AI builders look beyond accelerators to CPUs, packaging and faster execution
Reporting links agent workloads to renewed server-CPU demand, while Google is reportedly considering AMD's CPU and packaging work for a future TPU. Anthropic's proposed Decart deal would add software reported to accelerate model execution, underscoring a broader infrastructure race.
Artificial Intelligence··Morning
Agent pipelines put the server CPU back in demand
Agentic AI pipelines are restoring pressure on the ordinary server CPU alongside accelerators. AMD vice-president Madhu Rangarajan told IEEE Spectrum that 7 of the 8 stages in a realistic agentic pipeline run entirely on the CPU. The demand already shows up in supply: Intel has cut client CPU production in favour of server parts and says it is sold out of server CPUs through at least the end of the year, while AMD has set its server CPU forecast 2 times higher. 2 research papers cited in the Spectrum piece put numbers on the bottleneck: scheduling changes cut end-to-end latency by up to 1.8 times under sustained load, and adding CPU cores reduced time-to-first-token latency by roughly 1.5 to 7 times. Arm and Qualcomm have announced CPUs aimed at agentic work, Nvidia has given priority to its Vera CPU, and Amazon Web Services has reportedly seen waiting times for CPU server capacity climb as AI workloads press on its cloud. The accelerator still matters; the orchestration around it now competes for silicon as well.[2]
A reported tenth-generation TPU looks to AMD packaging and cores
On the accelerator side, market chatter reported by SemiAnalysis and covered by Tom's Hardware has Google working with AMD on a version-ten TPU. Analysts frame the attraction as packaging know-how, including advanced packaging and SoIC work, plus the CPU cores that reinforcement-learning workloads keep asking for; the arrangement would be AMD's first real part in a custom AI accelerator. Google has built 9 accelerator generations with Broadcom as the silicon designer, so it does not need AMD for a conventional TPU design. The report also fits a trend visible in Google's own systems: its inference-oriented TPU 8i machines pair 1 Axion processor with every 2 accelerators, where the seventh generation used 1 Xeon for every 4. Neither Google nor AMD has confirmed the project. If the chatter holds, the next TPU generation is being sketched less as a pure accelerator and more as a package that keeps host cores close to the matrix engines agents keep waking. The report remains unconfirmed market talk short of a signed supply contract.[1]
Anthropic reaches for software that speeds model runs
Software sits in the same race. Calcalist reports that Anthropic and Israeli startup Decart have exchanged advanced drafts on a deal valuing Decart at about 7 billion dollars, paid mainly in Anthropic shares, with signing expected by early September. Nvidia offered a higher valuation, yet Decart's founders and lead investor Sequoia Capital preferred Anthropic. Decart was founded in September 2023, employs about 100 people and has raised 450 million dollars; its software is reported to let AI models run up to 8 times faster, and the company also builds real-time video generation with robotics uses. It was valued at about 4 billion dollars in a recent round and works on a project basis without an established recurring revenue stream. If the purchase completes it would give Anthropic its first development centre in Israel. CPU scarcity, hybrid TPU packaging talk and a proposed execution-software acquisition all widen the infrastructure contest beyond a single accelerator brand: builders are buying time on every layer that keeps multi-step AI work moving.[3], [2], [1]
