Anker has begun promoting a new "Thus" chip built for its wireless earbuds line, and the headline figure is a 150x jump in on-device AI compute versus prior earbud silicon. The company attributes the gain to a compute-in-memory (CIM) design rather than a simply faster traditional processor. That's a real, well-documented class of chip architecture — but a 150x marketing multiplier from any hardware vendor deserves the same scrutiny applied to any other first-party performance claim, whether it's a GPU maker touting a generational leap or the cache-driven gains debated in Intel Nova Lake-S 22-Core vs AMD Ryzen 7 5800X.
What Compute-in-Memory Actually Changes
Traditional, von Neumann–style chip designs keep memory and compute as separate blocks connected by a bus. Every multiply-accumulate operation in a neural network inference pass means shuttling data back and forth across that bus, and in power-constrained silicon, that data movement — not the arithmetic itself — is usually the dominant energy cost. Compute-in-memory designs perform the multiply-accumulate step inside or immediately adjacent to the memory array, cutting the number of round trips data has to make.
This is not a new idea. Research labs and startups have pursued analog and digital CIM designs for edge AI for close to a decade, largely because the approach suits the extremely tight power budgets of hearables, wearables, and sensors. What would be genuinely new, if Anker's claims hold up under scrutiny, is a mainstream earbud shipping the approach at consumer scale rather than in a lab prototype or a niche microcontroller line.
Why "150x" Needs an Asterisk
Multiplier claims like this almost always compare a narrow, favorable workload against a narrow, unfavorable baseline — for example, a specific keyword-spotting or noise-suppression model run on the new chip against the same workload run inefficiently on a general-purpose DSP from a previous product generation. That's a legitimate way to show directional improvement, but it is not the same as saying "this chip is 150 times faster at AI in general," and vendors rarely disclose the exact workload, the baseline part, the numeric precision (INT8, INT4, or analog), or the batch size behind the headline number.
The pattern is familiar to anyone who follows manufacturer slide decks. Memory and chip vendors do the same thing when promoting capacity or bandwidth jumps — see the scale of the production claims discussed in Samsung and SK Hynix Plan $590B Chip Push as Memory Prices Climb — and the number that ships in a press release is rarely the number an independent lab reproduces a quarter later.
What a Real Benchmark Would Need to Show
For a 150x figure to be comparable across products, independent testing would need to publish:
- The exact model(s) run (a specific noise-cancellation or wake-word network, not "AI workloads" generically)
- Precision and quantization used on both the Thus chip and the baseline part
- Power draw during inference, not just peak throughput
- Latency at a fixed accuracy target, since a faster-but-less-accurate result isn't a fair trade
- The identity and generation of the baseline chip being compared against
None of that has been published by a third party as of this writing. Outlets that typically stress-test vendor performance claims on desktop and mobile silicon have not put out earbud-scale compute-in-memory benchmark data, and earbud chips in general receive far less independent scrutiny than GPUs or CPUs — there's no easy, repeatable way to run a comparable AI workload on a sealed earbud the way there is on a PC part like the ones compared in Ryzen 7 5700X vs 5800X for a 2026 Gaming Build.
How This Fits the Broader "AI Chip Everywhere" Trend
Anker's push mirrors a broader pattern of companies baking dedicated AI silicon into products that never had it before. Google's approach of integrating models directly into silicon, covered in Google's 'Frozen v2' Chip Bakes Gemini Into Silicon, and Nvidia's strategy of funding AI startups to keep them tied to its hardware stack, detailed in NVIDIA Bankrolls AI Startups to Tighten Its Chip Grip, both reflect the same underlying incentive: AI-branded silicon commands attention and often a price premium, whether or not the workload genuinely benefits from it.
It's also worth separating two comparisons that get conflated in casual coverage of chips like this. A comparison between an earbud's power-sipping CIM chip and a desktop discrete GPU or a data-center accelerator is not analytically useful — the two parts operate at power budgets separated by three or four orders of magnitude and serve entirely different workloads. A milliwatt-class earbud chip should be benchmarked against other milliwatt-class hearable and wearable silicon, not against a discrete GPU or training accelerator. Any chart that puts an earbud chip and a desktop-class GPU on the same performance axis is a red flag for the same reason a chart comparing across completely different socket and TDP classes, like Ryzen 7 5800X vs Ryzen 7 5700X, would need heavy caveats — the numbers aren't answering the same question.
Table: Reading Vendor AI-Chip Claims Critically
| Question to ask | Why it matters |
|---|---|
| What is the baseline part? | A 150x gain over a weak baseline is a different claim than 150x over the prior flagship |
| What workload was measured? | Keyword spotting, active noise cancellation, and generative on-device tasks have very different compute profiles |
| Was power draw held constant? | A throughput claim without a power number is incomplete for a battery-powered product |
| Has any outlet reproduced it? | First-party numbers without third-party reproduction are marketing until proven otherwise |
| What's the comparison class? | Comparing earbud silicon to desktop or data-center parts is not a like-for-like test |
If the Technology Delivers
Setting the unverified multiplier aside, the underlying premise is worth taking seriously. Vendors pursuing genuine compute-in-memory designs typically pitch two real benefits once memory bandwidth and thermal headroom cooperate: meaningfully more complex on-device models without a proportional battery hit, and lower round-trip latency for real-time tasks like noise cancellation, since inference doesn't have to leave the device. Both are legitimate engineering goals for a 10mm-class form factor with no active cooling. Broader memory-technology investment, like the moves detailed in Memory Prices Climb as Samsung and SK Hynix Pour $590B Into Chips, also feeds into how much bandwidth compact edge-AI chips can eventually draw on.
But "the architecture can plausibly deliver this" and "this specific chip has been shown to deliver this" are different claims, and only one of them is currently supported by public evidence. Not every "why does this chip exist" question has a straightforward marketing answer — the lockout logic explained in Sega's TMSS Explained: The Lock-Out Chip Behind the Genesis is a reminder that a chip's stated purpose and its actual function in a product don't always match on first read.
The Practical Takeaway
For buyers evaluating AI-enabled earbuds today, prioritize real-world battery life, noise-cancellation quality reported by outlets that run controlled listening comparisons, and how quickly promised AI software features actually ship — over a single unverified compute multiplier with no disclosed workload or baseline. Apple's current earbud line, including the AirPods 4 (details), remains a reasonable feature-for-feature reference point while independent Thus chip benchmarks are pending.
Citations and sources
As of publication, no independent outlet has published third-party verified compute-in-memory benchmark data for Anker's Thus chip — the workload, baseline part, precision, and power figures behind the 150x claim remain undisclosed by the vendor. This piece is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.
