Six stories worth an extra minute this morning. One of them is a robot market that turns out to be a research market, and one is an accounting structure that lets the biggest buyers of AI hardware keep the bill somewhere other than their own balance sheet.
About 7,000 humanoids sold worldwide last year, and many of them aren’t working
The industry figure for 2025 is roughly 7,000 autonomous humanoid robots sold globally for industrial and professional use. That’s the whole planet, across every vendor, in a year when humanoids were the most photographed machines in technology.
The caveat inside the number matters more than the number. A large share of those units went to research buyers: universities, corporate labs, integrators building demos. They were bought to be studied, not to move a pallet. Strip those out and the population of humanoids doing paid work somewhere is small enough to fit in a couple of large factories.

- Gartner CIO & IT Executive Conference 2026, September 21–23, Sheraton São Paulo WTC Hotel, São Paulo
- Breakbulk Americas 2026, September 22–23, George R. Brown Convention Center, Houston
- Nova Future Summit 2026, September 28–30, Napa
- Networking Field Day 2026, October 6–9, San Jose
- ONUG AI Networking Summit 2026, October 28–29, Penn District, New York
- FYUZ 2026, November 3–5, The Westin Seattle, Seattle
- ITC Vegas 2026, September 29–October 1, Mandalay Bay, Las Vegas
- Sidoti Small-Cap Virtual Conference: September 23-24, Online
- Status Summit 2026: October 7, Location TBA
- CreatorIQ Connect 2026: October 13, Los Angeles, California
That isn’t an argument the category fails. Early industrial robots sold in similar volumes before the automotive lines standardised on them. It’s an argument about timing. When you see a chart of humanoid unit growth, ask how many of those units are research platforms, because a research platform doesn’t re-order at scale and doesn’t prove an ROI case to the next buyer.
The iPhone 18 Pro moved its brain off the logic board
The teardown finding that stands out: the SoC no longer sits on the logic board. It’s been relocated to manage heat, which is the clearest signal yet that thermals, not transistor count, are the binding constraint on what a phone can run locally.
- Tech Digest, September 21, 2026: China’s Stock Tax Take Jumps 80%, SoftBank Lines Up $10B of Debt for OpenAI
- Apple’s Siri Home Hub, Raspberry Pi’s AI Skeptic, CXMT’s New LPDDR5X and Polymarket’s Laundering Scare
- Tech Digest: September 17–18, 2026
- Top 10 Emerging Technologies in 2026
- The World Economic Forum and Forrester Can’t Agree on What Counts as Emerging Technology in 2026
- Snap’s AI Glasses, Faraday Future’s Robot Push, and Fresh AI Funding Lead the Sept. 16-17 Tech Wire
- Bending Spoons Buys Miro at a 90% Discount
- Morning Tech Digest, September 10, 2026: Chinese AI Chip Prices Up 20% to 50% on HBM Costs, Nasdaq’s $100 Million Kraken Bet
- Apple Watch Series 12 and Ultra 4: The Hard Part of Audio Intelligence Is Everyone Not Wearing the Watch
- Apple iPhone 18 Pro: The Base Price Rose $100, the Top Storage Step Rose $600
On-device models are the reason. Sustained inference produces a very different heat profile from the burst loads phones were designed around, and you can’t solve it with software scheduling forever. Physically separating the hot component from the dense board is a structural answer to a structural problem, and it’s the kind of change that tends to stick for several generations.
The variable aperture gets credit for design and a penalty for repair. It’s described as thoughtfully engineered and difficult to service, which is the recurring trade in modern phone hardware. Every moving part added to a camera stack is another module that fails as an assembly rather than a component.
Residual value guarantees are quietly reshaping how AI capex gets reported
Large technology buyers are making more use of residual value guarantees to keep AI infrastructure spending off the balance sheet, with chip suppliers in a position to support their customers’ purchases through the same mechanism.
Here’s what a residual value guarantee does. Someone promises that the equipment will be worth a stated amount at the end of the term. That promise is what lets the arrangement be treated as a lease rather than an outright purchase, and it’s what moves the asset and the debt out of the headline numbers.
Two things follow. First, reported capex understates real compute commitments, so the gap between what these companies say they’re spending and what they’ve actually contracted for widens. Second, the guarantee has to be made by somebody, and when the supplier is standing behind the residual value of its own hardware, the vendor is carrying customer risk. That’s a familiar structure. It worked for years in telecom equipment until the residuals didn’t hold.
The question to keep asking is what a two-year-old accelerator is assumed to be worth in these agreements, and who eats the difference if the assumption is generous.
Amazon blocked Meta’s shopping agent, and the reasons are the interesting part
Amazon says it has blocked Meta’s Muse assistant from accessing Amazon.com to shop on users’ behalf. The stated grounds are terms of service violations, security risk, and merchant consent.
Merchant consent is the argument with legs. Amazon’s position is that a third-party agent buying on a customer’s behalf inserts itself between the seller and the buyer without the seller agreeing to it. Whether that holds up is a genuinely open question, and it’s going to get asked repeatedly over the next two years as every large consumer platform ships an agent that wants to transact somewhere else.
Underneath the legal framing is a commercial one. Retail media is a very large business built on controlling what a shopper sees at the moment of purchase. An agent that goes straight to checkout skips all of it. No platform with an ad business is going to let that traffic through voluntarily, whatever the terms of service say.
South Korea will let AI researchers serve in corporate labs again
Seoul has reopened a programme allowing AI specialists to satisfy their military obligation by doing research in company labs instead of conventional service.
For a country with mandatory conscription and a semiconductor industry competing for the same small pool of people as American labs paying in dollars, this is industrial policy wearing a defence uniform. Military service at the wrong age can remove a researcher from the field during exactly the years that matter most, and for the best candidates it’s a reason to build a career elsewhere.
Watch how narrow the eligibility is. If it’s restricted to a handful of large firms, it becomes another advantage for the incumbents in a market that already concentrates talent heavily.
The Clarity Act died in the Senate over what stablecoins pay
The crypto market structure bill collapsed in the Senate after heavy lobbying on one side and bank opposition on the other, with the fight centred on whether stablecoin issuers can pass yield to holders.
That detail is the whole thing. A dollar sitting in a checking account earns close to nothing for the depositor and a good deal for the bank. A tokenised dollar paying interest competes directly with that, and at scale it pulls cheap deposits out of the banking system, which raises the cost of funding every loan a bank writes.
Banks understood the stakes precisely and lobbied accordingly. Framing it as a crypto regulation fight makes it sound like an argument about digital assets. It’s an argument about deposits, and it will come back in the next session under a different bill number.