Tag: #asic9 results
Samsung Electronics has secured a significant ASIC production order from Meta, strengthening its position as a key manufacturing hub for AI giants like Tesla and Anthropic. The company's backlog is approaching 50 trillion KRW.
Anthropic, a leading AI research company, is reportedly shifting focus toward developing its own AI chips after 5 years of operation. This move aims to secure hardware sovereignty for their large language models.
Intel CEO Pat Gelsinger announced that x86 architecture will remain the dominant force in data centers by 2030, driven by its reliability and evolution for AI agent workloads. The company also introduced the Xeon 6 'Clearwater Forest' processor and a new rack-scale platform strategy.
Broadcom reported Q1 FY2026 revenue of $19.3B, up 29% YoY, driven by AI segment at $8.4B exceeding expectations. AI growth fueled by Google TPU shipments; Q2 guidance at $22B total with AI at $10.7B. VMware integration complete, debt ratios normalized.
TrendForce forecasts global top 8 CSPs' capex exceeding $710B in 2026, up 61% YoY, fueled by AI server and infra investments. Firms include Google, AWS, Meta, Microsoft, Oracle, Tencent, Alibaba, Baidu. Shifting to Nvidia/AMD GPUs plus ASICs for AI efficiency.
Taalas introduces hardware that etches LLM weights and architecture directly into silicon chips, achieving 16,000 tokens/second and under 1ms latency without HBM. They claim 60-day model-to-ASIC turnaround, LoRA support, and upcoming larger models. Built by 24 engineers with $30M, targeting low-latency AI applications.
Broadcom faces increasing pressure in the AI chip market as it struggles to compete with Nvidia. The company is also dealing with supply chain diversification strategies from major clients like Google.
Marvell's ASIC single cards underperformed in AI. Company pivots to interconnect tech to re-enter AI market. Raises guidance, emphasizes scalable cluster capabilities.
ASICs are gaining in inference due to lower costs and 1/10th GPU power consumption, suiting fixed models like Groq's 25 supported ones. GPUs dominate training, especially LLMs, due to flexibility and Nvidia ecosystem reliance. Big clouds partner with Marvell/Broadcom for ASIC development.