The neutral gateways and cheap models you chose to keep your options open spent this week being bought, repriced, and shut down.
Developers pick a model gateway for one reason. It stops any single company from owning the path between their app and the models they call. So the news that Stripe agreed to buy OpenRouter for over $7 billion, in "AI model gateways: Stripe buys OpenRouter for over $7 billion," is not really a funding headline. It is a change of landlord. OpenRouter routes one API call to more than 400 models and reached 8 million users by staying vendor neutral. That neutrality now answers to a payments company. The toll booth you used to route around someone else's lock-in is now owned by someone.
Hold that thought, because it repeated all week in different clothes.
If you built on DeepSeek because it was the cheap frontier option, this week it installed a rush-hour meter. "DeepSeek ends the cheap ride with peak pricing up to 1,100% higher" split its V4 line into peak and off-peak rates, with output on the new V4 Pro tier climbing to $3.96 per million tokens at peak from a flat $0.87. Capability went up. The flat, predictable discount that drove adoption went away, and the second change is the one that lands on next month's invoice.
If you wired Google's Imagen 4 into production, it went dark on August 17, months after the date was flagged, and the replacement costs more and needs a code rewrite. "Google Imagen 4 API shuts down, forcing a costlier migration" is the fine print of every managed API made literal. The model you rent gets retired on the vendor's calendar, not yours.
Notice what these three have in common. In each case a team made a sensible choice to stay flexible, and the flexibility was removed from the outside. The gateway got bought. The cheap model got a peak rate. The rented image model got a deprecation date. No code changed on the buyer's side, and the buyer's position got worse anyway.
The middle is where the money moved. For two years the contest was over who had the best model. This week the contest was over who owns the layer between you and the model. "Nvidia coding deal: $6B to license Poolside's models and hire its staff" pulls a two-year-old coding startup's models and more than 100 of its engineers into the company that already sells the chips. "Harvey Tenet: legal AI startup ships its own model, built on Kimi K3" has the legal startup that used to resell OpenAI and Anthropic access now owning the layer that touches the client's documents. "Google gets the option to buy $12.2B of Marvell in a custom-chip deal" ties a chip supplier to Google's own stock. Different corners of the stack, one direction. The pieces are collecting into a shrinking set of owners.
Even the raw materials moved the same way. "RAM crunch: DDR5 prices climb 500% as AI eats the memory supply" traces consumer memory up roughly 500 percent in a year as Samsung, SK Hynix, and Micron steer production toward the high-bandwidth memory accelerators need. A cost that fell almost every year for two decades now rises on every capex announcement. If your plan assumed hardware gets cheaper while you wait, that assumption expired quietly.
Here is the claim worth slowing down for. In "Anthropic revenue: first operating profit on $11.5 billion quarter," the company told investors it turned its first operating profit on preliminary second-quarter revenue above $11.5 billion. The number is real and large. The word doing quiet work is adjusted, which sets aside stock compensation and the training runs that built the models earning today's money. A self-funding lab is genuinely better news for anyone building on its API, because it lowers the odds a model gets repriced or pulled for the sake of survival. Read the adjusted profit with the same suspicion you would read any vendor's best-case math, then notice it points where everything else this week points. The companies that own the layer you depend on are the ones with the leverage to set its price.
Who this puts on the clock. If you buy, deploy, or operate AI, the thing you lost this week was not a specific tool. It was the assumption that a good architecture, chosen once, keeps you safe from the vendor. Vendor neutrality is not a property of your design anymore. It is a property of a market that is consolidating, and it degrades whether or not you touch your code.
That changes what dependency means. It stops being a diagram you draw once and becomes something you test on a schedule, the way you test backups. The useful question is no longer "am I multi-model." It is "which single vendor, if it doubled its price or closed an endpoint next month, could I not route around by Friday."
For buyers and operators, run a dependency drill next week. Take your highest-volume model call, stand up one live fallback you could switch to in an afternoon, and measure cost per successful task at the vendor's peak rate, not its sticker rate. DeepSeek's new peak pricing and the Imagen 4 shutdown are the templates for how the squeeze arrives, so treat the drill as fire practice, not paperwork, and put the finding in front of whoever signs the invoice.
For sellers, consultants, and software teams, there is a concrete offer in this. Package a 30-day dependency scorecard. Map every model call in a client's stack, flag the single-vendor chokepoints, build the fallback routing and the permission boundaries around it, and hand back a short list of the endpoints that would hurt most if they moved. This week supplied the demand. Every buyer who read the OpenRouter headline is now wondering which of their own neutral layers has a landlord.
The old operating principle was that you manage lock-in at design time by choosing open, cheap, or neutral components. This week's evidence retires it. Those components are being bought and metered by the same few companies, so lock-in is now something you manage continuously, by keeping a tested exit for every dependency that matters. Neutrality you cannot exercise on a Friday afternoon is just a slower kind of lock-in.
The week in one line: Vendor neutrality is no longer a design decision you make once; it is a fallback you have to test on a schedule, or it quietly stops being true.
Sources this week: Stripe buys OpenRouter, DeepSeek raises API pricing, Imagen 4 shutdown, Nvidia licenses Poolside, Harvey Tenet, Google's Marvell warrant, DDR5 prices climb 500%, Anthropic's first operating profit