No low-cost U.S. on-device AI chips: $100 vs $25

AMD’s January 5 CES 2026 launch of its Ryzen AI Halo developer platform gave the on-device AI debate a concrete new data point: a compact PC, built on Ryzen AI Max+ chips, that AMD said can run models with up to 200 billion parameters locally. The claim circulating on social media this week — that there are no low-cost U.S. on-device AI chips and that the cheapest U.S. parts start around $100 while Chinese and Taiwanese alternatives cost about $25 — is harder to verify from public listings. What the market does show is a split between high-end U.S. developer systems and a wider field of lower-cost embedded processors and modules sold through distributors. ### Where does the 200B-parameter AMD example come from? AMD said in a January 5 press release that its new Ryzen AI Halo developer platform was built with Ryzen AI Max+ processors and could run “up to 200 billion parameter models locally.” The company said the system includes up to 128GB of unified memory and up to 60 TFLOPS of RDNA 3.5 graphics. AMD described the product as a developer platform, not a low-cost phone or embedded chip. The launch material positioned it for desktop-class local AI work in a compact form factor, which is a different market from mass-market handset or IoT system-on-chips. ### Is there public evidence that U.S. edge AI chips start at $100? (ir.amd.com) Qualcomm’s Dragonwing QCS6490 product page lists the processor’s capabilities — including up to 12 dense TOPS, industrial and IoT use cases, and development hardware — but not a public chip price. Lantronix and Tria both sell systems built around QCS6490, again without posting a low standalone silicon price on the product pages surfaced in search. (ir.amd.com) That matters because the viral comparison appears to mix unlike things: bare chips, system-on-modules, single-board computers and developer kits. Public distributor listings that are easy to verify tend to be for finished boards or modules, which carry memory, storage, power and support costs on top of the processor itself. ### What do the public price checks actually show? Octopart showed Rockchip RK3588 pricing from several distributors at roughly $52.50 to $74.45 per chip in small quantities, based on a page crawled this week. (qualcomm.com) Mouser also listed RK3588-based single-board computers from Mixtile at $229 and $369, and a DFRobot ROCK Pi 5B configuration at $303.75. Mouser listed Hailo-8 AI acceleration modules at $149 to $169 and a Hailo-10H starter kit at $212.50. (lantronix.com) Those are Israeli products rather than U.S. silicon, but they show that publicly listed edge AI hardware often lands well above the $25 figure once it is packaged as a module or kit. NXP’s i.MX 95 family page and Texas Instruments’ edge AI page also advertise edge AI processors and software stacks, but the surfaced pages do not publish simple per-chip prices. (octopart.com) That makes any broad claim about a uniform U.S. floor price difficult to prove from open sources alone. ### Why does the comparison get messy so fast? Phones and embedded devices buy silicon in high volume and often on confidential terms. Distributor pages usually show engineering quantities, modules or boards, not the contract prices handset makers or appliance manufacturers negotiate directly with chip vendors. (mouser.com) The result is that a social-media claim can be directionally plausible without being cleanly provable. Public sources do support one narrower point: U.S. companies such as AMD, Qualcomm, NXP, Intel and TI are visible in higher-end PCs, industrial edge platforms and developer systems, while many low-cost ARM boards and embedded parts sold through distributors come from Asian suppliers. (nxp.com) ### What should readers watch next? AMD said the Ryzen AI Halo platform is scheduled to launch in Q2 2026. The next useful evidence will come from retail pricing for shipping systems, plus any public bill-of-materials or distributor listings for Qualcomm, NXP, TI and Rockchip-based edge AI modules aimed at lower-cost devices. (theoutpost.ai) (amd.com)