Phones keep losing cameras. Sony just spent ¥747 billion.

Hey {{first_name|Investor}} -

The average smartphone shipped with 2.18 rear cameras in the second quarter of last year, down from 2.37 the year before. That's 13 straight quarters of decline from the peak in early 2021.

Omdia, which does the counting, puts part of the cause on AI photography. Software now does with one lens what used to take three.

Waymo's newest driving system runs 13 cameras and 4 LiDAR units. The generation before it ran 29 cameras and 5. Alphabet ($GOOGL) cut the sensor count by 42% and the cost by more than half, and the system got better.

Tesla runs its Optimus robot on 8 cameras and no LiDAR at all.

So the eyes are getting fewer. And on August 11, Sony ($SONY) and TSMC ($TSM) signed a ¥747 billion agreement to build image sensors together in Kumamoto, Japan.

Both of those are true at once, and holding them together tells you where the money in this layer actually sits.

The two companies are setting up Advanced Vision Semiconductor Manufacturing Corporation in Koshi City, Kumamoto. Sony puts in about ¥465 billion, part cash and part existing fab assets moved across in a company split. TSMC puts in about ¥282 billion in cash, paid in stages as demand shows up.

That's roughly ¥747 billion, or about $4.7 billion at the exchange rate on the day. Sony controls it and consolidates it. Volume production starts in 2029, at a planned 10,000 300mm wafers a month, which is one mid-sized fab rather than the sprawling campus the headline number tends to suggest.

You'll see $6.3 billion in most of the coverage. That figure came from a Nikkei estimate published August 9, two days before anything was signed, and it estimates eventual total project spend rather than the contracted amount. The difference between the two is an expansion tranche that depends on Japanese government support nobody has committed yet.

Two more things the coverage filled in on its own. The process node hasn't been disclosed by either company; both releases say "advanced manufacturing process technology" and stop there. Neither has the ownership split. Sony's language is "sole controlling shareholder," the capital ratio works out to 62/38, and the widely quoted 60/40 traces back to that same pre-announcement Nikkei piece.

One more, because it trips up almost everyone writing about this. TSMC already runs a chip venture in Kumamoto called JASM, which TSMC controls and Sony owns 6% of. This new company is a separate thing: different city, different control, different product. Both happen to point at 2029, which is most of why the two keep getting merged into one story.

A modern image sensor is two or three separate chips bonded face to face and sold as one part. Copper pads on each surface get pressed together until the stack behaves like a single chip, which the industry calls hybrid bonding.

Once you see it as a stack, the business gets easier to read, because the three layers earn very differently.

Layer 1, the pixel. The top chip catches light. Photodiodes, one per pixel, turning photons into a charge. This is the layer people mean when they say camera, and it's the windowpane of the thing. Making good glass is hard. Making good glass is also something a growing number of companies can now do, which is why the pressure at the bottom of the price ladder keeps building.

Layer 2, the logic underneath. The second chip does arithmetic. It reads each pixel out, runs it through an analog-to-digital converter that turns an amount of light into a number, holds results in memory, and increasingly runs a small AI model before anything leaves the package.

This layer scales with the manufacturing process the way any other chip does. Smaller transistors mean more converters running side by side, which means faster frames and less of the wobble you get when a sensor reads its rows one at a time. More memory on the die means the whole frame can be grabbed at once. More logic means the noise reduction and the exposure blending happen before the data ever travels.

This is the layer Sony just bought a partner for.

Think of the wiring and the fuse box behind a wall. Nobody photographs it. It decides what the house can run.

Layer 3, the decision. What leaves the package. Either a full picture, which somebody else then has to process, or a short note that says person, here, now.

Every millimeter the decision moves closer to the pixel takes power and delay out of the system, because shifting data off a chip and across a board burns far more energy than the arithmetic itself, which is the single physical fact the entire layer 3 argument rests on. On paper that's the layer with the best economics, and I'll come back to what it has actually produced.

The two layers want opposite factories.

A pixel layer wants a large, mature, heavily tuned process where the whole priority is how cleanly each photodiode collects light. Sony's fabs are built for exactly that, and Sony has spent 20 years making them better at it.

A logic layer wants small transistors, and small transistors are TSMC's business, which is the one part of this arrangement that needed no explaining to anybody.

So Sony kept the part competitors struggle to copy and rented the part it was falling behind on. The division of labour is in the announcement: Sony leads technology development and product design, TSMC supplies the process and the manufacturing.

The other half of the structure is money. TSMC's ¥282 billion absorbs part of the capital and part of the process risk on a bet that doesn't produce a wafer until 2029.

Any company claiming to benefit from a trend has to clear two separate bars, and they fail at different places.

Gate one: does the number of billable units go up?

Gate two: does the price per unit hold, and does the company keep a margin on it?

Gate one is the one everybody runs, and in this layer it fails almost everywhere.

Phones went from 2.37 rear cameras to 2.18, 13 quarters running. Waymo went from 29 cameras to 13. Optimus uses 8 and nothing else.

Robots are supposed to fix this, so run the arithmetic on that too. Omdia expects general-purpose robot shipments above 446,000 units by 2030, pulling something over 2.2 million cameras along with them. Smartphones consumed 4.19 billion image sensors in 2025 alone.

The entire robot camera market at the end of the decade comes to about 0.05% of one year of phones. Swap in Bank of America's much larger robot forecast, which runs 1 to 1.2 million units by 2030 against Omdia's 446,000, and the answer barely moves.

Anyone selling robotics as a volume replacement for smartphones is doing the arithmetic wrong, and it isn't close.

Which leaves gate two carrying the entire case.

A layer where unit counts fall and value per unit rises rewards a different kind of holding than a layer that grows on volume. You're underwriting whether a company can keep raising what it charges for the same physical object, year after year, while somebody cheaper works on the low end.

That's checkable in the filings every quarter, which makes it the kind of position you can hold through a drawdown without watching it daily.

It also means the headline number is usually the wrong one to read.

The rest of this issue is for Insiders.

Below the line I run gate two on Sony's own filings, where revenue and operating profit point in opposite directions next year for a reason Sony explained on its call and almost nobody quoted.

I also run it on two LiDAR companies in the same end market that printed opposite answers in the same year. One grew units 141% and watched gross margin fall. The other grew revenue 56% and put gross margin up 400 basis points. Nothing about their volume told you which was which.

Plus the customer concentration nobody puts next to the CapEx, the patent that will get written up as Sony entering humanoid robotics, and the 7 things I'm watching with dates attached.