The FT says OpenAI's annualized revenue is about $50B, not the ~$70B everyone was passing around. CNBC and CNN's follow-ups say it's gross vs net, not lost sales.. Accountants, please calm down.

Oracle fell 5.5% anyway, about $24B gone in a day over what's basically a footnote.

Except it isn't one for Oracle. OpenAI signed up last year to buy roughly $300B of computing from it over about five years, starting 2027. That's ~$60B a year from a customer bringing in $50B. Basically a $6k/month lease on a $5k/month salary, plus a promise the raise is coming.

The only Oracle insider who bought stock this year paid $139.35 on Sep 29. It closed at $135.69. Welcome to the club.

Counting problem, or the first real crack in the AI spending story?

No position. Just popcorn.

The memory and storage names were the trade of 2026, and on paper they still are. SanDisk is up about 600% this year, Seagate about 190%, Western Digital about 135%.

But all three topped in June. Western Digital is down almost half from its high, Seagate and SanDisk 25-30%. On Tuesday Seagate fell 9% and WD 7%, and yesterday neither got it back.

Micron held up much better, down roughly 12% from its high after an 87% gross margin quarter.

Meanwhile Nvidia, the laggard of the group back in September, closed at a record on Tuesday.

Who's still holding these?

Delta reports Q3 Friday before the open.

Berkshire sold all four of its US airlines in 2020 at a loss. This year, under Abel, it went back in. 39.8M Delta shares in Q1, another 17.5M in Q2. That's 57.3M shares, about 8.7% of the company.

It's small for Berkshire, under 2% of the portfolio. But you don't end up with 8.7% of an airline by accident.

The stock closed at $83.66 yesterday, about 11% below where it ended June, so the stake is worth around $4.8B now against $5.4B on the June 13F.

What would you want to hear tomorrow to think this was a good call?

The 30-year touched 5.62% this week, the highest since 2002, and closed Tuesday at 5.59%.

Forward earnings yield (1 / forward P/E) for the 10 biggest US companies:

- Micron 13.9%

- Nvidia 5.3%

- Broadcom 4.9%

- Meta 4.4%

- Berkshire 4.4%

- Alphabet 3.9%

- Microsoft 3.9%

- Amazon 3.8%

- Apple 2.8%

- Tesla 0.5%

Only Micron clears the long bond, and it reports today. For the other nine, growth is the whole argument.

Is 5.6% risk-free finally enough to change what you hold?

I took every profitable US company over $10B, 486 of them, and compared earnings yield (earnings divided by price) with the 10-year at 5%.

335 of them, 69%, earn less per dollar of price than the Treasury pays. Among the 50 largest it's 44. The median P/E is 26.7, which works out to an earnings yield of 3.75%.

The obvious answer is that earnings grow and a coupon doesn't. Fair enough, that's the whole equity bet. But it does mean you're paying a real premium today for growth you still have to get.

Does the gap change anything for you, or has stocks-vs-bonds stopped being a comparison people actually make?

On 3 August the CEO, Mike Mahoney, bought 186,240 shares on the open market at about $48. That's roughly $9M of his own money. Three directors bought in the same few weeks.

These weren't plan trades or option exercises. Plain open-market purchases. People sell stock for a dozen reasons, but they generally buy it for one.

Worth saying they're already underwater. The stock is about 10% below where he paid.

What actually happened to BSX this year? I know the chart better than I know the story, and I'd rather hear it from someone who follows medtech

Nobody checks whether analyst ratings actually work, because to score one you need the share price on the day of the call, and hardly anyone keeps that. I've been storing them since 2011. 149,587 of them, with the firm named on each.

So I scored them. Every call gets compared against the median stock in the same universe over the same window, starting the same day. That way survivorship bias sits on both sides and cancels. Medians throughout, because a few ten-baggers otherwise drag the average all over the place.

Horizon	Upgrades	Beat rate	Downgrades	Beat rate
1 week	+0.06%	50.9%	−0.06%	49.0%
1 month	+0.14%	50.8%	+0.03%	50.2%
3 months	−0.00%	50.0%	+0.26%	51.1%
6 months	+0.72%	51.8%	+0.22%	50.5%
12 months	+0.93%	51.3%	+0.64%	51.0%

The beat rate never leaves 49-52%. Doesn't matter if you hold it a week or a year. At 3 months upgrades land on 50.0% with zero median excess, which is about as clean a coin flip as market data gets.

Downgrades are no better in reverse. If a downgrade meant anything the number should stay negative the whole way across. It turns positive at 3 months and stays there.

Averages are doing a lot of work in this industry. UBS upgrades average +32.7% excess. Their median is −0.6% and fewer than half beat the benchmark. A few enormous winners carrying a typical call that lost. All firms together: +10.1% mean, +0.9% median. Next time someone quotes a track record at you, ask for the median.

Best and worst by median at 12 months: Stifel +4.8%, Jefferies −3.9%.

What this can't tell you: companies that got downgraded and then delisted have no price 12 months later, so they drop out of the sample. That makes downgrades look better than they were. The honest version is that among companies which survived, a downgrade told you nothing.

One thing that nearly caught me. First time round I anchored the benchmark to the start of the month instead of the day of the call. Calls happen on any day, so a call made on the 28th was being compared against a market window that opened on the 1st. Different week entirely at short horizons. It gave me downgrades underperforming 0.30% in week one on a 46.4% beat rate, which looked like a real announcement effect and I nearly wrote it up that way. Anchored properly it's −0.06% and 49.0%. Nothing there.

It's my own data, so you can't go and check the rows, which I know is a fair objection. The method is specific enough to redo on a different universe if anyone wants to.

I compiled the Q2 2026 13F filings of 46 well-known value managers (Berkshire, Appaloosa, Himalaya, Baupost, Fundsmith, Oakmark, etc.) to see what they actually did as a group. A few things stood out:

**They split hard on Alphabet (GOOGL).** It was the single most divisive stock of the quarter — 8 managers bought or added it, while 12 sold or trimmed. Buffett himself increased Berkshire's stake ~45%, while Terry Smith (Fundsmith) cut his ~40%.

**The consensus BUY was quality/healthcare, not tech.** Danaher (DHR) was bought by 8 managers, Thermo Fisher by 6, UnitedHealth by 5.

**Buffett tilted toward housing.** New D.R. Horton position, added to Lennar (+30%), while trimming Capital One (-58%) and Nucor (-52%).

**Tepper went risk-on.** Appaloosa opened new Apple, Broadcom, and SpaceX positions and added to Meta (+55%) and Baidu (+87%).

Across all 46 managers: ~$1.1T combined, 213 new positions, 199 full exits.

Curious what people make of the Alphabet split — is Buffett early, or are the sellers taking profits after a big run?

Moderna insiders unloaded $47M of stock in the 180 days prior to it tripling. Mostly autopilot - and the CEO of Moderna came away with an extra 252,469 shares in his $28.7M "sale."

MRNA shares shot up from $62.96 to $174.38 from the market close on August 19. Pull up any tracker of insider transactions for the previous 180-days: 0 transactions by insiders-tens of millions in selling-signals a strong bear market.
I'm looked through every one of the 36 Form 4s filed in the time period. Sales on the open market totaled $47.1M-but 8 of the 9 forms were completed with the 10b5-1 checkbox highlighted-that's the form for the scheduled, automatous sales that takes the decision-making authority away from the individual. Instead of making the selling decisions himself on 15 May, President of Moderna, Stephen Hoge, executed a scheduled 53,336 share sale at $48.40.

Here's why:

Same thing happens on 15 June (53,336 shares at $51.37) and 15 July (53,336 shares at $67.60). The same amount of stock was sold every 15 th of the month on exactly the same share count down to the last share.

The form 4 for the CEO's 'sale' is a good example of how to use them to tell the signal. Bancel's $28.7M of sales were made on August 5 based on stock options set to expire on August 10; the shares would have been worthlessly forfeited if they were not utilized on or before that day. He purchased 751,715 options at a $19.15 strike price. To make his purchase and pay taxes on that transaction, he sold 499,246 shares and consequently netted an increase in his total Moderna holdings from 6,187,791 shares to 6,440,260. Despite being sold at an average $57 on August 5, Moderna's share price shot up to $174.38 on August 19.
There was one transaction with the 10b5-1 checkbox unchecked: Chairman of Moderna Noubar Afeyan sold $434,000 of shares on May 21. This made up only .9% of the reported insider sales.
Takeaway: Most insider selling is routine – tax burdens, needed cash, expiring option assignments, and more. Insider purchasing on the other hand always sends an informative signal – there is only one real reason to buy equity. Every transaction is readily available to look through for yourself in EDGAR-look at the 10b5-1 box at the very top of each transaction, and the footnote at the bottom.

So, why were insiders still selling if they knew what was happening?
There are three possibilities which often accumulate.
They probably had no idea of what was occurring. Clinical trial results are typically made accessible to a small group of scientists a few days in advance of the official publication, which is essentially the whole point of making a clinical trial blinded. On May 15, when the president exercised a portion of his allotted equity for that month, there may have been absolutely no evidence at that point to make any educated guess about any future change in stock price.

Even if they would have had insight at that moment, they already had their plan set in motion. This aspect always trips beginners up: actually canceling a sale under the 10b5-1 agreement because you've gained some privileged knowledge is a red flag and may subject you to investigation. Making the predetermined sale was, therefore, the less suspicious decision of the two possibilities.
And that this happens time after time, as all sales indicate... Most stock compensation is in the form of option grants or restricted stock. In the case of the CEO, Bancel received stock options at a strike price of $19.15 several years ago-there isn't a purchase side, as in you cannot 'purchase' part of your paycheck; what can be done, however, is to liquidate an equity gain to use as income. Taking 200 roughly equivalent market caps into consideration-over 80% of them demonstrated that they had been seeing insider sales in their last 180 day transaction record. A 4 out of 5 signal cannot accurately report specifics.

This brings us back to what seems important, and the last major lesson of this report: it is now virtually constant that insiders will be selling shares of large cap companies; those are merely an automatic byproduct of business. It is an insider purchase transaction-an event that takes place far less frequently-that sends a meaningful signal to investors.

Moderna insiders unloaded $47M of stock in the 180 days prior to it tripling. Mostly autopilot - and the CEO of Moderna came away with an extra 252,469 shares in his $28.7M "sale."

MRNA shares shot up from $62.96 to $174.38 from the market close on August 19. Pull up any tracker of insider transactions for the previous 180-days: 0 transactions by insiders-tens of millions in selling-signals a strong bear market.
I'm looked through every one of the 36 Form 4s filed in the time period. Sales on the open market totaled $47.1M-but 8 of the 9 forms were completed with the 10b5-1 checkbox highlighted-that's the form for the scheduled, automatous sales that takes the decision-making authority away from the individual. Instead of making the selling decisions himself on 15 May, President of Moderna, Stephen Hoge, executed a scheduled 53,336 share sale at $48.40.

Here's why:

Same thing happens on 15 June (53,336 shares at $51.37) and 15 July (53,336 shares at $67.60). The same amount of stock was sold every 15 th of the month on exactly the same share count down to the last share.

The form 4 for the CEO's 'sale' is a good example of how to use them to tell the signal. Bancel's $28.7M of sales were made on August 5 based on stock options set to expire on August 10; the shares would have been worthlessly forfeited if they were not utilized on or before that day. He purchased 751,715 options at a $19.15 strike price. To make his purchase and pay taxes on that transaction, he sold 499,246 shares and consequently netted an increase in his total Moderna holdings from 6,187,791 shares to 6,440,260. Despite being sold at an average $57 on August 5, Moderna's share price shot up to $174.38 on August 19.
There was one transaction with the 10b5-1 checkbox unchecked: Chairman of Moderna Noubar Afeyan sold $434,000 of shares on May 21. This made up only .9% of the reported insider sales.
Takeaway: Most insider selling is routine – tax burdens, needed cash, expiring option assignments, and more. Insider purchasing on the other hand always sends an informative signal – there is only one real reason to buy equity. Every transaction is readily available to look through for yourself in EDGAR-look at the 10b5-1 box at the very top of each transaction, and the footnote at the bottom.

So, why were insiders still selling if they knew what was happening?
There are three possibilities which often accumulate.
They probably had no idea of what was occurring. Clinical trial results are typically made accessible to a small group of scientists a few days in advance of the official publication, which is essentially the whole point of making a clinical trial blinded. On May 15, when the president exercised a portion of his allotted equity for that month, there may have been absolutely no evidence at that point to make any educated guess about any future change in stock price.

Even if they would have had insight at that moment, they already had their plan set in motion. This aspect always trips beginners up: actually canceling a sale under the 10b5-1 agreement because you've gained some privileged knowledge is a red flag and may subject you to investigation. Making the predetermined sale was, therefore, the less suspicious decision of the two possibilities.
And that this happens time after time, as all sales indicate... Most stock compensation is in the form of option grants or restricted stock. In the case of the CEO, Bancel received stock options at a strike price of $19.15 several years ago-there isn't a purchase side, as in you cannot 'purchase' part of your paycheck; what can be done, however, is to liquidate an equity gain to use as income. Taking 200 roughly equivalent market caps into consideration-over 80% of them demonstrated that they had been seeing insider sales in their last 180 day transaction record. A 4 out of 5 signal cannot accurately report specifics.

This brings us back to what seems important, and the last major lesson of this report: it is now virtually constant that insiders will be selling shares of large cap companies; those are merely an automatic byproduct of business. It is an insider purchase transaction-an event that takes place far less frequently-that sends a meaningful signal to investors.

Moderna insiders unloaded $47M of stock in the 180 days prior to it tripling. Mostly autopilot - and the CEO of Moderna came away with an extra 252,469 shares in his $28.7M "sale."

MRNA shares shot up from $62.96 to $174.38 from the market close on August 19. Pull up any tracker of insider transactions for the previous 180-days: 0 transactions by insiders-tens of millions in selling-signals a strong bear market.
I'm looked through every one of the 36 Form 4s filed in the time period. Sales on the open market totaled $47.1M-but 8 of the 9 forms were completed with the 10b5-1 checkbox highlighted-that's the form for the scheduled, automatous sales that takes the decision-making authority away from the individual. Instead of making the selling decisions himself on 15 May, President of Moderna, Stephen Hoge, executed a scheduled 53,336 share sale at $48.40.

Here's why:

Same thing happens on 15 June (53,336 shares at $51.37) and 15 July (53,336 shares at $67.60). The same amount of stock was sold every 15 th of the month on exactly the same share count down to the last share.

The form 4 for the CEO's 'sale' is a good example of how to use them to tell the signal. Bancel's $28.7M of sales were made on August 5 based on stock options set to expire on August 10; the shares would have been worthlessly forfeited if they were not utilized on or before that day. He purchased 751,715 options at a $19.15 strike price. To make his purchase and pay taxes on that transaction, he sold 499,246 shares and consequently netted an increase in his total Moderna holdings from 6,187,791 shares to 6,440,260. Despite being sold at an average $57 on August 5, Moderna's share price shot up to $174.38 on August 19.
There was one transaction with the 10b5-1 checkbox unchecked: Chairman of Moderna Noubar Afeyan sold $434,000 of shares on May 21. This made up only .9% of the reported insider sales.
Takeaway: Most insider selling is routine – tax burdens, needed cash, expiring option assignments, and more. Insider purchasing on the other hand always sends an informative signal – there is only one real reason to buy equity. Every transaction is readily available to look through for yourself in EDGAR-look at the 10b5-1 box at the very top of each transaction, and the footnote at the bottom.

So, why were insiders still selling if they knew what was happening?
There are three possibilities which often accumulate.
They probably had no idea of what was occurring. Clinical trial results are typically made accessible to a small group of scientists a few days in advance of the official publication, which is essentially the whole point of making a clinical trial blinded. On May 15, when the president exercised a portion of his allotted equity for that month, there may have been absolutely no evidence at that point to make any educated guess about any future change in stock price.

Even if they would have had insight at that moment, they already had their plan set in motion. This aspect always trips beginners up: actually canceling a sale under the 10b5-1 agreement because you've gained some privileged knowledge is a red flag and may subject you to investigation. Making the predetermined sale was, therefore, the less suspicious decision of the two possibilities.
And that this happens time after time, as all sales indicate... Most stock compensation is in the form of option grants or restricted stock. In the case of the CEO, Bancel received stock options at a strike price of $19.15 several years ago-there isn't a purchase side, as in you cannot 'purchase' part of your paycheck; what can be done, however, is to liquidate an equity gain to use as income. Taking 200 roughly equivalent market caps into consideration-over 80% of them demonstrated that they had been seeing insider sales in their last 180 day transaction record. A 4 out of 5 signal cannot accurately report specifics.

This brings us back to what seems important, and the last major lesson of this report: it is now virtually constant that insiders will be selling shares of large cap companies; those are merely an automatic byproduct of business. It is an insider purchase transaction-an event that takes place far less frequently-that sends a meaningful signal to investors.