AMD has agreed to acquire Taalas, while Anthropic has confirmed plans for an in-house silicon team focused on running AI models.
The Register reported that AMD plans to acquire Taalas, a Toronto startup developing what it describes as model-specific integrated circuits for AI inference, with the deal expected to close in the fourth quarter subject to regulatory approval. Separately, Ars Technica reported that Anthropic has confirmed it is hiring a custom silicon team to design chips for running Claude.
The common thread is inference, the stage where a trained model generates answers or actions for users. That is different from training, which is the expensive process of building the model in the first place. Both reports point to companies trying to tailor hardware more closely to that serving stage, where speed, cost per token and hardware availability all matter.
What AMD is buying with Taalas
According to The Register's account, Taalas takes an unusual approach: instead of storing model weights in high-bandwidth memory, or HBM, it etches those weights directly into silicon. In simple terms, the chip is built around a particular model rather than acting as a more general-purpose processor.
The publication says Taalas divides the chip into a mask-ROM recall fabric, where the model weights are etched, and an SRAM recall fabric, where KV caches and fine-tuning adapters are stored. Readers do not need the circuit detail to grasp the practical point: the design is intended to make token generation much faster by dedicating the hardware to a known model layout.
The Register said Taalas unveiled a first test chip, HC1, in February on TSMC's 6nm process. It reported initial benchmarks of 16,960 tokens a second serving Meta's Llama 3.1 8B model, and described that February claim as 48 times faster than Nvidia GPUs and 8.5 times faster than Cerebras accelerators at the time. Those figures are reported as Taalas's initial benchmark claims, not as independent verification across wider workloads.
The next chip, HC2, is described as due this summer with support for 20 billion parameters per chip. The Register added that AMD appears likely to pair Taalas-derived accelerators with its Instinct-based Helios racks, with GPUs handling prompt processing while Taalas-based hardware handles token generation. That architecture is presented as an implication of AMD's plans rather than a formally launched product.
The speed comes with a clear trade-off
The same design that may improve inference speed also limits flexibility. The Register noted that once these chips are deployed, customers are effectively tied to that model. Changes larger than something like a LoRA adapter would require a chip re-spin, which the report says is expensive and time-consuming.
That matters because AI models keep changing. The article says new models are arriving on a nearly monthly basis, which makes model-specific hardware easier to justify only when a customer is confident it will keep serving the same model for long enough. Taalas argued, as quoted in the report, that updates do not require a complete restart of the chip design and that only two metal layers need changing, but that still leaves a practical deployment constraint.
This helps explain who may adopt such hardware first. The Register suggested the technology will largely be deployed by AI model developers, infrastructure providers and a limited number of inference providers. It also noted that OpenAI, Anthropic and Meta are major Instinct customers, which could matter if AMD tries to combine its GPU systems with Taalas-derived accelerators after the acquisition closes. The deal is expected to close in the fourth quarter, subject to regulatory approval.
The article also links the hardware to test-time scaling, a technique it describes as letting a model "think" for longer before responding. More reasoning time can improve accuracy, but it also uses more tokens and increases latency. If inference becomes faster and cheaper per token, model developers may have more room to use that technique, though that remains a possible implication rather than an announced product plan.
Anthropic is building its own silicon team, but not abandoning other chips
Ars Technica reported that Anthropic has confirmed plans to design its own hardware for Claude. The immediate evidence is operational rather than product-level: the company is hiring for a custom silicon team, including roles such as silicon engineer and technical programme manager.
Anthropic's spokesperson told multiple outlets, as cited by Ars Technica, that the company will still use a "multi-chip approach". In plain English, that means Anthropic is not replacing outside hardware with a single in-house design. Instead, it plans to use its own chips alongside hardware from other companies.
The report gives two reasons AI providers are pursuing this route. One is dependence on Nvidia, which Ars says remains a strategic vulnerability in a market where demand for compute still exceeds available capacity. The other is performance: designing chips and models together could produce better results than treating the model and the hardware as separate layers.
Anthropic says its teams will co-design hardware and models side by side. That is a more direct form of vertical integration, where one company controls more of the stack from model design to the systems that run it. But the timeline is still early. Ars Technica cautioned that because Anthropic is still hiring key team members, it will be some time before the company or its users see any benefit.
The same report places Anthropic in a broader group. It says OpenAI recently announced a custom inference chip called Jalapeño developed with Broadcom, that Google has long run its models on its own hardware, that Meta has designed and deployed its own chips, and that Mistral is reportedly exploring similar work. In this article, those examples matter as context for Anthropic's move, not as a separate set of verified product comparisons.
Why it matters now
Taken together, these announcements suggest that some large AI companies no longer see general-purpose accelerators alone as the full answer for serving models at scale. AMD's move is the more concrete one: it has announced a deal for a company with test silicon and published benchmark claims, although that acquisition has not closed yet. Anthropic's move is earlier-stage, confirming intent through hiring rather than through a shipped chip.
For IT teams and business leaders, the immediate lesson is less about buying custom chips tomorrow and more about where supplier strategies are heading. The sources support strongest near-term relevance for model developers, infrastructure providers and large inference operators, not for every enterprise. The attraction is clear in the reporting: faster token generation, lower serving costs and less dependence on a single external hardware supplier. The hesitation is just as clear: model-specific silicon reduces flexibility, can require chip redesigns when models change, and in Anthropic's case remains at the hiring stage rather than deployment.
A likely implication is that AI infrastructure choices may become more segmented. General-purpose GPUs may remain important where organisations need flexibility across changing models, while model-specific hardware could be more attractive where workloads are stable enough to justify the engineering trade-off. What remains open is how quickly these custom approaches move from promising announcements and test results into widely used production systems.