NVIDIA Highlights a Shift From Speed to “Autonomous Operations” in Advertising at Cannes Lions

Cannes Lions, the annual meeting place for branding and creative tech, is also becoming a showcase for how businesses plan to use artificial intelligence to run marketing operations with less manual work. In a session built around AI infrastructure and partner deployments, NVIDIA said the industry has moved beyond simply adopting AI for faster execution, and is now focused on whether the underlying systems can support “autonomous operations” at the speed and scale modern advertising requires.

NVIDIA’s framing is anchored in a production constraint that adtech vendors know well: many decisions happen inside tight time windows. For advertisers and ad platforms, serving ads and recommendations across billions of daily transactions depends on inference, the step where models generate outputs, that is accurate, fast, and affordable enough to run at scale. The company argued that real-time bidding and optimization are where those requirements show up most clearly.

One example NVIDIA highlighted is AWS’s “cohesive stack” for adtech, combining cloud infrastructure, foundation models, and NVIDIA GPU-accelerated computing. The company said AWS provides a production-ready reference implementation that lets adtech players run AI-powered bidding directly inside auctions, using NVIDIA Triton Inference Server to perform deep learning inference within real-time auction windows. In that setup, NVIDIA said adtech companies can move from rules-based decisioning toward AI models for bid price optimization, audience activation, and deal scoring as part of the live auction pipeline.

Measurement is another area where NVIDIA positioned AI as moving beyond reporting to proving. The company pointed to Alembic, which it described as a “causal AI” platform that focuses on identifying what marketing initiatives actually drive business outcomes, rather than only documenting what happened. NVIDIA said Alembic models true causation across channels, markets, and audiences, and argued that this requires data processing at a scale that can handle enormous, fast-changing datasets without oversimplifying them into correlation-based assumptions.

NVIDIA said its hardware and software are used to expand the size and scope of those causal models. The company cited Alembic scaling its causal AI models using NVIDIA DGX Vera Rubin NVL72 systems, enabling what it described as analysis across more variables, larger simulations, and quantification of the true drivers of growth. NVIDIA also said Alembic plans to be the first causal AI company to use NVIDIA DGX Vera Rubin SuperPODs for enterprise-scale causal modeling, with the goal of giving executives a single source of “unbiased truth” about what drove outcomes and where capital may be wasted.

The deployments described at Cannes Lions also leaned heavily on “where the data already lives.” NVIDIA said Alembic’s inference runs on private supercomputing infrastructure inside Equinix data centers, keeping AI workloads local. World Wide Technology was also cited for extending that approach into secure and regulated environments. NVIDIA characterized the combination as an enterprise AI stack designed for executives and data leaders who are accountable for capital decisions, reflecting a common tension in enterprise AI between experimentation and governance.

Elsewhere in the event’s partner lineup, Criteo was presented as an example of continuous retraining for recommendation quality at scale. NVIDIA said Criteo trains AI on billions of shopper timelines and that speed translates into model quality. NVIDIA added that Criteo achieved a roughly 2x speedup in model training on NVIDIA Blackwell GPUs, driven by NVIDIA cuEmbed, and that the efficiency frees approximately 17,000 GPU hours per year, with further scaling planned.

For creative and media workflows, NVIDIA described an emerging pattern of AI agents as “digital coworkers” that handle long-running tasks across planning, execution, and optimization. But the company also emphasized that enterprise deployment requires controls, including safety guardrails, auditability, and role-based permissioning. NVIDIA said its Agent Toolkit, which includes NVIDIA NemoClaw blueprints and the NVIDIA OpenShell secure runtime, is intended to provide that trust and control layer.

One partner example of those agent-based workflows was Higgsfield AI, which NVIDIA described as a production platform for generating marketing images and video. NVIDIA said Higgsfield offers “Supercomputer agents” that manage a marketing automation lifecycle, from campaign ideation and planning through creative production and posting, plus autonomous campaign optimization in a single interface. NVIDIA said the platform orchestrates large language models alongside 35+ image, audio, and video models, and that Higgsfield’s proprietary Soul and Soul 2.0 models run on NVIDIA Blackwell architecture. NVIDIA added that specialized subagents powered by NVIDIA Nemotron open models run continuously inside each campaign, while Nem oClaw and OpenShell are being integrated to provide the enterprise “trust layer,” with campaigns for nearly 400 Fortune 500 companies created on the platform.

NVIDIA also connected multimodal understanding, meaning models that can process both visual and textual information, to infrastructure needs for content-aware advertising. It highlighted ’s “Moment Match Engine,” which NVIDIA said evaluates indicates across video frames and media assets to interpret scenes, objects, and products, then recommend ad creative based on those moments. NVIDIA attributed to over 10x improvements in processing speed and efficiency after adopting NVIDIA’s Nemotron 3 Nano Omni open model, and noted that on the MediaPerf open benchmark for AI video understanding, Nemotron 3 Nano Omni delivered the highest throughput and lowest inference cost among models evaluated, including open and closed sources.

Several other claims in NVIDIA’s description depend on partner execution details that are not independently quantified in the post itself. For example, NVIDIA does not provide documentation of end-to-end business lift for each deployment, nor does it release performance benchmarks beyond specific training speedups and the single benchmark comparison on MediaPerf. The company also does not describe how often causal modeling results are updated, what governance processes are used internally by each enterprise customer, or how the “autonomous operations” approach is constrained in practice by safety and audit requirements beyond the general statement about guardrails.

Still, NVIDIA’s Cannes Lions message is clear: the next competitive layer in adtech may be less about individual model capability and more about orchestration, inference efficiency, and enterprise governance that can withstand real-time constraints and compliance expectations. What to watch next is whether these partner demos translate into measurable production rollouts, including whether causal measurement systems become a standard part of marketing budgeting and whether agent-based creative and optimization workflows expand beyond pilot campaigns into repeatable operating models across different regulated sectors.