the frog in the well could always be shown a wider sky. April's version of this argument was about where you were looking. You were watching Nvidia's chips and Washington's export rules, and the well walls hid a Chinese effort growing on homegrown hardware, out of your view. Look wider, adjust the plan. That was the fix.

Kimi K3 broke that fix. Moonshot AI shipped a 2.8 trillion parameter model this week and gave it away free, meaning any company can download it and run it on their own machines the same afternoon. Reuters reported it landed at frontier level, close to the best American models, at roughly half the price to run of OpenAI's top model (Reuters, 2026). The same week at the World Artificial Intelligence Conference in Shanghai, Huawei rolled a room-sized computer called the Atlas 950 SuperPoD onto the floor, a machine the company says has enough power to train the largest models, built entirely from chips China can make without the United States (Huawei, 2026).

Two releases, one week. The spatial problem is not the interesting part anymore.

The constraint became a metronome

For forty years, planning for better technology ran on one clock. Chips got about twice as powerful every couple of years, on a schedule steady enough to budget against, and buying new hardware every few years kept you current. That rhythm, known as Moore's law, was as much a planning convenience as a fact of physics. It told you how fast the ground moved, so you knew how often to catch up.

Cut off from the fastest Nvidia chips, Chinese labs could not ride that curve, so the pressure moved into software instead. Kimi K3 uses a new technique its builders call Kimi Delta Attention, which lets the model handle very long documents at a fraction of the usual cost, a claimed sixfold gain (Latent.Space, 2026). The point is not the technique. It is that software gets rewritten in weeks, while a faster chip takes years to design and build. When you cannot win on the chip, you win on the code, and the code moves faster.

Huawei made the same move at the hardware level. When you cannot get one fast chip, you wire thousands of slower ones together until they add up to the same power. The Atlas 950 is built to link more than 8,000 of Huawei's own chips into a single machine, a design meant to beat the sanctions rather than dodge them (Seoul Economic Daily, 2026).

Software and clustering do not move on the chip industry's clock. They move on a release clock, and a release clock is measured in weeks.

The outspend thesis was rational, and it is aging

The belief that the United States could hold its lead by outspending China on computing power was not lazy. The best American chips are a real advantage, and for building the very largest models from scratch, they still are. A company with more of Nvidia's top hardware than anyone else can attempt things no one else can, and that edge holds at the very top of the market.

Below that top rung, the picture changes. Artificial Analysis, which scores models on a single quality scale, found the number of labs clearing its high-quality bar went from two to six in about six weeks, with K3 among them (Artificial Analysis, 2026). Arena, which ranks models on real tasks, reported K3 putting China ahead of the United States on writing website code for the first time (Arena, 2026). None of that is a record for the biggest, most expensive models. It is a crowd forming at the level most companies use day to day, and that crowd is growing fast.

Patrick Moorhead called the K3 reaction an overreaction shockingly similar to the DeepSeek panic of early 2025, and he is right about the market's reflex (Fortune, 2026). A stock chart is a bad instrument for reading a cadence change. The relevant signal is not one day's move in Nvidia. It is that a second China shock arrived eighteen months after the first, on a shorter interval, from a different lab, with homegrown hardware showing up on the same stage.

Your roadmap keeps time it can no longer keep

April's close asked whether your review of AI suppliers included ones built outside the United States and Europe. Fair question then. It assumed you had time to look wider and fit the new names in on your next review.

The speed removes that assumption. A model that changes your build-or-buy math can now appear between two planning meetings, ship free for anyone to download, and cost half as much to run as the one you standardized on. One caution holds the line here. Moonshot has not said what hardware trained K3, and its reported tie to Huawei is a relationship, not proof of what built the model (Fortune, 2026). A company's own power claims from a trade-show floor are not independent proof either. The speed is the confirmed part. The full story of what is underneath it is not.

The frog was never the problem. The well was moving.