Pinterest today announced a planned $4 billion commitment to Amazon Web Services through 2031 - the largest infrastructure deal in the company's history - covering AI model training, inference, and platform infrastructure for a visual discovery platform that serves more than 600 million monthly users worldwide.

The agreement, disclosed on June 4, 2026, extends a partnership between Pinterest and AWS that dates to 2010. Together, the two companies have built one of the largest-scale data lakes on AWS infrastructure. The renewed deal is designed to support what Pinterest describes as the next phase of growth across three technical domains: AI model training, model inference, and broader platform infrastructure.

A 16-year partnership enters a new phase

Pinterest's relationship with AWS predates the current wave of AI investment by over a decade. What began as a cloud hosting arrangement has, according to Pinterest, matured into a joint engineering effort at significant scale. The $4 billion figure, spread across the period ending in 2031, represents a formalization and expansion of that dependency - not a new relationship started from scratch.

The timing is notable. Pinterest reached 600 million monthly active users in the third quarter of 2025, its ninth consecutive quarter of record user growth, with revenue expanding 17% year-over-year to $1.049 billion in that period alone. The platform's Q1 2026 results, reported on May 5, 2026, showed 631 million monthly active users and $1.008 billion in revenue, an 18% year-over-year increase. Serving a user base of that size, with AI features operating across every query, requires compute infrastructure at a scale that demands long-term planning and committed capacity agreements.

AWS custom silicon at the center

The technical core of the deal involves two specific chip families developed by Amazon. Pinterest plans to use AWS Trainium to host and run the large language models and vision-language models that power personalized visual search and AI-assisted discovery on the platform. Trainium is Amazon's custom chip designed specifically for machine learning training workloads, built to reduce the cost and time required to train large-scale models compared to general-purpose GPU alternatives.

The second hardware component is AWS Graviton. According to Pinterest, Graviton already powers roughly a third of its compute infrastructure. The new agreement expands that footprint further, with Graviton handling more of the systems that support discovery for hundreds of millions of users every month. Graviton processors use an Arm-based architecture and are designed to deliver better price-performance ratios for a broad range of cloud workloads - meaning general compute tasks, web serving, and the inference-side delivery of AI model outputs to users.

The combination of Trainium for training and Graviton for inference workloads reflects a deliberate hardware strategy. Training a large model and then running it at scale for hundreds of millions of daily users are computationally distinct problems. Trainium addresses the first; Graviton addresses the second. Using Amazon's custom silicon for both stages reduces reliance on third-party GPU suppliers and, according to the companies, delivers infrastructure efficiency gains that matter at Pinterest's scale.

"Pinterest is heavily investing in AI to make discovery more personal, visual, and actionable for the hundreds of millions of people who use our platform every month," said Mat Madrigal, chief technology officer at Pinterest. "This expanded commitment with AWS gives us the compute flexibility, hardware optionality, and infrastructure efficiency to accelerate our AI vision for the next generation of visual discovery on Pinterest. This strategic partnership will help accelerate AI innovation at Pinterest, improving both our consumer experience and advertiser performance by advancing our proprietary models and our use of open-source models."

Dave Brown, senior vice president of AWS Compute and ML Services, added: "AWS is the best place to do AI at this scale, and we're committed to helping Pinterest's teams move faster and think bigger - benefiting users all over the world."

The Taste Graph and the models it requires

Pinterest's underlying AI system is its proprietary Taste Graph, a multimodal AI structure trained on billions of images, videos, and text. The Taste Graph powers recommendation systems across the platform, mapping connections between user interests and content to generate personalized discovery feeds. Its role extends beyond organic content discovery - the same signals feed Pinterest's advertising products, including Performance+, the automated campaign suite the platform launched in October 2024.

Performance+ early results showed a 10% improvement in cost per acquisition for most advertisers, with campaign creation time cut by 50% through reduced required inputs. By Q1 2026, Performance+ accounted for 30% of Pinterest's lower-funnel revenue. The Taste Graph is also central to Promote a Pin, the simplified self-serve tool launched March 24, 2026, which uses the graph for automatic targeting without requiring advertiser expertise.

The compute demands of maintaining and improving the Taste Graph - retraining it against new data, running inference across every user session, and updating recommendation signals continuously - are precisely the workloads that the expanded AWS deal is designed to support.

More recently, Pinterest launched Pinterest Assistant, described in the deal announcement as bringing multi-turn conversational discovery to its visual search experience. The assistant is powered by open-source vision-language models optimized for scale. It had been in beta rollout since Q4 2025, with Pinterest's CEO Bill Ready emphasizing deliberate pacing in the expansion of its availability. Pinterest Assistant represents a shift from static retrieval toward interactive, dialogue-based discovery - a user can refine a search through conversation rather than reformulating text queries manually.

From transformer models to Kubernetes

Pinterest's AI architecture has shifted substantially in recent years. According to the announcement, the company has evolved from traditional retrieval methods - essentially matching queries to content through keyword or feature similarity - to transformer-based generative models. Transformers are the architecture underlying most of the large language models that have achieved broad commercial adoption since 2022, and their application to visual discovery requires handling both image and text data within unified model structures. That is what vision-language models do: they process images and text together, enabling the system to understand what a user means when they pin a home decor image alongside a typed aesthetic description.

Running transformer-based generative models at Pinterest's scale places different demands on infrastructure than running the simpler retrieval systems they replace. Transformer inference is computationally heavier per query, meaning that serving 600 million monthly users generates substantially more compute load than equivalent traffic through an older retrieval system.

Alongside the AI-specific investments, Pinterest is using the expanded agreement to continue a separate infrastructure modernization effort. The company is transitioning from traditional Amazon EC2-based environments to a Kubernetes-based architecture running on Amazon Elastic Kubernetes Service (EKS). According to Pinterest, this migration is expected to improve developer velocity, operational reliability, and infrastructure efficiency across the global platform.

Kubernetes is an open-source container orchestration system that automates the deployment, scaling, and management of software containers. Moving workloads from EC2 instances - which require more manual provisioning and configuration - to EKS simplifies operations at scale, allows faster deployment cycles, and improves the efficiency of resource allocation across a large fleet of services. For a company running AI inference continuously across a global user base, those operational gains compound over time.

Infrastructure context for the marketing community

The deal carries direct implications for advertisers using Pinterest's platform. AI model quality determines recommendation relevance, and recommendation relevance determines whether advertising inventory performs. Pinterest's shift toward Trainium-powered model training means the visual search and personalization models underpinning ad targeting will be retrained and improved on infrastructure built for exactly that workload.

Pinterest has positioned itself as a commerce destination since at least 2024. The tvScientific acquisition, announced December 11, 2025 and closed in Q1 2026, extended Pinterest's audience data and Taste Graph signals into connected TV advertising. An early home furnishings partner reported a 190% increase in incremental audience reach and a 159% increase in incremental sales after layering Pinterest audience data onto CTV campaigns, though these figures came from a single disclosed partner in early testing. CFO Julia Donnelly confirmed in the Q1 2026 earnings call that infrastructure costs in Q2 2026 would reflect the full-quarter impact of tvScientific and investment in GPU capacity, suggesting that infrastructure spending is already accelerating ahead of today's announcement.

Pinterest's January 2026 restructuring reduced approximately 780 roles while explicitly reallocating resources toward AI functions. The AWS deal provides the cloud infrastructure counterpart to that internal headcount shift - fewer engineers managing legacy systems, more compute running AI workloads.

For practitioners using Pinterest's ad products, the infrastructure investment matters most at the level of targeting precision and ad relevance. The Taste Graph determines which users see which ads, and the accuracy of those signals depends on how frequently and effectively the underlying models are retrained. AWS Trainium, according to Amazon's public documentation on the chip family, is designed to make that retraining faster and less expensive than equivalent GPU-based approaches - which in practice means Pinterest can iterate on its models more frequently, potentially improving ad performance at a pace that pure GPU infrastructure would not support as cost-effectively.

The Graviton expansion - from roughly one-third of compute infrastructure to a larger share - affects the inference side. Every time a user loads a Pinterest feed or searches for a product, the platform runs inference across its recommendation models to generate a personalized result. Faster, cheaper inference translates directly into lower per-query costs, which at 600 million monthly users amounts to a substantial line item. More efficient inference also enables Pinterest to run larger, more capable models without proportionally increasing infrastructure costs, which is the practical argument for custom silicon over commodity cloud compute.

PPC Land has tracked AWS's role in advertising infrastructure, reporting in November 2025 on Project Rainier - AWS's 500,000-chip Trainium2 cluster - where Trainium2 had reached full subscription status and become a multibillion-dollar business growing 150% quarter-over-quarter. Pinterest's expanded use of Trainium places it among the wave of technology companies committing to Amazon's custom silicon as an alternative to Nvidia GPU-dominated training infrastructure.

Scale and competitive positioning

The $4 billion figure over roughly five years implies roughly $800 million per year on average in AWS cloud commitments, though the actual distribution across the period was not disclosed. Pinterest's annual revenue in 2024 was $3.65 billion, meaning the infrastructure commitment represents a substantial ongoing proportion of the company's revenue base. The commitment signals confidence in the company's revenue trajectory and in the centrality of AI workloads to its product strategy.

Pinterest is not alone in making large cloud commitments for AI infrastructure. The combination of a decade-long existing relationship with AWS, the specific adoption of Trainium and Graviton rather than general-purpose GPU infrastructure, and the parallel Kubernetes migration marks this deal as an engineering-led infrastructure strategy rather than a straightforward capacity purchase. Whether Trainium-based training delivers the cost and performance advantages Pinterest requires to compete on AI model quality will become clearer as the platform's recommendation and advertising products develop through the remainder of 2026 and into 2027.

Timeline

Summary

Who: Pinterest, the visual discovery platform (NYSE: PINS), and Amazon Web Services (AWS), a division of Amazon.com. Statements are attributed to Mat Madrigal, Pinterest's chief technology officer, and Dave Brown, AWS senior vice president of Compute and ML Services.

What: A planned $4 billion cloud services commitment from Pinterest to AWS, running through 2031. The deal covers AI model training using AWS Trainium chips, inference and general compute using AWS Graviton processors (which currently run approximately one-third of Pinterest's infrastructure), and a migration from EC2-based environments to a Kubernetes architecture on Amazon EKS. The deal also supports training and deployment of the vision-language models powering Pinterest Assistant and the Taste Graph recommendation system.

When: The announcement was made on June 4, 2026. The commitment covers cloud services through 2031. The AWS partnership itself dates to 2010.

Where: Pinterest is headquartered in San Francisco, California. AWS infrastructure operates globally. Pinterest's platform serves users across international markets, with Europe and Rest of World regions growing faster in proportional terms than the US and Canada in recent quarters.

Why: Pinterest's shift from traditional retrieval models to transformer-based generative models - including multi-turn conversational discovery through Pinterest Assistant and multimodal visual search - requires substantially more compute for both training and inference than legacy systems. The custom silicon available through AWS Trainium and Graviton is intended to deliver the price-performance efficiency needed to run those workloads at scale across 600 million monthly users, while the Kubernetes migration reduces operational overhead for the engineering teams managing that infrastructure.