Berkshire Hathaway filed its second quarter 13F on August 14, 2026, and the coverage settled on one line: Greg Abel sold Warren Buffett's Amazon and bought Google.
Buffett initiated the Alphabet position himself and said so in July. The terms he accepted are unlike any deal Berkshire has ever done.
What happened
August 14, 2026. Berkshire disclosed 105,979,600 Alphabet shares worth $37.76 billion at June 30, against a total portfolio of $299.25 billion. That is 12.6 percent of the book and its third largest holding, behind Apple at 22.0 percent and American Express at 17.1 percent. The position grew 83 percent in the quarter, an addition of 48,144,587 shares. Amazon does not appear at all (13F-HR, SEC, filed August 14, 2026). Berkshire was also a net buyer of equities after fourteen straight quarters of net selling, close to $20 billion worth, with cash down from a record $397.4 billion to $365.5 billion (CNBC, August 14, 2026).
August 14, 2026. Hours later, Nvidia was reported to have cut the financial backstop behind OpenAI's Ohio data centre campus from $250 billion to less than $120 billion. It would cover only the first phase of roughly five gigawatts and decide on the rest later. The change was made to answer investor concern about how much risk the chipmaker was carrying in order to support demand for its own chips, and its shares had fallen 5 percent when the original $250 billion figure was first reported (The Wall Street Journal, August 14, 2026).
June 2, 2026. Alphabet priced an $84.75 billion equity raise, upsized from the $80 billion announced the day before. Public buyers paid $355.1982 for Class A and $351.8018 for Class C. Buyers of the mandatory convertible preferred took a 6.25 percent coupon running to May 2029. Berkshire's $10 billion came in a concurrent private placement at $351.81 for Class A and $348.20 for Class C, a discount of about one percent, with no preferred and no warrant attached (Alphabet Form FWP, SEC, June 2, 2026).
July 22, 2026. Alphabet spent $44.924 billion on property and equipment in the second quarter against $39.069 billion of cash from operations. Free cash flow came in at negative $5.855 billion, its first quarterly outflow in close to twenty years, though the trailing twelve month figure is still positive at $53.273 billion (Alphabet second quarter results, SEC exhibit 99.1). On the earnings call the company lifted full year capital expenditure guidance to a range of $195 billion to $205 billion, from the $180 billion to $190 billion it had given in April.
December 8, 2025. Todd Combs left Berkshire for JPMorgan (CNBC, December 8, 2025). Amazon was never Buffett's own pick. When the stake first surfaced in 2019 he told CNBC that one of the two managers in his office had bought it, and Combs is the one reported to have run it since.
The exit is visible line by line in the filings. Berkshire held 10,000,000 Amazon shares at September 30, 2025. At December 31, 2025, the close of Buffett's final quarter as chief executive, it held 2,276,000, a cut of 77.2 percent. At March 31, 2026, Abel's first quarter, the position is gone, in the same filing that drops Visa, Mastercard, Domino's Pizza and UnitedHealth (13F-HR filings, SEC, November 14, 2025 through May 15, 2026).
Why these belong together
Put them in order and the succession story stops working.
The Amazon sale started under Buffett, not Abel, and it finished as housekeeping. Four other positions were closed alongside it. The thing those five names have in common is the manager reported to have picked them, and he had resigned in December.
The Alphabet purchase was Buffett's own idea, and he put it plainly to CNBC on July 15, 2026. "I initiated it," he said. He was just as plain about who signs. "I am not doing anything that he doesn't approve of. He's not doing anything I don't approve of. We talk all the time, but he is the decider."
So the chairman sourced the idea and the chief executive carried it. That is close to the inverse of the story that ran, in which a new man arrives and rebuilds the portfolio.
His reason was not artificial intelligence. Buffett said he had made a mistake by not buying Alphabet years earlier, having watched Google's advertising business work from the inside as an early customer through Geico, and having decided at the time that he could not pick the winner. He has been describing that error publicly since 2018. The position is a man correcting it.
Two things he said in the same interview cut against the purchase. He does not like Alphabet as well as at least four or five other businesses Berkshire already owns. And he named the spending itself as the open question: "The real question with Google and all of its competitors now, because they're all laying out hundreds of billions, and that's real money."
Berkshire committed $10 billion to a company its chairman ranks below several things he already held, in a business whose capital requirements he had just called the real question.
Why the cheque was this size
Buffett's stated reason explains the company. It does not explain the size or the shape of the trade, and what follows is my reading rather than anything he said.
When you are sitting on $397 billion in cash, the hard part stops being which business is good. It becomes which one will absorb $10 billion in a single ticket without the price running away from you.
A company raising $84.75 billion of equity in one week is one of the very few places on earth where that trade exists at all. It also explains the ranking problem. Buffett did not buy his best idea. He bought the best idea that came in his size.
Berkshire has always charged for this, and this time it did not
Berkshire has been the buyer of size before, and it has never done it for free.
| Deal | Year | Structure |
|---|---|---|
| Goldman Sachs | 2008 | $5 billion preferred at 10 percent, plus warrants |
| General Electric | 2008 | $3 billion preferred at 10 percent, plus warrants |
| Bank of America | 2011 | $5 billion preferred at 6 percent, plus warrants on 700 million shares |
| Occidental Petroleum | 2019 | $10 billion preferred at 8 percent, plus a warrant on 80 million shares at $62.50 |
| Alphabet | 2026 | $10 billion of common stock. No preferred, no coupon, no warrant. |
Occidental is the exact parallel. The same $10 billion, also written to fund a specific corporate need, also negotiated directly with the issuer. In 2019 that cheque earned an 8 percent perpetual dividend and a warrant on 80 million shares. In 2026 it earned a price about one percent below what public buyers paid in the concurrent underwritten sale.
Ordinary buyers in that same deal took 6.25 percent a year on the preferred until 2029. Berkshire could have asked for that structure. It has asked for it every other time. It took the common instead.
That is the tell. When Berkshire takes preferred and warrants, it is renting out its balance sheet and expects to be paid for the rental. When it takes plain common at a one percent concession, nobody is paying it to show up. It simply wants the shares.
The other balance sheet moved the other way
Set the two August 14 facts side by side, because they are the same trade seen from opposite ends.
Berkshire committed $10 billion to help fund Alphabet's compute. Nvidia cut its backstop behind OpenAI's compute by more than half, because its own shareholders did not want it carrying that much of the risk.
Both are balance sheets of enormous size. Both are financing the same buildout. In one week one stepped forward and the other stepped back, and the one that stepped back is the company whose chips the whole buildout exists to buy.
The financing is still being assembled, just further from the chipmaker. Nvidia announced this month that it is partnering with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR on compute financing platforms, aiming to deploy more than $500 billion of outside capital. What changed on August 14 is not the amount of money going in. It is who takes the first loss.
What this does to my own cash test
On August 12 I published that the AI buildout is being financed by somebody else's balance sheet, and that the market was reading that transfer as reassurance. Two days later the filing named one of those balance sheets.
Buffett had already said the difficult part out loud in July. The hundreds of billions are the real question. He wrote the cheque anyway, which is either the flaw in my test or the flaw in his, and it is worth being honest that I cannot yet tell which.
The names I rated do not look better for it. What follows is the part the filings did change, and I would rather write it down than pretend they changed nothing.
Here the filings corrected something I had assumed. The easy version of this story is that both hyperscalers are burning cash on AI, so the swap cannot have been about cash. Read the actual statements and they are not in the same position at all.
Alphabet's second quarter free cash flow was negative, but across twelve months it still converted $53.273 billion. Amazon has gone further. Its trailing twelve month free cash flow is negative $7.604 billion, against positive $18.184 billion a year earlier, a swing the company attributes to a $66.1 billion year over year increase in net spending on property and equipment (Amazon second quarter results, SEC exhibit 99.1).
One is a quarter into the crossover. The other is a full year past it. Berkshire sold the one that has gone cash negative over twelve months and bought the one that has not, which is the opposite of what the shared burning cash framing would predict. I am not claiming that was the reason. Buffett gave a different reason and I take him at his word. But the cash test does separate these two, and it separates them in the direction the trade went.
The rest of the market is arriving at the same place I did, and faster than I expected. Alphabet, Amazon, Meta and Microsoft are now projected to spend $740 billion between them on AI computing infrastructure in 2026, and about $1 trillion in 2027, increasingly funded with long dated debt. Alphabet carries $27 billion of bonds maturing in 2056 or later, Amazon roughly $32 billion, Meta about $21 billion. Maria Llerena of Domini Impact Investments put it to Bloomberg on August 16, 2026: "Capex has largely been self-funded from cash flows, but now it seems to be tipping over into the companies raising debt."
That is the August 12 observation, arriving as consensus four days later. The thirty year Treasury yield sitting near its highest level since 2007 is what that shift costs.
What separates Alphabet from the four buildout names I reviewed is not the cash statement either. It is what sits behind the spending. Alphabet's capital expenditure is a decision taken against a business that generated $174 billion of operating cash flow in the twelve months to March 31, 2026, and a cloud backlog above $460 billion with roughly half expected to convert inside 24 months. It also holds assets with no connection to search: 551.2 million SpaceX shares at June 30, worth about $94.2 billion at that day's price of $170.86, against the $900 million it put in during 2015 (Reuters, August 14, 2026). CoreWeave's spending is an obligation taken against a $98.8 billion order book where only 36 percent converts inside two years.
One of those can stop spending tomorrow and still print money. The other cannot stop at all. My test measured the cash and missed that difference, and the fix is to measure whether the spending is reversible, not only whether it is covered.
What it does not tell you
A 13F is a photograph of long US listed equity positions on June 30, already six weeks stale when it lands. It does not show what was sold in July or August. It shows nothing of bonds, foreign listings, or anything held short.
The filing came out after the close on Friday, August 14. No session has traded since. Whatever the market makes of it does not exist yet, and anyone describing a reaction today is describing something that has not happened.
The case for Alphabet is also not clean. Gemini 3.5 Pro, its flagship model, is delayed with no announced date, which has left investors questioning the roadmap in commercially important areas like AI coding (Bloomberg, August 13, 2026). A company spending $200 billion a year on compute eventually has to turn it into models people choose over the alternatives.
The $10 billion placement bought roughly 28.6 million shares at the disclosed prices. Berkshire added 48.1 million. Where the other 19.5 million came from, and at what average price, is not in this filing.
Berkshire has never said why it closed Amazon, Visa, Mastercard, Domino's Pizza and UnitedHealth inside a single quarter. The link to Combs is an inference the press has drawn and I am drawing it too. It fits the timing and it fits who is reported to have picked those names, but it is not a stated reason and it should not be read as one.
Buffett said he initiated the Alphabet idea and that Abel is the decider. Which of the two settled on $10 billion, on a private placement rather than open market buying, and on common rather than preferred, is not something either man has said. I am reading the structure, not a statement.
Buffett's remark about preferring four or five other businesses was not a ranked list and he named none of them. It cannot be checked.
And none of this scores Berkshire. Being early, large and famous is not the same as being right. Whether $10 billion into a company spending upwards of $200 billion a year on compute was a good decision gets settled by the return on that capital expenditure, and that takes years to observe. I have the identical problem with my own committee: seven sessions of ratings on the record, not one of them yet resolved against an outcome.
BagsCapital manages no outside capital and holds no position in Alphabet, Amazon, Berkshire Hathaway, Nvidia, or any listed equity. August 16, 2026.