Meta will spend "well over $100 billion this year alone" on infrastructure, according to Santosh Janardhan, the company's Head of Infrastructure, in a YouTube interview published on October 8, 2026. Speaking to creator Tom Shaw, Janardhan said Meta has contracted 6.5 gigawatts of nuclear power through 2035, designs separate custom chips for ranking and recommendations, for large language models, and for AI training and inference, and expects its future data center fleet to split into two or three distinct templates. The video description states that it "is a part of a paid partnership project with Meta."

In Short

A senior Meta executive who runs the company's data centers explained in a YouTube interview how Meta builds the computers behind Facebook, Instagram, WhatsApp and its AI products, and said the company will spend well over $100 billion this year doing it. Most of that money comes from advertising, so the scale of this building programme shapes the ad systems that marketers buy from every day. The interview was paid for by Meta, so its figures are best read alongside what Meta has told investors, which in places is more precise.

The interview and its framing

The video, titled "The man spending $100 billion on AI in 2026," was published on Tom Shaw's YouTube channel on October 8, 2026, according to the page details. The capture of the page reviewed for this article shows 1,483 views and 59.7 thousand subscribers to the channel. The interview runs to about 18 minutes, split into 15 chapters, from "Meta is more than a software company" to "Training vs inference infrastructure."

The description identifies the guest as "Santosh Janardhan, Head of Infrastructure at Meta." In the conversation itself, Shaw refers to him only by his first name, as "Santosh, the head of infrastructure at Meta." The full name and title used in this article come from the video description.

That description also contains the line that matters most for reading what follows: "This video is a part of a paid partnership project with Meta." Shaw's questions are open and descriptive. There is no challenge on cost, energy prices, water use or local opposition to data centers, topics PPC Land has covered extensively over the past year. The video is, in effect, a brand deal in which the brand is also the interviewee. The figures Janardhan gives are spoken claims, attributed to him below, and are set against Meta's published disclosures where those exist.

From enabler to differentiator

Janardhan's account of the Meta Compute Initiative begins with a timing claim. "There was this fundamental realization, I want to say, sort of late last year," he said, that "infrastructure had gone from being this enabler of things like I was talking about to being a strategic advantage."

Given the October 2026 publication date, "late last year" places that realization in late 2025. Meta publicly introduced Meta Compute in January 2026, according to Axios. By the first quarter of 2026, Meta was deploying more than 1 gigawatt of custom silicon developed with Broadcom as part of the initiative, alongside chips from AMD and systems from Nvidia.

"Infrastructure now, especially in the age of AI, is at a point where it can make a company or break a company," Janardhan said. "If you don't have the compute power you want, if you don't have the horse power of the capacity you want, you are constrained from doing anything meaningful or at least scaling it."

He described the initiative's origins in terms of ambition rather than shortage. "What we realized is that while we were big, our ambitions were much bigger," he said. He added that Meta "wanted three legs to the stool," though the transcript does not record him naming the three legs at that point in the interview.

Janardhan also framed Meta as something other than a social media company. "Meta is as much a infrastructure and a hardware company as it is a software company, as it is a content company as it is about connecting people," he said. He put the user base at 4 billion people, and described "a lot of magic that happens behind the scenes for all of the 4 Billion people to interact with each other at the same damn time."

A gigawatt, two ways

Shaw asked how to visualize a gigawatt. Janardhan gave two reference points. "The microwave in your kitchen runs at 700W," he said, before offering the other end of the scale: "A gigawatt is enough power to run the city of San Francisco."

The microwave comparison implies that a gigawatt equals about 1.43 million kitchen microwaves running at once, a PPC Land calculation. The San Francisco comparison is Janardhan's own and was not sourced in the interview. PPC Land has previously used a different yardstick: Meta's Prometheus cluster in New Albany, Ohio, targeted at over 1 gigawatt, uses roughly the energy of 750,000 homes running continuously.

The scale matters because Meta's sites are now measured in gigawatts. In February 2026, Meta broke ground on a 1 gigawatt data center in Lebanon, Indiana, with an investment of over $10 billion. Hyperion, in Richland Parish, Louisiana, is targeted at 2 gigawatts by 2030. And on July 28, 2026, Meta disclosed a venture with BlackRock to develop a 1GW data center in El Paso, Texas.

BYOP and the grid

Where does that electricity come from? "Hopefully a lot from the US," Janardhan said. "Our grid is constrained."

He described an industry shift he called BYOP, or "bring your own power." A company building a data center can connect directly to a dedicated power source, he explained, but Meta's stated preference is to add generation to the grid and draw from it. "It is better to do it to the grid then the grid benefits," he said. "I get the redundancy of the grid as well." The grid, he said, "can do failover, direct me to different power sources," and Meta is then not "a net negative to the grid."

This matches language in Mark Zuckerberg's letter published on August 10, 2026, which said Meta builds its own energy generation where it invests and in some cases supplies surplus energy to local grids. PPC Land noted at the time that the 6,500-word letter never mentions advertising once.

The nuclear claim, against the record

The most specific energy figure in the interview concerns nuclear power. "Meta is a single biggest purchaser of nuclear energy in the world that is not a government," Janardhan said. "6.5GW of nuclear power is what we have contracted through 2035, single largest purchaser of of a non-governmental nuclear power," he said in the nuclear chapter of the video.

He called it "clean power" that "doesn't have sort of a sort of a exhaust or sort of carbon emissions," adding: "So it's a big deal."

Meta's own announcement in January 2026 was framed differently in two respects. When Meta disclosed agreements with Vistra, TerraPower and Oklo on January 9, 2026, the deals were described as supporting up to 6.6 gigawatts of current and new generation by 2035, according to Axios. That is 0.1 gigawatts more than the figure Janardhan gave, and it is a capacity that the plans "support" rather than a volume fully contracted at present. Part of that generation depends on reactors from TerraPower and Oklo that do not yet operate.

The superlative also differs. At the time, Joel Kaplan, Meta's chief global affairs officer, said the deals made Meta "one of the most significant corporate purchasers of nuclear energy in American history," according to Axios. In the paid interview, that became "single biggest purchaser of nuclear energy in the world that is not a government." Neither the video nor its description offers a source for the global ranking.

Earlier nuclear commitments are part of the same total. In June 2025, Meta signed a 20-year agreement with Constellation for the Clinton Clean Energy Center in Illinois, covering about 1.1 gigawatts according to Axios.

Factories of intelligence

Janardhan returned several times to a single metaphor. "Data centers are the modern day equivalent of factories," he said. "Back in the day, when you thought about factories in the Industrial Revolution, you needed power." Now, he argued, "the product is a sort of digital, sort of intellect that comes out."

Phones and laptops, in this account, are only access points. "Your phone is the conduit," Janardhan said. "The engine of the internet is a data center." Later he sharpened the image: "If data centers are the factories of today, the product is intelligence."

He compared the current build-out to the transcontinental railroad and the US highway system - "huge infrastructure build out that fundamentally changed sort of the country for years, decades to come." He also acknowledged that Meta is not alone: "a bunch of companies are building out" similar capacity.

That is visible in the numbers. Alphabet guided to $175 billion to $185 billion of 2026 capital expenditure in February, and the range has since risen to $195 billion to $205 billion. In a September interview, writer Cory Doctorow put the AI sector's gross annual revenue at about $50 billion against roughly $1 trillion of spending, a contrast absent from the Meta interview.

Why Meta designs its own chips

The most technical part of the conversation concerned silicon. Janardhan described Meta as "fully vertically integrated," meaning: "We design and build our own data centers. We design and build our hardware, our chips, our software, our models, which also means we can customize each one of them to each other."

The practical consequence, he said, is that a compromise in one layer can be absorbed in another. "If I take a shortcut in any one of this, I can compensate in a different layer," Janardhan said. His example: a custom chip that "is running hotter. Physically hotter." Because Meta controls the building, "I can go and do a different cooling," he said, or "put the racks further apart" and increase airflow. "The point is, I now customize the data center for the chip versus the other way around."

Why build a chip at all, given partnerships with Nvidia, AMD and Arm? Janardhan's answer was about general-purpose design. A merchant chip vendor, he said, serves automotive, mobile, internet and financial customers at once, so "I will make a chip that is the amalgamation of all those use cases." A company at Meta's scale can instead build a chip "the only function of which is to ensure that it does the job, I want it to do, nothing more, nothing less."

Separate silicon for ranking and for LLMs

Then came the detail most relevant to advertisers. "I have different chips, even internally, a different one for ranking and recommendations, news feed and a different one for LLMs, which runs sort of primarily AI workloads," Janardhan said. Within AI, he added, Meta has "an inference chip which is different than a training" chip.

Janardhan did not name any of the chips. PPC Land reported in April 2024 that the second generation of Meta's MTIA accelerator doubled compute and memory bandwidth over the first and was already deployed to serve Meta's ranking and recommendation models on Facebook and Instagram. Ranking and recommendation is the same machinery that decides which ad appears in which feed. In the fourth quarter of 2025, Meta doubled the GPU cluster used to train its GEM ads ranking model.

Portfolio, not dependency

Janardhan was explicit that custom silicon does not replace suppliers. "When you run sort of supply chain at scale at the numbers that we run, what you realize is that you cannot take a dependency exclusively on a single partner or a single vendor," he said. Non-delivery of a component by one partner "should not mean a hiccup to us as a company."

He described three benefits of the mix: "we have a strategic independence. We have sort of a cost leverage. And we also have vertical integration." No volumes, prices or ratios between Meta-designed and third-party chips were given.

The $100 billion figure

Janardhan listed what he considers the ingredients of competitive AI: compute, data ("Meta has a lot of data in and around things"), and researcher talent. He then added a fourth: "you need a very healthy balance sheet."

"This stuff is expensive," he said at the 15:09 mark of the interview. "We're going to be spending well over $100 billion this year alone." He repeated the point - "Think about it. $100 billion. A lot of money" - and concluded: "We can afford it. Meta, the company has very healthy balance sheets."

"Well over $100 billion" is a conservative description of Meta's own guidance. Meta began 2026 with capital expenditure guidance of $115 billion to $135 billion, raised it to $125 billion to $145 billion after the first quarter, and narrowed it to $130 billion to $145 billion in July. Capital expenditure, including principal payments on finance leases, was $19.84 billion in the first quarter and $31.08 billion in the second. That makes $50.92 billion in the first half, leaving roughly $79 billion to $94 billion implied for the second half if Meta lands inside its range, a PPC Land calculation.

The phrase "very healthy balance sheet" also needs context. In the second quarter of 2026, Meta's free cash flow fell to $784 million from $8.55 billion a year earlier, and long-term debt issuance rose $24.91 billion. Cash and marketable securities stood at $90.26 billion against long-term debt of $83.66 billion. Meta gave no capital expenditure outlook for 2027, with chief financial officer Susan Li describing planning as "highly dynamic."

The balance sheet Janardhan cites is built almost entirely on advertising. Second-quarter 2026 advertising revenue was $59.36 billion out of total revenue of $60.80 billion, about 97.6%. The word "advertising" does not appear anywhere in the interview transcript.

Two templates for the next decade

Asked what Meta's data center fleet will look like in five to ten years, Janardhan said he was "pretty sure we'll have at least 2 or 3 templates or different kinds of data centers."

The first is the training campus. "The needs of AI training is that you need big, interconnected data centers co-located in a single site," he said. A certain share of the fleet "will end up being this massive co-located data centers, probably gigawatts in scale."

Inference is different. "Inference is, is the act of serving sort of a question you have of an AI model," Janardhan said. It "tends to be short, tends to be bursty, tends to be geographically divided depending on the use case."

His examples set out how latency requirements would shape the network. A user of Meta's glasses asking "what am I looking at?" needs a reply at once: "You want an answer instantaneously." By contrast, a Meta AI user requesting "an itinerary for a ten day tour of Europe" will accept a delay: "If that takes five seconds, you'll wait." Depending on the use case, inference capacity might be "globally distributed," "near the user," running on "an edge network," or still in central data centers. The edge model resembles the way a CDN caches web content close to users, but for AI responses rather than static files.

Janardhan gave no capacity split between training and inference, and no timeline for the distributed inference sites.

Why this matters for the marketing community

For advertisers, the interview describes the supply side of the auctions they buy into. Meta's ad products, from ranking to automated campaign tools, run on the same infrastructure Janardhan oversees, and he confirmed that Meta builds dedicated silicon for ranking and recommendations separate from its LLM chips. When Meta's models for ad delivery get larger, as GEM did when its training cluster doubled in late 2025, the hardware comes out of this budget.

The financial loop is also direct. Advertising pays for almost all of it. With second-quarter free cash flow at $784 million and capital expenditure running above $31 billion a quarter, the spending Janardhan describes is funded by current ad revenue and new debt. That raises the stakes for Meta's ad pricing and ad load in 2027, a year for which Meta has declined to give a capital expenditure figure.

There is a second reading for publishers and media buyers. Meta chose to tell this story through a sponsored creator interview rather than an earnings call, where analysts can push back. Some figures in the paid video - the 6.5 gigawatts and the "single biggest" nuclear claim - are rounder or stronger than the versions Meta put on the record in January. The difference is small in gigawatts and larger in framing.

Finally, the interview signals where AI products will physically sit. If inference for glasses and assistants moves closer to users while training concentrates in gigawatt campuses, the cost of serving an AI answer, and therefore of any advertising attached to it, will depend on how much of that edge network Meta builds.

Timeline

Summary

Who: Santosh Janardhan, Head of Infrastructure at Meta, interviewed by YouTube creator Tom Shaw in a video produced as a paid partnership with Meta.

What: Janardhan said Meta will spend "well over $100 billion this year alone," has contracted 6.5GW of nuclear power through 2035, builds separate custom chips for ranking and recommendations, for LLMs, and for training and inference, and expects two or three data center templates as training and inference diverge. Meta's own January 2026 disclosure described up to 6.6GW by 2035, and its 2026 capex guidance stands at $130 billion to $145 billion.

When: The video was published on October 8, 2026, the date used for this article, taken from the YouTube page details.

Where: On Tom Shaw's YouTube channel, discussing Meta's data center sites across the United States, including Ohio, Louisiana, Indiana and Texas.

Why: Janardhan argued that infrastructure has moved from an enabler to a factor that "can make a company or break a company" in AI. For advertisers, the same infrastructure runs Meta's ad ranking systems, and advertising revenue, about 97.6% of Meta's second-quarter 2026 total, is what pays for it.