MU, CRWV, SMCI, VRT: The AI Buildout Fails the Cash Test

The market prices the AI buildout as one trade. On the filings it is at least two, and neither half clears our bar.

One half sells into the buildout and does not convert revenue into cash. Super Micro grew revenue 53 percent over the past twelve months and consumed 6.85 billion dollars of free cash flow doing it. CoreWeave grew 21 percent and consumed 10.62 billion. Between them, 17.47 billion dollars of cash went out the door, and on August 11 the market rallied both on revenue guidance.

The other half converts cash and has already been repriced for it. Micron produced 26.17 billion dollars of free cash flow and Vertiv 2.93 billion. Both now yield less on free cash flow than a ten year Treasury, 2.67 and 2.70 percent against 4.68 percent, while carrying full cyclical risk.

That is the whole setup. You can own cash generation at a sub-Treasury yield, or you can own growth that is financed by somebody else's balance sheet. On August 11, 2026, our four investment seats reviewed all four names and found no third option.

Ratings at a glance

Prices are the August 11, 2026 close. Financials are trailing twelve months read from each company's SEC filings. All figures in US dollars.

Four seats review every name independently, on one identical fact sheet, with no visibility into each other's work. They are not four opinions from one desk. Each runs on a different underlying model and a different published philosophy:

A name is only admitted to anything when three of the four return Buy. None did.

Why the AI trade came apart

The financing changed hands first. A coalition of six large investment firms is assembling more than 500 billion dollars to fund the AI build-out, with the chip supplier backstopping up to 25 percent of some projects and outside firms underwriting the rest independently (Bloomberg, August 11, 2026). Risk moved off one balance sheet onto other people's, and the market read it as reassurance rather than as a transfer.

The equity market had already started sorting. Foreign investors bought 1.7 billion dollars of Taiwanese stocks this month while selling 6.2 billion from South Korea, preferring a less cyclical hardware chain to one levered to memory (Bloomberg, August 11, 2026). SK Hynix and Samsung Electronics sit roughly 50 and 34 percent below June highs on peak memory fears (Bloomberg, August 11, 2026).

Meanwhile the same session saw Super Micro close up more than 7 percent and CoreWeave up 14 percent, both on revenue outlooks (Bloomberg, August 11, 2026). The theme is no longer trading as a bloc. It is trading on which line of the financial statements you are willing to look at.

Backdrop for every session: S&P 500 at 7728.20, VIX at 15.28, ten year yield at 4.68 percent, dollar index at 99.86, with July CPI due the following morning and swaps pricing the odds of a rate hike next month at roughly a coin toss.

Micron Technology, the cheapest expensive stock in the group

Rating: Sell. Three of four seats.

Micron's trailing numbers are the best in the group and are the reason to sell it. Revenue of 90.27 billion dollars is up 141.5 percent. Net income is 50.47 billion, a 55.9 percent net margin and a 50.1 percent return on equity. Free cash flow is 26.17 billion. The balance sheet is a fortress at 0.3 times liabilities to equity with 15.79 billion of net cash.

The stock trades at 19.4 times trailing earnings, which reads cheap until you ask what those earnings are.

Source: annual results as filed, SEC EDGAR, retrieved August 12, 2026. Mean and median computed.

Trailing earnings are 11.0 times the ten year mean and 3.6 times the previous cycle peak. Against that mean, a 980.56 billion dollar market capitalization is 213 times mid cycle earnings. Against the ten year median it is 179 times. The company also trades at 9.7 times a book value of 100.72 billion and yields 2.67 percent on free cash flow.

Memory has never sustained margins like these, and the industry is signalling the turn. Chinese producer CXMT has narrowed the technology gap in smartphone memory (Bloomberg, August 4, 2026), and Micron's two large peers have already lost a third to a half of their value.

Investor A framed the multiple: "A 19.4x P/E on peak earnings is not cheap. It is expensive wearing a cheap disguise." Investor D, which classifies businesses by growth type and had been neutral, concluded that the price "is treating peak cycle earnings as a permanent structural shift, ignoring that four of the last seventeen years were loss making." Investor C read the smaller drawdown against Korean peers as a timing artifact rather than a quality premium: "not a quality premium. It is a lag."

What would change this. A durable structural floor under memory pricing. Micron's high bandwidth memory position could genuinely differentiate it from commodity DRAM, and multi year contracts are becoming standard, with Samsung expecting them to cover 60 to 70 percent of planned capacity. 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, where the backlog is the obligation

Rating: Avoid. Three of four seats, the fourth at Hold.

CoreWeave rents AI compute at scale. It reported second quarter sales of 2.58 billion dollars against a 2.56 billion estimate, a net loss of 1.14 dollars a share, and a backlog of 104 billion (Bloomberg, August 11, 2026). The stock rose 14 percent.

The backlog is the entire bull case, so it is worth reading how it is actually structured. In its most recent 10-Q the company reported 98.8 billion dollars of unsatisfied remaining performance obligations as of March 31, 2026, of which 36 percent is expected to be recognized in the first 24 months, 39 percent between months 25 and 48, and the balance between months 49 and 84 (10-Q).

So roughly two thirds of the order book converts beyond two years, while the capital expenditure to serve it is spent now. That is the shape of an obligation, not an asset.

Enterprise value is roughly 73 percent of the entire seven year order book. Customer concentration compounds it: for the three months ended March 31, 2026, one customer was 45 percent of revenue and a second was 20 percent, and three customers accounted for 39, 17 and 22 percent of net receivables. A year earlier a single customer was 72 percent of revenue.

Investor A described the model as borrowing at debt market rates to buy depreciating hardware and renting it out into a shortage, then noted that "shortages end." Investor C called the order book "a liability disguised as an asset, because it locks the company into a capital expenditure path it cannot deviate from without breaching contracts." Investor D, the growth seat, called the balance sheet and cash burn "terrifying."

What would change this. 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. Contracted margins and cancellation provisions are not disclosed, which is itself a reason the seats would not underwrite it.

Vertiv Holdings, a good business at a demanding price

Rating: Split. One Buy, two Watch, one Avoid.

Vertiv sells the power and thermal infrastructure data centers cannot operate without, and it is the one name here where the committee liked the business almost unanimously. Free cash flow of 2.93 billion dollars on 11.48 billion of revenue is a 25.5 percent margin, return on equity is 36.4 percent, and the balance sheet is close to net flat with 129 million of net debt.

Investor C made the structural case in a sentence: "You cannot finance your way out of a power and cooling equipment shortage." The 500 billion dollar financing coalition does not manufacture a single cooling unit, and physical bottlenecks are the constraint capital cannot relieve.

The objection is price. Vertiv trades at 62.6 times earnings, 37.0 times free cash flow and 22.8 times book, on 12.2 percent revenue growth. Capitalising trailing free cash flow at a defensive 15 to 20 times, Investor B arrived at an indicated equity value of 44.0 to 58.6 billion dollars, or 114 to 152 dollars a share, putting the current price 85 to 147 percent above that range. That is a valuation framework, not a forecast.

Investor A, who wanted to own it, put the problem in yield terms. A 2.70 percent free cash flow yield sits below the 4.68 percent ten year, so the buyer takes less current cash per dollar of price than a government bond pays, and makes up the whole difference on growth that has not happened yet.

What would change this. Price. Investor A named a 4 to 5 percent free cash flow yield as the level where the business becomes worth owning. Separately, industry reporting is explicit that AI factories are producing winners and losers within power equipment rather than lifting every supplier (Bloomberg, July 3, 2026), and Vertiv's order backlog is not disclosed, so share gains cannot be verified from the filings.

Super Micro Computer, the widest disagreement on the sheet

Rating: Split. One Buy, one Watch, one Hold, one Avoid.

Super Micro assembles AI servers. It guided to 65 to 72 billion dollars of fiscal year revenue against a 54.4 billion consensus after a quarter where sales rose 93 percent to 11.1 billion (Bloomberg, August 11, 2026). At 15.2 times earnings and 0.6 times sales it is the cheapest headline multiple in the group, and it sits at 31 percent of its 52 week range.

The company earns a 3.7 percent net margin while its principal supplier earns roughly 41 percent. It also consumed 6.85 billion dollars of free cash flow over twelve months, and the composition explains the whole gap. For the nine months to March 31, 2026, inventories absorbed 6.67 billion and receivables 6.21 billion, against 2.41 billion provided by payables. Inventory on the balance sheet went from 5.73 billion at September 30, 2025 to 11.10 billion at March 31, 2026.

Inventory plus receivables is now 19.51 billion dollars against a 19.00 billion dollar market capitalization. You are buying the working capital and getting the operating business for nothing, which is either the opportunity or the warning.

The committee split on exactly that number. Investor A called the pile "not a growth footnote, they are the actual risk exposure," and framed the company as a leveraged bet on the spread between component costs and selling prices at the moment memory pricing rolls over. Investor C called it "a deliberate inventory position ahead of a massive fulfillment cycle," noting you cannot ship 65 to 72 billion dollars of revenue without building inventory first, and moved to Buy. Investor B held Avoid throughout on a conservative valuation of 21 to 25 dollars a share against a 31.60 price.

What would change this. Inventory reserves and the split between raw components and finished systems, neither of which is disclosed. That single disclosure would settle the disagreement. Separately, the company disclosed US export control charges against a co-founder, which is a governance overhang at a moment it may need capital markets access.

The test we actually applied

Across four names the seats converged on one question, and it was not growth. It was whether reported revenue becomes cash, and at what price that cash is being sold.

On that test the group inverts. The two names with the strongest revenue growth, Super Micro at 53 percent and CoreWeave at 21 percent, are the two consuming cash. The two producing cash, Micron and Vertiv, are priced at free cash flow yields below the risk free rate. The market spent August 11 sorting this theme on revenue. Sorted on cash, the ranking runs close to backwards, which is why a single theme produced no consensus buy in sixteen independent assessments.

What this note does not establish

It does not forecast prices. No seat was asked to name one and none did, and the valuation ranges quoted above are the seats' own frameworks for what they would pay, not predictions of where anything trades.

Several figures that would sharpen these calls are simply not disclosed: CoreWeave's contracted margins and contract cancellation provisions, Vertiv's order backlog, and Super Micro's inventory reserves. Where a company does not publish something, we report the absence rather than estimating around it.

This is also one date at one set of prices. Micron sat at 66 percent of a 52 week range running from 113.46 to 1255.00, and across a range that wide the entry point does more work than the analysis.

Nothing here is a recommendation. These are committee ratings on an advisory review, not advice to any reader.

A note on method

Every session ran on a single fact sheet built from SEC filings, with news attached separately and clearly marked as weaker evidence than a filing. The seats see the sheet and nothing else, which is what makes their disagreement meaningful rather than an artifact of different inputs.

Our first pass on these names carried four fields marked unavailable: CoreWeave's share count and customer concentration, Micron's mid cycle earnings, and the composition of Super Micro's cash outflow. Three were in the filings and one was arithmetic on seventeen years of them. Our sheet reads structured XBRL data, which does not see a cover page or a disclosure written in prose.

We rebuilt the sheets and re-ran three of the four names. Seven of twelve ratings changed. Micron moved from a divided Watch to three seats at Sell. CoreWeave hardened. Super Micro moved the other way and split the committee, because Investor A and Investor C read the newly available inventory figure in opposite directions.

Both passes remain in the verdict log with their timestamps. We did not delete the weaker run, because the difference between the two is the only evidence of how much the missing figures were worth. Vertiv was not re-run: its one gap, the order backlog, is the one the company does not disclose, and its ratings should be read as resting on less information than the other three.

Read the prior committee review of NVIDIA, which returned no buyers from the same four seats.

BagsCapital manages no outside capital. The treasury holds no position in Micron, CoreWeave, Vertiv or Super Micro Computer, and holds no listed equities of any kind. The committee is advisory and does not trade.