The largest disclosed customer commitment of the agentic-AI era is a CPU deal. Meta Platforms (NASDAQ: META) and Amazon Web Services, the cloud arm of Amazon (NASDAQ: AMZN), jointly announced on April 24, 2026 that Meta will deploy “tens of millions of Graviton cores” — the Arm-based server CPUs AWS designs in-house at Annapurna Labs — to run “agentic AI” workloads at production scale. Both companies called it “one of the largest Graviton deployments” in AWS history. The chip generation is Graviton5, a 192-core Neoverse V3 part on TSMC’s 3-nanometer process. The use case, in the language both press releases share, is “real-time reasoning, code generation, search, and multi-step task orchestration.” That is the continuously running inference layer behind every Facebook, Instagram, and Reels surface.
The cost arithmetic landed five days later. On its Q1 2026 call, Meta raised 2026 capex guidance to $125 billion to $145 billion and recorded a $107 billion one-quarter step-up in contractual commitments from multi-year cloud and infrastructure deals. The AWS Graviton commitment closed after the Q1 books shut. It sits on top of the $107 billion, not inside it. Combined 2026 capex across Meta, Alphabet, and Amazon, per The Next Web’s tally of the Q1 prints, is roughly $650 billion. The AWS deal sits inside that envelope.
The compute moat the discourse misses
Matt Garman, CEO of Amazon Web Services, amplified the deal in a LinkedIn post on May 2:
“Meta is deploying tens of millions AWS Graviton cores to support that shift, one of the largest Graviton deployments in Amazon Web Services (AWS) history. These are production systems that reason, plan, and operate in real time at global scale… This is what the next generation of AI infrastructure looks like.”
Garman runs the cloud the customer is buying from. The framing is sales copy and we treat it as such. The number is not. “Tens of millions of cores” is the verbatim figure both AWS and Meta used in their releases, and the figure Andy Jassy repeated in his Q1 2026 chips-business commentary, saying the deployment “allows Meta to run the CPU-intensive workloads behind agentic AI with the performance and efficiency they need at their scale.” Three principals, one number. That much is disclosure.
What it sits on is the story. Our piece earlier today on Maria Rua Aguete’s Omdia data laid out the demand side: Meta captures roughly 70 percent of global social-ad revenue, with four platforms collectively above 90 percent. Meta’s Andromeda ads-retrieval engine, disclosed in a December 2024 Meta Engineering blog at “a meaningful increase of model capacity (10,000x)” on NVIDIA Grace Hopper plus Meta’s MTIA accelerator, is the technical engine behind Advantage+ automation. Meta’s release does not name Andromeda. But the workload Garman describes — production systems reasoning, planning, and operating in real time — matches the public profile of that stack. Meta’s stated reason for the AWS deal, in the words of Head of Infrastructure Santosh Janardhan, is that “no single chip architecture can efficiently serve every workload.” Translation: the moat is wide enough that one of the world’s largest data-center operators is buying CPU capacity from a competitor to keep it deepening at the rate the auction wants.
Where Pichai’s complaint and Zuckerberg’s capex agree
The other Big Three Q1 2026 calls confirm the same observation from other seats. Sundar Pichai told Alphabet (NASDAQ: GOOGL) investors the company is “compute constrained in the near term” and that “our cloud revenue would have been higher if you were able to meet the demand.” Alphabet raised 2026 capex to $180 billion to $190 billion and disclosed a Google Cloud backlog of roughly $462 billion. Mark Zuckerberg said Meta is “rolling out more than one gigawatt of our own custom silicon” with Broadcom alongside “significant” AMD and NVIDIA capacity. CFO Susan Li attributed Q1 infrastructure-cost growth on the same call to “higher depreciation, data center operating costs, and third-party cloud spend.” Jassy, on Amazon’s print, named the Meta deal and added that AWS’s chip business is on a $20 billion-plus run rate growing at triple-digit year-over-year percentages, with Graviton at 98 percent of the top 1,000 EC2 customers.
Three CEOs, three seats, one read. Agentic-AI inference is binding the Big Tech compute stack at the same time, and each is paying to stay on the curve. Pichai prescribes own-the-silicon. Zuckerberg prescribes own-and-rent. Jassy prescribes rent-from-us. The prescriptions diverge; the diagnosis is identical. The independent ad-tech cohort pitching agentic AI as the open-internet alternative is not in the conversation about who can afford the substrate.
What the “independent” pitch is competing against
Our piece last week on the 2026 ad-tech consolidations argued the word “independent” had decayed into a marketing label, and that buyers should treat the four 2026 deals (Trade Desk Ventura, Mediaocean–Innovid, Viant–TVision, the IAB Tech Lab PGC seating) as smaller walled gardens, not the structural alternative to Google, Amazon, and Meta. The Aguete data was the demand-side number underneath that argument. The Garman commitment is the supply-side compute number. One structural read across three stories: Meta’s compounding ad-revenue advantage runs on auction-efficiency improvements that only first-party data plus continuously running CPU and GPU capacity at hyperscaler scale can fund, and the hyperscaler scale is now disclosed.
The strongest version of the counterargument is real and worth conceding. Agentic-AI ad-tech doesn’t have to match Meta’s compute to compete; it can differentiate on auditability, on cross-publisher reach, on the procurement questions the gardens deflect. That’s the case Jeff Green has been making for Ventura, and it isn’t wrong as far as it goes. The hedge: “as far as it goes” stops at the workload class itself. When the auction runs continuously and the inference is real-time, the cost-of-compute curve becomes a structural barrier, not a procurement variable a clean room can route around. The independents’ pitch in 2026 is auditability. The hyperscalers’ counter is that auditability is a service, and the service runs on their cores.
We argue the post-ATT pattern repeats. GDPR was supposed to dilute platform concentration; Meta substituted first-party data and got bigger. iOS App Tracking Transparency was supposed to dilute it; Meta substituted compute for signal via Andromeda and got bigger. TikTok was supposed to dilute it; Meta took the format and the surface and got bigger. Each reset added a layer of capability the next-tier players couldn’t match without the same capital base. The Graviton commitment is the next layer of the same gradient, the first one with a hyperscaler-customer disclosure attached.
What buyers do with this is the real call. Procurement frameworks built for “Meta vs. open programmatic” stop scaling when the open programmatic substrate is competing against $125 billion to $145 billion in annual capex from a single buyer. The honest pitch nobody in the cohort has given yet: differentiation has to be something hyperscalers can’t replicate quickly. Pricing arbitrage they haven’t gotten to yet doesn’t qualify. The next print that tests our read is The Trade Desk (NASDAQ: TTD)‘s Q1 2026 call on May 7, the largest pure-play DSP entering the agentic-era earnings cycle, the first real chance for forward commentary that names the moat instead of pitching around it. Meta’s own Q2 capex print, due late July, is the second test. If $125 billion to $145 billion holds, the gradient holds. We don’t expect it to slip.
