AI Data Centers and Traditional Data Centers Run the Same Workloads Now

AI Data Centers and Traditional Data Centers Run the Same Workloads Now

Nobody converted these buildings to do something new. The things they already did got rebuilt as AI. Here is what Meta and Google disclosed in 2026, what happened when Meta needed five buildings' worth of AI capacity in a hurry, and what the Meta Terms of Service says about your side of the arrangement.

Someone left four lines on one of my posts. No data centers, period. No digital ID. No digital currency. No mass surveillance.

They typed it on a phone, inside an app, and it traveled to a building full of servers and came back before it appeared on a screen.

I am not interested in dunking on that person, because there is a real concern buried in the bumper sticker. So here is the actual work. What these buildings do. What hyperscale means. And why the boundary between AI infrastructure and ordinary internet infrastructure stopped being meaningful, measured by workload rather than by branding.

A note on what this covers

This article argues the convergence entirely on use cases and workloads: what the buildings do, not what they consume. I've handled the energy and water fights separately, in "A Tale of Two Curves" and "Will AI Data Centers Drain America Dry?", and none of that material is repeated here. Every figure below traces to a source at the bottom, most of them Meta's and Google's own earnings calls and engineering posts. One framing device gets flagged on arrival: the "three eras" structure is my teaching device for documented facts, not an industry-standard taxonomy.

1 trillion parameters, roughly, in the ads ranking model Meta serves in under a second, distilled from GEM

2.5 billion monthly users shown Google AI Overviews, whether or not they wanted, read, or noticed them

0 occurrences of the words train, training, model, or models in the entire 5,460-word Meta Terms of Service

5 production data centers Meta emptied, buildings serving live traffic, to build one AI supercluster

What Is the Difference Between a Data Center and a Hyperscale Data Center?

A data center is a building that holds servers, storage, and networking equipment. The rack of gear in a machine shop's back office in Bay City is a small one. Nothing about the category requires it to be enormous.

Hyperscale is a size and ownership label, not a different technology. Synergy Research Group, now part of TechInsights, builds its hyperscale numbers by tracking the footprint of roughly 21 of the world's largest cloud and internet firms, spanning SaaS, IaaS, PaaS, search, social media, e-commerce, and gaming. That is the definition: a set of operators big enough to build at industrial scale. As of mid 2026, those operators controlled about 48 percent of worldwide data center capacity, on track for 67 percent by 2031, while enterprise on-premise capacity fell from 56 percent in 2018 to 32 percent.

I ran a school for eight years, so here is the analogy that lands for me. A one-room schoolhouse and a consolidated district high school are both schools. Same purpose, same function, radically different scale. Nobody argues the high school is a different species of institution. Hyperscale is the high school.

One honest caveat before we go further, because the simplest version of this argument does get overstated. Buildings designed for dense accelerator racks are engineered differently from buildings designed for general-purpose CPUs. Meta uses direct-to-chip liquid cooling for its GPUs while keeping air cooling for other equipment in the same facility, which tells you the hardware requirements genuinely diverge. The claim worth defending is not that the buildings are identical. It is that they belong to one fleet, run by the same companies, doing the same list of jobs.

Three Eras of What These Buildings Actually Do

The clearest way to see the convergence is to look at how the work itself changed.

Era One: Retrieval and Rules

Keyword indexes. Chronological feeds. PageRank. A query came in, the system looked something up in an index, applied ranking rules, and returned it. This ran on ordinary CPUs, and it is what most people still picture when they hear "server farm."

Era Two: One Small Model Per Job

Then every individual task got its own machine learning model. A spam classifier. A face detector for photo tagging. A click-through prediction model. A separate ranking model for Feed, another for Reels, another for Stories, another for Marketplace. Dozens of narrow, task-specific models, each trained and tuned independently.

This was already AI. It just did not get called that, because it worked quietly and nobody had a chatbot to sell. Meta's advertising business ran this way for most of two decades.

Era Three: One Large Model, Distilled Into Everything

The current architecture inverts era two. Instead of many small independent models, there is one very large foundation model that learns broadly, and its knowledge gets transferred down into the individual product surfaces.

Meta's own stack is the textbook case. GEM, the Generative Ads Recommendation Model, sits at the top. Meta's engineering team describes it as training at LLM scale on several thousand of the latest-generation GPUs. GEM is too computationally expensive to serve ads directly, so its learning is transferred through knowledge distillation into Lattice, the centralized ad ranking architecture, and into Andromeda, the retrieval engine. Lattice exists specifically to consolidate what used to be fragmented per-surface models, so that a conversion insight learned on Reels can immediately improve Feed.

That is the whole story in one system diagram. The narrow models that used to run each product independently got replaced by a foundation model that feeds all of them.

What Meta's Data Centers Are Actually Doing

Working from Meta's own earnings calls, engineering blog, and executive interviews through 2026, here is the product list, surface by surface.

Instagram Feed and Reels ranking. Every public Reel and Feed post is automatically processed through a large language model and analyzed across dimensions from topic to tone, with those signals feeding the ranking systems. On the Q2 2026 earnings call, Mark Zuckerberg called it a key building block toward greater personalization and said Meta is extending it to more of Facebook.

Video classification and summarization. Handled by Muse, a separate model family applied to video across surfaces. Muse Image and Muse Video are the generative variants, positioned by Zuckerberg as a third content source on Instagram alongside friends and creators.

Ads retrieval. Andromeda, running on a custom NVIDIA Grace Hopper Superchip SKU that Meta co-designed, plus Meta's in-house MTIA silicon. Andromeda alone enabled what Meta describes as a 10,000x increase in the complexity of the models used for ads retrieval, the stage that narrows tens of millions of candidate ads down to a few thousand.

Ads ranking. An adaptive ranking model of roughly one trillion parameters, distilled from GEM, served at sub-second latency.

Threads recommendations. Llama-based recommendation systems.

Translation and dubbing. AI dubbing live in nine languages, with hundreds of millions of people watching AI-translated video daily.

Algorithm controls. The Instagram "Your Algo" page, where users write natural-language instructions to tune their own recommendations.

The assistant. Meta AI across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban glasses, at more than a billion monthly users.

Safety and integrity. Automated detection systems, described in Sections 1.5 and 1.6 of the Terms of Service as artificial intelligence and machine learning.

Read that list and try to draw a line separating the AI half from the traditional half. Feed ranking is the oldest product Meta has. It is now a large language model workload. The ads business, which predates the generative AI boom by fifteen years, runs the largest model in the building. All of it runs against more than three billion daily active users.

There is a detail here that kills the "AI chips are for chatbots" assumption entirely. Matt Steiner, Meta's VP of Monetization Infrastructure, Ranking and AI Foundations, explained that recommender workloads have a different compute-to-memory ratio than a standard LLM GPU, which is precisely why Meta built MTIA, its own custom silicon. Retrieval is memory-bound at Meta's scale, so Andromeda got its own co-designed Grace Hopper variant. The traditional workload was important enough to justify purpose-built AI hardware.

And it works, which is why it keeps expanding. Meta reported that first quarter 2026 ranking improvements drove a 10 percent lift in Reels time spent on Instagram and more than 8 percent growth in global Facebook video time, the largest quarter-over-quarter gain in four years. In the second quarter, Instagram time spent grew double digits year over year, and the largest Reels ranking update the company has shipped moved overall Instagram sessions by 15 basis points.

What Google's Data Centers Are Actually Doing

Google's seventh-generation Tensor Processing Unit, Ironwood, is generally available to Google Cloud customers. Internally, the same silicon powers Gemini, Search AI Overviews, and YouTube's recommendation stack.

That single sentence settles the argument. YouTube recommendations are about as conventional a hyperscale workload as exists on the internet, predating the generative AI boom by more than a decade. It now runs on the same accelerators, in the same buildings, as the chatbot everyone argues about.

Search AI Overviews. More than 2.5 billion monthly users as of Google I/O 2026.

Search AI Mode. More than 1 billion monthly users roughly a year after launch, merged with AI Overviews at I/O 2026 into one experience running on Gemini 3.5 Flash.

The Gemini app. More than 950 million monthly active users as of the Q2 2026 earnings call.

Google Workspace. Gemini embedded in Gmail, Docs, Sheets, Slides, Drive, Meet, Chat, and Vids, with Gemini 3 Flash as the default model for most interactions since March 2026.

Chrome. Gemini in Chrome, expanded to additional languages and regions in June 2026, on by default for Workspace customers unless an admin disables it.

Android. Circle to Search available on more than 580 million devices.

Google Cloud. Gemini Enterprise at more than 8 million paid seats across 2,800 companies, with 13 million developers building on Google's generative models.

Two things about the search numbers deserve care, because this is where commentary usually gets sloppy.

The 2.5 billion AI Overviews figure and the 1 billion AI Mode figure are not the same kind of number. AI Mode is a deliberate choice: people opened a conversational tab. AI Overviews counts everyone who saw an AI-generated summary appear above the results, whether or not they wanted it, read it, or noticed it. Most of those people never decided to use AI. They searched for a plumber.

The second thing is what AI Mode does to the work happening underneath. Google describes a technique called query fan-out, where the system splits a question into related subtopics, runs multiple searches across them, and synthesizes a response. One typed question becomes many searches plus a generation step. The interface looks like the search box people have used since 2001. What happens behind it does not resemble 2001 at all.

The Buildings Themselves Move Between Categories

If shared hardware is not convincing enough, the physical evidence is blunter.

At the AI Infra Summit in Santa Clara, Yee Jiun Song, Meta's VP of Engineering for Infrastructure Foundation, described what the company did when it needed AI capacity faster than it could build new sites. Meta emptied five production data centers, buildings that were serving live consumer traffic, and consolidated the hardware into a single supercluster.

"We moved thousands of racks between data centers as part of this operation. And in order to do this quickly, we had to redesign the loading docks in our data centers, build brand new robots to move these 1,000-pound racks, and even design crateless packaging for these racks to speed up the moves."

Yee Jiun Song, Meta, AI Infra Summit, September 2025

Meta also quadrupled the networking in those buildings, replaced hundreds of meters of fiber, and dug new trenches to connect the five buildings to each other.

Those five buildings were traditional data centers on a Monday. They were an AI supercluster once the racks stopped moving. Meta's global director of data center engineering, Alan Duong, told Data Center Dynamics the design philosophy plainly: these are 20 to 30 year facilities, and Meta builds retrofittability into them so the design can be modernized when the workload changes. The company treats the label as a phase a building passes through.

The accounting says the same thing. Neither Meta nor Alphabet publishes an AI data center capital expenditure line separate from general technical infrastructure. Alphabet splits its disclosure only between servers and facilities. Meta reports companywide capital expenditure. Not because either is hiding anything, but because the split would be arbitrary. There is no meter that tells you which rack was doing AI this afternoon.

The Other Half of the Bargain: What You Already Agreed To

Everything above is what the buildings do. Here is what makes it possible, and it is the part nobody reads.

The Meta Terms of Service currently in effect runs 13 pages and roughly 5,460 words. At average adult reading speed, that is about 23 minutes. Facebook has more than three billion daily active users. The number who have spent 23 minutes with this document is, charitably, small.

Section 1.9 is titled "Ensuring access to our services," and it names the buildings directly:

"To operate our global services and enable you to connect with people around the world, we need to transfer, store and distribute content and data to our data centers, partners, service providers, vendors and systems around the world, including outside your country of residence. The use of this global infrastructure is necessary and essential to provide our services."

Meta Terms of Service, Section 1.9

The commenter said no data centers, period. The agreement they accepted says the data centers are necessary and essential, and that the data travels outside the country. Both parties agree on the facts. They disagree about whether it was a good trade.

Section 1.6 goes further and names the technology, committing Meta to use and develop advanced technologies such as artificial intelligence, machine learning systems, and augmented reality. Anyone who accepted these terms agreed, in writing, to a service built on AI.

What else is in there:

The license. Section 3.3 grants Meta a non-exclusive, transferable, sub-licensable, royalty-free, worldwide license to host, use, distribute, modify, run, copy, publicly perform or display, translate, and create derivative works of content you post. You keep ownership. They keep permission.

Deletion takes longer than people assume. Content sent to trash begins deleting after 30 days. It may take up to 90 days to delete content after the process starts, and up to another 90 days to clear backups and disaster recovery systems. If someone else used your content under that license and has not deleted it, the license continues until they do.

Your face is ad inventory. Section 3.3 states you give Meta permission to use your name, profile picture, and information about actions you have taken next to or in connection with ads and sponsored content, "without any compensation to you."

Liability is capped at $100. Section 4.3 limits Meta's aggregate liability to the greater of $100 or what you have paid the company in the past twelve months. For a user who has never bought an ad, that is $100.

Disputes go to California. Section 4.4 requires claims to be resolved exclusively in the US District Court for the Northern District of California or a state court in San Mateo County, under California law.

Automated collection runs one direction. Section 3.2 prohibits you from collecting data from Meta products by automated means without permission, including while logged in. Nothing in the document constrains Meta's collection of yours.

The Training Permission Is Not in This Document

Here is the part that surprised me, and it took a full-text search of the agreement to find. The words "train," "training," "model," and "models" do not appear anywhere in the Terms of Service. Not once in 5,460 words. There is no clause stating that your content is used to build Meta's AI systems.

That permission lives in a different document: Meta's Privacy Center page titled "How Meta uses information for generative AI models and features." It is referenced from the Terms in a single line near the end, and it is considerably blunter than the Terms are.

On what gets used, the page lists training sources including "information shared on Meta Products including things like posts or photos and their captions, and your interactions with features that are part of AI at Meta."

On who is covered, it goes further than most people assume:

"Even if you don't use our Products or have an account, we may still process information about you to develop and improve AI at Meta. For example, this could happen if you appear anywhere in an image shared on our Products by someone who does use them or if someone mentions information about you in posts or captions that they share on our Products."

Meta Privacy Center, How Meta uses information for generative AI models and features

You can have never created a Facebook account. A relative posts a photo from Thanksgiving. You are in scope. That is Meta's own stated example, not an inference.

On how long it is kept, compare the two documents side by side. The Terms give content deletion a clock: 30 days in trash, up to 90 days to delete, up to another 90 to clear backups. The generative AI page gives training data no clock at all. It says Meta keeps training data "for as long as we need it on a case-by-case basis," and may keep it further to protect its own or others' interests or to comply with legal obligations. Deleting the post does not pull it back out of the model.

And the page contains its own version of Section 1.9, which brings this back to where the article started: "Meta operates a global infrastructure to develop and improve AI at Meta."

One scope detail deserves care, because it is easy to misread in Meta's favor and easy to overstate against it. The page offers a request form to everyone, user or not, but the right is written narrowly: requests relate to "their personal information from third parties" being used for AI development. That is not a right to object to your own Meta posts being used. Meta extends that broader objection right in the European Union and the United Kingdom under GDPR, with a May 27, 2025 deadline for existing content. There is no equivalent right in the United States.

None of this is hidden. Meta has disclosed AI training publicly since 2023. The point is narrower and, I think, more useful: the document people mean when they say "I agreed to the terms" does not contain the training permission, and the document that does is two clicks further out and mentioned once.

Then the Terms Met the Ad System

Two more changes connect these documents directly to the workloads described above.

Meta announced on October 1, 2025 that beginning December 16, 2025, interactions with Meta AI across Facebook, Instagram, Messenger, and WhatsApp would be used to personalize content recommendations and advertising. The company said more than a billion people use Meta AI monthly. There is no opt-out. Users in the European Union, United Kingdom, and South Korea are excluded, because their privacy laws prohibit it. Everyone else is in. Dozens of consumer and privacy organizations filed complaints with the Federal Trade Commission.

Meta also trains its models on public posts, comments, and photos from adult accounts. European users had until May 27, 2025 to object. In the United States there was never an opt-out to miss. In July 2026 the company briefly shipped a Muse Image feature that let anyone tag a public Instagram account in a prompt and generate images using that person's public photos, with eligible accounts included by default and without notice. Meta discontinued the generation function in under a week.

The pattern is not a conspiracy. It is a business model executing normally. The products exist to hold attention. Holding attention requires knowing people well. Knowing people well requires data and models. The terms make the data available. The models make the ranking possible. The ads pay for all of it.

Where I Was Wrong, and What the Stronger Argument Is

When I replied to that comment, I wrote that social media "is a form of digital ID" and "exclusively uses digital currencies for shopping." I want to correct that on the record, because loose claims lose arguments even when the underlying point is sound.

A Facebook account is not a government-issued digital identity and carries none of the legal force of one. Marketplace transactions happen in cash in driveways across Saginaw County every weekend. Both lines were rhetorically satisfying and factually shaky, and a hostile reader would have taken them apart in one reply.

The stronger version needs no exaggeration. A large language model reads every public Instagram post before deciding whether anyone sees it. A trillion-parameter model picks which ad reaches you. Two and a half billion people get an AI-written answer at the top of a search page most of them never asked for. All of that sits in earnings calls, engineering blogs, and a terms of service document you can read in 23 minutes.

What This Means If You Run a Business Here

Most small businesses in the Great Lakes Bay Region tell me some version of "we haven't gotten into AI yet." That is almost never true, and believing it is the expensive part.

You already run on AI systems, as a customer of them. If you buy Meta ads, a foundation model decides who sees your creative. If your leads come from Google, an AI-generated answer sits above your listing. Your marketing performance is already downstream of these systems whether or not you have ever opened a chatbot.

That changes what optimizing means. Andromeda works backward from historical engagement, ad copy, creative, and format rather than starting with advertiser-defined audiences. Meta consolidated learning across surfaces so old exploits stop working. The playbook of manual audience micro-targeting has less and less to act on, and creative quality has more.

Price the platforms honestly. Meta and Google are not free. You pay in data, attention, and a license you already granted. That belongs in the decision when you weigh owned channels like email and your own website against rented ones.

Read Sections 1.9, 3.3, 4.3, and 4.4 once. Twenty-three minutes for the whole document, far less for those four. You will negotiate harder over a $2,000 vendor contract than you did over the one governing your customer relationships.

Stop treating AI as opt-in. The choice available is which AI you use deliberately, not whether AI touches your business.

The Bottom Line

The distinction between an "AI data center" and a "regular" one has stopped describing anything useful, because the jobs are the same jobs. Search, feed ranking, ad targeting, translation, spam filtering, and photo organization were always machine learning problems, and they now run on foundation models sharing silicon with the chatbots. The buildings move between categories in a single rack migration, and the companies themselves do not account for the two as separate things, because there is nothing separate to account for.

The commenter and I want mostly the same thing: an arrangement where the tradeoffs are visible and consented to rather than buried in Section 3.3. We disagree about whether refusing to name the tradeoff helps anyone see it. If you use Facebook, Instagram, Google Search, YouTube, Gmail, or Chrome, you are not deciding whether to use AI. You are deciding whether you use it on purpose.

Wondering what your business actually agreed to across its own tool stack? Find out where you stand: take the free AI Readiness Assessment. Two minutes, no sales pitch, and you get a clear picture of your next step.

Sources & Further Reading

Meta Workloads and Infrastructure

1. Meta Platforms, Q2 2026 earnings call transcript, July 29, 2026. Source of the every-public-post LLM processing disclosure, "Your Algo," Muse positioning, and Reels ranking update figures.
2. Meta Platforms, Q1 2026 earnings call transcript, April 29, 2026. Source of the Reels and Facebook video time-spent lifts.
3. Meta Platforms, "2026: AI Drives Performance," About Meta, January 2026. Source of the nine-language AI dubbing figure.
4. Engineering at Meta, "GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model," August 3, 2026.
5. Engineering at Meta, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine," December 2024.
6. Meta for Business, "AI Innovation in Meta's Ads Ranking Driving Advertiser Performance." Source of the 10,000x retrieval complexity figure.
7. Chipstrat, interview with Matt Steiner, VP of Monetization Infrastructure, Ranking and AI Foundations at Meta, April 2026. Source of the trillion-parameter ads model and the MTIA compute-to-memory rationale.
8. Search Engine Land, "Inside Meta's AI-driven advertising system: How Andromeda and GEM work together," February 2026.
9. Yee Jiun Song, Meta VP of Engineering, Infrastructure Foundation, AI Infra Summit, Santa Clara, September 2025. Source of the five-data-center consolidation account.
10. Data Center Dynamics, "How Meta redesigned its data centers for the AI era," interview with Alan Duong.

Google Workloads and Infrastructure

1. Alphabet, Q2 2026 earnings call and investor materials, July 22, 2026. Source of the Ironwood, Gemini app, and Gemini Enterprise figures.
2. Sundar Pichai, Google I/O 2026 keynote, May 2026. Source of the AI Overviews and AI Mode user figures.
3. Google Workspace Updates blog, Gemini in Chrome availability expansion, June 2026.
4. Google Workspace Help, "Google Workspace with Gemini," updated August 2026.

Industry Definitions and Capacity

1. Synergy Research Group, "Hyperscale Operators to Account for 67% of all Data Center Capacity by 2031," April 2026. Source of the hyperscale definition and capacity shares.

Terms, Privacy, and AI Training

1. Meta Platforms, Terms of Service, effective January 1, 2025. facebook.com/terms. Word counts and section quotes from a full-text extraction of the current PDF.
2. Meta Privacy Center, "How Meta uses information for generative AI models and features." facebook.com/privacy/genai. Source of the training sources, non-user processing, and retention quotes, all verbatim.
3. TechCrunch, "Meta plans to sell targeted ads based on data in your AI chats," October 1, 2025.
4. Meta Newsroom, "Making AI Work Harder for Europeans," April 14, 2025. Source of the EU/UK objection right and deadline.

Companion Articles from Understanding Your AI

1. Will AI Data Centers Drain America Dry?
2. A Tale of Two Curves
3. The AI Energy Crisis That Isn't: What the Data Actually Says
4. Does AI Really Use 10 Gallons of Water Per Image?

Where does your team actually stand with AI?

Stop guessing. The free AI Readiness Assessment takes a few minutes and gives Tim what he needs to point you at your highest-value next step, whether or not you ever hire us. We serve Midland, Bay City, Saginaw, and Michigan's Great Lakes Bay Region with research-backed AI training and fractional AI leadership.

Start the Free AI Readiness Assessment | Book a 20-Minute Call