It's happening...

Executive summary

Matthew Berman reviews Anthropic’s S-1, portraying the company as a highly focused enterprise AI business with unusually rapid customer and efficiency growth. Its core strategy is to build leading AI models, sell them to businesses, and use revenue, customer feedback, and model-assisted development to improve future models.

The S-1 also describes material risks: dependence on compute and key employees, customer concentration, uncertain AI safety, and a governance structure that gives founders substantial voting control.

Strategy and operating playbooks

• Focused enterprise strategy: Berman describes Anthropic’s approach as concentrating on making high-performing models and selling them for enterprise use, rather than expanding into a broader range of consumer products.
• Growth flywheel: The S-1’s “close the loop” concept links model capability, adoption, and investment:
  1. Better models support more use cases.
  2. More use cases drive customer adoption and sales.
  3. Greater usage provides feedback and supports further investment in models and compute.
  4. Claude also assists with research, training, and testing of subsequent models, potentially speeding development.

Berman characterizes this as a recursive self-improvement dynamic. The S-1 excerpts he discusses say Claude is increasingly used in development; they do not establish that the process is fully autonomous or without human involvement.

• Compute allocation: Anthropic says it manages compute from multiple suppliers as a flexible pool, directing capacity toward the workload expected to be most productive at the time, such as serving customers or training models.
• Multi-vendor infrastructure: The company says its systems work across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, reducing reliance on a single compute platform.
• Talent and productivity: Anthropic presents its mission and focus as tools for recruiting and retaining technical talent. Berman also points to Claude’s use inside the company as a way to increase employee productivity.
• Efficiency and pricing: The S-1 says lower per-task costs can encourage greater use. Berman relates this to Jevons’ paradox: making a resource cheaper or more efficient may expand total consumption enough to increase overall revenue.
• Sponsored product example: Berman promotes Kylon as a shared workspace for people, AI agents, and tools, with common context and human review before work ships. This is presented as a way to reduce duplicated or missed work across disconnected agents.

Business metrics cited

The figures below are reported in the video as S-1 disclosures or estimates; some are rounded.

Metric Figure cited Customers spending at least $100,000 annually About 11,900 as of September 30, 2026 Customers spending at least $1 million annually More than 1,900 Customers spending at least $10 million annually 250 Growth in $100,000-plus customers 20 in 2023; 200 in 2024; 1,500 in 2025; nearly 12,000 in 2026 Net revenue retention 900%, as reported by Berman Opus 5.5 cost per task vs. Opus 5 Approximately 40% lower Revenue growth vs. compute and infrastructure expense growth Revenue grew 1,300% year over year; expenses grew 300% for the three months ended June 30, 2026 Compute efficiency 5.2× improvement from Q1 2025 to Q2 2026 Revenue per employee More than $6.6 million trailing 12 months, versus $1.4 million as of September 30, 2025 Operating margin From –221% in Q1 2025 to +19% at the later period discussed Capital raised $141.1 billion in aggregate as of June 30, 2026 Cash and cash equivalents $112.4 billion as of June 30, 2026 Short-term investments $19.3 billion as of June 30, 2026 Claude’s share of code contributions More than 90% of codebase contributions in September 2026, according to the S-1 as described Internal Claude usage Monthly value reportedly exceeded monthly personnel expense, excluding stock-based compensation, in April 2026

Berman interprets the capital, cash, and investment figures as evidence that Anthropic has spent relatively little of the capital it raised. He also says the company’s revenue per dollar of compute improved enough that compute investment was generating multiple dollars of revenue.

Products, use cases, and market opportunity

• Enterprise coding: The video presents coding as a central commercial use case and a source of customer feedback that can help improve models.
• Beyond software: The S-1 discusses possible applications in drug discovery, disease detection, personalized education and advice, power-grid and battery design, robotics, manufacturing, logistics, and materials science.
• Physical-world agents: The S-1 refers to AI agents operating lab and manufacturing equipment, including microscopes, liquid handlers, and robotic arms. Berman says Anthropic has established a physical wet lab controlled by AI.
• Market estimate: The International Data Corporation estimate cited in the video projects $22.5 trillion in cumulative global economic impact from AI investment in solutions and services by 2031. Anthropic says the impact could be greater. This is an impact estimate, not company revenue.

Governance, risks, and trade-offs

• Public-benefit structure: Anthropic is described as a public benefit corporation. The S-1 says it contributed 846 million Class N shares—5% of fully diluted shares as of June 30, 2026—to the Anthropic Public Benefit Fund. Its co-founders have pledged 80% of their personal Anthropic equity to philanthropic causes.
• Founder voting power: The S-1 says a special Class F share, held through an LLC whose initial members are the seven co-founders, will carry 50.1% of total voting power. The structure also gives the founders significant influence over board composition. Berman argues that this concentrates control and limits ordinary shareholders’ ability to influence key decisions.
• Compute and supply risk: Anthropic warns that it may not secure enough compute, may allocate it inefficiently, or may face competitors able to outbid it or secure priority chip and cloud capacity. The company also says it has previously underestimated compute needs, constraining product development and deployment.
• Revenue concentration and growth uncertainty: Large enterprise contracts are a major strength, but reliance on a relatively small number of high-spending customers creates concentration risk. The S-1 also cautions that rapid growth may not continue indefinitely.
• AI safety and reliability: Disclosed risks include harmful or misaligned model behavior and the possibility that errors or biases in current models propagate into future systems. Anthropic also says it may not detect acceleration in model capabilities until after it has occurred.
• Agent economics: More agentic workloads may consume substantially more compute than ordinary chat. Anthropic warns that compute demand could therefore rise faster than the revenue associated with those workloads.
• Regulatory and geopolitical exposure: The S-1 discusses political and regulatory uncertainty, as well as dependence on critical inputs and infrastructure. Berman highlights the broader risk that control of compute and energy could concentrate economic and political power.
• Electricity commitments: Anthropic says it has publicly committed to covering consumer electricity-price increases caused by its data centers, including related infrastructure costs.
• Safety and deployment trade-off: The S-1 reportedly says Anthropic delayed pre-training a new frontier base model for more than a month to strengthen security measures, while also describing extensive internal use of Claude in development.

Practical business takeaways

• Build a reinforcing cycle between product quality, customer adoption, feedback, and reinvestment, while tracking whether growth is profitable at the level of compute and infrastructure.
• Treat compute as a portfolio of capacity across vendors, and allocate it according to expected marginal value.
• Evaluate price reductions by their effect on usage and total revenue, not just revenue per task.
• Monitor customer concentration alongside headline enterprise-customer growth.
• Plan for the higher compute intensity of agentic products; usage growth alone does not guarantee improving margins.
• Balance faster AI-assisted development with testing, security, and checks against errors propagating into later models.
• Scrutinize governance and shareholder rights separately from operating performance; strong growth does not resolve questions about who controls the company.

Presenters and sources

• Presenter: Matthew Berman.
• Primary source discussed: Anthropic’s S-1 filing.
• Other sources cited or quoted: Sam Altman, in an interview excerpt about the risks of recursive self-improvement; the International Data Corporation, for its estimate of AI’s economic impact.
• Sponsor featured: Kylon.