PayPal’s historic BigQuery migration: why moving 300+ petabytes is really an AI strategy (not just a database project)

When a company says it migrated “more than 300 petabytes” of analytics data with “zero downtime,” my default reaction is to check whether my coffee has been replaced with an energy drink. Then I read the details, and the story gets even more interesting: PayPal’s leadership is explicitly framing an enormous data-warehouse consolidation as the prerequisite for its next era of generative AI products.

That’s not marketing fluff (well… it’s not only marketing fluff). Gen AI systems are famously allergic to fragmented, stale, poorly governed data. Retrieval-augmented generation (RAG), agentic commerce, fraud modeling, personalization, risk scoring—these all get dramatically harder when your “single customer view” requires a treasure map and three different SQL dialects.

In a new post on the Google Cloud Blog, PayPal SVP & Global Head of Data, AI & ML Technology Mani Iyer and Sr Director Data Analytics Vaishali Walia describe how PayPal consolidated its analytics footprint—spanning legacy Teradata, Hadoop, and multiple cloud warehouses—culminating in a move of analytics to BigQuery. The post is dated February 26, 2026, which (conveniently for all of us) is also the day this topic started ricocheting through my group chats. citeturn1view0

The big number: 400 PB fragmented, 300+ PB migrated

Let’s anchor on the figures PayPal itself published:

  • Roughly 400 petabytes of data spread across about a dozen siloed systems, a situation PayPal attributes to both scale limitations and acquisitions (including Venmo and Braintree). citeturn1view0
  • With help from Google Cloud Consulting, PayPal says it migrated more than 300 petabytes, decommissioned around 25% of workloads, and did so with zero downtime and “no impact to customers.” citeturn1view0
  • The endpoint: a unified analytics platform in BigQuery, framed as a foundation for PayPal’s gen AI roadmap. citeturn1view0

“300 petabytes” can be a numbing number, so here’s the journalist’s translation: if your org has ever debated whether to partition a 2 TB table, PayPal is playing a different sport. At this size, the “data platform” is not a backend service—it’s a geopolitical entity.

Why this matters now: gen AI punishes data sprawl

PayPal’s post reads like a classic modernization narrative—until you notice what’s driving the urgency. It’s not just cost. It’s not just “faster queries.” It’s the realization that gen AI is a data product, and data products don’t thrive in a balkanized analytics landscape.

PayPal’s leadership says fragmentation limited personalization and made it difficult to unify experiences across brands (a small business might use PayPal for online sales and Venmo for local transactions, for example). citeturn1view0 That’s the customer-facing angle. The AI-facing angle is harsher:

  • Training data freshness is critical for many AI use cases (fraud patterns change, merchant behavior changes, seasonal effects happen, regulations shift).
  • Feature engineering gets messy when features are defined differently across platforms.
  • Governance becomes a bottleneck when every system has its own access model, lineage story, and metadata quality.
  • RAG and enterprise search require high-quality indexing and consistent identifiers across domains (customers, merchants, devices, sessions, risk signals).

PayPal doesn’t claim it solved every AI challenge on earth, but it does claim the migration produced 16x fresher data for model training and improved feature engineering through “instant access to clean, governed data.” citeturn1view0 That “freshness” claim is especially notable, because it’s the kind of metric teams track when they care about model behavior—not just BI dashboards.

BigQuery as the landing zone: the architecture choice behind the headlines

PayPal lists familiar reasons for choosing BigQuery: fully managed, cloud-native, independent scaling of compute and storage, SQL accessibility, and native AI integration. citeturn1view0turn2search6

For context, Google describes BigQuery’s architecture as a separated storage layer and compute layer, which can operate independently. That separation is one reason BigQuery is often positioned as a fit for large-scale analytics where workloads vary heavily over time. citeturn2search6

The other key part is AI proximity. Google has spent the last couple of years tightening the coupling between BigQuery and Vertex AI so teams can run ML inference and even call foundation models from within BigQuery workflows—often via SQL constructs—reducing the need for custom glue code or large-scale data movement. citeturn2search1turn2search4

That coupling isn’t just “nice to have” for gen AI; it addresses a recurring enterprise failure mode: the AI team prototypes something using a copy of data, the governance team has a small heart attack, and then the project stalls. Running more of the workflow “where the data lives” is a practical attempt to keep things governable.

The Teradata factor: the migration everyone fears

PayPal specifically calls out its “now sunset Teradata system” and suggests that, by some measures, the effort may be one of the largest data migrations in history—highlighting what it describes as the world’s largest Teradata deployment. citeturn1view0

Teradata has been a heavyweight in enterprise analytics for decades. If your company grew up in the era when “data warehouse” meant “big iron, big licensing, big governance meetings,” chances are Teradata was either the platform—or the platform you were trying to replace while promising you’d do it “safely and gradually.” PayPal’s decision to consolidate away from a multi-platform world that included Teradata, Hadoop, Redshift, Snowflake, and “various other systems” is a reminder that even digital natives can accumulate mainframe-scale complexity. citeturn1view0

Also, let’s be honest: migrating off Teradata isn’t just a database project. It’s an organizational change project. It forces you to confront:

  • SQL dialect differences and function parity
  • Workload scheduling patterns
  • Cost attribution and chargeback models
  • Access control conventions
  • Data quality rules that existed only in someone’s 12-year-old stored procedure

The PayPal post doesn’t get into the gory technical edge cases (sadly for people like me who read migration runbooks recreationally), but it does emphasize four success factors: alignment, discovery/lineage, strategy, and execution with automation plus dashboards—plus FinOps embedded through the process. citeturn1view0

FinOps wasn’t a footnote—it was a prerequisite

PayPal says it integrated FinOps into the migration with clear visibility of consumption and performance. citeturn1view0 In a cloud migration of this scale, that’s not optional unless you enjoy surprise invoices and executive “alignment meetings” that feel like performance art.

The FinOps Foundation defines FinOps as an operational framework and cultural practice aimed at maximizing business value from cloud, enabling data-driven decisions, and creating financial accountability via collaboration between engineering, finance, and business teams. citeturn4search0

In other words: when you move analytics to a consumption-based model, you need mechanisms that let teams move fast without turning cost into a taboo topic. The most mature FinOps programs don’t just optimize spend; they help decide where higher spend is justified because the business value is real.

What PayPal claims it gained: speed, freshness, fewer vendors

PayPal lists several outcomes worth pulling apart:

  • 2.5x to 10x faster queries, including complex data science queries. citeturn1view0
  • 16x fresher data accessible for model training. citeturn1view0
  • Vendor consolidation: reducing data infrastructure vendors from four to one.
  • Eliminated data duplication between platforms. citeturn1view0

These are the kinds of results that tend to compound. Faster queries aren’t just a developer convenience; they increase experimentation velocity. Fresher data isn’t just an engineering metric; it reduces the lag between “something changed in the world” and “our models and analytics acknowledge reality.”

And vendor consolidation—while it can create platform concentration risk—can also simplify governance, reduce integration overhead, and make it easier to enforce consistent controls. In heavily regulated spaces like financial services, reducing the number of analytics surfaces where sensitive data can leak is often a quiet win.

Gen AI innovation: what becomes possible with unified analytics

PayPal is careful not to announce a specific “PayPalGPT does your taxes” product in this post. Instead, it frames the unified data foundation as enabling a range of AI-powered experiences, including:

  • Predictive fraud prevention
  • Personalized financial insights for merchants
  • Seamless payment experiences that adapt to customer patterns
  • More intelligent risk assessment to expand access to underserved communities

Those are broad categories, but they map neatly onto where AI has real leverage in payments: risk, trust, personalization, and workflow automation. citeturn1view0

Fraud: the arms race doesn’t pause for your data migration

Payments fraud is a live-fire domain: attackers iterate quickly, and defenders need strong signals, fast feedback loops, and consistent identity resolution across touchpoints. A unified analytics platform can help here in mundane but crucial ways:

  • Reduce time to build and backtest features
  • Make label generation and event correlation less painful
  • Improve monitoring for model drift
  • Enable near-real-time analytics that can feed scoring systems

This is also where “freshness” matters. If data that reaches feature stores or training sets is stale, models learn yesterday’s fraud patterns—great for historians, less great for chargeback rates.

Personalization and merchant insights: the “single view” problem

PayPal’s ecosystem spans consumer payments, merchant processing, and brands with their own histories. Braintree and Venmo, for instance, arrived via acquisition (Braintree acquired Venmo in August 2012). citeturn3search0 PayPal then acquired Braintree in 2013 (which brought Venmo along for the ride), a move widely discussed in payments industry history. citeturn3search1

Whenever you stitch together products with different data models, you inherit identity mismatches and schema drift. The practical consequence is that personalization becomes expensive and slow—exactly what PayPal cites as a problem when trying to unify views for small businesses using multiple PayPal-family products. citeturn1view0

Gen AI can help teams explain trends, summarize business health, or draft customer support responses—but it still needs a coherent underlying record of “what happened” and “to whom.” If customer and merchant entities aren’t resolved consistently, the AI layer becomes a very confident storyteller who has mixed up the characters.

Industry context: every fintech is chasing AI, but data gravity wins

The broader financial services sector is in a race to deploy gen AI in ways that are safe, explainable enough for compliance, and economically viable at scale. PayPal’s own scale is enormous; public financial reporting and analysis often cites hundreds of millions of active accounts and processing volumes in the trillions of dollars annually (figures vary by fiscal year and reporting definitions). citeturn3news14

At this size, AI strategy is inseparable from infrastructure strategy. If your analytics foundation can’t reliably serve governed data to teams, you can still do AI—but you’ll do it in pockets, with duplicated pipelines and inconsistent controls. That’s the “shadow AI” problem: lots of pilots, few platform-level capabilities.

Agentic commerce is coming for checkout flows (and data foundations will decide who wins)

PayPal’s Google Cloud post mentions “agentic commerce” as part of the “future possibilities” it can imagine with a unified data platform. citeturn1view0 That phrase is not theoretical in PayPal’s world.

On October 28, 2025, PayPal announced a partnership with OpenAI to support Instant Checkout and adopt the Agentic Commerce Protocol (ACP), aiming to bring PayPal-powered checkout into ChatGPT experiences starting in 2026. citeturn4search1turn4news12

PayPal also describes “agentic commerce services” as solutions that support product discovery, cart management, and payment processing via natural-language interactions—essentially making commerce programmable for AI agents. citeturn4search2

Here’s the connective tissue: if agentic commerce becomes a mainstream interface, the winners will be companies that can serve accurate, real-time, policy-compliant product and transaction information to AI systems. That requires clean catalogs, consistent merchant identifiers, reliable telemetry, and fast analytics feedback loops. In other words, it requires the kind of unified data foundation PayPal is arguing for.

The migration playbook: what other enterprises can steal (legally)

PayPal’s post provides a high-level playbook that’s worth translating into practical advice for other large organizations contemplating similarly terrifying modernization efforts.

1) Alignment: make it enterprise-wide (or don’t bother)

PayPal says stakeholder alignment was the first hurdle and made the migration an enterprise-wide priority. citeturn1view0 In practice, this means incentives and timelines have to be aligned across:

  • Data engineering
  • Analytics/BI
  • Security & privacy
  • Finance
  • Product teams who “own” the data

If any one of those groups can effectively veto or delay the program without consequence, you’ll end up with a forever-migration—an oddly common phenomenon.

2) Discovery and lineage: dependency graphs are your real scope

PayPal calls out detailed inventories of data, workloads, streams, and lineage to map dependency graphs. citeturn1view0 This is where migrations often fail: teams underestimate the number of downstream consumers, undocumented ETL jobs, and critical “temporary” tables that have been powering monthly reporting since 2017.

Lineage work can feel like archaeology, but it’s the difference between a controlled cutover and an incident where your CFO discovers the “new platform” doesn’t reconcile to the old one.

3) Strategy: lift-and-shift vs. modernization isn’t binary

PayPal frames strategy choices in terms of lift-and-shift versus modernization, along with security principles, governance guardrails, and consumption tracking. citeturn1view0

In most large migrations, you end up with a spectrum:

  • Lift-and-shift for low-value workloads you just need off the old platform
  • Partial refactors where SQL changes are required for performance or compatibility
  • Full modernization for pipelines that will become AI feature generation or near-real-time analytics

The trick is to decide intentionally, not accidentally.

4) Execution: automate everything, instrument everything

PayPal emphasizes automation and live dashboards to monitor progress. citeturn1view0 That sounds obvious until you’ve seen a migration tracked in a spreadsheet called “final_final_v7.xlsx.”

At petabyte scale, you need programmatic progress tracking, repeatable validation, and rollback plans that assume something will go wrong. Not because your engineers aren’t good, but because physics is undefeated.

Security, privacy, and compliance: the unspoken center of gravity

One of the more impressive claims in PayPal’s write-up is “zero downtime” and “no impact to customers” during a migration involving extraordinarily sensitive financial and identity-adjacent data. citeturn1view0

PayPal doesn’t detail its security architecture in the post, but in financial services the constraints are well known:

  • Access controls must be fine-grained and auditable
  • Encryption and key management are non-negotiable
  • Data residency and regulatory requirements vary by geography
  • Operational resilience requirements push teams toward phased cutovers and dual-running

These constraints are part of why “just move to the cloud” is rarely a simple rewrite. The bigger point: if PayPal is betting its gen AI future on this data foundation, then governance and safety controls aren’t add-ons—they’re the enabling feature.

What to watch next: the second-order effects

Migrations of this magnitude usually produce benefits in waves. The first wave is consolidation and performance. The second wave is platform reuse: teams build new capabilities because the friction is gone.

Here are a few plausible next chapters—framed as what to watch for, not what PayPal has promised:

  • Richer internal data products: standardized “gold” datasets for merchants, consumers, risk, and disputes.
  • AI-assisted analytics: more natural-language-to-SQL tooling and governed LLM access patterns where prompts and outputs can be audited.
  • Operational analytics feeding real-time decisioning: tighter coupling between observability, risk systems, and business metrics (PayPal has previously written about streaming analytics modernization too). citeturn0search0
  • Agentic commerce instrumentation: if PayPal commerce flows expand into AI surfaces like ChatGPT, measuring conversion, fraud, and user experience in those new channels becomes an analytics problem—again.

A reality check: big cloud consolidations have trade-offs

No migration story is complete without acknowledging what can go wrong:

  • Platform concentration risk: fewer vendors can mean fewer failure domains, but also fewer escape hatches.
  • Cost surprises: serverless doesn’t mean free; it means you need disciplined governance (hello, FinOps).
  • Skill transitions: teams must adapt to new tooling, new performance patterns, and new operational practices.
  • Data product debt: if you move messy data faster, you can end up with a faster mess.

PayPal’s leadership argues the trade-offs are worth it because the alternative—trying to scale and unify on-prem and multi-warehouse analytics while the gen AI ecosystem is accelerating in the cloud—would have been too slow and too expensive. citeturn1view0

Bottom line

PayPal’s BigQuery consolidation is a reminder that, in 2026, gen AI isn’t merely “model selection.” It’s data readiness at industrial scale. If your company wants to build AI features that are accurate, safe, and responsive to real-world changes, you have to solve the unglamorous problems first: data fragmentation, lineage, governance, and cost accountability.

PayPal is betting that migrating 300+ petabytes and simplifying its analytics stack is the foundation for everything that comes next—fraud prevention, personalization, merchant intelligence, and agentic commerce that turns chats into checkouts. The hilarious part is that the “historic database migration” may end up being remembered less for the migration itself, and more for the products it finally made possible.

Sources

Bas Dorland, Technology Journalist & Founder of dorland.org