Washington spent the spring of 2025 banning the Nvidia H20 and the summer unbanning it, and moved almost no silicon into China either way. Not a slowdown. Close to nothing. Beijing told its own AI firms to skip the chip even after the export licenses came through in August, and Nvidia turned around and told its suppliers to stop building it because the orders were not there. That is what a policy eating itself looks like, the people writing the rules and the people they are aimed at both playing a longer game than the press releases admit.
Three things landed across this summer, and on their own they read like a chip story, a laptop story, and a car story. Stacked up, they sketch something less comfortable: the United States making huge bets on software, autonomy, and AI platforms while the hardware and policy scaffolding underneath those bets quietly gives way.
The H20 is the chip Nvidia built to live one inch under the line. It takes the Hopper architecture, strips out the training-grade compute the October 2022 export rules cap, and keeps the memory capacity and bandwidth that matter for inference. The result runs large models well and trains them slowly, on purpose, and ships to China without a license. I took the hardware apart in detail when the reversal hit; the short version is that it is a deliberate piece of legal engineering, not a crippled accident.
Then April happened. Commerce decided that even inference silicon at H20 scale was feeding Chinese AI, especially with Chinese labs getting better at wringing training out of weaker parts, and it pulled the license. AMD’s MI308 got caught in the same net. Nvidia booked a $4.5 billion charge against H20 inventory it could no longer sell, lost another $2.5 billion in revenue it could not ship that quarter, and guided to roughly $8 billion gone from the next one. Real money, and it stung.
Three months later Commerce blinked and started waving the licenses through again, and then the deal got strange. By August, Nvidia and AMD had agreed to hand the US government 15% of their China AI-chip revenue in exchange for those licenses, a cut with no precedent anywhere in the history of US export control. A license is supposed to govern where a chip goes, not skim the register on the way out. The 15% reads like a device for having it both ways, reopen China for the headline and price the deal high enough that not much actually moves. The president waved off the security worry by calling the H20 “old,” which is true against Blackwell and beside the point commercially, because most deployed AI compute is inference and the H20 is still good at it. I walked through the licensing mechanics separately when the ban first lifted.
China’s own AI firms were quietly told to pass on the H20, and that was arithmetic, not pride. The April ban proved a US license can vanish on a few days’ notice, which turns any chip you pour a data center around into a single point of failure you do not control. Huawei’s Ascend 910C, built on SMIC’s 7nm-class line, was landing somewhere in the neighborhood of the H20 on the inference workloads Chinese labs actually run. So the real choice is a foreign chip with revocable supply against a domestic chip with secure supply, at roughly comparable inference performance, before you even add the 15% surcharge. By late August, Beijing was steering firms away and Nvidia was telling Amkor and Samsung to halt production. The net flow stayed near zero.
The scale of the buildout makes the whole fight look small. Microsoft alone committed about $80 billion to AI data centers in fiscal 2025, more than half of it in the US, and the big hyperscalers together are spending at a pace that exceeds that, with credible forecasts putting AI infrastructure spend into the trillions by the end of the decade. Against that, the slice of Chinese inference revenue the H20 fight was actually about runs maybe single-digit billions a year. The policy may have cost more than the prize. And the contradiction beneath it was never resolved. You cannot deny China advanced compute and let Nvidia sell it inference-capable chips at the same time, and the whole saga is the administration swinging between those two incompatible goals without ever picking one.
While the chip fight burned, Qualcomm was quietly assembling the layer underneath the next device era, and it started by buying a research lab. On April 1, 2025 it acquired MovianAI, the generative-AI group spun out of Hanoi’s VinAI, founded by former DeepMind scientist Hung Bui. Price undisclosed. The buy tells you what Qualcomm thought it was short on. Its Snapdragon silicon had carried an NPU capable of on-device LLM inference since the 8 Gen 3, so silicon was never the problem. What it lacked was the research muscle to wring real generative performance out of that NPU, the software that drags hardware toward its ceiling, and VinAI had spent years building exactly that. It fits the same own-the-software-layer logic driving Qualcomm’s edge push: a capability buy aimed at a specific hole, not a headcount grab.
I am a declared Snapdragon Insider, so weight my enthusiasm accordingly, but the X Elite spread by mid-2025 is hard to wave off. The flagship runs twelve Oryon cores and a 45-TOPS Hexagon NPU on TSMC’s 4nm process, clearing Microsoft’s 40-TOPS Copilot+ floor, and by summer it was shipping across Surface, Dell’s XPS 13, HP’s EliteBook, Lenovo’s ThinkPad and Yoga lines, Samsung’s Galaxy Book, and Asus’s Vivobook. That kind of spread across consumer, prosumer, and enterprise is escape velocity for an Arm-on-Windows push the industry had written off a few years earlier.
Qualcomm pushed the same NPU logic into industrial territory at Computex 2025, partnering with Advantech, one of the biggest industrial-IoT hardware makers around. The pitch is Snapdragon compute modules running real-time inference at the edge for quality control and predictive maintenance, the fanless, low-latency, wide-temperature work the NPU was actually born doing. This is far closer to the chip’s native habitat than the PC ever was, and the automotive and IoT lines were already posting records while everyone watched the laptops.
Microsoft put the new Snapdragon Surfaces into business and enterprise channels on July 14, starting around $1,650, and enterprise procurement is a different beast than consumer sales. IT wants Intune and Autopilot and BitLocker working, long-term driver servicing, and the SAP-Salesforce-Adobe stack running without falling over, and by mid-2025, the Prism emulation layer in Windows 11 had gotten solid enough that most of it did, the holdouts shrinking to legacy 32-bit apps and kernel-mode drivers. An enterprise win at Surface scale is the validation Qualcomm has chased since the first Arm-on-Windows misfire back in 2017. That platform leadership is real and American. The fab it all depends on is in Taiwan, on TSMC’s 4nm line, which is the same structural hole that runs under the H20 fight, just dressed in nicer clothes.
Cars are where the contradiction gets loudest, because Washington killed its EV consumer incentive in the same season it staked its autonomy future on a fleet you could park in a single lot. The One Big Beautiful Bill, signed July 4, wiped out the federal EV tax credit the 2022 climate law set up: $7,500 on a new EV, $4,000 used, up to $40,000 commercial, all gone for anything bought after September 30, a date that lines up with the federal fiscal year rather than anything about the car market.
It was bending behavior before it even bit. Dealers reported a rush of buyers trying to lock the credit before the September cutoff, Tesla openly told people to “YOLO” a purchase ahead of the deadline, and that kind of pull-forward just borrows sales from next year. Once the credit lapses, the average EV’s price gap over a comparable gas car widens by roughly the $7,500 the subsidy used to paper over, and analysts were already modeling US EV sales sliding toward a million units within a couple of years, well down from where 2025 was tracking.
While Washington pulled the consumer incentive, Tesla was pushing out its Austin robotaxi geofence, which had gone from about 20 square miles at the June 22 launch to roughly 80 by early August. The fleet doing the driving is a couple dozen modified Model Ys, and the architecture is Tesla’s defining wager: eight cameras doing all the work, no LiDAR or radar, none of the pre-built HD maps everyone else leans on, running FSD on the in-house HW4 computer. Camera-only is the whole bet. The case for it is cost and simplicity, since a LiDAR unit runs hundreds to thousands of dollars and adds moving parts. The case against it is that LiDAR measures depth directly while a camera has to infer it, and bad depth guesses are a leading cause of autonomous crashes.
Waymo is the only rigorous public benchmark, and the numbers are sobering for anyone running a leaner stack. It has driven tens of millions of fully driverless miles and published safety data showing dramatically fewer injury-causing crashes than human drivers, and it gets there with a redundant stack of cameras, LiDAR, radar, and HD maps where several systems have to fail at once for a safety-critical error to slip through. Tesla’s fleet of a few dozen, months old, has not driven enough to produce a number that means anything next to that. The geofence expansion is the real test, because pushing into streets the network was never specifically validated on is exactly where a camera-only system either earns trust or does not. I am not going to settle Waymo versus Tesla here, that is three more posts on its own. What matters right now is the architecture, not the fleet count, and camera-only is still a contested bet rather than a settled one.
China is the other half of the car story. It sold north of eleven million EVs in 2025, somewhere around half its new-car market, against roughly 1.3 to 1.4 million in the US at under ten percent. That gap is not about taste or chargers. It is manufacturing cost. Chinese makers like BYD were down around $60 to $75 per kWh at the cell level by mid-2025, against $80 to $100 for US and European builders, and that feeds straight into the sticker. BYD’s Seagull, a real urban EV, sells in China for around $10,000, a number no US automaker touches even with the $7,500 credit fully intact. Kill that credit and you widen the gap at the exact moment Chinese makers are pushing sub-$15,000 EVs into Southeast Asia, Latin America, and Europe.
Three breaks, and they rhyme. The export controls leak a little more every quarter Huawei closes the gap. The platform layer is American right up until you ask where the wafers come from, which is Taiwan. And the car market is trading structural ground to Chinese builders for what amounts to a fiscal-year accounting win. Under all of it is the same reflex: bet on the software, on the autonomy, on whatever the spreadsheet likes this quarter, and assume the hardware and the slow grind of industrial policy will sort themselves out. Each of those bets might even be right. Tesla’s camera gamble could pay off; Qualcomm’s platform play probably does; the chip controls made real sense back when China had nowhere else to turn. It is making all three at once, while the floor under each one quietly rots, that should worry you. Win every last one of these bets and you can still end up in 2030 owning the software for machines somebody else actually builds. And powers. And sells back to you cheaper than you can make them yourself.