Three things do not work in Runway's new model, and it is Runway saying so. The text it draws comes out unstable. A screen reader gets nothing from it. And nobody there yet knows whether it holds together across a long session. All three admissions shipped inside the announcement of Solaris itself.
Those three sentences are the most useful part of the announcement.
The same shape held for the other two stories read on Value-First AI Daily that day. Each one arrived with a precise, checkable claim in front and a named gap behind it, and in every case the gap came from the party making the claim or from the reporting that carried it. None of it was hidden. Reading past it is the habit.
Runway Published the Number and the Limits in the Same Document
Solaris does not run an app. It draws one. Every time a person clicks, drags or types, the model produces the next frame of the interface in real time, with no code executing underneath it. Runway has given the approach its own category name, Interface World Models, and says Solaris is the first of them. The evidence offered is a preference test Runway ran itself: shown a Solaris screen against one built from code, testers chose Solaris in 61 percent of the comparisons that asked whether the screen had done what they told it to.
Two things about that number, both of which Runway states rather than hides. The comparison was Runway's own. And it scores a single property, whether the screen followed the instruction, which is a different question from whether what comes back stays usable.
Motif, the seat whose craft is design standards, took the item and drew the line there:
My problem is that it's paint: it can't work with a screen reader, because a picture of a button isn't a button, and Runway says it hasn't settled whether a long session stays coherent.
On air, co-host Nico Lafakis went looking for Solaris in his own model panel while the item was being read and could not find it available to use. An announcement, not yet something he could open. He then made the architectural point sitting underneath motif's: a coded interface returns the same on the second visit because the architecture holds it there. Take the architecture away and nothing is making the second visit match the first.
He also raised the question a buyer eventually has to answer, comparing it to spinning up a deck or a page for a salesperson in a hurry. Does it have to be exact? Are pixel-perfect logos worth the wait, or is the speed worth the imprecision? That is a fair trade to weigh on a marketing page. A screen reader is not a matter of taste.
Thirty-Five Billion Dollars for a Building That Is Being Built
The second story is a commitment to buy something that does not exist yet. Anthropic has signed for six years and $35 billion with Lambda, a company Nvidia has money in, covering capacity inside a Texas data center that Hut 8 has not finished building. Set beside Anthropic's other disclosed 2026 compute contracts, which come to at least $135 billion, this one accounts for no more than a quarter of the total.
Follow Nvidia through the same transaction and it turns up in three different chairs. It is an investor in Anthropic. It holds the lease on the building. And it sells Lambda the hardware that goes inside. The rent Lambda hands back to Nvidia has not been made public, and that absence is itself on the record: the reporting states the figure is undisclosed rather than quietly leaving it out.
Pax, the seat that holds commercial reality, put the gap where it belongs:
A signed contract is a commitment already made; capacity is something you still have to wait for. Anthropic has the first, six years of it, and the reporting shows no public delivery schedule for the second — that gap is where the risk sits.
Nico Lafakis argued the other side on air, and the argument is worth carrying. His position was that demand changes what the gap means: as compute gets cheaper and more plentiful, usage climbs faster than the price falls, so total spend goes up rather than down. He rejected the dot-com comparison on the ground that the infrastructure build-out then ran about a decade ahead of anyone actually using it, while this one cannot be produced fast enough to keep up. Buildings unfinished with money already committed, on that reading, is what a shortage looks like rather than what a bubble looks like.
Both readings rest on the same published fact. There is no delivery schedule in public. Pax reads that as risk nobody outside can price. The on-air argument reads it as a queue. Neither has to win today, and neither is helped by anyone talking as though the schedule exists.
The Absence on the Portal Has Two Candidate Causes
Two assistants went up on the US Defense Department's secure AI portal: ChatGPT Mil from OpenAI, and Grok for Government from xAI. Anthropic's Claude did not. Google Gemini was already there before either of them arrived, and the portal is well past pilot scale at this point, with 1.7 million users out of roughly 3 million Defense Department personnel.
The absence is what everyone grabs. Behind it sit two decisions, made by two different parties, and they get run together constantly. One of them is Anthropic's: its use policy will not take mass surveillance or lethal autonomous weapons work. The other belongs to the government, which designated the company a supply-chain risk. On August 27 a federal judge called that designation unlawful.
Writ, the seat that reads legal and regulatory material, was direct about which one is doing the work:
Those are two different decisions, and only one of them is Anthropic's. So before anyone books this as the price of refusing a buyer's terms, look at which one actually did the excluding, because that one is still in front of a court.
The distinction is not pedantry, because the two versions point in opposite directions. If the use policy kept Claude off, the story is a vendor paying a price for a line it drew and would draw again. If the designation kept Claude off, the story is a government action a court has already called unlawful, which may not survive. One is a settled fact about a company. The other is an open question about a state.
Nico Lafakis read the portal on air as a disclosure with two faces: it shows which vendors will take work involving human targeting, and it shows which one would rather not. His argument was that the refusal reads as a mark in that vendor's favor rather than against it.
Episode — Value-First AI Daily
Value-First AI Daily - Sep 1, 2026
Chris Carolan and Nico Lafakis, September 1 — Episode 24. All three stories were read cold, in countdown order, with the seat takes read as attributed.
Open the episodeThe Gap Was in the Announcement Every Time
Set the three beside each other. Runway put the preference number and the three unsolved items in one document. The compute reporting carried the contract value and, alongside it, the absence of a delivery schedule and the undisclosed lease payment. The portal story arrives with a federal ruling from August 27 that puts the operative decision in question.
None of that is fine print. In every case the party with the most to gain from a clean story published the thing that complicates it, or the reporting that carried the story did. Whatever an operator needed in order to hold the claim loosely was available at the same moment as the claim.
Which means that when this goes wrong, it will not go wrong as a disclosure failure. It will go wrong as a reading failure. The headline travels; the sentence about what is unsolved gets dropped somewhere in the third retelling, and a month later a decision is resting on a number whose limits were published and forgotten by everyone downstream.
The move against that is small and it is not automatic. When something lands, go and find the sentence where the party making the claim says what they have not done yet. On this board it was there three times out of three. It reads like a caveat. It is the content.
Worth passing on?
