OpenAI Seeks $20 Billion at a $1.25 Trillion Valuation Instead of $30 Billion at $1.4 Trillion

According to the Financial Times and the Wall Street Journal, OpenAI has revised its plans for a new funding round. Instead of $30 billion at a $1.4 trillion valuation, the company is now seeking about $20 billion at $1.25 trillion. So far it has raised roughly $11 billion, and some investors have reduced their commitments amid a selloff in AI assets. This is not a burst bubble, but it is a telling signal: even for the market's most prominent player, money is no longer unlimited or free.

Expensive capital hits both the lab and the clouds

The problem is not limited to one company. According to the WSJ, the neoclouds CoreWeave, Nebius and Nscale have in recent weeks effectively been shut out of the cheap bond market, as lenders demand tougher terms. It is the same story seen from the other side of the balance sheet. A lab raises equity to train and run models, while a cloud operator borrows to buy GPUs and build sites. If both kinds of money get more expensive, the whole chain slows down.

For the market, this marks a change of regime. For the past couple of years, chips and electricity were considered the main constraint. Now the WSJ names a shortage of cheap money as the new bottleneck. AI infrastructure can be built quickly, but the more expensive the loan, the tougher the requirements for capacity utilization and payback periods. Operators with a high share of debt will have a harder time than those funded from operating cash flow.

Altman asks people not to keep DRAM at home

Another thread of the day concerns hardware. In a conversation with Tom's Hardware about running models locally, Sam Altman said that scarce DRAM is used more efficiently in data centers than in home PCs, where, by his account, it sits idle 95% of the time. From a utilization standpoint, the argument is clear: memory in the cloud is loaded more densely.

But the thesis also has a commercial subtext. Memory is scarce, cloud providers compete for it with the consumer market, and a call to keep it in data centers benefits those who build those data centers. For developers who plan to run models locally, it is a reminder that memory remains the main bottleneck, and it won't get cheaper.

Water, politics and data centers in Arizona

The price of centralization is visible beyond financial reports. The Verge describes the pushback against data centers in drought-prone Arizona: the state already has more than 170 data centers, and new projects are running into protests over water. The most contested case is Project Blue near Tucson. It has been permitted to pump up to 31 million gallons of groundwater a year after the plan to use reclaimed water was dropped. Activists are marching under the slogan "NOT ONE DROP FOR DATA."

The statistics here are mixed: most of Arizona's water goes to agriculture. But data centers have become a convenient political target. They are conspicuous and noisy, they consume a lot of power, and they are associated with wealthy outside companies. For developers, this means that securing water and electricity is becoming a public process rather than a technical formality.

The supply chain for IDE extensions

One more story concerns security. Anthropic said that an attacker, using access to the Cursor marketplace, published a modified copy of the Claude Code extension that the company did not create. It is a small story next to capex in the hundreds of billions, but it shows a risk: the more often developers install agentic tools directly in their IDE, the more valuable the extension supply chain becomes. Swapping in a fake "programmer's assistant" can potentially give access to repositories, tokens and workflows. The practical takeaway is simple: verify the extension's publisher, limit token permissions in the development environment, and do not install agentic plugins from unverified sources.

Who stands to win

The essay The Commanding Heights argues that as models become commodities, lasting wealth will go not to the labs but to those who control distribution, payments, infrastructure and dollar stablecoins. For now this is a scenario, not a fact, but it fits the picture of the day: if models get cheaper while infrastructure and capital get more expensive, the winner is whoever controls the channel, the meter and the cash register. Today's AI market increasingly looks like an ordinary capital-heavy industry, where what matters is the cost of borrowed money, access to resources and permits from local authorities.