Publicis Buys LiveRamp for $2.4B: AI's Data Power Play

So, What Actually Happened?

Monday morning the desks are still cleaning up a weekend that handed strategy teams three named operating tracks at once. Publicis agreed to acquire LiveRamp, pulling identity, consent, and addressable data into a single ad-agency spine built for agent-driven marketing. Babcock & Wilcox locked a $2.4B reactor-class order with Base Electron, and the AI buildout just put a forgotten power vendor back in the bid stack. We scanned 190,000 articles this week so you don't have to. Meanwhile DayOne filed for the first dual US-Singapore IPO for a data-center operator, naming Asia as a co-headline market for AI-capacity capital alongside Wall Street.

The Bottom Line: When an agency holding company buys an identity-resolution platform, when an old-economy power vendor lands a billion-dollar AI-driven order, and when a data-center operator files in two markets on the same week, the AI operating layer is not waiting for the next model release. It is locking down the data, the power, and the listings before the next earnings cycle. The strategy lead walking into Tuesday with a one-page map of who owns which of those three layers in their own stack runs the next quarter. The rest will be reading from a Q1 deck.

The Tracks That Matter

1. Publicis Buys LiveRamp To Build The Agent Data Spine

The clearest agentic-marketing signal of the weekend sits on a Publicis press release most CMOs will skim as agency-holding-company news. Publicis Groupe is acquiring LiveRamp, the identity-resolution and data-collaboration platform that sits inside hundreds of brand stacks, with the explicit thesis of building a ”smarter-agent” data spine: a single, consented, addressable customer layer that an AI agent can act against, not just a CDP that humans query.

The pairing matters because it sharpens what ”agent-ready data” really means. A CMO can buy a fashionable agent product from any vendor in a week, but if the underlying customer data is fragmented across a CDP, a DSP, a CRM, and three pixel partners, the agent ends up doing parlor tricks instead of operating spend. Publicis just priced the data-spine assumption at acquisition scale. The agency model itself is being rewritten: not ”we run the campaign for you” but ”we own the identity, consent, and audience layer your agents work on top of.”

The strategic implication: the CMO and head of data just gained a ”named agent-grade data foundation” line on the operating scorecard that did not exist on Friday. The question is no longer ”which agent platform do we trial.” It is whether the agent has a clean, consented, persistent identity layer to act on, or whether it is renting one back from the same holding company that runs the brand's media plan.

Here's what works: Ask the CMO, CIO, and head of data together: for our top three agent-driven workflows (paid acquisition, lifecycle email, in-product personalization), do we have a named owner for the identity and consent layer that those agents operate against, or is it spread across four vendors with no single audit owner? Publicis just put a price on that question for the segment.

2. Babcock & Wilcox Books $2.4B Order As AI Reawakens Old Power

The most underweighted energy signal of the weekend sits on a brief about Babcock & Wilcox jumping on a $2.4B Base Electron contract backed by Applied Digital's AI-data-center push. A 159-year-old combustion-and-boiler company most CFOs do not have on their vendor list just got a contract larger than its entire 2024 market cap, because the AI data-center buildout needs every megawatt of dispatchable, on-site power it can find.

The contrast that sharpens the read is what this same vendor was selling five years ago: legacy industrial steam plants and small-modular nuclear concepts most utilities had quietly shelved. The ”AI infrastructure is built by hyperscalers plus a couple of neoclouds” mental model that most procurement teams still walk into renewal with does not have a row for ”159-year-old combustion vendor signs a multibillion-dollar power-supply contract because the data-center workload outran the grid.” It does now.

The strategic implication: the head of platform, the COO, and the CFO just gained a ”named on-site power line” on every new AI workload assessment. The question is no longer just ”which cloud region.” It is whether the workload has a behind-the-meter or co-located power partner with a named delivery timeline, and what the cost-per-megawatt-hour curve looks like against the standard rate card. The energy stack is now part of the AI stack, and old vendors are back in the bid.

Here's what works: Ask the COO and head of platform together: for our top three power-heavy AI workloads (training runs, inference at scale, retrieval-heavy agent pipelines), is there a named on-site or behind-the-meter power partner on the file, or is the answer ”we assume the grid will hold”? The AI buildout just made that a board-grade question.

---

Try It Yourself

3. DayOne Files First Dual US-Singapore IPO For Data Centers

The sharpest capital-markets signal of the weekend sits on a News.com.pk brief on DayOne's dual-listing plan most CIOs will scroll past as Asia-finance commentary. DayOne, the GDS Holdings-affiliated data-center operator, is filing simultaneously in New York and Singapore, naming both the US deep-pool capital market and the Asia AI-demand market as co-headline venues, not the usual ”list in NY, ADR-listing in Asia later” sequence.

The contrast that sharpens the read is what most data-center listings have looked like the past decade. A single-market IPO is the default; a dual listing of an AI-infrastructure operator at the same scale on opposite sides of the world signals that the demand side, the buyer pipeline of hyperscaler and enterprise AI capacity contracts, no longer fits inside one regulatory perimeter. Coatue is backing it. The ”AI capacity is a US-headquartered conversation” framing that quietly shaped most 2024 procurement decks just got its first named cross-listing counter-reference.

The strategic implication: the CFO and head of platform just gained a ”named Asia-region AI capacity option” line on the procurement scorecard. For two years, the answer to ”where do we put the next AI workload” has been ”us-east-1 or us-west-2.” When a Singapore-listed AI-data-center operator goes public next to a New York listing, the procurement chart gets a third and a fourth named region with public-markets-grade transparency on cost and capacity.

Here's what works: Ask the CFO, head of platform, and chief data officer together: for our top three latency-sensitive or sovereignty-sensitive AI workloads, do we have a named Asia-Pacific or sovereign-region option on the file, with a cost line and a contracted-capacity number? The capital markets just priced the optionality question for the segment.

4. South Korea Pivots From Build AI To Monetize It

The Korea-specific signal of the weekend sits on a Financial News Korea brief about ”Now It's Time to Monetize AI” most US-focused execs will treat as regional color. The piece names a posture shift inside the Korean tech sector: from the 2022-2025 era of ”build the LLM and the foundation model” to a 2026 chapter explicitly framed as ”monetize what we built,” with global Big Tech flocking south to plug into Korean infrastructure, semiconductor, and enterprise-AI demand.

The pairing matters because it lines up with what South Korea's national R&D budget did two weeks ago when AI was named the first-position priority. Build-then-monetize is a national policy now, not just a vendor pitch. Most US-headquartered enterprise AI strategies still have one row labelled ”Asia” with a single named partner. The Korean signal is that the row should fork into ”semiconductor supply,” ”model and chip codesign,” and ”enterprise distribution partner,” each with a named owner.

The strategic implication: the CIO, head of strategy, and CFO just gained a ”named Korea-specific AI partner per stack layer” line on the operating scorecard. For two years, the answer to ”what's our Korea strategy” has been ”we ship through a regional reseller.” After Korean infrastructure starts pricing monetization at scale, the question becomes whether your AI roadmap has a named Korean chip, model, or distribution partner on the file before the next architecture review.

Here's what works: Ask the CIO and head of strategy together: for our top three AI-capital and AI-supply-chain assumptions, is there a named Korean counter-reference (chip vendor, foundry partner, distribution partner) on the file, with a cost line and a delivery commitment? Korea just retired the ”ship-through-a-reseller” answer.

5. Malaysia Issues Three Data Protection Guides Ahead Of The Region

The quietest regulatory signal of the weekend sits on a Hogan Lovells brief about Malaysia's three new data protection risk management guides most GCs in North America will not see all week. The Malaysian Personal Data Protection Department issued formal guidance on data protection impact assessments, data breach notification, and data protection officer appointments, sliding Malaysia ahead of much of Southeast Asia on operating-level data-AI guardrails.

The contrast that sharpens the read is the pattern across the same fortnight: Connecticut passed AI employment rules, Colorado walked back its own AI Act, and the global cyber agencies issued SBOM-for-AI guidance. Then a mid-tier Asia-Pacific jurisdiction lands three formal data-AI guidance documents in a single drop. The ”regulators will catch up eventually, mostly in Europe and a couple of US states” framing that anchored most 2025 AI compliance plans just got its first named ASEAN-jurisdiction counter-reference.

Here's what works: Ask the GC, chief privacy officer, and head of regional ops together: for our top three operating jurisdictions in Asia-Pacific, do we have a named DPIA, breach-notification, and DPO file ready to satisfy Malaysian-style guidance, or are we still working off a 2024 GDPR checklist with regional notes? Malaysia just priced the gap.

6. Google Quietly Routes Users To A Hidden Gemini 3.2

The most under-reported model-strategy signal of the weekend sits on a LetsDataScience brief about Google routing users to a hidden Gemini 3.2 model. Without an announcement, without a press cycle, Google is silently routing a slice of consumer and enterprise traffic to a higher-capability Gemini variant that has not been formally launched. The model substitution is invisible to the user, and (more importantly) invisible to anyone benchmarking model performance against the published version.

The contrast that sharpens the read is what most enterprise AI procurement teams assume about a model contract. They assume the model named in the SLA is the model serving the traffic. Silent routing breaks that assumption. The ”we standardised on model X at version Y for our regulated workflows” CISO posture just got a named counter-reference from a hyperscaler. If the model behind your contract can be substituted without notice, every audit-trail, every red-team result, every behavioural eval was run against a moving target.

The strategic implication: the CIO, CISO, and chief compliance officer just gained a ”named model-version-pinning clause” line on every AI vendor contract. The question is no longer ”which model are we paying for.” It is whether the contract gives us a stable named version, a notification window before substitution, and a re-evaluation right when the routing changes underneath us.

Here's what works: Ask the CIO and CISO together: in our top three regulated AI deployments, is there a named model-version-pinning clause in the vendor contract with a notification-and-re-eval right, or are we paying for ”Gemini, OpenAI, Claude” as if those were stable products? Google just put the question on the procurement table.

7. Bad Data Just Killed Another Pilot, And Yours Is Next

The most uncomfortable operating signal of the weekend sits on an Australian Financial Review piece on bad data derailing AI. The piece names the pattern most boards do not have a row for: AI pilots are not failing on the model layer, the agent layer, or the GPU layer. They are failing on the data layer underneath, the same one most CIOs deprioritised in 2023 to chase the LLM headlines. Three years later, the proof-of-concept-to-production gap is mostly a data-quality gap, not a model-quality gap.

I have been DJing data warehouses long enough to know what this sounds like. It sounds like a turntable with a worn needle: the track is on, the speakers are loud, but the high end is gone and the crowd can hear it. You can pay for a better mixer (a better model) all day. If the source vinyl is scratched, the room thins out. The CIO still expecting the agentic-AI vendor to fix data quality at runtime is going to be the one explaining to the board why the third pilot in a row stalled at production.

Here's what works: Ask the CIO, chief data officer, and head of platform together: for our top three named AI pilots, is the data-quality file (lineage, completeness, freshness, schema-drift monitor) green, yellow, or red, and is the owner the same person who owns the AI pilot or a different person two layers away? The AFR just retired the ”we will clean the data after the pilot” answer.

Signal vs. Noise

🟢 Signal: Data Governance. Data governance climbed in real influence on the wires Sunday into Monday morning while the generic ”AI” label kept growing in volume but lost ground in named operating reach. Buyers are quietly moving the AI conversation from the model layer back down to the data layer, the same one most coverage stopped watching in 2024.

🔴 Noise: Regulatory Compliance As A Single Topic. Regulatory compliance pulled more than a hundred article mentions across the weekend wires but its real operating influence dropped sharply, because the conversation already split into named sub-tracks (employment AI rules, SBOM-for-AI, jurisdiction-specific DPIA guides, model-version pinning). Anyone still tracking ”AI regulation” as a single signal is reading from a 2024 keyword filter while GCs split it into five named files.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

Publicis bought LiveRamp to wire an agent-ready identity layer, Babcock & Wilcox locked a $2.4B power order for AI data centers, and DayOne filed a dual-market IPO for AI capacity, all inside one 72-hour window.

Each desk reads these as unrelated stories. The marketing trades cover Publicis. The power-and-utilities wires write up Babcock & Wilcox. The capital-markets press files the DayOne IPO. Read them on the same morning and a different picture emerges: three of the load-bearing layers underneath every enterprise AI roadmap (the data spine, the power supply, the regional capacity) just got publicly named price tags in the same news cycle. The ”AI strategy is a model-and-vendor selection problem” framing that anchored most 2025 board decks just got three concrete counter-references at once, each from a different sector.

The strategic move on Tuesday is mapping which of those three layers (data, power, capacity) has a named owner inside your own stack, and which one still says ”TBD” or ”owned by the platform team in general.” The weekend just priced the gap.

By The Numbers

• Babcock & Wilcox's stock jumped on a $2.4 billion contract with Base Electron, backed by Applied Digital, A 159-year-old combustion vendor lands a contract larger than its 2024 market cap because AI data centers need on-site power. The ”grid will hold” assumption is being repriced.

• Physical AI device shipments are projected at 145 million units by 2035, The category most enterprise CIOs do not yet have a line item for (humanoid, embodied, agent-controlled hardware) is being priced as a nine-figure unit market a decade out. Hardware roadmaps that are LLM-only will look thin by the next capex cycle.

• 77% of workers report AI has increased their workload, per a government-transformation report, Named workforce-impact stat against the ”AI frees up time” narrative. The change-management line item in the AI budget continues to be underpriced.

• 88% of heavy AI users report burnout symptoms, Named operating cost that does not appear in any vendor ROI deck. The ”AI productivity dividend” framing is being repriced inside the workforce.

• Allianz cut AI-driven insurance claims processing from days to minutes, Named operating result at one of the largest European insurers. The ”we will benchmark AI ROI after a 2026 pilot” CFO posture keeps getting lapped.

• Malaysia issued three formal data protection risk-management guides in a single drop, DPIA, breach notification, and DPO appointment, all formalised inside one ASEAN-tier jurisdiction. The ”EU plus a few US states” mental map of AI regulation is now genuinely out of date.

• See what's rising across the 190K-article corpus this week →

Deep Dive: The AI Stack Just Named Its Load-Bearing Layers Out Loud

Every DJ knows the moment in a set when you stop fighting the room and start listening to it. The crowd is telling you something. The drum that worked at 1am is dragging at 2am. The bassline that filled the floor an hour ago is now hitting underneath an empty space where dancers were. You can keep playing the headline track or you can pull the working tracks that the room is actually moving to. This weekend the AI economy did the second. The model lab funding rounds and the policy theatre took the night off. The data, power, and capacity layers underneath came up in the mix.

The Data-Spine Track

Publicis acquired LiveRamp for one explicit reason: the agency holding company that owns the identity, consent, and addressable-audience layer underneath an enterprise brand is the one who gets to sell agentic marketing as a managed service. Not the agent vendor. Not the LLM lab. The data-spine owner. The CMO still negotiating a fashionable agent contract without renegotiating the identity layer is going to find out in two quarters that the agent is doing parlor tricks against a fragmented CDP. Publicis just put a real acquisition price on the data-spine moat.

The Power Track

Babcock & Wilcox locked in a $2.4 billion power contract because Base Electron and Applied Digital cannot build the AI data-center capacity they have already pre-sold without dispatchable, on-site, behind-the-meter power. The ”AI capacity comes out of the regional grid like any other workload” assumption shaped most 2024-2025 enterprise procurement decks. After a 159-year-old combustion vendor lands a contract larger than its market cap, the assumption is fully retired. The COO who walks into the next architecture review without a named power-partner row on the AI scorecard is going to spend the rest of the year explaining to the CFO why the next training run got de-prioritised by the cloud region.

The Capacity Track

DayOne's dual US-Singapore IPO filing is the capital-markets equivalent of saying: the AI-capacity buyer pipeline is no longer a single-market conversation. Coatue and the data-center desks in both New York and Singapore are pricing it as a two-market story at the same scale. The ”us-east-1 or us-west-2” answer to every workload-placement question is one IPO closer to being insufficient for any enterprise that takes data-residency, latency, or sovereignty seriously. The CFO still treating AI capacity as a US-region budget line is going to be the one updating the cost-by-region table the day after the dual listing prices.

What Actually Works

1. Put a named owner on the data-spine layer per top customer-data workflow. Every agent-driven workflow (acquisition, lifecycle, in-product) gets one named owner for identity, consent, and addressable audience. Publicis just priced the data-spine reference for you.

2. Put a named on-site or behind-the-meter power partner on every power-heavy AI workload. Every training, inference-at-scale, or retrieval-heavy pipeline gets a named power partner with a delivery date and a cost-per-megawatt-hour line. Babcock & Wilcox priced the power reference for you.

3. Map a named Asia-Pacific or sovereign-region capacity option per latency- or sovereignty-sensitive workload. Every workload that has a data-residency clause or a sub-100ms latency target gets a named Asia or sovereign-region option with a public-grade cost number. DayOne priced the capacity reference for you.

4. Pin model versions in every regulated AI vendor contract. Every regulated AI deployment (finance, health, employment, claims) gets a named model-version clause with a notification window and a re-eval right. Google just priced the routing-substitution risk for you.

The headlines will come back. They always do. The model labs will print another round, another benchmark, another safety paper. The room is still moving. The operator who walks into Tuesday with the data, power, capacity, and version-pinning rows already on the dashboard is the one mixing for the rest of the year. The one who waits for the headline track is going to play to a thinner floor by Q3.

What's Coming

The First Named US Agency Holding Company With An Agent-Grade Identity Platform On Its P&L

Publicis-LiveRamp is the trigger. The next move is the first peer agency holding company (Omnicom, WPP, Interpublic, or Dentsu) publicly disclosing a named agent-grade identity platform on its earnings call, with a contribution-to-revenue line and a named brand-side reference customer. That filing is probably one to two cycles out, and the CMOs who already drafted a named data-spine owner on their side absorb the news as routine.

The First Named On-Site Power Partner Listed In A Top-50 Enterprise's AI Architecture Filing

Babcock & Wilcox's $2.4B order is the trigger. The next move is the first non-utility, non-hyperscaler enterprise (financial services, pharma, logistics, defence) disclosing a named on-site or behind-the-meter power partner in an SEC filing or analyst-day deck, with a megawatt number and a delivery date. That disclosure is probably one to two cycles out, and the COOs who already have a named power-partner row absorb the news as routine.

The First Named Hyperscaler Contract With An Explicit Model-Version-Pinning Clause

Google's silent Gemini 3.2 routing is the trigger. The next move is the first publicly disclosed hyperscaler AI contract, almost certainly in a regulated industry (finance, healthcare, defence), that names model-version pinning, substitution notification, and re-eval rights as contract terms. That disclosure is probably one to two cycles out, and the CISOs who already redlined version-substitution clauses absorb the news as routine.

For Your Team

Strategic purpose: Tuesday is the day this weekend's data-spine, power, capacity, and model-routing signals get translated into one Stack-Layer Owner Map before the next architecture review. The work is one named owner per load-bearing layer: the identity-and-consent layer, the on-site power layer, the Asia or sovereign capacity option, and the model-version-pinning clause. Everything else is commentary.

Tuesday's meeting prompt: ”If Publicis just bought a data-spine company to wire its agents on Sunday, if Babcock & Wilcox just got a $2.4B AI-power order from a 159-year-old vendor, if DayOne just filed a dual US-Singapore IPO for data-center capacity, and if Google is silently routing users to an unannounced Gemini variant, who in this room owns the named scorecard across our data layer, our power layer, our capacity layer, and our model-version layer, and is that owner one person, or four people who have never been in the same room?”

The Load-Bearing Layer Framework:

1. One named owner per load-bearing layer. CMO and head of data co-own the agent-grade identity layer. COO and head of platform co-own the named power partner. CFO and head of platform co-own the Asia or sovereign-region capacity option. CIO and CISO co-own the model-version-pinning clause.

2. Named agent-grade identity layer. Every agent-driven workflow gets a named owner for identity, consent, and addressable audience. Publicis priced the data-spine reference for you.

3. Named on-site or behind-the-meter power partner. Every power-heavy AI workload gets a named partner with a delivery date and a cost-per-megawatt-hour line. Babcock & Wilcox priced the power reference for you.

4. Named Asia-Pacific or sovereign-region capacity option. Every latency- or sovereignty-sensitive workload gets a named non-US-region option with a public-grade cost number. DayOne priced the capacity reference for you.

5. Named model-version-pinning clause. Every regulated AI deployment gets a contract clause with version pinning, substitution notification, and re-eval rights. Google priced the routing-substitution risk for you.

Share-worthy stat: A 159-year-old combustion vendor just signed a $2.4B AI-power contract larger than its own 2024 market cap. Drop that on the next AI strategy review and the ”AI capacity is a hyperscaler conversation” framing reframes itself in 30 seconds.

Go deeper: Track the data, power, and capacity signals in real time →

The Track of the Day

”Agentic AI readiness is identity readiness.”
, Okta, 2026 Businesses at Work Report

Today's set: ”Sunday Morning” by Maroon 5, cued at the 2am slot when the headline acts have left the booth and the working stack is carrying the room. The model-lab funding round was loud this weekend. The Publicis-LiveRamp data spine, the Babcock & Wilcox AI-power order, the DayOne dual-market IPO, the Malaysian data-protection guides, and the silent Gemini 3.2 routing were the tracks that actually moved the floor. The operator who waits for the headline track while the data, power, capacity, and version-pinning layers keep shipping is going to play to a thinner room by Q3. The one who walks into Tuesday with all four rows already on the dashboard is headlining the rest of the cycle.

Yves Mulkers, your data DJ, mixing 190,000 articles into the tracks that actually matter.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: May 18, 2026 | Curated by Yves Mulkers @ Ins7ghts