Add the numbers up. Amazon's Q1 capex came in at $44.2 billion, up 77% year-on-year, with a fiscal-2026 guide that lands the full year at roughly $200 billion. [1] Microsoft's FY26 guide is approximately $190 billion. [2] Google's 2026 capex guide is approximately $185 billion. [3] Meta raised its capex range to $125-145 billion at the high end of $145 billion. [4] That is the five-firm cohort that the paper has been calling Mag-5 minus the abstainers: Amazon, Microsoft, Google, Meta, and Oracle (whose Stargate-related commitments add a sixth at $50 billion). For 2026, the deploying cohort spends roughly $715 billion on AI infrastructure. In 2025, the same cohort spent $410 billion. The year-over-year increase is $305 billion.

Now subtract the abstainers. Apple's H1 FY26 capex prints at approximately $4.3 billion, with a full-year guide of roughly $13 billion — a number that has held flat since FY24. [5] Berkshire Hathaway's deployed AI capex, as Greg Abel made plain in his first Saturday Q&A in Omaha, is essentially zero. The full-year Berkshire capex line continues to be dominated by Burlington Northern, MidAmerican Energy, and Berkshire Hathaway Energy infrastructure. None of it is labeled AI. None of it is positioned to be. [6]

The paper's Saturday account that Apple bought back $100 billion into a memory-inflation guide framed Apple as the only Mag-7 vote against the capex regime. Berkshire's record $397.4 billion cash position, also disclosed Saturday, completes the abstainer pair. The five-to-two split is the cohort's editorial vote. The five say AI is the once-in-a-generation infrastructure event. The two say there are not enough shovel-ready, return-on-invested-capital projects to justify deploying capital at the scale being asked.

What MSM has emphasized: each name in isolation. Bloomberg covers Berkshire as a continuity story. 247WallSt frames Amazon-vs-Microsoft as a divergence. The Motley Fool treats Apple as a buyback story. [7][8] The cohort split — the structural argument that two members of the same cohort are voting differently than the other five — is not in the corporate-desk reporting. What X has emphasized: the split. Markets-tracking accounts have been pairing the Berkshire cash and the Apple buyback against the Meta bond raise and the Amazon $200 billion guide as a single narrative since Friday. The X argument is plain: in a cohort week, the abstention is the news.

The numbers under the deployment side are not just about scale. They are also about how the deployment is being financed. Amazon's Q1 free cash flow collapsed from $26 billion (year-ago) to $1.2 billion. [1] Long-term debt grew from $65.6 billion to $119.1 billion. The capital structure shift is not a footnote. Meta priced $25 billion in investment-grade bonds across up to six tranches on May 1; the orderbook closed at approximately $96 billion, four times oversubscribed. [9] The bond market is buying the Meta capex thesis at a yield that suggests it does not believe the equity-side discount the same balance sheet is taking. Meta stock fell 9.5% on the print. The bond market and the equity market are pricing the same balance sheet incompatibly.

That incompatibility is what Apple's $100 billion buyback and Berkshire's $397 billion cash both pre-empt. Apple is not financing growth; it is returning cash to shareholders at $0.27 per share, plus the buyback authorization, against an FY26 capex line that has not moved. Berkshire is not financing growth either. Berkshire is also not deploying. Abel's "doesn't mean you need to deploy all your capital" sentence on Saturday is the cleanest articulation of the abstainer thesis the cohort has produced. [6] If you do not see a return at the scale being asked, you do not deploy. You wait. You hold cash. You buy back stock. You collect insurance float and wait for asset prices to come to you.

The argument the deployers are making is structural. AI infrastructure is the data-center, fiber, power, and silicon stack that the next generation of compute will run on, and there is a window — five years, possibly seven — in which the firms that build the stack will own the rents on every dollar of compute that runs through it after. That is the Microsoft 10-Q math, where 45% of $625 billion in commercial RPO is tied to OpenAI. [10] It is also the Amazon AWS math, where the AI-services share of the $124 billion run-rate is rising every quarter. The deployers believe they are buying optionality on a rents-bearing infrastructure. The abstainers believe they are watching a capex bubble.

What the cohort week shows is that both reads are operating inside the same set of firms. Microsoft, Google, Amazon, Meta, and Oracle are deploying. Apple and Berkshire are not. That is not a difference in information. It is a difference in conviction. The information set is shared. Apple sits on the same data Microsoft does about AI demand. Berkshire's underwriting horizon is longer than any of theirs. Greg Abel knows what GEICO sees in claims data. Tim Cook knows what App Store consumption looks like. The two abstainers are choosing not to deploy with the same data.

The 5:2 minority controls more shareholder yield optionality than the 5:7 majority controls deployment optionality. Apple alone authorized $100 billion in buybacks Friday — more than Meta and Microsoft combined intend to return to shareholders this year. Berkshire's $397 billion cash position is the largest single pool of corporate dry powder in American history. The cohort split is therefore not a "two stragglers" story. It is a "two cash-distributors versus five capex-deployers" story. The yield-and-deploy split is the way 2026 will look in retrospect when the next memory-inflation cycle prices in.

The arithmetic is hard. $305 billion of incremental capex must produce returns. The deployers' shareholders have implicitly accepted that, with the bond market underwriting more of it than the equity market. The abstainers' shareholders have implicitly rejected it. By 2027 it will be measurable which read was right.

For now, on Sunday, the cohort splits five-to-two. The number to remember is 715. The other number to remember is 397.

-- THEO KAPLAN, San Francisco