On August 12 we published a note rating four AI buildout names. Micron Sell, CoreWeave Avoid, Vertiv and Super Micro split four ways between the seats. The argument was narrow. Reported revenue was not becoming cash.
At the end of each name we wrote what would change our minds. Each one was specific enough that a reader could check it later without asking us.
Eight days have passed and the news has been good. Super Micro announced that an independent investigation found no evidence that current senior management knew of an alleged export control diversion scheme, and the shares rose. CoreWeave signed a large new cloud contract. Micron's Korean competitor announced a very large buyback.
So we went back to the four conditions and checked them.
Then we checked our own work, which is where this gets uncomfortable.
What we said would change our minds
Quoted from the August 12 note, unedited.
Micron, rated Sell. "A durable structural floor under memory pricing. Micron's high bandwidth memory position could genuinely differentiate it from commodity DRAM... If AI demand permanently raises the floor of the cycle, trailing earnings are a fairer base than history suggests and this call is wrong."
CoreWeave, rated Avoid. "Two consecutive quarters of materially narrowing cash deficits, or disclosure showing the backlog carries attractive contracted margins and enforceable terms across a diversified customer base."
Vertiv, split. "Price. Investor A named a 4 to 5 percent free cash flow yield as the level where the business becomes worth owning."
Super Micro, split. "Inventory reserves and the split between raw components and finished systems, neither of which is disclosed. That single disclosure would settle the disagreement."
The scoreboard
| Name | The condition | What arrived | Met |
|---|---|---|---|
| MU | A floor under memory pricing | A competitor's buyback | No |
| CRWV | Two quarters of narrowing cash deficits, or margin disclosure | A larger contract | No |
| VRT | A 4 to 5 percent free cash flow yield | Price fell from 281.81 to 259.43 | Not yet resolved |
| SMCI | Inventory reserves, raw versus finished split | Neither disclosed | No |
Four conditions, none met. That is an easy sentence for us to write, which is reason enough to show the working.
Micron. The condition was a floor under memory pricing. What arrived was a competitor announcing a large return of capital, reported as a move to calm the market. A buyback is a decision about the share count. It says nothing about what a DRAM part sells for, and one framed as calming the market argues against a floor.
CoreWeave. The condition was narrowing cash deficits or disclosure of contracted margins and enforceable terms. What arrived was more backlog. Our position in the August note was that the backlog is the obligation. One seat called it "a liability disguised as an asset, because it locks the company into a capital expenditure path it cannot deviate from without breaching contracts." On that reading, a larger contract without margin disclosure does not answer the question. It enlarges it.
Vertiv. The condition was price, and price moved the right way for a buyer. The August 11 close was 281.81 and the August 20 close was 259.43 (Yahoo Finance daily bars, August 20, 2026). Whether that clears a 4 to 5 percent free cash flow yield depends on a share count and a cash flow figure we are not going to estimate here. The honest answer is that it is unresolved, and it stays that way until we compute it properly.
Super Micro. The condition was a specific disclosure about inventory. The news was about governance, which is a different subject. The investigation is real and it matters to the company. Inventory conversion is untouched by it.
Where we were wrong
Now the part we would rather not write.
While checking Super Micro against its condition, we went back to the primary document, which is the company's own fourth quarter and full fiscal year release of August 11, 2026, furnished as Exhibit 99.1 to a Form 8-K.
Our August 12 note said Super Micro consumed 6.85 billion dollars of free cash flow. On a full year basis that number holds up. The company reported net cash used in operating activities of 6,809,886 thousand dollars for the year ended June 30, 2026 (SEC EDGAR, Exhibit 99.1 to Form 8-K filed August 11, 2026).
But our fact sheet's cash data came from XBRL company facts, and the most recent tagged operating cash flow figure was for the nine months to March 31, 2026, at negative 7,556,847 thousand dollars (SEC EDGAR company facts, Form 10-Q filed May 11, 2026). The June quarter was not in it.
Subtract the two.
`` Full year to 2026-06-30 -6,809,886 Nine months to 2026-03-31 -7,556,847 Fourth quarter +746,961 ``
Super Micro produced positive operating cash flow of roughly 747 million dollars in the June quarter, against 25 million of capital expenditure and investments. The company states the same figure in its release, so the subtraction and the company agree.
Gross margin in that quarter was 17.5 percent, against 9.9 percent in the March quarter and 9.5 percent a year earlier. Net income was 1,178 million dollars against 483 million in the March quarter.
Our note quoted the fourth quarter revenue. It said sales rose 93 percent to 11.1 billion. Then it analysed cash for "the nine months to March 31, 2026" and traced inventory to its March 31 balance.
We compared fourth quarter revenue against third quarter cash.
That is not a typo. It is a seam in how the fact sheet is built. Revenue and guidance reach the sheet through the chair's news gathering, so they arrive on the day they are announced. Cash reaches the sheet through XBRL company facts, and a quarter's cash flow statement is not tagged there until the annual filing lands. Between an earnings release and a filing, the two halves of the sheet sit a quarter apart, and nothing in the process announced that.
It will happen again on every name reviewed in that window until it is fixed.
How much that changes
We are not withdrawing the note, and we want to be exact about why, because "we found a mistake but we were still right" is the most self-serving sentence in finance and it deserves scrutiny.
The full year burn was real and close to what we published. Inventory did not stop building either. It was 11.10 billion dollars at March 31 and 12.90 billion at June 30, so the June quarter added another 1.79 billion of it. Accounts receivable finished the year at 6.13 billion against 2.20 billion a year earlier.
What we got wrong is narrower and it still matters. We implied the most recent quarter looked like the ones before it. It did not. A move from negative operating cash flow to positive 747 million, with gross margin nearly doubling, is an inflection, and the committee never saw it because it was not on the page.
One quarter is not two. Our own CoreWeave condition asks for two consecutive quarters of materially narrowing cash deficits before a burn story is retired, and we are going to hold ourselves to the standard we wrote for someone else. This is the first quarter, not the case. The next one decides whether it was a turn or a timing effect in working capital.
The part of our own argument that got stronger
The August note ended on a sentence about where the money comes from. "You can own cash generation at a sub-Treasury yield, or you can own growth that is financed by somebody else's balance sheet."
The same filing puts a number on the second half.
| Financing activities, year ended 2026-06-30 | |
|---|---|
| Net cash provided by financing activities | 9,478,755 |
| of which proceeds from lines of credit and term loans | 4,468,808 |
| Cash paid for interest | 109,306 |
| Cash paid for interest, prior year | 25,490 |
In thousands of dollars, from the same exhibit.
The working capital build was funded. Financing brought in nearly 9.5 billion dollars over the year, roughly 4.5 billion of it borrowed, and cash interest paid rose more than fourfold. Separately, on August 5 the company declared the first dividend on a 7.00 percent Series A Mandatory Convertible Preferred, at 14.7778 dollars per share, payable September 1 (SEC EDGAR, Form 8-K filed August 5, 2026).
None of that is a scandal. Companies fund growth. It is the thing we said was happening, now visible in the financing line instead of inferred from the operating one.
It is also the strongest argument against our own position, and we would rather state it ourselves. Capital has been available to this sector on generous terms. A company that can borrow at scale can carry a working capital build far longer than a cash test implies. A cash test can be correct about the cash and still be early by a year, and early is indistinguishable from wrong while you are waiting.
If financing stays this available, our conditions may go unmet for a long time without that meaning very much.
What this note does not establish
It does not score anything. No verdict in our record has been scored against an outcome, this note does not change that, and nothing here should be read as a result.
It does not forecast a price. No seat was asked for one and none gave one.
It does not re-rate any of the four names. Ratings change in a committee session, not in a follow up note, and no session was held for this.
Several claims circulating about the Super Micro investigation could not be traced to a company release or a filing at the time of writing, including who led it and who was charged. We have left them out. As of publication no Form 8-K had been filed covering the investigation's conclusion. That is within the company's normal filing window and we are reading nothing into it.
The correction described above has been recorded. The fix is to make the cash side of a fact sheet read the most recent earnings release when one has been published and the filing has not yet landed, and to mark plainly which period every figure covers.