Prices out of control, capex under strain, open models as a shield

With the Muse Glimmer launch, under an Apache 2.0 license, what changed was Meta's behavior, not the market's reading of it. Meta had signaled openness since 2024 while shipping proprietary licenses full of conditions; Apache 2.0 closes that gap. What still holds is our distinction: open weights are not open source. The license frees up use of the model, not the training data or the code. Adopters gain sovereignty over infrastructure and less vendor dependence, not transparency.

Gemini 3.7 Flash's price cut comes with two conditions that weren't in the headline of the announcement. Simon Willison was one of the few to note that the price is introductory and doubles on January 1, 2027, and found the logic strange: who plans today to use this model five months from now, if the previous one lasted three weeks? The second condition is the removal of the cheapest reasoning tier, precisely the one that made high-volume, simple, repetitive tasks viable. The published price sheet is a floor with an expiration date, and real cost went up for part of the use cases.

And Apple closes out the argument for this dilemma. After two years treating the model as an input it could buy from third parties, the decision to develop its own model for China suggests a strategic conclusion: where regulation is the bottleneck, whoever doesn't control the model also loses control over the channel.

The 53% who can't bring business context into their systems and the 77% who say that context is what determines quality, read together, describe a company that knows where the problem is and can't attack it. The same Alteryx survey suggests why: 37% say AI strategy sits within IT and 38% say delivery does too, while business units are left to just define requirements, at 30%. The knowledge the AI needs sits in one part of the company, and the system is being built in another. It's not a technology problem, it's an organizational design problem. One caveat applies: the survey is sponsored by a vendor whose product claims to solve exactly the problem it measures, so it's worth reading as direction, not as a precise measurement.

Ethan Mollick, the voice we track on this dilemma, didn't publish this week, but the contrast with what he's been repeating helps place the survey. His advice is about the individual: pick a model, pay for the subscription, and hand an agent a real task from your real job. The survey describes what happens when the whole organization tries the same move: what was personal learning becomes a governance problem, because now what each person carried in their head has to be written down, versioned, and audited.

On the seller side, the week's other two facts, OpenAI's cybersecurity models on Bedrock and Gemini positioned for agentic workflows, point in the same direction. We've moved from "a model you consult" to "a system that acts inside your environment," and that takes the discussion out of the IT department and into the boardroom, because what's now at stake is who answers for what the machine did.

What's new in the Uber announcement isn't the technology, it's the deal structure. Pony.ai supplies the self-driving, Uber supplies demand and billing, and a local operator owns the fleet and handles maintenance, cleaning, and charging. Three companies, three roles, three blocks of capital. That matters because a fleet consumes cash and doesn't scale the way software does, and separating whoever puts up the capital from whoever puts up the technology is what unlocks growth.

Anthropic's talks to acquire Decart, the week's second fact here, point to the input that bottlenecks the category. Physical autonomy runs less into an algorithm problem and more into real-world data, expensive and slow to collect, and a simulated environment is a way around that. A frontier lab buying a world-generator says something about where the constraint is being attacked.

The comparison with humanoids is this week's takeaway, and it explains why the two subjects belong to the same dilemma. The robotaxi moved forward partly because it split capital, technology, and operations across three parties. Much of the humanoid pipeline still concentrates technology, hardware, and market development in the same manufacturer, meaning designing the robot, building it, training the software, and finding the customer, all at once. Less a robotics problem than a capital-structure one, and that's what sets the pace.

This week's real finding is methodological. Three measures of the same phenomenon circulated and they don't match: $1.5 trillion in purchase commitments across six companies, per the FT; another $1.5 trillion in leases, per Goldman Sachs; and $1.65 trillion in off-balance-sheet obligations across five companies, excluding Nvidia, per Nikkei Asia's review of SEC filings. None of them is wrong, because none of them is measuring the same thing: they differ in the number of companies covered, the type of instrument, and the cutoff date. A purchase commitment is a signed contract to pay later; an unstarted lease is rent for a building that doesn't exist yet. Neither is debt in the sense most people read off a balance sheet, which is why they end up in the footnotes. What the three numbers say together, despite the divergence, is one thing: future commitments are growing much faster than present cash.

Alphabet shows that gap within a single quarter. Purchase commitments jumped from $332.4 billion to $811 billion, and in the same period the company posted its first-ever negative quarterly free cash flow since going public, while Meta's fell 91%. That confirms a reading we'd already flagged: strong trailing-twelve-month cash doesn't measure resilience when commitments already signed for coming quarters exceed the cash the operation generates. SpaceX repeats the pattern at smaller scale: AI revenue already overtook space-operations revenue in the second quarter, but the first half closed with $3.47 billion in operating cash against $28.48 billion in investment. And the cost of covering that gap rose in the same week: the US Treasury sold 30-year bonds at 5.216%, the highest rate in roughly 25 years, while Broadcom fell 6% and Wall Street swapped the debate over how big the investment is for the debate over how it's being financed. More expensive money makes any long-term build-out costlier, and data centers are today among the most capital-intensive infrastructure in the economy.

On the other side, the market is still willing to pay for the future, and Anthropic's planned IPO is the measure of that. Anthropic is heading toward what would be the largest IPO in history, with three caveats the headline doesn't carry: the $2 trillion projection comes from investors, not the company; the meetings were described as preliminary and without a discussion of numbers; and the FT itself notes growth slowed in June. Ben Thompson added an argument about financing on the 11th: Nvidia keeps finding new ways to help its own customers raise the money they use to buy Nvidia's chips.

SMIC's case says less about SMIC and more about the shape of the bottleneck. Anyone who runs a factory recognizes the dynamic: a full plant is the point where price stops being negotiated by the customer and starts being set by the supplier. The detail is that SMIC doesn't even make advanced AI chips. It's full because demand overflowed into common components, the power-management chips and sensors that nearly every piece of electronics uses. The bottleneck moved from the top of the supply chain down into the middle of it.

DeepSeek's price hike is the cleanest case this week. The company became known in 2025 for pricing well below its American competitors. The pressure behind the increase doesn't look purely commercial: the cheap model attracted volume far beyond what the company's compute fleet can handle, and the price is starting to reflect capacity scarcity. The detail that reinforces this reading is that the price varies by time of day, which is how you manage a queue, not how you manage margin. In the same week Google cut its price in half, DeepSeek raised its own. The difference isn't willingness to compete, it's who has the compute fleet installed to sustain the fight.

One note on the service that's up to 14 times faster, which OpenAI unveiled on the 13th. Ben Thompson argued in May that agentic inference would change the economics of the industry and that speed would matter less, since without a human waiting, the response can take longer. This week's offering bets on the opposite, because there are applications where waiting makes the use case unviable. Both readings are probably right, and the market may be splitting into two inference regimes, the patient and the impatient, with different economics. A theme worth tracking, not a closed conclusion.

The Astra pause is one of the clearest cases yet of a frontier lab halting internal activity because a model may have hit the top of its own risk scale. On the company's scale, "Critical" means being able to find and exploit flaws in well-protected systems on its own, without human help. The assessment is preliminary, according to OpenAI itself.

In Sanders's letter, the method matters more than the ask: he didn't propose a new principle, he demanded the companies live up to commitments they themselves had already published. It's the first time a lab's voluntary commitment has become a tool for political pressure, and that changes the calculus for whoever publishes commitments next.

There's a layer the coverage missed that changes the nature of the episode. On June 4, Anthropic itself published "When AI Builds Itself," signed by Marina Favaro and Jack Clark, arguing the world should have the option to temporarily pause frontier model development. The condition is decisive: the company argued a pause would only make sense if coordinated across multiple countries and companies, with verifiable rules, and warned that a unilateral slowdown would leave everyone less safe, because it would hand an advantage to whoever cares least about safety. Sanders is asking each company, separately, for exactly what Anthropic said, two months earlier, doesn't work separately. This isn't Congress versus industry: it's two diagnoses that agree on the risk and disagree on the instrument.

Z.ai's claim closes out the week on the racing side. The company says it has come close to Anthropic's most advanced model on cybersecurity, beating it on one vulnerability-identification benchmark while trailing well behind on attack construction. These are figures from the vendor itself about a competitor's model, without independent verification, and should be read that way. What matters isn't the scoreboard: it's that frontier cyber-capability became a sales pitch in the same week it became a reason to hit the brakes.

And a note on proximity: in the same week it paused a model over cyber risk, OpenAI moved frontier cyber-capability closer to market. These aren't contradictory positions, one deals with capability that isn't yet well understood how to measure, the other with capability already assessed and fenced in, but the gap between the two shows just how narrow the line is that the industry is walking.