Taiwan’s org-chart smuggling, Nvidia’s Perplexity cheque, and a $103,265 H-1B proposal
Monday’s stack after the plumbing Brief. Export controls showed up inside Nvidia’s Taiwan unit, the supplier started shopping a customer above $30 billion, and the six-figure visa fee came back as a proposed rule.
In this Brief
- Taiwan charges nine over AI servers dressed as a local install
- Nvidia is shopping Perplexity; the cheque looks like demand finance
- DHS puts $103,265 in the Federal Register as a proposed H-1B fee
- Pew: a tenth of the sampled web, a third of dated post-ChatGPT pages
- Software hiring is climbing, and the openings are senior
Taiwan’s org-chart smuggling
Setup: Ars Technica reports Taiwan prosecutors charged nine people Monday over illegal exports of high-end AI servers to China. The group includes one employee of Nvidia’s Taiwan unit, identified by AP only as manager Chang, and two Super Micro Taiwan employees. Charges include breach of trust and document forgery.
The scheme dressed 130 Super Micro B300 servers as if they were installed at a rented Taiwan facility. Seventy-four reached Chinese customers: 50 through Indonesia, 16 direct, eight via Japan then Hong Kong. Customs blocked the other 56.
Prosecutors are seeking the five-year maximum for several of the nine, including Chang.
This follows the March U.S. arrest of Super Micro co-founder Wally Liaw, accused of funneling $2.5 billion in restricted servers since 2024.
Super Micro told AP two former employees were arrested after its cooperation.
Nvidia spokesperson Patrick Rutherford said the company will work with Taiwan authorities.
Take: The border still has a stamp. The alleged failure is the signature that said the rack was already in use. Chang is the name on the indictment. The 56 servers that never left Taiwan still needed someone inside to bless them as installed.
Earnings are Wednesday. The org chart is already in the charging documents.
The stamp is what customs never sees.
Longer cut: Taiwan Charges Nine in Alleged Nvidia-Super Micro AI Server Export Scheme.
See full breakdown at @Aaron_Harme on X.
Nvidia’s Perplexity cheque
Setup: The Information reports Nvidia has been discussing an investment that would put Perplexity above $30 billion, with a technology licensing arrangement also on the table, The Next Web writes. Neither company has commented. Stake size and other participants are unreported. It is talks.
The last reported Perplexity mark was roughly $20 billion. More than $30 billion would be a jump of over 50 percent. Nvidia is already on the earlier cap table, with Jeff Bezos among prior investors. Perplexity was among the first named takers for Nvidia’s Vera CPU.
Take: Nvidia puts equity into a customer that already buys its silicon, and may also license tech that sits outside a normal sales relationship. Demand gets financed upstream. The browser war with Google and OpenAI still has to earn the multiple.
Term sheets sign later. The valuation talk is already doing work.
Longer cut: Nvidia is shopping Perplexity. The cheque looks like demand finance..
See full breakdown at @Aaron_Harme on X.
The $103,265 H-1B proposal
Setup: Bloomberg Law reports a DHS proposed rule that would make a $103,265 fee permanent on cap-subject H-1B petitions. It is a proposal, not in effect, with a 30-day comment window and a chance of being finalized by year end.
Trump’s temporary proclamation fee of about $100,000 expires in September. Boston Judge Leo Sorokin ruled in June 2026 that the proclamation fee was an unlawful tax Congress never authorized. The First Circuit is reviewing.
The proposed fee would not apply to people already in the U.S. on student visas, or to renewals. The annual cap remains 65,000 plus 20,000 advanced-degree visas. Pre-proclamation fees typically ran $2,000 to $5,000.
Take: The illegal-tax ruling did not kill the number. They put it in the Federal Register. Until then the workaround is already on paper: keep the fee, change the statute it hides behind.
The number survived the courtroom as a proposed rule.
See full breakdown at @Aaron_Harme on X.
Pew’s AI web
Setup: Pew Research ran nearly 490,000 English pages from Common Crawl through Open Pangram. In a July 2026 snapshot of 10,000 pages, 10 percent show significant signs of AI authorship or editing.
Among pages in that snapshot that carry a publication date after ChatGPT’s Nov. 30, 2022 launch, the share is over one-third, about 35 percent. That 35 percent is dated pages, not a claim about the whole internet.
Pew’s methodology notes only about 10 to 15 percent of pages even have a detectable publication date, so the post-ChatGPT cut is a newer slice. By domain in 2026 samples: roughly 10 percent of .com pages, 4.6 percent of .org, around 1 percent of .edu and .gov. Versus 2023, em dashes about doubled, Oxford commas rose 63 percent, and “it’s not just X, it’s Y” nearly tripled.
Take: This is a pile statistic from one model’s scores. Old pages still dilute the 10 percent, and new dated pages are where the third shows up.
Train the next model on the open web and you are increasingly training it on last year’s model. Search, citation, and the dated record are the first places that shows up.
The open web is filling with text that looks like the models that trained on it.
See full breakdown at @Aaron_Harme on X.
Software hiring is senior
Setup: Business Insider reports that as of Q1 2026, nearly 70 percent of U.S. software-development job postings on Indeed were senior-level, up from 55 percent in early 2019. That was the highest senior share among more than 40 sectors Indeed analyzed. Postings are still below pre-pandemic levels, but they were up roughly 15 percent in June versus early 2025.
Indeed Hiring Lab economist Sneha Puri: AI tools can increasingly handle the routine, well-defined coding tasks that used to go to junior engineers. That shifts demand toward workers who can direct AI output, catch its mistakes, and make the judgment calls AI cannot.
Take: The rebound is real. The ladder under it is thinning. Today’s entry-level workforce becomes tomorrow’s senior-level experts. If employers want senior talent capable of managing AI later, they need junior roles now. Pair that with a six-figure H-1B proposal and the pipeline argument is no longer hypothetical.
See full breakdown at @Aaron_Harme on X.
Also noted
- Alibaba’s Hong Kong shares fell as much as 10 percent Monday after the AI placement, Business Insider reports. Michael Burry moved his BABA stake to JD.com. Days later Wan3.0 opened wider at $0.20 per second 1080p versus Veo 3.1’s $0.40 standard tier, The Next Web writes. Covered at @Aaron_Harme.
- Nvidia says Groq 3 LPX racks are in full production after the $20 billion asset deal. Artificial Analysis: 3,400 tokens per second on Gemma 4 31B with a 100,000-token input. The Register flags that as a possible best-case; larger mixture-of-experts scaling is unproven. Covered at @Aaron_Harme.
- SoftBank’s record ~¥1 trillion (~$6.3 billion) retail bond, framed around OpenAI and AI capex, went out Monday morning at @Aaron_Harme.
Sources
- Ars Technica – Taiwan AI server smuggling indictment
- AP via ABC – Taiwan charges nine
- Bitter Fool – Taiwan org-chart long take
- The Information – Nvidia Perplexity talks
- The Next Web – Nvidia Perplexity $30B talks
- Bitter Fool – Nvidia shopping Perplexity
- Bloomberg Law – H-1B $103,265 proposed rule
- Pew Research – How much of the internet is written with AI
- Business Insider – Software hiring is senior
- Business Insider – Burry dumps BABA after AI share sale
- The Next Web – Alibaba Wan3.0
- CNBC – Nvidia Groq 3 racks
- The Register – Groq 3 LPU benches
- @Aaron_Harme on X
