Berkshire's Tiny 'Secret' Portfolio Dumped Alphabet and Broadcom — While the Main Fund Quietly Added $17B to Alphabet

The Oct. 7 framing of a Berkshire 'secret portfolio' refers to New England Asset Management (NEAM), a wholly-owned investment subsidiary inside General Re, which Berkshire Hathaway acquired in 1998. NEAM files its own 13F, so its $863M in equities never shows up inside Berkshire's main $299B equity filing — and that filing gap is what makes the book 'secret' to anyone who only reads [brk-b]'s primary disclosure.

The structural point matters for interpretation. NEAM runs separate accounts for insurance clients under fiduciary duty; its portfolio decisions are made by its own investment committee, not by Buffett, Abel, or Weschler. When NEAM trimmed Alphabet Class A by 48% in Q2 and cut Broadcom by 97% (from a 6,605-share position to ~198 shares), that is a NEAM trade — not a Berkshire main-fund vote on AI. Treating it as a [brk-b] tell is the same mistake as reading a subsidiary insurance float decision as Buffett's stock-picking signal.

Berkshire's Q2 2026 13F, filed Aug. 14, 2026, shows the parent fund ended a 14-quarter net-selling streak with $19.8B of net equity purchases — $23.5B bought against just $3.7B sold. Alphabet absorbed the bulk of that bid: about 24.5M Class A shares ($7.9B) and 23.6M Class C shares ($7.6B) added during the quarter, plus a separate $10B private placement bought directly from Alphabet at the issuer level — roughly 60% of the 48.1M shares added came through that block.

The end-Q2 stack is now 78.79M GOOGL shares ($28.16B) plus 27.19M GOOG shares ($9.61B) — a combined $37.77B Alphabet position that is the third-largest issuer holding in the portfolio, behind only Apple ($65.95B) and American Express ($51.28B). On a combined-share-count basis, Berkshire lifted its Alphabet stake ~83% in a single quarter, the largest single-name increase in the filing.

The pattern matters more than the print. NEAM trimmed AI exposure inside an $863M book; Berkshire main simultaneously added AI exposure inside a $299B book. A 'smart money is selling' thesis requires the larger book to be the seller — and it isn't.

The Q2 13F tape does show a genuine exodus from Alphabet among other marquee names, just not from Berkshire. Bill Ackman's Pershing Square fully exited Alphabet in Q2: the 311,726 Class C shares and 32,376 Class A shares left over from a Q1 95% cut were sold out, leaving zero Alphabet exposure in the $19.5B, 14-position portfolio. Q2 new buys rotated into Netflix, [Visa], [Mastercard], [S&P Global] and the new Pershing Square USA vehicle — not into AI infrastructure.

Tiger Global / Cubed cut its Alphabet stake 45% in Q2, H&H (Duan Yongping) trimmed GOOG 47%, and Atreides cut GOOGL 12%. On the buy side, [Baillie Gifford] lifted GOOGL ~44% and Stanley Druckenmiller's Duquesne opened a fresh $120M GOOGL stake. The Q2 split is unusually clean: most of the publicly visible sellers are macro/tactical managers, most of the visible buyers are long-duration quality holders.

Michael Burry's Scion — in what the manager has flagged as his final 13F as he deregisters the firm — disclosed 50,000 put options on Palantir with $912M notional and 10,000 puts on Nvidia. He shed the December 2026 puts on both names but kept 2027-dated PLTR puts and direct short equity in both stocks, alongside disclosed shorts in Oracle, Nebius and Caterpillar. Burry's stated view is that AI-related names trade at dotcom-era multiples on circular financing.

The Alphabet/Broadcom exit trade has two transmission paths. For Broadcom, the customer base is hyperscaler ASICs and AI-networking silicon — Alphabet's TPU, Meta's MTIA, Microsoft's Maia and the broader Top-of-Rack Ethernet switch market. NEAM's 97% AVGO trim and Burry's broader AI skepticism would, if it spread to other large allocators, slow the AI-capex order book at Taiwan Semiconductor and cap AVGO's multiple. Today, the order book is intact: hyperscaler 2026 capex guides remain up double digits, and AVGO's networking backlog is sold out into 2027.

For Alphabet, the supply chain is internal — Google controls its own TPU silicon (with Broadcom as the principal co-developer), runs its own data centers, and increasingly layers Gemini across Search, Cloud and Android. A 48% NEAM trim and an Ackman exit do nothing to Google's $80B-plus 2026 capex run rate; [brk-b]'s $17B Q2 buy, if anything, validates the capex story to other institutional holders.

Between now and mid-November, two data points will tell us whether the NEAM/Ackman/Tiger trim is the start of an AI-leader de-rating or a one-quarter tactical move. The first is Berkshire's Q3 2026 13F, due 45 days after Sept. 30 — a second consecutive net buyer quarter with another large Alphabet print would force a re-rating of the 'Buffett rotates out of AI winners' narrative; a flat or trimming quarter would vindicate it. The second is Pershing Square's Q3 13F, which will show whether Ackman rotated back into Alphabet or stayed in Netflix/Visa/Mastercard.

Over a 1–3 year horizon, the structural question is whether the AI-capex curve flattens. Burry's 2027-dated PLTR puts and his Nvidia short express the bear case: AI revenue compounds but capex compounds faster, so the integrated AI vendors compress to commodity multiples. The bull case — the one Berkshire main is voting for with $17B of incremental Alphabet in a single quarter — is that Search, Cloud and Android monetize the new inference layer with operating leverage that hardware capex cannot erode. Q3 prints will move the tape; the Q3 prints will not settle that debate.