Token吃掉三成工资,硅谷AI账单失控了

marsbitPublished on 2026-07-06Last updated on 2026-07-06

Abstract

硅谷AI应用正呈现两极分化:一方面,研究机构SemiAnalysis通过大模型将内部Token支出控制在员工薪资的30%以内,实现了生产效率的数倍提升,重写了专业服务的成本结构。英伟达CEO黄仁勋更是积极推动AI工具普及,视Token为新时代的生产资料。 另一方面,Uber、微软等巨头却因AI使用量激增而面临预算严重超支、成本失控的困境。有研究表明,在许多场景下,雇人仍然比使用AI更经济,且过度使用可能导致运维事故。 关键在于未来的成本趋势。通过软件优化(如使吞吐量提升14倍)和硬件迭代(如新一代芯片性能可达旧款的数十倍),大模型推理成本正快速下降。Anthropic等公司的毛利率不降反升,预示Token价格将持续走低。 当前全球科技公司一边大幅增加AI资本开支,一边持续裁员,但AI对宏观经济的直接影响尚不明显。这类似于基础设施建设的阵痛期:先投入铺设“管道”,应用和效益才会随之涌现。对于企业而言,选择是尽早利用成本下降的趋势拥抱AI转型,还是等待观望。

每百万Token只花0.99美元。

这是SemiAnalysis——硅谷最硬核的半导体研究机构——自己账单上的真实成本。

但更炸裂的是这个数字:内部大模型Token支出,已经占到员工总薪资的30%。

听起来不少——但反过来算,这笔钱买到的产出,过去得靠好几倍的人力成本才能覆盖。人均每月吞掉近50亿个Token,是Meta人均水平的5倍以上,核心贡献者月消耗更突破1000亿。

原本需要初级分析师花几个小时搞定的Excel模型转换、财报图表制作,如今几分钟内完成,只需要几美元。

SemiAnalysis自己的评价一针见血:这不是10%的效率提升,而是专业服务业的单元经济正在被重写。

研究公司、对冲基金、律所——所有靠人脑吃饭的行业,Token支出占到薪资的两三成,只是时间早晚的问题。

英伟达CEO黄仁勋比谁都急。

今年GTC大会上他直接放话:一个年薪50万美元的工程师,年底Token消费不到25万美元?

「我会彻底抓狂。」

他打算给英伟达每个工程师发相当于半年工资的Token预算,还要让7.5万名员工搭配750万个AI智能体一起干活。

不用AI?老黄说,这跟芯片设计师坚持用纸和铅笔没区别。

Token已经不是工具了,它正在变成新时代的「生产资料」。

但硅谷的另一半,正在为AI账单抓狂

有意思的是,就在SemiAnalysis用Token省下真金白银的同时,硅谷的巨头们正因为AI账单焦头烂额。

Uber是最经典的案例。

去年底公司向5000名工程师推广Claude Code,还搞了排行榜——用得越多,排名越高,内部竞争直接拉满。

结果太成功了:2月工程师使用率32%,3月就飙到84%,到了4月,95%的工程师每月都在用AI,70%的提交代码来自AI生成,而全年预算——已经花完了。

CTO说「要从头重做预算」。后来更狠——Bloomberg曝出,Uber给每位员工设了每月1500美元的Token上限,超了要特批。

但COO Andrew Macdonald在播客里说了句大实话:AI使用量确实在涨,但它跟消费者功能创新之间的联系......目前还看不到。

微软的情况更魔幻。上个月《The Verge》曝出,微软正在取消大部分Claude Code许可证,转向自家的GitHub Copilot CLI。

原因很简单:花钱的速度比产出的速度还快。

英伟达应用深度学习副总裁Bryan Catanzaro在今年4月说得更直接:「对我的团队来说,计算成本远远超过了员工的成本。」

MIT 2024年的研究:在以视觉为主要工作内容的岗位中,只有23%的场景下AI自动化在经济上划算。

剩下77%的情况下,雇人比用AI便宜。

甚至还有工程师吐槽AI智能体在使用中「毁掉了他的数据库和网络」——他称之为「过度使用」的代价。

天价预算、使用失控、翻车不断——硅谷正处在AI经济学最撕裂的阶段。

一边是技术带来前所未有的生产力,一边是账单以同样前所未有的速度膨胀。

成本塌缩才刚刚开始

但SemiAnalysis的核心论点是:别盯着今天的价格看,成本塌缩才刚开始。

先看软件端。

在B300上运行DeepSeek R1,通过wideEP、disagg与MTP三层纯软件优化,单GPU吞吐量能从baseline的1000 tokens/秒飙升至14000 tokens/秒——14倍提升,纯靠代码。

再看硬件端。

最优化配置的GB300 NVL72吞吐量是H100的17倍,切到FP4精度直接拉到32倍。

Opus 4.7的标价是输入5美元/百万、输出25美元/百万,看起来不便宜。

但由于智能体工作负载的输入输出比高达300:1,加上90%以上的缓存命中率,实际混合成本被压到了0.99美元。

连标价的五分之一都不到。

把软件和硬件叠在一起,一个结论很难回避:大模型的毛利率扩张,不是一次性的定价巧合,而是结构性的趋势。

Anthropic今年的ARR从90亿美元冲到了440亿以上,毛利率从38%飙到了70%以上——Token变便宜了,但卖Token的人反而更赚钱了。

Gartner今年3月的报告佐证了这一点:到2030年,万亿参数大模型的推理成本将比2025年下降超过90%

SemiAnalysis的判断很明确:如果你想预估2027年Token价格,答案就一个字——降。

钱花了,然后呢?

这正是当下AI最撕裂的地方:全球科技公司今年AI资本开支已宣布7400亿美元,比去年暴增69%;同一时间,科技业裁员速度已超去年全年。

钱在狂烧,人在被裁,但Goldman Sachs首席经济学家说了句大实话——AI对经济的实际影响,到目前为止基本为零。

这不是AI不行,而是每一轮基础设施革命都要经历的阵痛:先烧钱建管道,再等水流过来。

电网如此,互联网如此,AI也不例外。

差别只在于,这一次管道铺设的速度,和水流过来的速度,都是上一代人没见过的量级。

SemiAnalysis已经站在水流过来的那一边了——30%的薪资换来了数倍的产出杠杆,而成本曲线还在急剧下降。

至于其他公司:是现在蹚水过河,还是等对岸的人已经建好了城再追。

参考资料:

https://x.com/SemiAnalysis_/status/2070915305858007345

本文来自微信公众号“新智元”,作者:ASI启示录,编辑:所罗门

Trending Cryptos

Related Questions

Q根据文章,SemiAnalysis机构内部大模型的Token支出占员工总薪资的比例是多少?

A30%。

Q文章中提到的Uber公司为员工设定的每月Token使用上限是多少美元?

A1500美元。

Q文章中,英伟达CEO黄仁勋认为一个年薪50万美元的工程师,年底Token消费应该达到多少才算合理?

A至少25万美元。

Q根据Gartner的报告预测,到2030年,万亿参数大模型的推理成本将比2025年下降多少?

A下降超过90%。

Q文章指出,目前AI对经济产生的实际影响程度如何?

A根据高盛首席经济学家的观点,到目前为止基本为零。

Related Reads

MSTR Discloses Sale of 3,588 Bitcoins, Stock Price Drops Over 5% at One Point During Trading

MicroStrategy, the world's largest corporate holder of Bitcoin, has significantly shifted its business model. Between June 29 and July 5, the company sold 3,588 bitcoins for approximately $216 million to fund quarterly dividends for its preferred stock. This marks its largest-ever Bitcoin sale and signals a strategic pivot: Bitcoin is transitioning from a "buy-and-hold" reserve asset to a liquidity management tool for the company. This move follows a recent authorization allowing Bitcoin sales when equity fundraising is less attractive. The announcement contributed to a more than 5% intraday drop in MicroStrategy's stock price, while Bitcoin fell to around $61,800—below the company's average holding cost of roughly $75,700. The sale represents a major departure from MicroStrategy's long-standing "never sell" commitment, which saw its first minor breach in May with a $2.5 million sale. The latest, hundred-times-larger transaction underscores growing financial pressures. Analysts note the company faces about $1.5 billion in annual preferred dividend obligations, far exceeding cash flow from its software business. As of July 5, MicroStrategy holds 843,775 bitcoins. Its current operational logic involves buying Bitcoin during favorable financing conditions and selling portions to cover dividends when needed, creating a flexible capital management cycle amidst a challenging market environment.

华尔街日报2h ago

MSTR Discloses Sale of 3,588 Bitcoins, Stock Price Drops Over 5% at One Point During Trading

华尔街日报2h ago

Q-Day Countdown: Will Quantum Computing End Cryptocurrencies?

Quantum Computing's Threat to Cryptocurrency: A Countdown to Q-Day Quantum computing, specifically Shor's algorithm, poses a fundamental threat to the public-key cryptography (e.g., ECDSA, RSA) that secures blockchain networks like Bitcoin and Ethereum. This critical juncture, known as Q-Day, is estimated to occur potentially within the next 5-15 years. The core vulnerability stems from the public and immutable nature of blockchains. Assets in addresses where the public key is already exposed on-chain (e.g., spent outputs) are at direct risk, as a sufficiently powerful quantum computer could derive the private key. This threatens the very trust model of cryptocurrencies. The response lies in Post-Quantum Cryptography (PQC)—algorithms like lattice-based ML-DSA and hash-based SLH-DSA, which are resistant to quantum attacks. NIST has standardized key PQC algorithms (FIPS 203, 204, 205), providing a migration path. However, the primary challenge is not technical but socio-economic and involves complex governance: * **Bitcoin's** path is constrained by its conservative ethos. Migrating requires a soft-fork to new address types, facing hurdles like significantly larger signature sizes and, most critically, the divisive governance question of how to handle at-risk legacy UTXOs without violating core principles. * **Ethereum** is pursuing a "cryptographic agility" strategy, with a multi-layered roadmap. It leverages account abstraction for user accounts and is developing compressed hash-based signatures (e.g., leanXMSS) for its consensus layer, aiming for a full-stack upgrade over time. In conclusion, quantum computing does not spell an instant end for cryptocurrency but initiates a critical countdown. The industry has a limited "engineering comfort window" to orchestrate a coordinated, ecosystem-wide migration to PQC. The ultimate bottlenecks are the immense coordination efforts and governance decisions required for this foundational transition.

marsbit3h ago

Q-Day Countdown: Will Quantum Computing End Cryptocurrencies?

marsbit3h ago

Trump, the President Who Knows Best How to 'Trade Stocks'

Former US President Donald Trump reported a record-breaking $2.2 billion in personal income for 2025, the highest annual income ever disclosed by a sitting president. This figure, from a 927-page government ethics filing, represented a 3.5-fold increase from his $600 million income in 2024 and boosted his net worth to $6.5 billion. The primary drivers were cryptocurrency (64% of income, approximately $1.4 billion) and real estate (26%, approximately $575 million). His crypto earnings stemmed largely from the launch of his personal meme coin, $TRUMP, generating over $600 million in licensing fees, and substantial profits from the WLFI token and its parent company. Despite a sluggish property market, his Mar-a-Lago resort and associated golf clubs saw revenue surges of 50% and 27%, respectively, attributed to their use as venues for presidential events. Trump's financial disclosure also revealed an unprecedented level of stock market activity, with over 22,000 trades executed in 2025, averaging 87 trades per market day. Media analyses noted several instances where significant trading coincided with major policy announcements, such as proposed tariffs, raising questions about potential conflicts of interest. While the White House stated these trades were handled by a family-managed trust fund and not Trump directly, critics highlighted this as a departure from the blind trusts traditionally used by presidents post-Watergate. The report has intensified debate over the commercialization of the presidency. Supporters view it as a success story of a businessman-president, while critics argue it demonstrates an unprecedented conversion of public influence into private wealth, with policy decisions potentially linked to personal financial gains. The controversy centers on whether Trump's earnings represent innovative entrepreneurship or a fundamental conflict of interest, sparking renewed calls for stricter ethics reforms in US governance.

marsbit3h ago

Trump, the President Who Knows Best How to 'Trade Stocks'

marsbit3h ago

Countdown to Q-Day: Will Quantum Computing End Cryptocurrencies?

The article explores the existential threat quantum computing poses to cryptocurrencies and the urgent need for "post-quantum" migration. It outlines that quantum computers, through Shor's algorithm, could break the elliptic-curve cryptography (ECC) underlying blockchain security, potentially allowing private keys to be derived from public keys. The core challenge is not a lack of post-quantum cryptography (PQC) standards—like NIST's ML-KEM and ML-DSA—but the immense complexity of upgrading entire ecosystems before "Q-Day" (when quantum computers become capable of such attacks, estimated around 2035-2045). Key points include: * **Bitcoin's** risk is concentrated in legacy UTXOs with exposed public keys (e.g., early P2PK outputs). Migration faces massive hurdles: PQC signatures are much larger, increasing transaction size and cost, and the governance dilemma of handling un-migrated assets threatens its "code is law" ethos. * **Ethereum's** strategy focuses on "cryptographic agility," using Account Abstraction for user accounts and developing compressed hash-based signatures (like leanXMSS with SNARK aggregation) for consensus. Its migration is a complex, full-stack overhaul of execution, consensus, and data layers. * The "security debt" is enormous. The comfortable engineering window for a coordinated, ecosystem-wide upgrade is only 5-8 years. High-value infrastructure (exchanges, bridges) may face pressure before mainnet protocols. In conclusion, quantum computing is not an instant "doomsday" event but a forcing function for systemic change. Bitcoin's ultimate test is social consensus and property rights governance, while Ethereum's is technical complexity. Failure to migrate in time could lead to a fundamental re-pricing of crypto assets.

链捕手4h ago

Countdown to Q-Day: Will Quantum Computing End Cryptocurrencies?

链捕手4h ago

Trading

Spot

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of S (S) are presented below.

活动图片