a day after I argued that the interesting problem had stopped being spatial and become temporal, The Washington Post ran two charts that put a coordinate on it. One shows cumulative downloads of open AI models. The American line and the Chinese line climb together through 2024 and into 2025, then cross in August 2025, and the gap has been widening since. The other shows the share of US token traffic on OpenRouter, a marketplace for running open models, moving from 16 percent Chinese at the start of this year to roughly half by the last week of June.
The crossover is not a forecast. It is a printed line with a date on it.
The outspend thesis has a measured crossover point now
In the Moonshot piece I called the belief that America could hold its lead by outspending China on compute rational but aging. That was a claim about direction. The Post supplied the odometer. When the two download curves cross and then separate, you are no longer arguing about whether the pattern is real. You are reading how fast it moves after the crossing, and it is accelerating, not settling.
Notice which measurement crossed. Cumulative downloads and marketplace token share are usage, not capability. Nobody in the Post's reporting claims a Chinese lab built a bigger, better model than the top American one. The frontier-from-scratch lead holds, exactly where I said it would. What crossed is the part underneath the frontier, the everyday inference that most companies actually run, and that is the part that pays vendor invoices.
The named buyers are no longer startups you can wave off
I have been logging the on-the-record switchers for months: Airbnb on Qwen, Pinterest on open weights, Shopify testing Qwen for its vendor agents. The Post adds the name that closes the "just scrappy startups" defense. AT&T, whose chief technology officer says the company auto-switches to cheaper models for tasks where the cutting edge is not needed. He draws one line, no Chinese models for national-security work, and routes everything else on price. That is not an ideology. That is a procurement officer doing the math out loud, at carrier scale.
The founder the Post follows makes the cadence point better than my prose did. He spent almost $12,000 running Claude himself in May, moved to Z.ai's GLM-5.2 after the export controls on Anthropic's Fable model, and cut his personal AI spend by more than half. His reason for not switching back was plain: he sees no argument for paying top dollar for a model he considers inferior for his work. Read that as a sentence about switching cost. It has collapsed to nearly zero. The thing that used to lock you in, the pain of moving, is the thing that just disappeared.
The chart cuts both ways, and the footer knows it
Two facts in the same report keep this from being a victory lap for open weights. Alibaba has started shipping closed, proprietary versions of Qwen, and Reuters reports Beijing is weighing limits on overseas access to homegrown models. I wrote three weeks ago that two governments can freeze your model and one router will not save you. The Post's own reporting now carries the Beijing half of that in a single paragraph. The download curve that crossed in August can be throttled from the top by the same government that let it climb.
So the crossover is confirmed and the roadmap risk is confirmed in the same article. Both things are true. The company that read the first chart and standardized on a Chinese open model to cut inference cost now owns the second risk, a frozen asset if the next version never ships to them.
The download curves crossed in August 2025 and your organization did not decide to be on either side of that line. Your inference bill did. Pull one month of AI spend and sort it by task, not by vendor. For the workloads that do not touch the frontier and do not touch regulated data, what percentage are you still paying premium closed-model rates on out of habit, and how much of that is a decision you made versus a default your router picked for you?
Sources
Waldvogel, Miriam. "America's Hottest AI Models Face Powerful Foreign Rivals." The Washington Post, 19 July 2026, washingtonpost.com.
Ovide, Shira, and Kevin Schaul. "American AI Models Are Losing Ground to Lower-Cost Foreign Alternatives." The Washington Post, 19 July 2026, washingtonpost.com.
ATOM Project. "Analysis of HuggingFace Download Data," as reported by The Washington Post, 19 July 2026.
OpenRouter. US Traffic Share Data, as reported by The Washington Post, 19 July 2026, openrouter.ai.
Prior coverage: "Moonshot Shipped a Frontier Model on a Clock Your Procurement Can't Match." shashi.co, 18 July 2026, shashi.co.
Prior coverage: "Two Governments Can Freeze Your AI Model. One Router Won't Save You." shashi.co, July 2026, shashi.co.
Prior coverage: "The AI Price War Just Made Model Choice a Board-Level Decision." shashi.co, June 2026, shashi.co.
