Two things got bigger this week, and one of them was not the technology. The money hit new extremes: a hundred billion dollars behind a single data center, a chip startup doubling in value in a month, and a leading lab showing its first profit. At the same time the boundaries got sharper, as governments began sorting the world into AI camps and one of the biggest products started sorting its users by age. Here is what mattered.
Nvidia put up to $105 billion behind one OpenAI data center
On August 17, Nvidia said it will invest $1.5 billion in SB Energy, a SoftBank subsidiary developing a data center campus for OpenAI in Pike County, Ohio, and will backstop the first phase with up to $105 billion in financing. The site, called Ports-Pike, is planned to start at 4.25 gigawatts and scale toward 8, with Nvidia as the sole supplier of compute. It sits on land the federal government once used for uranium enrichment. [TechCrunch, Axios]
Our take: the headline is the number, but the signal for everyone else is what it says about cost. The frontier now runs on power plants and multi-year construction, which is exactly why almost no business should try to own that layer. The move is to rent it. You reach the same frontier models through an API for cents, and you spend your effort on the part Nvidia cannot build for you: your data, your workflow, your customers. We build on rented compute by default and keep the switching cost low, so when the price of intelligence drops again, and it will, you are not tied to one provider.
Anthropic reported its first profitable quarter, with an asterisk
Anthropic reported its first profitable quarter, with more than $11.5 billion in revenue for the period and positive adjusted operating income, according to figures first reported by Bloomberg. The numbers are preliminary and unaudited, and the company has not said which costs it adjusted out. Fuller, audited results are expected in an IPO prospectus later this year. [Forbes, CNBC]
Our take: read the word "adjusted" carefully. Positive adjusted operating income is genuine progress, and it can still leave out large costs like the money spent training the next model, so it is not the same as net profit. That habit is useful far beyond AI labs. When any vendor or partner shows you a flattering number, ask what the adjusted figure removed before you rely on it. When we model the cost of an AI system for you, we count the whole bill: usage, retries, the human review around it, and the upgrade you will want next year. A number that only looks good once the hard parts are taken out is not one you can plan against.
A chip startup doubled to $21 billion in a month
Etched, a startup building chips designed to run AI models rather than train them, raised $700 million at a $21 billion valuation, roughly double its value from a month earlier. Jane Street led the round and is also its first customer, having taken delivery of hardware for its own trading systems. Etched is one of several companies trying to undercut Nvidia on the cost of running models in production. [TechCrunch, Etched]
Our take: the bet here is on inference, the cost of running a model every time someone uses it, not the one-time cost of training it. That is the bill a business actually pays, and it is where the next round of savings is coming from. You do not need to follow the chip race to benefit from it. The practical move is to design systems that can shift to cheaper hardware and cheaper models as they arrive, rather than wiring today's provider into everything. We keep that flexibility in from the start, so a drop in the cost of running models turns into a lower bill for you instead of a rebuild.
The US moved to make 35 countries pick an AI side
Reuters reported on August 14 that the US State Department drafted a letter to the 35 countries that signed its "AI Opportunity Statement" in June, telling them that membership in a new US initiative called Pax Silica cannot sit alongside "duplicative initiatives," an oblique reference to China's World Artificial Intelligence Cooperation Organization, which Xi Jinping announced in July. The draft warns that "to be part of everything is to be part of nothing." [CNBC, The Next Web]
Our take: the AI stack is becoming political, from chips to cloud regions to the models themselves. Most businesses will never receive such a letter, but the same fault lines can show up quietly in your vendor contracts and in where your data physically sits. The sensible step is to know where your AI actually runs and where your data is stored, and to keep enough portability that a policy change on either side does not strand you. We design with that in mind, favoring setups you can move, so a shift in the rules is an inconvenience rather than a crisis.
OpenAI started guessing how old its users are
On August 18, OpenAI began rolling out ChatGPT for Teens, a separate experience for users it estimates to be 13 to 17, using an age-prediction system to route them there automatically. The teen version adds stronger safety limits, holds back romantic or emotionally dependent conversation, and gives parents more controls. Adults wrongly flagged as minors can verify their age to restore full access. [OpenAI, Axios]
Our take: the quiet story is age prediction. OpenAI is now estimating how old its users are and changing the product based on the answer, a preview of how AI will adjust itself to whoever is on the other end. The lesson for any business is that the same assistant can and should behave differently for different audiences. A tool that talks to your staff does not need the same limits as one that talks to the public, and one aimed at the public needs guardrails you can defend. When we build an assistant for you, we set those boundaries deliberately: what it can say, who it can say it to, and when a person steps in.
The one thing to remember
This was a week of very large numbers and very clear lines. One data center drew a $105 billion commitment, a lab showed its first profit once the awkward costs were set aside, and a chipmaker doubled in value in a month. In the same days, governments began sorting the world into AI camps and one of the biggest products started sorting its users by age. The pattern underneath is the one we keep returning to: the capability is racing ahead, and the value now sits in the boundaries you put around it. Knowing what your AI runs on, what it truly costs, and who it is allowed to talk to is quickly becoming the real work.