Amazon Designs Custom AI Chips for Fire TV

In an interview with CNBC, Panos Panay said, "We do make our own end-to-end silicon for the devices that we ship," identifying the Echo Show 8, Echo Show 11, and various Fire TV models as current examples (CNBC). Amazon introduced its AZ3 and AZ3 Pro chips in October 2025; independent reporting at the time described them as powering a sensor-fusion platform called Omnisense, which combines audio, ultrasound, Wi-Fi radar, accelerometer, and camera data, and improved wake-word detection by more than 50% on the Echo Dot Max. Separately, Taiwanese supply-chain analyst Ming-Chi Kuo said in an X post, reported by Yahoo Finance, that Amazon has selected Alchip as its exclusive back-end design and testing partner for a broader shift to in-house processors covering Kindle, Fire TV, and Alexa-enabled devices including Blink and Ring. Kuo said the transition would begin in 2027 and could eventually reach about 40 million annual shipments of Amazon-designed processors, and linked the move partly to cost discipline, noting Amazon's trailing-twelve-month free cash flow fell 95% year-over-year to about $1.2 billion as of Q1 2026 amid heavy AI infrastructure spending (Yahoo Finance, citing Ming-Chi Kuo).

The Alchip and 40-million-unit figures come from a single analyst's account, not an Amazon disclosure, so specifics such as exact chip families, timing, and volumes should be treated as reported projections rather than confirmed plans. That single-source caveat does not apply to the AZ3/Omnisense details or Panay's CNBC quote, which describe products already shipping today. Taken together, the pattern is consistent with what device makers with in-house silicon programs, such as Apple and Google, have already done: move AI inference on-device to cut latency and cloud compute costs while giving product teams more control over the accelerator roadmap.

Teams shipping ML models to Amazon's device ecosystem should plan for multiple silicon targets rather than a single reference platform: current AZ3/AZ3 Pro hardware already requires vendor-specific quantization and compilation work, and a broader in-house chip rollout would extend that requirement to more product lines over time. Cross-architecture CI, hardware-in-the-loop thermal and power validation, and versioned over-the-air model delivery become more important as the device fleet running local inference grows.

Confirmation from Amazon itself, rather than analyst reporting, on the Alchip partnership and 2027 timeline; any developer-facing SDK or compiler toolchain Amazon releases for AZ3-class silicon; and whether the free-cash-flow pressure Kuo cites shows up in Amazon's own quarterly disclosures as a stated rationale for the custom-silicon push.