DEV Community: Amber D.

The Real Cost of AI: What Anthropic’s Explosive Growth Reveals About the Future of Software

Amber D. — Mon, 26 Oct 2026 19:26:04 +0000

AI is scaling faster than almost any technology before it. But behind every intelligent response lies a costly infrastructure, an uncertain business model, and a question every developer should be asking: Who can afford to build the future?

Artificial intelligence has a remarkable ability to make difficult things look easy.

A developer asks an AI assistant to debug a function, and seconds later, a plausible solution appears. A few prompts later, an application has a working prototype. Code that once took hours to write can now take minutes to generate.

From the user's perspective, the experience feels almost magical.

But magic is a poor business model.

Behind the interface are data centers, specialized chips, electricity, networking infrastructure, research teams, and enormous financial commitments. The more capable AI becomes, the more infrastructure may be required to train and serve it at scale.

On October 26, 2026, Anthropic's public IPO filing offered a revealing look at this reality. The company behind Claude disclosed extraordinary revenue growth alongside substantial operating costs, ambitious infrastructure commitments, and risks that extend beyond financial performance.

According to Reuters' report on Anthropic's financial disclosures, the company's revenue reached $16.29 billion in the first half of 2026, while its operating loss for that period was $693 million. Its reported net loss was much larger, at $80.7 billion, largely because of a non-operating accounting item involving convertible bonds.

These numbers tell a story more complicated than either an AI success story or a warning about an impending bubble.

They reveal an industry trying to turn one of the most consequential technological breakthroughs of our time into a sustainable business.

And developers have a front-row seat.

1. AI Has a Scaling Problem, Not Just a Coding Problem

For decades, software development followed a relatively familiar economic pattern.

Write the software once. Distribute it to millions of users. The marginal cost of serving each additional user can become very small.

AI complicates that equation.

Traditional software generally executes predefined instructions. Generative AI systems must perform computation to produce outputs, and that computation consumes resources every time a model is used.

A user asks a question. A model processes tokens. A coding agent examines a repository, generates changes, runs tests, and perhaps repeats the process several times. Each additional step can consume more compute.

At small scale, the costs may seem manageable. At millions of users and billions of model interactions, they become an important part of the business model.

This creates a fundamental tension: the product becomes more valuable as more people use it, but increased usage can also increase the cost of delivering it.

For AI companies, growth is therefore not automatically equivalent to profitability.

The critical questions include:

How much does it cost to serve each customer?

How frequently do customers use the product?

Can pricing cover the underlying inference and infrastructure costs?

Does the product generate enough economic value for customers to keep paying?

Can the company improve efficiency faster than demand and model complexity increase?

These questions matter just as much as benchmark scores or impressive demonstrations.

For developers building AI-powered applications, they matter even more directly. An application that works beautifully in a demo can become economically unsustainable when real users begin relying on it.

The next stage of AI engineering will require developers to optimize not only for correctness and latency, but also for cost per useful result.

2. Explosive Revenue Does Not Automatically Mean a Sustainable Business

Anthropic's reported growth is extraordinary.

Reuters reported that its revenue for the three months ended June 30, 2026, increased 1,368% compared with the same period a year earlier.

That kind of growth signals substantial demand. Businesses are finding practical applications for AI, particularly in areas such as software development, knowledge work, and automation.

But rapid growth can conceal important questions about long-term economics.

Consider two hypothetical AI services.

Service A generates $1 million in monthly revenue and spends $900,000 delivering its product.

Service B generates $10 million in monthly revenue but spends $12 million delivering it.

Service B is growing at a much larger scale, yet it has a more difficult immediate profitability problem.

Revenue tells us how much customers pay. It does not, by itself, tell us whether the underlying product economics work.

There is another important distinction in Anthropic's disclosures: its reported $80.7 billion first-half net loss should not be interpreted as an $80.7 billion operating loss or equivalent cash burn. The figure was largely driven by a non-operating accounting item related to convertible bonds. The reported operating loss was $693 million, a materially different measure.

This distinction matters because financial headlines can easily obscure what is happening inside a business.

For developers, the equivalent mistake is confusing product activity with product value.

A chatbot can attract millions of conversations without solving sufficiently important problems. An AI coding tool can generate impressive amounts of code without reducing the total time required to deliver reliable software.

The more meaningful metric is not how much AI a product produces.

It is how much useful work it accomplishes relative to its total cost.

3. The Developer's New Job Is to Manage Intelligence as a Resource

AI coding tools are changing the economics of software development.

A developer can delegate boilerplate generation, ask an agent to investigate a bug, or have a model propose an implementation across multiple files. These capabilities can reduce friction and accelerate experimentation.

However, they also introduce a new layer of engineering decisions.

Which model should handle a task? How much context does it need? Is an autonomous agent worth the additional cost? When should a workflow stop retrying? How can the application detect that an apparently convincing answer is wrong?

These are not merely prompt-engineering questions. They are systems-design questions.

Imagine a coding application that sends every request to the most expensive model available. It might perform impressively, but the business could struggle to sustain its costs.

Now imagine a more deliberate architecture:

A lightweight model classifies the request.

A stronger model handles complex reasoning when necessary.

Deterministic tools perform calculations, database queries, and other tasks that do not require generative reasoning.

Automated tests validate generated code.

A human reviews changes that carry significant consequences.

This approach does not assume that more intelligence is always better. It treats intelligence as a resource that should be allocated according to the task.

That principle is becoming increasingly important as developers build agentic systems capable of taking multiple actions over extended workflows.

An agent that makes ten unnecessary tool calls is not necessarily more intelligent than one that solves the problem in two.

An agent that generates hundreds of lines of code is not necessarily more productive than one that produces a small, correct patch.

The goal is not to maximize model activity. The goal is to maximize successful outcomes.

4. Infrastructure Is Becoming Part of the Application Architecture

One of the most important implications of Anthropic's disclosures is the scale of investment required to compete at the frontier of AI.

The company has committed hundreds of billions of dollars to infrastructure and computing arrangements over multiple years, according to Reuters' related reporting on its IPO disclosures.

These commitments help explain why AI development is not simply a matter of hiring talented researchers and writing better algorithms.

Compute capacity, access to advanced chips, data-center availability, energy, and cloud partnerships can all influence what a company is able to build and deliver.

This has consequences for application developers.

When you choose an AI provider, you are not selecting only a model. You are also making a decision about the infrastructure, pricing structure, service availability, and commercial dependencies behind your application.

That dependency deserves architectural attention.

Consider the following questions before building a product around a single provider:

What happens if API prices increase?

Can the application switch to another compatible model?

How will it behave when a provider experiences an outage?

Can frequently requested answers be cached safely?

Can some tasks be handled by smaller models or conventional software?

How will you monitor usage and enforce spending limits?

Not every application needs a sophisticated multi-provider architecture. Abstraction layers and fallback systems also introduce maintenance costs.

But understanding the trade-offs early can prevent expensive redesigns later.

The lesson is simple: AI infrastructure is no longer an implementation detail. It is part of the product's long-term strategy.

5. The Business Model May Matter as Much as the Model Itself

The AI industry is often discussed as a competition between increasingly capable models.

Which system writes better code? Which one reasons more accurately? Which one performs best on a benchmark?

These comparisons are useful, but incomplete.

A commercially successful AI product must connect technical capability to a sustainable revenue model.

Anthropic's business includes consumption-based revenue, in which customers pay according to usage, as well as subscription products. These models create different incentives and different economic challenges.

Consumption-based pricing can align revenue with usage, but it also requires careful cost management. Subscriptions offer more predictable billing for customers, but heavy usage can put pressure on margins if the price does not adequately reflect the resources consumed.

Enterprise customers introduce another set of considerations: procurement cycles, security requirements, support obligations, reliability expectations, and competition from alternative providers.

For developers, this suggests that the best AI product is not necessarily the one with the most sophisticated underlying model.

It may be the one that solves a narrow, expensive problem reliably enough that customers are happy to pay for it.

An AI tool that saves an engineering team several hours every week can have a clear value proposition. An AI feature that adds novelty without improving a meaningful workflow may struggle to justify its costs.

This is an opportunity for independent developers and small teams.

You do not need to train a frontier model to build a valuable AI product.

You need to understand a problem, identify where AI provides a measurable advantage, and design a system that delivers that advantage efficiently.

6. Safety Is an Engineering Cost, Not an Optional Feature

There is another dimension to Anthropic's disclosures that deserves attention.

The IPO materials describe risks associated with increasingly powerful AI systems, including misuse, security vulnerabilities, and outputs that could be interpreted as deceptive, coercive, or manipulative.

These concerns are not separate from the economics of AI.

Safety failures can create financial losses, legal exposure, reputational damage, and operational disruption. In some applications, they can also cause direct harm to users.

For developers, this means that evaluation and security must be part of the development process rather than activities added after deployment.

A practical AI engineering workflow should consider:

Input validation: Prevent malicious or unexpected inputs from bypassing application assumptions.

Permission boundaries: Give agents only the tools and access they actually need.

Output validation: Verify generated results before allowing them to trigger important actions.

Testing: Evaluate systems against realistic failure cases, not just successful demonstrations.

Monitoring: Track quality, latency, cost, and unexpected behavior in production.

Human oversight: Require review for consequential actions where automation cannot be trusted to operate independently.

A model can be impressive and still make mistakes. An agent can complete a task and still take an unsafe path to reach the result.

Reliable AI engineering requires both capability and control.

And as AI systems become more autonomous, the cost of ignoring that distinction will grow.

7. What Developers Should Do Now

The financial realities emerging from the AI industry do not mean developers should avoid AI.

They mean developers should approach it with the same discipline they apply to any other technology choice.

Here are five practical principles worth adopting.

Measure cost per successful task

Do not focus exclusively on tokens, requests, or model latency.

Measure what it costs to complete a task correctly, including retries, tool calls, verification, and human intervention.

Choose the smallest adequate model

Use the model that meets the task's quality requirements. Reserve more capable and expensive systems for work that genuinely benefits from them.

Build evaluation into development

Create representative test cases and measure correctness, consistency, and failure rates before deploying changes. A cheaper model is not a bargain if it introduces expensive errors.

Design for changing economics

Keep provider dependencies manageable, establish usage limits, and understand your options if pricing or availability changes.

Build products around outcomes

Customers rarely care how many tokens an application consumes. They care whether it saves time, reduces errors, improves decisions, or enables something they could not previously do.

The strongest products will connect technical sophistication to measurable value.

Conclusion: The Future Belongs to Useful Intelligence

Anthropic's financial disclosures capture a defining moment in the development of artificial intelligence.

Demand is growing rapidly. Revenue is scaling. Companies are making extraordinary investments in computing infrastructure. At the same time, the costs, risks, and uncertainties of building frontier AI systems are becoming harder to ignore.

For developers, this is not a reason to become pessimistic.

It is a reason to become more precise.

The early excitement around AI encouraged experimentation, and that experimentation remains valuable. But the next phase will demand more than impressive demos and rapidly generated code.

It will demand systems that are reliable, economically viable, secure, and genuinely useful.

We should expect model capabilities to continue evolving. We should also expect competition, changing prices, new architectures, and continuing debates about how much investment the industry can sustain.

No single financial filing can settle the question of whether today's AI investments will ultimately produce returns that justify their costs.

What it can do is remind us that technological progress and business sustainability are different challenges.

A powerful model is an achievement. A useful application is a product. A sustainable AI business requires both.

For developers, the opportunity is not simply to use more AI.

It is to build better software by understanding where intelligence creates value, where conventional engineering remains superior, and how to combine the two responsibly.

The future of AI will not be determined only by who builds the smartest models. It will also be determined by who learns to turn intelligence into something people can depend on—and afford.

Source: Reuters — Anthropic's soaring sales, steep losses put AI frenzy to IPO test, October 26, 2026.

This article is an independent analysis inspired by the reported financial disclosures, not a reproduction of the Reuters article.

The Government Says Web Dev Jobs Will Grow 5%. The Footnote Is the Interesting Part.

Amber D. — Fri, 23 Oct 2026 18:59:58 +0000

The U.S. Bureau of Labor Statistics updated its Occupational Outlook Handbook entry for web developers in August 2026, and the headline looks reassuring. The profile is now titled “Web Developers and Digital Designers,” and it projects 5% employment growth from 2025 to 2035. That beats the 3% average for all occupations, though it trails the 6% projected for computer occupations overall.

I think the footnotes and the structure of the numbers say more than the headline does, and I’d like to hear whether they match what you’re seeing.

The numbers, briefly

Combined category: about 220,100 jobs in 2025, a median wage of $99,520, and 11,300 net new jobs by 2035.

Web developers alone: about 87,400 jobs, a median of $92,650, 4% growth, and only 3,300 net new jobs over the decade.

Web and digital interface designers: about 132,700 jobs, a median of $104,000, 6% growth, and 8,000 net new jobs.

Openings: about 13,600 a year across the combined category, mostly from people leaving the occupation or the workforce.

For comparison, the median wage across all U.S. occupations is $50,980, so this is still well-paid work.

Observation 1: “Growth” and “opportunity” are different things

The combined category adds about 1,130 net jobs a year, against 13,600 openings a year. By my arithmetic, that means roughly 92% of openings come from churn, not expansion. People retire, drift into other roles, or move into management.

For developers specifically, 3,300 net new jobs over ten years is small. The field isn’t shrinking, but it isn’t the hypergrowth story many of us were sold. If you’re deciding whether to enter it, “will there be openings?” and “will the field be bigger?” are separate questions with different answers.

Observation 2: The designers outnumber and outgrow the developers

I didn’t expect this one. The designer side of the profile is about 50% larger than the developer side by headcount, projected to grow faster, and paid more at the median. The top 10% of designers earned over $201,550, versus over $162,290 for developers.

We tend to treat “web dev” as synonymous with “writes code,” but by BLS’s count, the people who own look, feel, and usability are the bigger and faster-growing group. Is that your experience? Or do you think the lines are blurring so much (design systems, Figma-to-code, designers shipping CSS) that the categories no longer separate cleanly?

Observation 3: The AI paragraph is short and carefully worded

The Job Outlook section says demand should rise as e-commerce expands. Then it adds a caveat: better tooling and AI for web development may soften growth. Those tools could make workers more productive, and they could let people in other occupations handle basic web development tasks.

That second point is the one I keep thinking about. It’s less “AI replaces developers” than “the floor of what counts as doing web dev moves, and people outside the profession can now reach it.” The marketer who ships their own landing page, or the ops analyst who builds an internal dashboard, never shows up as a web developer in BLS data, but they absorb work that used to land on someone’s desk.

If that’s right, the squeeze isn’t on the whole profession evenly. Basic brochure sites, simple CRUD, and template work are most exposed. Performance, accessibility, security, integration, and the judgment about what to build are less so.

Observation 4: The page itself shows how fast this field moves

A few details in the handbook read as slightly out of date. It lists Flash as a multimedia tool employers want developers to know, and it describes the coding work in terms of HTML and XML. Nothing about frameworks, TypeScript, accessibility standards, cloud deployment, or performance budgets appears in the skills description.

That isn’t a criticism of BLS. Labor statistics describe occupations, not tech stacks. But it’s a useful reminder that the official picture of this job lags the real one, and that the line about developers needing to keep up on new tools throughout their careers is the most accurate sentence on the page.

Observation 5: The statistics miss a lot of us

About 18% of web developers and 13% of designers are self-employed, and the BLS wage survey doesn’t cover self-employed workers. Freelancers, agency owners, and indie builders are in the employment counts but not the pay figures. So the median salary describes people with employers, and the picture of independent work is thinner than the headline suggests.

What I’d like to discuss

I’d rather not turn this into a “will AI take our jobs” thread, since we’ve all read that one. These are the questions I’m curious about:

Does 4% for developers match what you’re seeing? Are you seeing hiring slow, hold steady, or shift toward different skills?

Where is the floor moving? Which tasks that used to justify a dedicated web developer now get done by someone else, with or without AI?

Which skills would you bet on? If basic implementation gets cheaper, what holds its value: accessibility, performance, security, systems thinking, or something else?

Is “web developer” still the right unit? Do you identify as a web developer, a software engineer, a product engineer, or something else, and does it matter for how you think about your future?

For people early in their careers: the handbook says employers range from requiring a bachelor’s degree to accepting demonstrated ability. Which of those is true where you’re looking?

I’m not predicting anything here. I’m curious what people with actual hiring and job-hunting experience think the official numbers are missing.

Source:

U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, “Web Developers and Digital Designers,” last modified August 27, 2026.