大家好 我是为什么叫QQ

Hello everyone, I'm WhyCallQQ.

Anthropic上半年的净亏损是807.48亿美元

Anthropic posted a net loss of $80.748 billion for the first half of the year,

可到了第二季度 它又拿到9.74亿美元经营利润

yet in the second quarter, it generated $974 million in operating profit—

第一次交出盈利的经营季度

marking its very first profitable operating quarter.

把这两条新闻放在一起

When you put these two headlines side by side,

很多人已经不知道该看哪条

many folks are at a total loss on what to make of it.

我的建议是 先别争它赚没赚钱

My advice: don't get caught up arguing whether they're 'in the black' or not.

把利润表 现金流量表和长期算力合同分开看

Separate the income statement, the cash flow statement, and long-term compute commitments.

程序员每天用的Claude Code

Claude Code, which developers use every single day,

恰好坐在这三张表的交叉口

sits right at the crossroads of these three financial statements.

企业为任务付费 Anthropic为推理付账

Enterprises pay for tasks, Anthropic foots the bill for inference compute,

工程师则要为结果验收

and engineers are left to verify the final code deliverable.

这笔账能否持续 比上市当天的股价更值得跟

Whether this math holds up is far more worth tracking than the IPO day stock price.

这期我不猜股价 只用工程方法检查收入质量

In this episode, I won't speculate on valuation—instead, we'll apply an engineering lens to revenue quality,

单位任务成本和失败风险 先把时间钉住

unit task cost, and failure risks. Let's first nail down the timeline.

2026年10月26日 Anthropic公开提交S-1注册文件

On October 26, 2026, Anthropic publicly filed its S-1 registration statement,

计划在纳斯达克使用ANT代码

planning to list on Nasdaq under ticker ANT.

发行价区间和发行股数仍然空着 上市还没有完成

The price range and share count remain blank—the IPO is not yet finalized.

媒体报道的11月中旬

Media buzz about mid-November,

以及1.5万亿到2万亿美元估值 都属于发行预期

as well as $1.5 trillion to $2 trillion valuation figures, are purely offering expectations.

你可能会问 这跟自己有什么关系？

You might ask: why should developers care about this?

招股书第一次把模型公司的收入

Because this S-1 marks the very first time an AI lab has laid out revenue,

算力费用和客户使用情况放到同一份公开文件里

compute infrastructure expenses, and customer usage profiles in a single public filing.

以后评估AI编程工具 除了跑分和试用体验

Going forward, evaluating AI coding tools won't just rely on leaderboards and subjective trials—

还多了一份经营数据

we now have hard operational business metrics.

招股书里的数字各有日期

The numbers in the prospectus cover distinct windows:

半年 季度和前九个月不能混排

half-year, quarterly, and nine-month figures cannot be mixed together.

后面的比较 我都会说出期间 收入曲线很陡

In the comparisons ahead, I'll explicitly state the period. The revenue curve is steep:

2023年全年4100万美元 2024年3.86亿美元

$41 million in full-year 2023, $386 million in 2024,

2025年45.85亿美元

$4.585 billion in 2025,

2026年上半年 收入达到162.86亿美元

and in the first half of 2026, revenue surged to $16.286 billion.

半年数据已经超过上一年全年

Six months of revenue already blew past the entire prior year,

但两个期间不能直接当成同口径增速

though different time spans can't be treated as direct apples-to-apples growth rates.

再看钱的来路

Next, let's look at where the money actually comes from.

企业按Token使用量购买模型服务

Enterprises purchase model API access billed on token usage,

个人用户购买Pro和Max订阅

while individual developers buy Pro and Max subscriptions.

Claude Code进入开发流程后

Once Claude Code embeds into developer workflows,

模型调用就跟真实工作绑在了一起

model invocations become directly tethered to real business work.

这也能解释 编程工具为何会跟

This explains why coding tools now appear side by side

云平台出现在同一张成本表里

with cloud platforms on the exact same infrastructure bill.

调用越深入业务流程 账单越接近生产资源

The deeper calls embed in core workflows, the more the invoice resembles production compute.

你觉得企业续费更看重模型排名 还是任务交付？

Do you think enterprises renew based on leaderboard rank, or task delivery?

把你的判断打在公屏上

Drop your verdict in the comments.

第二季度 Anthropic收入约116亿美元

In Q2, Anthropic generated approximately $11.6 billion in revenue,

算力和基础设施费用约56亿美元

against compute and infrastructure costs of about $5.6 billion.

公司披露的每美元算力对应收入达到2.08美元

The company reported $2.08 in revenue per dollar of compute expense—

2025年第一季度还只有0.40美元

up from just $0.40 in Q1 2025.

同期调整后经营利润率为19%

Adjusted operating margin for the quarter came in at 19%,

GAAP经营利润为9.74亿美元 这里要管住联想

with GAAP operating income of $974 million. But hold back wild extrapolations:

2.08美元只描述收入与算力费用的比例

that $2.08 only reflects the ratio between top-line revenue and compute spend;

19%也不能代替净利率

and 19% operating margin is not net profit margin—

人员 销售和其他费用仍在账上

headcount, SG&A, and R&D still sit on the ledger.

对工程师来说 这组数据提供了一个观察角度：

For software engineers, this data offers a crucial perspective:

模型能力提升之外

beyond raw model intelligence gains,

推理优化 缓存和调度也能改变商业结果

inference optimization, prompt caching, and intelligent scheduling directly flip unit economics.

你在自己的项目里也能做同类核算：

You can perform the exact same accounting in your own team:

把每次调用费用归到任务

attribute every token call to a specific Jira ticket or task,

再看收入或节省工时能否覆盖它

then verify whether business value or saved engineering hours actually cover it.

很多人会觉得

Many people assume that

订阅价格不高 AI编程应该是薄利多销

because subscription fees look modest, AI coding must be low-margin high-volume,

企业账单更像云服务

yet enterprise invoicing actually behaves just like cloud infrastructure.

读代码 生成补丁 运行工具

Reading codebases, authoring patches, invoking tool loops—

每一步都可能消耗Token

every single step burns through tokens.

Claude Code能进入采购清单

Claude Code enters enterprise procurement contracts

靠的是能否嵌进现有流程

primarily because of how cleanly it embeds into existing pipelines.

招股书把Claude Code放在企业产品体系里

The S-1 positions Claude Code right inside its core enterprise suite,

也把工具调用

grouping tool use,

上下文管理和Agent运行环境放在同一层

context management, and agent execution environments onto the exact same architectural tier.

采购方买的逐渐变成一条任务流水线

Buyers are no longer paying for chat completions—they're procuring an automated task assembly line.

企业采购还会检查系统接入 结果验收和预算报警

Procurement audits check auth integrations, acceptance verification, and budget alerts.

模型排名只负责打开评估的大门

Leaderboard rankings merely get you in the door for evaluation.

Demo里写出函数只算开始

Spitting out a function in a flashy demo is just step zero;

能在权限 测试和审查约束下完成工单

delivering tickets within strict permissions, testing, and PR review constraints

才有持续使用量

is what drives sustained, recurring token consumption.

Agent跑一次任务 账本上往往有好几行

Running a single agent task often writes multiple line items to the ledger:

读取仓库占用输入Token

ingesting the repo consumes input tokens;

规划和写补丁产生输出Token

planning and authoring diffs burns output tokens;

测试失败后还要带上下文重试

failing CI tests triggers retries dragging along historical context.

子Agent并行能压缩等待时间 也会增加调用量

Parallel sub-agents compress wall-clock latency, but multiply total call volume.

你可以直接给每次任务挂一个trace ID

You should attach a unique trace ID to every agent workflow,

记录缓存命中

logging prompt cache hits,

模型调用 工具耗时 重试次数和人工接管

model API latency, tool execution time, retry iterations, and human takeover points.

下次评审别只问模型聪不聪明 把失败路径打开看

At your next team retro, don't just ask if the model is 'smart'—inspect the failure traces.

很多时候 成本失控来自上下文反复膨胀

More often than not, cost runaway stems from explosive context ballooning

或者一个测试失败触发了整轮重来

or a single flaking test triggering a full-suite reboot.

这类日志还能指导模型路由

These telemetry logs can also inform smart model routing:

简单分类任务可以走低预算通道

simple triage can take a low-cost model channel,

复杂迁移再给更强模型留出空间

reserving the frontier model's budget for complex architectural refactors.

Anthropic披露的长期安排很醒目

Anthropic's long-term commitments in the filing are eye-popping:

微软相关支出至少449亿美元

at least $44.9 billion committed to Microsoft-related infrastructure,

Google约1145亿美元 Amazon至少1100亿美元

approx. $114.5 billion to Google, and at least $110 billion to Amazon.

期限横跨多年

These commitments span multiple years,

部分条款要求实际用量不足时仍需补足付款

with take-or-pay clauses requiring full minimum spend even if actual usage falls short.

合同也不能随意取消

Nor can these multi-billion-dollar pacts be casually canceled.

这些金额是未来的承诺 不能全塞进当期成本

These sums represent forward-looking promises, not immediate period expenses.

它们更像一份容量规划

They function much like long-range capacity planning:

需求增长够快 提前锁定算力能保障供给；

if demand surges, pre-booking compute guarantees cluster capacity;

需求不及预期 固定付款会压缩调整空间

if demand lags, fixed minimums crush financial maneuvering room.

工程团队做容量预留时也会遇到同一道题

Engineering teams face the exact same trade-off when sizing reserved instances—

只是Anthropic把金额推到了千亿美元级

only Anthropic scaled the stakes to hundreds of billions of dollars.

供应商同时是投资方和基础设施伙伴

The hyperscalers act simultaneously as investors, cloud hosts, and channel partners.

融资 采购和容量决策因此绑得更紧

Financing, procurement, and capacity bets are thus tightly intertwined.

做技术选型时 这层依赖也要登记

When doing vendor selection, make sure this structural dependency is factored in.

现在回到开场的807.48亿美元

Now let's revisit that opening headline: the $80.748 billion net loss.

里面约723.88亿美元来自可转换票据

Roughly $72.388 billion of that stems from convertible notes,

衍生负债和SAFE等工具的公允价值变化

derivative liabilities, and SAFE fair-value accounting adjustments.

公司估值上升 相关负债的会计计量也跟着变

As Anthropic's private valuation soared, accounting rules forced these liabilities to revalue upward—

它不是同额现金流出

it does NOT represent cash leaving the company's bank accounts.

上半年经营亏损为6.93亿美元

Operating loss for H1 was actually a much more modest $693 million,

经营现金流则为正39亿美元

while operating cash flow was positive $3.9 billion.

6月底 现金及等价物1124亿美元

As of June 30, cash and cash equivalents stood at $112.4 billion,

短期投资193亿美元

with an additional $19.3 billion in short-term liquid investments.

读这份报表 至少要把净亏损

When reading this S-1, you must examine the net loss,

经营结果和现金流动分成三行

the operating result, and the cash flow statement across three distinct lines.

少看一行 结论就会偏出去

Miss even one line, and your conclusion will be completely skewed.

正经营现金流也不保证未来每个季度都如此

Positive operating cash flow today doesn't guarantee every quarter will look the same:

收付款时间会影响单个期间

working capital timing creates quarterly swings,

长期合同还要继续兑现 后面还要看资本开支

long-term contracts must be fulfilled, and heavy CapEx looms ahead.

截至9月30日 超过1.19万家客户在过去十二

As of September 30, over 11,900 customers contributed at least

个月各自贡献至少10万美元收入

$100,000 in trailing 12-month revenue,

超过1900家达到100万美元

and more than 1,900 accounts topped $1 million.

前五大客户占前九个月收入约25%

The top five customers accounted for ~25% of revenue over the first nine months,

没有单一客户超过10%客户面确实在扩大

with no single account exceeding 10%—the customer base is undeniably broadening.

可招股书也提醒 很多大客户没有长期合同

Yet the prospectus candidly notes that many top customers have no long-term lock-in,

或者合同很快到期 可以随时减少支出

or hold contracts nearing expiry that allow dialling back spend at will.

客户数增长与收入稳定是两件事

Customer count expansion and revenue durability are two very different beasts.

看到这里你可能会想

Seeing this, you might wonder:

企业会为了切换便宜模型放弃现有工作流吗？

would an enterprise abandon an entrenched workflow just to switch to a cheaper model?

把你的看法打在公屏上

Drop your thoughts in the chat.

迁移成本这道题很值得争一争

The question of switching costs is well worth debating.

评估黏性时 可以看工作流接入深度

When measuring vendor stickiness, look at workflow integration depth,

工具迁移成本 以及替换模型后的验收表现

custom tool migration overhead, and test pass rates after swapping backends—

这些都比注册账户数更接近真实使用

these reflect real platform moat far better than raw registered accounts.

DoorDash在招股书客户案例里说

DoorDash highlighted in the filing's customer case studies

主要产品上线时间缩短超过一半

that major product release cycles were cut by more than half,

每月使用Cowork的员工超过7000人

with over 7,000 employees actively using Cowork monthly.

这是公司提供的客户案例

Keep in mind, this is vendor-provided customer marketing—

适合当线索 不宜当成独立生产率实验

treat it as a directional clue rather than an independent empirical trial.

我更在意它对结果的描述：

What stands out to me is how they frame the outcome:

团队关心产品更快上线 代码合并只是过程

engineering leadership cares about shipping faster; writing and merging code is just an intermediate step.

你自己的评测也可以这样做

You can design your internal benchmarks the exact same way:

固定一组真实工单 用同一仓库和验收标准

take a curated set of real production tickets against a pinned repo and identical tests,

对比人工流程与Agent流程

benchmarking human baseline workflows directly against AI agents.

记录从接单到合并的时间

Measure the end-to-end cycle time from assignment to merge,

再把返工和线上缺陷补回去 别停在代码生成速度

then account for rework and post-release defects—don't stop at raw token generation speed.

最好保留失败样本

Be sure to preserve failure traces:

只看成功案例 会漏掉Agent在陌生依赖

cherry-picking only successful PRs completely masks the hidden costs of agents stumbling on unfamiliar deps,

遗留代码和不完整需求上的额外成本

legacy code, and underspecified requirements.

一个够用的面板可以先放五项指标：

A solid engineering dashboard should track five core metrics:

任务合并周期 一次验收通过率 人工接管次数

Cycle time to merge, first-pass acceptance rate, human takeover frequency,

合入后缺陷率 每个成功合并任务的总成本

post-merge defect rate, and total cost per successfully merged task.

总成本要包含模型 工具 重试和审查工时

Total cost must aggregate model API spend, tool execution, retries, and human code review hours,

再除以成功合并的任务数

divided cleanly by the count of accepted PRs.

每周看分布 别只看平均数

Inspect weekly percentile distributions, not just naive averages:

少数超长任务可能吃掉大部分预算

a handful of runaway tasks might consume the lion's share of your token budget,

也可能恰好是最有价值的工作

or they might happen to be your most critical architectural breakthroughs.

把任务按缺陷修复 迁移 新功能分类

Bucket tasks into bug fixes, migrations, and greenfield features—

结论会清楚很多

the trade-offs will become instantly apparent.

你的团队如果只能加一个指标

If your team can only afford to track one single metric tomorrow,

我会先加人工接管次数

I'd start with human takeover count.

面板还要按模型版本留快照

Keep historical metric snapshots by model version:

升级后如果成功率提高 重试也增加

if a model bump raises pass rate but spikes retry churn,

你可以直接判断这次升级是否划算

you have the hard data to judge whether the upgrade is truly ROI-positive.

别凭一次体验做决定

Never make tooling decisions on gut feeling alone.

招股书用很长篇幅披露滥用

The prospectus devotes extensive space to risk factors, including misuse,

代码破坏 欺骗和极端伤害等风险

code sabotage, deception, and catastrophic harms.

风险披露描述的是可能发生的情景

Remember that SEC risk disclosures outline potential worst-case scenarios,

不能读成这些事件都已在用户环境发生

not that these catastrophic failures have actively breached customer environments.

落到工程上 安全要进入完成定义

From an engineering perspective, security must be baked into your Definition of Done.

Agent能不能直接读生产密钥 谁能批准写权限

Can the agent read production secrets? Who authorizes write permissions?

工具调用有没有日志 失败后能否回滚

Are tool calls audited? Can actions cleanly roll back on failure?

这些都该有默认答案

These questions demand hardened defaults.

模型能力越强 权限边界越要提前写清楚

The more capable the model, the stricter the permission boundaries must be drawn up front;

否则一次看似普通的自动化改造

otherwise, an innocent automation upgrade

可能把原本孤立的系统串成一条攻击路径

can chain previously air-gapped systems into an unbroken attack vector.

评测集里也应加入越权请求

Your evaluation harness must include privilege escalation prompts,

恶意仓库说明和带毒工具输出

poisoned repo instructions, and malicious tool mock outputs.

正常任务能跑通 只能说明功能可用

Passing standard happy paths only proves basic functionality;

生产准入还要过安全测试

production readiness demands adversarial resilience.

Anthropic的治理结构也很少见

Anthropic's corporate governance structure is also highly unconventional:

七位联合创始人通过一股

its seven co-founders hold a single share

Class F股票掌握50.1%投票权

of Class F super-voting stock representing 50.1% voting control—

这股没有经济权益；

with zero economic rights attached;

长期利益信托则能选出董事会多数席位

meanwhile, the Long-Term Benefit Trust elects the majority of the board of directors.

普通股东可以投资 却很难改变公司的路线

Public shareholders can invest, but hold virtually no sway over corporate direction.

程序员选供应商时也该保留自己的退出方案

When selecting foundational model vendors, engineers should always maintain an off-ramp exit plan:

把模型调用包在内部接口后面

wrap model APIs behind internal gateway facades,

保存可迁移的评测集

maintain portable golden evaluation suites,

给提示词和工具Schema做版本管理

and version-control all prompts and tool schemas.

这样遇到价格 容量或政策变化

That way, when pricing shifts, rate limits bite, or terms of service change,

你能切换实现 不必重写整条业务链

you can hot-swap backends without rewriting your entire product pipeline.

评测集尤其重要

A robust internal evaluation suite is your greatest asset:

它既能量化新模型是否达标

it objectively measures whether a newer model meets baseline quality,

也能减少迁移时靠感觉调提示词的成本

while eliminating the tedious guesswork of tweaking prompts by intuition.

切换时 你至少有数据可谈

When negotiating or migrating, you have authoritative data on your side.

这份招股书给程序员的动作清单很具体

This S-1 filing leaves developers with a crystal-clear action checklist:

给Agent任务补上成本日志

1. Attach token cost and latency telemetry to every agent ticket;

把测试和人工接管接入同一条trace

2. Wire test failures and human takeover events into unified distributed traces;

给生产权限加审批与回滚

3. Enforce strict human sign-off and rollback gates on all production actions;

再准备一套可跨模型运行的验收集

4. Curate an internal golden evaluation suite that runs agnostically across any model.

以后看AI编程新闻 可以先问四件事：

Next time you read AI coding hype, ask these four simple questions:

任务成了多少 花了多少

How many tasks actually merged? How much did it cost?

失败谁兜底 供应商能不能换

Who foots the bill when it breaks? How easily can you switch vendors?

如果你还没有完整平台

If you don't yet have an end-to-end evaluation platform,

先用一张表记录二十个真实任务 也能开始

simply tracking twenty real tickets in a shared spreadsheet is plenty to get started.

样本不必漂亮

Your initial dataset doesn't need to be fancy;

口径要能复现 下周复盘就会看到具体瓶颈

the metrics just need to be reproducible—and your next sprint retro will reveal exact bottlenecks.

代码能生成 工程要验收

Models can generate code, but engineering demands rigorous verification.

你的团队现在最缺哪项数据

Which metric is your engineering team currently missing most:

是任务成本 缺陷率 还是人工接管次数？

task cost, post-merge defect rate, or human takeover frequency?

评论区聊聊

Let's discuss in the comments.

我是为什么叫QQ 感谢观看 别忘了一键三连

I'm WhyCallQQ. Thank you for watching, and remember to like, share, and subscribe!