The claim circulating this week is that the world has decided China is the stabilizing power and the United States is not. That is not quite what the data says. The gap between what it says and what it is being asked to say is the more interesting object.
What the Survey Found
Pew released two linked reports on July 15. Across thirty-six countries and 42,151 respondents interviewed between February 8 and May 13, China is now viewed more favorably than the United States in most places surveyed; confidence in Xi Jinping now exceeds confidence in Donald Trump on the median. Pew's Laura Silver has said this is the first such inversion in roughly two decades of the Center's global polling.
The Canadian series is the cleanest illustration. In 2023, 57 percent of Canadians viewed the United States favorably and 14 percent viewed China favorably. In 2026 the figures are 33 percent and 44 percent. That is a ninety-point swing in relative position in three years, in the country with the longest undefended border and the deepest supply-chain entanglement with the United States on earth.
But the stability finding, the one carrying the headline, is narrower than the headline. It comes from a supplementary battery asked only in seventeen middle-income countries across Latin America, sub-Saharan Africa, the Middle East, and the Asia-Pacific. In those countries a median of 75 percent say the United States interferes in the affairs of other nations a great deal or a fair amount, against 45 percent for China; and in many of them, more people credit China with contributing to peace and stability. In South Africa, the share saying China contributes to global peace and stability rose from 47 percent in 2023 to 64 percent in 2026.
Three qualifications matter, and they are not decoration.
First, the finding is comparative, not absolute. It says China is rated higher; it does not say either is rated well. In Indonesia, the shares crediting both powers with contributing to global peace and stability declined. That is not a transfer of confidence. That is a net withdrawal of it, with a differential rate of decay.
Second, the United States still leads on the one question that has historically anchored the difference: whether a government respects the personal freedoms of its own people. The lead is shrinking, and shrinking from the American side rather than the Chinese. In Sweden, the share saying the United States respects its people's freedoms fell from 61 percent in 2021 to 27 percent now; comparable drops of twenty-five points or more appear in Canada, France, Germany, Italy, the Netherlands, South Korea, and Spain.
Third, and most usefully: the six countries that still rate the United States above China are India, Japan, the Philippines, South Korea, Israel, and Poland. In the Philippines, 81 percent call the United States a reliable partner against 42 percent for China. In Pakistan the numbers run the other way (84 percent for China, 36 percent for the United States).
Look at that list of six. Every one of them shares a border, a near-sea, or a live threat vector with the power they are declining to prefer. Sentiment toward a security guarantor varies inversely with exposure to the thing the guarantee is for. Publics that might actually need the guarantee are still buying it; publics that would only ever consume it as a public good are not.
That is a finding about what the survey measures. It measures the price of American reliability among people who are not currently purchasing it.
What a Survey Cannot See
Favorability is a sentiment variable. Reliability, in the sense that matters to a defense minister, is a throughput variable. They are correlated, but they run on different clocks, and this fortnight they diverged in public.
Interceptors. Before dawn on August 1, Russia sent roughly thirty-five missiles (twenty-seven of them ballistic) and 185 attack drones at Kyiv. Ukrainian air defense downed 154 drones, an interception rate above 80 percent. Of the twenty-seven ballistic missiles, it stopped one. Zelensky's explanation was not tactical; it was inventory. There are no interceptors for the Patriot systems. A second mass attack on August 5 killed at least seventeen.
The arithmetic underneath is not classified. Lockheed's PAC-3 line has been delivering on the order of 620 interceptors a year. A 2026 agreement targets roughly 2,000 a year by 2030. In the opening days of the Iran conflict, United States and partner forces reportedly expended more than 800 Patriot interceptors and hundreds of THAAD rounds; that is to say, something north of a year of production, consumed in under a week, in one theater. Washington has since committed 58.6 billion dollars against the shortfall. The money moves at the speed of appropriation. The missiles move at the speed of a supply chain with sole-source seekers and solid rocket motors, and they are queued behind Japan, South Korea, and a Europe restocking what it already gave away.
An alliance commitment denominated in a good with one production line and a five-year backlog is not a promise. It is a position in a queue. Every state in that queue can now compute its own place in it, which is a very different exercise from answering a favorability question.
Controllers. On July 30 the FBI and the EPA issued a joint advisory warning that malicious cyber actors were remotely manipulating operational-technology devices (programmable logic controllers governing water pressure and chemical dosing) at water and wastewater utilities. Roughly thirty Minnesota facilities were hit over a single Sunday night and Monday morning. A Minnesota Bureau of Criminal Apprehension memo assessed that the likely desired effect was loss of system pressure and potential contamination. At least seven states reported incidents; some outlets have since put the figure at twelve, and the count was still moving as of this writing. Some utilities issued boil-water notices; some dropped to manual operation. No contamination has been confirmed. NBC reported on August 4 that Iran was likely responsible, on the word of two people familiar with the incidents. At a cabinet meeting on July 31, the President publicly disputed the attribution.
Note the shape of the remediation. CISA's guidance is to remove exposed industrial control equipment from the internet and to ensure facilities can transition safely to manual operation. That is sound advice. It is also, precisely, a recommendation to un-adopt two decades of remote-management capability, and to retain the trained personnel who can run a plant by hand. It is adaptation by reversion. It is the only move available on a fortnight's notice, and it is negative-productivity by construction.
Tokens. The third measurement runs the other way and is easy to miss because the leaderboards obscure it. The frontier of model capability remains American; Anthropic and Google still hold the top of the intelligence indices. But OpenRouter, the largest neutral router of model traffic, shows Chinese-origin open-weight models moving from a rounding error to the majority of tokens actually processed, somewhere between late 2024 and mid-2026. Coding went from roughly 11 percent of that traffic in early 2025 to more than half by mid-2026, and Chinese models are disproportionately strong and cheap there. The Elo gap between open and closed models has compressed from roughly 150 points to roughly 30; the intelligence-index gap from thirteen points to six. This week MiniMax's latest went open-weight and runs on a single consumer GPU.
Holding the frontier and holding the throughput are different assets. The first is a claim about capability. The second is a claim about what the world's infrastructure will be built on top of in five years. Export controls were designed against the first.
Running It Through the Signature
Chokepoint-throughput concentration. Rising, and rising in the least visible place. The concentration that matters is no longer only rare earths and lithography; it is interceptor production, and it is the token layer. Both are single-digit-supplier markets. Both are now load-bearing for commitments that were made when the reserve was assumed to be deep.
Rising-narrative commitment. Rising, fast, and asymmetrically. Canada moved twenty-four points on American favorability in three years. Narrative commitment is the stickiest of the three variables; publics that have committed to a reading of American unreliability will discount subsequent American delivery, which means the throughput problem and the sentiment problem now compound rather than merely coexist. The six exposed holdouts are the countervailing evidence, and they are real. But their loyalty is a function of threat proximity, not of persuasion, which means it is contingent on the guarantee continuing to clear.
Institutional-adaptation velocity. Falling, or at best flat, and the water sector is the tell. The fastest available response to a fortnight-old attack on thirty facilities was to disconnect and revert. Meanwhile the attribution was contested at the political layer within seventy-two hours of the technical layer forming a view, which slows the operational response independent of who is right.
The Mismatch
Three clocks, three speeds.
Sentiment moves in months. Canada repriced the United States in three years; a survey window of thirteen weeks captured the crossing.
Threat moves in nights. Thirty water systems in one night. Twenty-seven ballistic missiles in one night. A year of interceptor production in one week.
Capacity moves in half-decades. Two thousand PAC-3s a year by 2030, if the line holds and the seekers arrive.
The institutions are keyed to the third clock and are being graded on the first two. That is the temporal mismatch in its purest available form, and it is why I do not think the Pew result is best read as a preference. A preference implies a choice between two functioning options. What the seventeen middle-income publics are reporting looks more like an inference: they are observing delivery rates and drawing a conclusion, and the conclusion happens to be legible as an opinion.
What Would Falsify This
I want to be honest about the ways this reading fails.
If PAC-3 deliveries beat the 2030 schedule materially (say, 1,200 a year by 2028), the chokepoint argument weakens considerably and the sentiment shift reverts to being a story about one administration rather than about structural capacity.
If the six exposed holdouts start defecting (South Korea and the Philippines are the ones to watch; Seoul already moved from twice-as-confident-in-Trump to parity in a single year), the proximity hypothesis fails and this is a broader unwind than I have described.
If the water-sector intrusions turn out to be a criminal or hacktivist campaign rather than state-directed, the adaptation-velocity read holds but the threat-clock read is inflated, and I will have overweighted it.
And if Chinese open-weight share on OpenRouter plateaus or reverses over the next four quarters, the token-layer argument was a snapshot mistaken for a trend. That is the one I am least confident about, because I am reading a single router's traffic as a proxy for global deployment, and it may not be.
The Number That Matters
The most quoted number from the Pew release is that more people now express confidence in Xi Jinping than in Donald Trump. It is a real number and it is not, I think, the important one.
The important one is that Kyiv stopped one missile out of twenty-seven, and everyone with a border problem did the arithmetic.
Reliability has quietly changed units. The reserve is gone. What is left is production rate — and production rate is now the unit in which credibility is denominated.
◆ End of dispatch — opening of Volume Three