All Topics
All Topics
Technology
Technology
AI
AI
Business
Business
Entertainment
Entertainment
News
News
Programming
Programming
Science
Science
Design
Design
Environment
Environment
Finance
Finance
Crypto
Crypto
Politics
Politics
Sports
Sports
Education
Education
Gaming
Gaming
Art
Art
Music
Music
Health
Health
Security
Security
Books
Books
Food
Food
Travel
Travel
Personal
Personal
Bluesky
Twitter
First reported by 36氪
8点1氪丨前华为天才少年吐槽DeepSeek面试,称面试官质疑他抄袭;苹果折叠屏已在量产;特朗普呼吁购买戴尔电脑,引发戴尔暴涨百亿美元

DeepSeek与智谱双双造芯:国产大模型从算法到芯片的全栈突围,推理芯片自研的范式革命

14h agoen
Read on happyrock.cloud

From the article

一、引言 2026年7月8日,两条消息几乎同时引爆科技圈:路透社曝出DeepSeek正在秘密推进自研AI推理芯片项目逾一年;The Information披露智谱AI正在评估定制AI芯片方案。48小时内,中国大模型第一梯队的两家代表企业,先后将触角伸向了算力最底层的硅基世界。 这不是孤立事件。从OpenAI的Jalapeño芯片到Anthropic的造芯计划,从谷歌TPU到亚马逊Trainium,全球头部AI玩家的竞争早已越过模型层,下沉到硬件最深处。当DeepSeek与智谱相继踏上这条重资产赛道,标志着中国AI产业正式从"算法优先"的上半场,迈入"软硬全栈"的下半场。 二、为什么是现在?三重驱动力的共振 2.1 推理成本的失控 算力成本是大模型公司最大的运营支出,占比普遍超过60%。训练是一次性的重投入,而推理是永不停歇的流水消耗——每一次用户提问、每一行代码生成、每一张图片输出,都在实时燃烧算力。 随着AI应用规模化铺开,行业算力需求的重心正快速从训练向推理倾斜。Trendforce预测,到2026年,ASIC(专用集成电路)的增长率将达到44%。英伟达GPU的毛利率长期维持在60%以上,高端AI芯片更是卖方市场。对于日均Token调用量数十亿级别的头部厂商而言,每Token成本下降一分钱,全年就能节省数亿元开支。 自研推理ASIC芯片虽然前期投入巨大(设计成本约5亿美元),但针对特定模型优化后,单位推理成本可比通用GPU降低30%-50%,功耗下降更为显著。当业务规模达到临界点,自研芯片的投资回报曲线就会翻转——这是一笔越用越划算的账。 # Python实现:自研推理芯片的成本效益分析 class ChipCostAnalysis : """自研推理芯片 vs 英伟达GPU的成本效益分析""" def __init__ (self): # 英伟达GPU方案 self . nvidia = { "chip_cost" : 35000 , # 单卡H100成本(美元) "power_w" : 700 , # 功耗(瓦) "throughput" : 1000 , # tokens/s "price_per_m_token" : 2.0 , # 每百万token推理成本(美元) "lifespan_years" : 4 , } # 自研推理ASIC方案(估算) self . asic = { "rd_cost" : 500_000_000 , # 研发成本(5亿美元) "chip_cost" : 8000 , # 单芯片成本 "power_w" : 250 , # 功耗 "throughput" : 2000 , # tokens/s "price_per_m_token" : 0.8 , # 每百万token推理成本 "lifespan_years" : 4 , } def break_even_analysis (self, daily_tokens: int) -> dict: """ 盈亏平衡分析 Args: daily_tokens: 日均Token调用量 """ days_per_year = 365 years = 5 # 每年的token处理量 annual_tokens = daily_tokens * days_per_year # GPU方案年度成本 gpu_annual_cost = (annual_tokens / 1_000_000 ) * self . nvidia[ "price_per_m_token" ] # ASIC方案年度成本(不含研发) asic_annual_operating = (annual_tokens / 1_000_000 ) * self . asic[ "price_per_m_token" ] # 累计成本对比 cumulative = [] asic_rd_remaining = self . asic[ "rd_cost" ] for year in range( 1 , years + 1 ): gpu_cumulative = gpu_annual_cost * year if year == 1 : asic_cumulative = asic_annual_operating + asic_rd_remaining else : asic_cumulative = asic_annual_operating * year + asic_rd_remaining savings = gpu_cumulative - asic_cumulative cumulative . append({ "year" : year, "gpu_cost" : round(gpu_cumulative, 2 ), "asic_cost" : round(asic_cumulative, 2 ), "savings" : round(savings, 2 ), "break_even" : savings >= 0 }) return { "daily_tokens" : daily_tokens, "annual_tokens" : annual_tokens, "cumulative" : cumulative, "years_to_break_even" : next( (y[ "year" ] for y in cumulative if y[ "break_even" ]), ">5 years" ) } def token_cost_comparison (self) -> dict: """每百万token成本对比""" gpu_cost = self . nvidia[ "price_per_m_token" ] asic_cost = self . asic[ "price_per_m_token" ] return { "gpu" : gpu_cost, "asic" : asic_cost, "savings_percent" : round(( 1 - asic_cost / gpu_cost) * 100 , 1 ) } # 分析不同规模下的盈亏平衡 daily_scenarios = [ 1_000_000_000 , 5_000_000_000 , 10_000_000_000 ] # 10亿/50亿/100亿tokens print( "=" * 70 ) print( "自研推理芯片 vs 英伟达GPU 成本效益分析" ) print( "=" * 70 ) analysis = ChipCostAnalysis() # 每百万token成本对比 cost_comp = analysis . token_cost_comparison() print( f " \n 每百万Token推理成本对比:" ) print( f " 英伟达GPU: $ { cost_comp[ 'gpu' ] : .2f } " ) print( f " 自研ASIC: $ { cost_comp[ 'asic' ] : .2f } " ) print( f " 成本节约: { cost_comp[ 'savings_percent' ] } %" ) for daily in daily_scenarios: result = analysis . break_even_analysis(daily) print( f " \n { '=' * 70 } " ) print( f "日均Token调用量: { daily : , } " ) print( f "年化Token量: { result[ 'annual_tokens' ] : , } " ) print( f "盈亏平衡时间: { result[ 'years_to_break_even' ] } " ) for yr in result[ 'cumulative' ]: status = "✅ 已盈亏平衡" if yr[ 'break_even' ] else "❌" print( f " 第 { yr[ 'year' ] } 年: GPU=$ { yr[ 'gpu_cost' ] : ,.0f } " f "ASIC=$ { yr[ 'asic_cost' ] : ,.0f } " f "差额=$ { yr[ 'savings' ] : +,.0f } { status } " ) 2.2 供应链主权 对于中国AI企业而言,造芯的紧迫性比海外同行更强。2022年以来,美国对中国AI芯片的出口管制持续升级——从A100/H100到H800/A800再到H20,每一次管制收紧都在压缩可用算力空间。今年4月美国一度禁止H20对华出口,虽然后续恢复但附加了15%收入上缴条件。
Continue reading on happyrock.cloud

You might also wanna read

Comments

Sign in to join the conversation.

No comments yet. Be the first.