance report, it is not possible to say what NVIDIA’s CEO meant with precision, nor which operational bottlenecks NVIDIA believes will improve. For now, the most defensible interpretation is that NVIDIA leadership is not treating the “token apocalypse” thesis as a reason to slow down AI workload growth, and instead expects the market to adapt through engineering, optimization, and scale.

One caveat is that AT&T’s stated targets, including the 70% to 80% longer-term range, are not accompanied in the available material by a timeline beyond the broad “over time” phrasing, nor by the conditions required to reach that target. The report also does not detail whether the shift involves changing model routing, procurement strategy, or performance and cost benchmarks, nor does it disclose how AT&T is measuring “total AI usage” across its systems.

Looking ahead, investors and industry watchers will likely focus on whether AT&T can operationalize the increase in reliance on OpenAI models without eroding unit economics, and whether NVIDIA’s leadership stance translates into measurable demand indicators for inference compute. The next clear checkpoints would be AT&T’s future AI-related disclosures, any follow-up commentary from NVIDIA leadership clarifying what “token apocalypse” fears they believe are overstated, and additional disclosures from AI vendors on cost-per-token and serving optimizations.