Edge AI.
Squeezed by the cloud.

If you bought a phone this year, you helped pay for someone else's data center.

AI data centers are buying so much DRAM and flash that memory prices are up more than 300% from a year ago, and phone makers are passing the bill along. IDC expects the average smartphone price to jump about 28% this year, while shipments fall by the most it has ever recorded.

That's bad news for anyone selling phones, and worse news for edge AI. Edge AI, or on-device AI, means running models on the device itself instead of in a data center: on phones, laptops, cars, cameras, robots, earbuds, and factory equipment. It's faster, more private, and it works without a connection. But it needs enough memory to hold the model, and memory is exactly what just got scarce. To hold prices, phone makers are shipping phones with less RAM, which limits what those phones can run locally.

That's the problem for edge AI stocks going into 2027: running AI on the device takes memory, and the cloud is buying all of it. The way through is to need less of it, or to sell into devices where memory isn't the expensive part.

This list of edge AI stocks starts with the companies built to ride out the squeeze, then moves off the phone: the chipmakers diversifying away from handsets, the physical-AI specialists, the tiny-device players, and one private name.

These two have structures that soften the memory hit, one through engineering and the other through how it gets paid.

Apple NASDAQ: AAPL

Apple has the clearest answer to the memory problem. Its new on-device model has 20 billion parameters, but it keeps the full model in flash storage and loads only the 1 to 4 billion it needs into DRAM at a time. That's a big model running on a phone without a phone-sized memory bill.

Apple's buyers can also absorb the cost better than most. IDC expects iOS shipments to slip just 1.3% this year while Android falls 24.3%, pushing Apple's share to a record. iPhone revenue grew 22% in the June quarter.

Arm NASDAQ: ARM

Arm designs the processor cores inside more than 99% of the world's smartphones. Its royalties rose 22% last quarter in a year when phone shipments are falling, helped by higher rates per chip on its newer designs.

That pricing power is the point. Arm is also reorganizing around three businesses, one each for cloud, physical and edge AI, so it gets paid whichever devices win.

These two built their businesses on phone chips. Now they're finding growth in other devices while the phone market shrinks.

Qualcomm NASDAQ: QCOM

Qualcomm still makes most of its money from phones, and that business fell 20% last quarter in what it called a "challenging memory and supply environment." Its automotive revenue rose 61% to $1.6 billion, the 23rd straight quarter of double-digit growth.

It's pushing further into physical devices. Qualcomm bought Arduino, the hobbyist and industrial hardware platform, last year, and it targets $40 billion of non-phone revenue by fiscal 2029. Qualcomm is a phone company trying to stop being one before the phone market forces the issue.

MediaTek TWSE: 2454

MediaTek is further along. Its edge-device business, which covers TVs, routers, PCs and other smart hardware, grew 26% last quarter and is now 53% of revenue. Its phone chip revenue fell 20%, and management blamed rising smartphone component costs.

MediaTek expects smartphone units to fall about 15% this year. It's also Nvidia's partner on a new Windows PC chip due this holiday season.

Cars, cameras and robots need AI that sees and acts in real time, and they can't wait on a round trip to the cloud.

Ambarella NASDAQ: AMBA

Ambarella makes the vision chips inside security cameras, car cameras, drones and robots. Edge AI revenue hit a record last quarter, and total revenue grew 13%. In May it signed a long-term agreement with Hanwha worth more than $800 million in potential revenue over a decade. It just released its first standalone AI accelerator as well. Management doesn't buy the memory itself, but it has warned that higher memory prices could cut customers' volumes.

Some edge devices run on milliwatts and a sliver of memory. For the companies that sell into them, the squeeze barely registers, which makes these two the edge AI stocks least exposed to it.

Ambiq Micro NYSE: AMBQ

Ambiq makes ultra-low-power chips for wearables, hearing aids and smart sensors. Sales grew 90% last quarter, its fifth straight quarter of growth. The CEO says the company is on pace to more than double sales in the second half, and that its main limit right now is supply. For most of this list, the constraint is customers' memory bills. For Ambiq, it's how fast it can get chips made.

CEVA NASDAQ: CEVA

CEVA licenses the designs for wireless and AI processing blocks that other companies build into their own chips. Its licensing revenue hit a three-year high last quarter, and a major AI and computing company it didn't name licensed its NeuPro-M AI engine. Devices shipping with CEVA technology grew 16% to 567 million in the quarter. CEVA gets paid on volume across thousands of small devices, not on the fate of any one phone.

Axelera AI

Axelera is a Dutch chip startup that raised more than $250 million in February. In September it started shipping Europa, an AI accelerator validated in Dell and Supermicro systems. Axelera started out selling edge chips and is now moving up into on-premises servers, which says where it sees the money right now.

Edge AI's next breakthrough won't be a faster chip. It'll be a model that runs well in 8 gigabytes of RAM instead of 12, because that's what phone makers will actually ship next year.

The work has already started. Apple keeps most of its model in flash and loads only what it needs. Google released quantized versions of its Gemma 4 models specifically "to dramatically reduce memory requirements." Mixture-of-experts designs run only a small part of the model for each word.

For years, edge AI progress was measured in bigger NPUs and more TOPS. In 2027 it gets measured in memory saved. The companies that design for scarce memory, in silicon and in software, will own the edge while the cloud keeps buying all the DRAM. That's the test to hold edge AI stocks to in 2027.