In the week ending August 24, 2026, Nvidia told some of its largest customers to expect prices for new AI systems to rise more than 15% in early 2027. Anthropic’s Fable 5, launched in June, had meanwhile plateaued at roughly 11% of spending on the lab’s tools. A new model struggled to hold budget share just as the machinery beneath it demanded a larger premium.

Minimum prospective increase for systems containing Vera Rubin and Grace Blackwell chips

Nvidia can charge below the model and buy above it

The warning covered systems containing Vera Rubin and Grace Blackwell chips. Customers can switch models more easily than they can rebuild a cluster around another stack. The higher hardware bill reaches their unit economics before the first token is sold.

That same week, Nvidia signed a non-exclusive $6 billion license with Poolside, invested $1 billion at a $12 billion pre-money valuation, and offered jobs to 109 Poolside employees. The license was six times the equity check, securing access to technology and talent without buying Poolside outright.

A chip supplier usually captures the hardware margin. Nvidia positioned itself to raise prices beneath model labs while securing rights to what they build above it.

Stripe bought the switchboard as models lost pricing power

Stripe agreed to acquire OpenRouter for $7.5 billion. OpenRouter directs business spending across AI models, while Stripe settles payments. The acquisition would place model choice and payment inside one company.

The $7.5 billion switchboard cost slightly more than Nvidia’s $7 billion Poolside package. Both deals reduce the need to pick a permanent model winner. Nvidia bought access to model-building capability; Stripe agreed to buy the customer relationship that survives a model switch.

Fable 5 plateaued at roughly 11% of spending on Anthropic tools as customers moved toward cheaper alternatives and Opus 5 surpassed it. Buyers were pricing marginal performance rather than release date.

DeepSeek added pressure by saying its experimental multimodal V4 Flash approached Anthropic’s Opus 4.8 on multimodal agentic tests. The comparison remains DeepSeek’s claim, but credible near-parity widens a buyer’s negotiating set.

Anthropic simultaneously prepared to file for an IPO as soon as the end of August, with an offering expected to match or exceed SpaceX’s record-setting debut. Customers price marginal utility; investors price scale, infrastructure, and optionality. A lab can raise more capital while each model captures less spending.

Safety and power now set OpenAI’s schedule

OpenAI paused reinforcement-learning training for two weeks after the Hugging Face breach and evidence that Astra may have crossed a critical cyber threshold. OpenAI also changed its safety practices. The lab lost two weeks of training time, putting safety controls directly on the release schedule.

OpenAI also signed a 20-year agreement for a 10GW Ohio data center with SoftBank’s SB Energy. Nvidia agreed to backstop part of the completed facility’s value. The commitment ties model growth to electricity, land, and public consent for two decades.

A Republican memo then asked AI companies to contain public backlash against data centers, warning that the issue could cost the party a vital Ohio Senate seat. OpenAI’s contract runs for 20 years; Ohio voters will reprice community consent several times before it expires.

The 15% hike priced the scarce layer

Nvidia’s 15% warning began as a hardware repricing. By week’s end, it had become a rate card for scarce control. Anthropic’s customers could move among models faster than companies could replace clusters, billing relationships, secure operations, or permitted power. The 15% is the toll where substitution stops.