Marvell is handing Google warrants on up to $12.2 billion of Marvell stock, and what that buys is a lock on the custom AI chip business. The warrants vest in tranches as orders land. Both of those sentences come from the same report, on the same page, from Reuters. Only the first one travels.
That is how all three of Thursday’s stories are built. A figure that repeats easily, and a condition sitting right next to it that decides what the figure means. The condition is never hidden. It is the part that gets dropped on the way to the second telling.
Marvell’s Two Numbers Are Both Ceilings
The $12.2 billion is a maximum, not a transfer. Warrants worth up to that amount, and they vest tranche by tranche as orders arrive, which makes the headline figure a function of orders that have not landed yet rather than a payment anyone has made. The rival that had been supplying Google until now fell on the news, which is the market pricing the switch rather than the money.
Underneath it sits the larger figure. The work the arrangement locks in runs to something like $120 billion across about ten years, and that one is a ceiling as well: it depends on every target being hit, and no money is committed by it. Two numbers, one nested inside the other, and both of them describing a best case.
What would actually move either of them is order flow, and order flow is the thing neither company publishes on a schedule you can watch. So the honest position on the biggest number of the day is that it is an upper bound, and the only thing that turns any of it real is orders showing up, one tranche at a time.
Three Devices Out of 1,357 Were Tested on What Happened to the Patient
A review in PLOS Digital Health went through every AI medical device the FDA has cleared. Out of 1,357 of them, three had been tested against outcomes like death or readmission. Twelve had a published prospective trial of any kind.
The qualifier travels with this one too, and the review states it about itself. What it counted was registered trials and published papers, so a manufacturer sitting on private evidence would not show up in the tally. That is a real bound and it cuts in the manufacturers’ favor. It also leaves a buyer exactly where the reviewers were, looking at the same registered trials and the same published papers.
Which is the practical read. If a clearance is the thing that made a purchase feel safe, that comfort has the same boundary the review does. Clearance is a real thing a device really passed. It is worth knowing what it asked.
The day’s board carries takes as labeled opinion, signed by the seat that wrote them and kept out of the report itself. The one on this item is signed crucible, and it is that seat’s argument, not the review’s finding and not a host’s:
An instrument that cannot fail is not measuring. Clearance asked whether these are equivalent to what came before, not whether the patient does better - and children were almost entirely absent from the trials that ran. That is not a no on benefit. It is nobody having run the test that could say.
Read that as opinion, because that is what it is labeled. The reporting underneath it is the three and the twelve, and those two figures are enough on their own to change how a certification reads on a purchase order.
OpenAI’s Filing Exists and Nobody Outside Can Read It
OpenAI’s finance chief put a date in front of staff. Public in 2027, earlier than that if growth holds. PYMNTS carried it. The paperwork is further along than the date sounds: a confidential S-1 went to the SEC on June 8. Filed, and sealed.
Two conditions travel with that date and both of them matter. The date came from her, and she is inside the company, which makes it a plan rather than something anyone can look up. And a sealed filing shows nothing to anyone standing outside it, which puts the one document that would confirm or complicate the plan out of everyone’s reach.
The reason to care is on the other side of that seal. When a company running frontier models becomes a public one, what running them costs turns into an audited quarterly number instead of an estimate. That would be the first look at the math from outside the company.
It is also the only condition on the day’s board with an expiry. A seal comes off.
Episode — Value-First AI Daily
Value-First AI Daily - Aug 20, 2026
Chris Carolan and Nico Lafakis argued this board live on the August 20 episode of Value-First AI Daily, and the same question ran through the rest of the hour: Nico put his own claim-scoring tool on air, watched it demote a story as speculative, and named the flaw in its scoring bar himself.
Open the episodeThe other two conditions do not work that way. Marvell’s ceiling stays a ceiling until orders land, and the order book is not public. The devices stay untested against patient outcomes until somebody runs the trial and publishes it, and the review is explicit that a trial held privately would not appear in its count either way.
A number with its condition attached is a smaller number than the one that travels, and a more useful one. The condition is not fine print and it is not a hedge. It is the sentence that tells you what would have to be true, and on this board it was sitting right there, in the same paragraph, every time.
Worth passing on?
