Alembic will become the first causal AI company to run NVIDIA DGX Vera Rubin SuperPODs for enterprise-scale marketing measurement, as a cluster of advertising technology companies prepares to demonstrate GPU-accelerated infrastructure at Cannes Lions in France this week.

On June 18, 2026, NVIDIA published details of a set of collaborations with six advertising and marketing technology companies - Alembic, Amazon Web Services, Criteo, HiggsField, KERV.ai, and Taboola - that are each bringing NVIDIA-powered infrastructure to the Cannes Lions International Festival of Creativity, scheduled from June 22 to 26 in Cannes, France. The announcements span four distinct problem areas: causal measurement, real-time bidding, full-lifecycle marketing automation, and video content intelligence.

The scope of the deployments is technical in nature. None of these are concept demonstrations. Each involves production-grade infrastructure decisions - the choice of a specific GPU generation, a particular inference server, an open library, or a runtime environment - that have direct consequences for the speed, cost, and accuracy of marketing systems running at enterprise scale.

What Alembic is doing with DGX Vera Rubin

Causal AI sits at the center of Alembic's deployment. The company positions itself around a specific measurement problem: not just reporting on what happened in a marketing campaign, but modeling what actually caused business outcomes. The distinction between correlation and causation in marketing data has occupied measurement teams for years, and causal modeling requires far greater computational resources than conventional attribution or analytics.

According to NVIDIA, NVIDIA DGX Vera Rubin NVL72 systems will enable Alembic to scale its causal AI models to analyze more variables, run larger simulations, and quantify the true drivers of growth across marketing investments. The NVL72 designation refers to a specific configuration within NVIDIA's DGX Vera Rubin line - a system built around NVIDIA's Blackwell-generation GPU architecture. Running causal models across every channel, market, and audience simultaneously demands the ability to process enormous datasets that are constantly changing. Alembic's inference runs on private supercomputing infrastructure inside Equinix data centers, which means AI workloads remain local to where enterprise data already resides - a configuration that addresses both latency and data governance requirements. World Wide Technology extends this to secure and regulated environments.

The SuperPOD configuration - clusters of NVL72 systems operating together - is what makes enterprise-scale causal modeling tractable. According to NVIDIA, Alembic will be the first causal AI company to use these systems for this purpose. For marketing executives accountable for capital allocation decisions, the practical output is a single source of what NVIDIA describes as "unbiased truth" on what drove business outcomes and where capital is being wasted.

The growing interest in causal measurement infrastructure reflects a broader crisis in marketing analytics. Research published in February 2026 showed that up to 75% of senior planning and analytics decision-makers at US brands and agencies report that advanced measurement approaches - attribution, incrementality testing, and marketing mix modeling - fail to deliver the rigor, timeliness, and trust needed to justify spending. Causal modeling represents an attempt to fix that failure at the infrastructure level.

AWS and Criteo: AI inside the auction

The second problem area is the auction itself. Programmatic advertising auctions operate in windows measured in milliseconds. The constraint is physical: any AI model that runs too slowly misses the auction entirely. The systems that NVIDIA and its partners are demonstrating at Cannes are designed to fit within those windows.

Amazon Web Services is presenting a production-ready reference implementation for AI-powered bidding inside auctions. According to NVIDIA, AWS is bringing cloud infrastructure, foundation models, and GPU-accelerated computing together into a cohesive stack for the ad tech industry - one scaled for the era of AI agents. The infrastructure targets demand-side platforms, supply-side platforms, and independent software vendors, offering them a path to move from rules-based decisioning to AI-powered models for bid price optimization, audience activation, and deal scoring, all operating directly within the live auction pipeline.

The key component enabling this is the NVIDIA Triton Inference Server, which according to NVIDIA delivers deep learning inference fast enough to fit within real-time auction windows. That technical detail matters. Rules-based bidding systems make decisions based on pre-defined logic. AI-powered models score each impression against a learned representation of value - a fundamentally different calculation that requires substantially more compute per decision. Triton makes that compute feasible at auction speed.

Criteo's deployment addresses a different part of the same problem: keeping product recommendations relevant as shopping behavior changes. Criteo operates one of the largest recommendation networks in digital advertising. Continuously retraining AI on billions of shopper timelines is the mechanism by which those recommendations stay accurate. Speed in retraining directly translates to quality in production.

Collaborating with NVIDIA, Criteo achieved approximately a 2x speedup in model training on NVIDIA Blackwell GPUs, driven by the NVIDIA cuEmbed open library. According to NVIDIA, that efficiency already frees roughly 17,000 GPU hours per year. The companies are now scaling the work further. For Criteo, the improvement is consequential given the competitive environment the company is navigating. Criteo's Q1 2026 results showed activated media spend exceeding $1 billion in a single quarter for the first time, while the company simultaneously manages the downstream effects of retail media client scope reductions. Efficiency at the model training layer reduces the operational cost of maintaining recommendation quality across its network.

The trajectory Criteo has been on with AI infrastructure includes the company's emergence as OpenAI's first and only formal programmatic partner inside ChatGPT, where AI-referred shoppers have been converting at nearly double the rate of traditional search referrals in categories like consumer electronics and home and garden. The cuEmbed-driven training improvements reported at Cannes extend the same infrastructure logic to recommendation quality across the broader retail network.

HiggsField: the full lifecycle in a single interface

The third problem area is the complete marketing workflow. Running a campaign from ideation through production, posting, and performance optimization has historically required multiple tools, multiple teams, and substantial manual coordination. HiggsField's Supercomputer is positioned as a system that handles all of those steps autonomously.

According to NVIDIA, HiggsField offers agents that manage the full marketing automation lifecycle - from campaign ideation, planning, creative production, and posting through autonomous campaign optimization - in a single interface. The platform orchestrates more than 35 image, audio, and video models, including HiggsField's proprietary Soul and Soul 2.0 models built on NVIDIA Blackwell architecture.

The agentic dimension is supplied by NVIDIA's Agent Toolkit. According to NVIDIA, the toolkit includes NVIDIA NemoClaw blueprints and the NVIDIA OpenShell secure runtime, which provide the controls - safety guardrails, auditability, and role-based permissioning - that make autonomous agents deployable in enterprise settings. Within HiggsField, NVIDIA Agent Toolkit software, including NVIDIA Nemotron open models, powers specialized subagents running continuously inside every campaign. NemoClaw and OpenShell are being integrated to provide what NVIDIA describes as the enterprise trust layer.

The practical consequence is that a campaign on HiggsField runs through a set of specialized subagents - one handling ideation, another production, another optimization - coordinated through a shared interface. According to NVIDIA, marketing campaigns for nearly 400 of the Fortune 500 companies are created on the platform.

Why does the agent control layer matter? The ad tech industry's broader push into agentic infrastructure has repeatedly surfaced the same question: what stops an autonomous system from making costly or brand-damaging decisions without human review? The NVIDIA OpenShell runtime is a direct answer to that question, establishing a secure boundary around what agents can do and keeping a log of what they have done.

Taboola: conversational AI with advertising built in

Taboola's application of NVIDIA GPU infrastructure focuses on DeeperDive, its AI answer engine. The product allows publishers to embed a generative AI search and answer experience directly on their own websites, drawing from the publisher's editorial archive rather than from a web-wide index. The commercial model depends on inserting contextually relevant advertising into AI-generated answer pages - turning reader questions into advertising inventory. For that to work at scale, the inference layer powering DeeperDive has to be fast and affordable enough to serve millions of daily queries.

According to NVIDIA, Taboola uses NVIDIA GPUs to power DeeperDive and is extending that infrastructure to AI platforms and chatbots so they can generate revenue from advertising.

Taboola has been developing DeeperDive since launching it in June 2025, initially with Gannett's USA TODAY Network and The Independent as launch partners. DeeperDive reached nearly 7 million monthly active users by April 2026 and expanded to French, German, Hebrew, Japanese, Korean, and Spanish, as the company built toward what CEO Adam Singolda has described as the largest ad-supported large language model for the open web. HuffPost UK adopted the system in April 2026 as search referral traffic continued to erode across the publishing industry.

The GPU infrastructure question for Taboola is cost per query. Publishers integrating DeeperDive do not pay for the service; they receive a share of the advertising revenue generated. That economics model requires inference to be cheap enough to remain profitable at scale across a network touching more than 9,000 publisher sites and approximately 600 million daily active users. Taboola's broader agentic trajectory, including the April 2026 launch of Realize+ and its Claude Skills integration, situates DeeperDive within a company that has been systematically moving its infrastructure toward autonomous, AI-native operations across both the advertiser and publisher sides of its business.

KERV.ai: understanding every frame of video

The fourth area is content intelligence - specifically, the ability to understand what is happening inside a video at the level of individual frames, objects, and scenes, in order to match advertising to contextually relevant moments.

KERV's Moment Match Engine evaluates signals across every video frame and media asset, understanding individual scenes, objects, and products, then provides content recommendations based on the visual and textual elements of an advertisement. According to NVIDIA, KERV.ai recently optimized its processing pipeline, achieving over 10x improvements in speed and efficiency using the NVIDIA Nemotron 3 Nano Omni open model.

The benchmark that gives this claim context is MediaPerf, an open benchmark for AI video understanding. According to NVIDIA, Nemotron 3 Nano Omni delivered the highest throughput and lowest inference cost of any model evaluated on MediaPerf, open or closed source. Ecosystem partners including PYLER, which uses NVIDIA DGX B200 systems, also adopted the model.

For advertisers, the practical consequence is a tighter connection between creative and context. KERV's system analyzes what each ad or media brief contains, who it resonates with, and which exact moment within content to target. That specificity - matching creative to the right frame, not just the right program - is what KERV positions as "improved engagement."

The infrastructure question behind all of it

What connects these six deployments is not a single product or platform. It is an infrastructure question: can the hardware running AI models for advertising keep pace with the volume, speed, and complexity that modern advertising requires?

US programmatic ad spending is expected to exceed $200 billion in 2026, according to EMARKETER. Every auction, every recommendation, every content classification decision within that market involves an AI model making a judgment in milliseconds. The transition from rules-based to model-based decisioning is well underway, but the infrastructure cost of running models at this scale is not trivial. The announcements at Cannes are, in part, NVIDIA's argument that its GPU systems are the answer to that infrastructure problem.

The specific technical choices on display are revealing. Criteo's 2x training speedup comes from cuEmbed, an open library for embedding operations - a narrow but high-impact optimization for recommendation systems that process billions of user timelines. AWS's reference implementation uses Triton Inference Server, which manages the model deployment layer so that auction-speed inference becomes a configurable infrastructure choice rather than a custom engineering project. Alembic's move to DGX Vera Rubin NVL72 SuperPODs is a bet on scale: that causal modeling across every channel and market simultaneously requires a level of compute that only dedicated supercomputing infrastructure can provide.

The agentic control layer is a second thread running through all of these announcements. HiggsField's deployment of NemoClaw and OpenShell reflects a point that the IAB Tech Lab's Agentic RTB Framework has been making since November 2025: autonomous AI systems running inside enterprise advertising workflows need safety guardrails, role-based permissioning, and auditability before they can be responsibly deployed at scale. The NVIDIA Agent Toolkit is one answer to that requirement at the infrastructure level.

WPP Media released a midyear forecast on June 16, 2026, projecting global advertising to reach $1.3 trillion in 2026, with AI investment named as the primary driver. That forecast arrived in the week immediately preceding the Cannes Lions festival, alongside a dense cluster of agentic product launches from Yahoo, Fox, Horizon Media, LiveRamp, Stagwell, and DoubleVerify - all within 72 hours of each other. NVIDIA's Cannes showcases are part of the same structural moment: hardware vendors, cloud providers, and ad tech companies converging on the same argument that GPU-powered, autonomous advertising infrastructure is now the primary axis of competition in the industry.

Whether that argument holds at the margins - whether the 2x training speedup translates to measurably better campaign outcomes, whether causal AI at SuperPOD scale changes what marketers actually decide - is the question that the next several quarters of results will answer.

Timeline

  • June 4, 2025 - Taboola announces the full commercial launch of Predictive Audiences with up to 270% conversion improvements for early adopters. PPC Land coverage
  • June 11, 2025 - Taboola launches DeeperDive, an AI answer engine designed to live directly on publisher websites, with Gannett and The Independent as launch partners. PPC Land coverage
  • November 13, 2025 - IAB Tech Lab releases Agentic RTB Framework version 1.0 for public comment, standardizing containerized agent deployment in real-time programmatic advertising. PPC Land coverage
  • January 5, 2026 - PubMatic launches AgenticOS, the first operating system built specifically for autonomous advertising execution. PPC Land coverage
  • February 5, 2026 - Criteo announces an Agentic Commerce Recommendation Service achieving up to 60% improvement in recommendation relevancy. PPC Land coverage
  • February 7, 2026 - IAB State of Data 2026 report finds up to 75% of marketers say attribution, incrementality, and MMM underperform on rigor, timeliness, and trust. PPC Land coverage
  • March 2, 2026 - Criteo announced as the first formal ad tech partner inside the ChatGPT advertising pilot. PPC Land coverage
  • April 10, 2026 - Taboola's DeeperDive reaches nearly 7 million monthly active users and expands to six languages. PPC Land coverage
  • April 23, 2026 - Taboola launches Realize+, an agentic AI system for open web performance campaigns, alongside Claude Skills integration. PPC Land coverage
  • April 28, 2026 - LiveRamp adds NVIDIA GPU infrastructure to its clean rooms, enabling up to 15x faster AI model training. PPC Land coverage
  • May 6, 2026 - Criteo reports Q1 2026 activated media spend exceeding $1 billion for the first time; Taboola reports Q1 revenue of $466.4 million (+9.1%). PPC Land Criteo coverage - PPC Land Taboola coverage
  • June 16, 2026 - WPP Media publishes midyear forecast projecting global advertising to reach $1.3 trillion in 2026, naming AI investment as the primary driver.
  • June 17, 2026 - DoubleVerify launches DV Neura, its cognitive AI engine, with agentic execution capabilities and a nearly 300x increase in content classification output. PPC Land coverage
  • June 18, 2026 - NVIDIA publishes details of six advertising and marketing technology partnerships to be showcased at Cannes Lions 2026 (June 22-26).

Summary

Who: NVIDIA, alongside six advertising and marketing technology partners - Alembic, Amazon Web Services, Criteo, HiggsField, KERV.ai, and Taboola - are the central parties. World Wide Technology and Equinix are named as infrastructure partners in the Alembic deployment. PYLER is named as a partner adopting Nemotron 3 Nano Omni for video understanding.

What: NVIDIA published details of GPU-powered infrastructure deployments across four distinct advertising problem areas: causal marketing measurement (Alembic, using DGX Vera Rubin NVL72 SuperPODs), real-time AI bidding (AWS, using Triton Inference Server; Criteo, using cuEmbed on Blackwell GPUs with a 2x training speedup and approximately 17,000 GPU hours freed per year), full-lifecycle marketing automation (HiggsField, using NVIDIA Agent Toolkit including NemoClaw and OpenShell), and video content intelligence (KERV.ai, using Nemotron 3 Nano Omni with over 10x speed improvements). Taboola is also named, applying NVIDIA GPU infrastructure to its DeeperDive AI answer engine.

When: NVIDIA's announcement was published on June 18, 2026. The Cannes Lions International Festival of Creativity, where these partnerships are being showcased, runs from June 22 to 26, 2026, in Cannes, France.

Where: Cannes Lions 2026 in Cannes, France. Alembic's inference runs on private supercomputing infrastructure inside Equinix data centers. HiggsField campaigns serve Fortune 500 companies globally. KERV.ai's MediaPerf benchmark is an open evaluation platform. Taboola's DeeperDive operates across its network of more than 9,000 publisher sites.

Why: The advertising industry's shift from rules-based to AI-powered decisioning requires infrastructure capable of running complex models at auction speed and enterprise scale. Demonstrating these deployments at Cannes Lions - the industry's principal annual gathering for senior marketing and media executives - allows NVIDIA and its partners to position GPU-accelerated computing as the underlying requirement for autonomous advertising operations at the scale the industry now demands.